1/*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright 1998, 2000 Marshall Kirk McKusick. 5 * Copyright 2009, 2010 Jeffrey W. Roberson <jeff@FreeBSD.org> 6 * All rights reserved. 7 * 8 * The soft updates code is derived from the appendix of a University 9 * of Michigan technical report (Gregory R. Ganger and Yale N. Patt, 10 * "Soft Updates: A Solution to the Metadata Update Problem in File 11 * Systems", CSE-TR-254-95, August 1995). 12 * 13 * Further information about soft updates can be obtained from: 14 * 15 * Marshall Kirk McKusick http://www.mckusick.com/softdep/ 16 * 1614 Oxford Street mckusick@mckusick.com 17 * Berkeley, CA 94709-1608 +1-510-843-9542 18 * USA 19 * 20 * Redistribution and use in source and binary forms, with or without 21 * modification, are permitted provided that the following conditions 22 * are met: 23 * 24 * 1. Redistributions of source code must retain the above copyright 25 * notice, this list of conditions and the following disclaimer. 26 * 2. Redistributions in binary form must reproduce the above copyright 27 * notice, this list of conditions and the following disclaimer in the 28 * documentation and/or other materials provided with the distribution. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE AUTHORS ``AS IS'' AND ANY EXPRESS OR 31 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 32 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 33 * IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY DIRECT, INDIRECT, 34 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, 35 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS 36 * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND 37 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR 38 * TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE 39 * USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 40 * 41 * from: @(#)ffs_softdep.c 9.59 (McKusick) 6/21/00 42 */ 43 44#include <sys/cdefs.h> 45__FBSDID("$FreeBSD$"); 46 47#include "opt_ffs.h" 48#include "opt_quota.h" 49#include "opt_ddb.h" 50 51/* 52 * For now we want the safety net that the DEBUG flag provides. 53 */ 54#ifndef DEBUG 55#define DEBUG 56#endif 57 58#include <sys/param.h> 59#include <sys/kernel.h> 60#include <sys/systm.h> 61#include <sys/bio.h> 62#include <sys/buf.h> 63#include <sys/kdb.h> 64#include <sys/kthread.h> 65#include <sys/ktr.h> 66#include <sys/limits.h> 67#include <sys/lock.h> 68#include <sys/malloc.h> 69#include <sys/mount.h> 70#include <sys/mutex.h> 71#include <sys/namei.h> 72#include <sys/priv.h> 73#include <sys/proc.h> 74#include <sys/racct.h> 75#include <sys/rwlock.h> 76#include <sys/stat.h> 77#include <sys/sysctl.h> 78#include <sys/syslog.h> 79#include <sys/vnode.h> 80#include <sys/conf.h> 81 82#include <ufs/ufs/dir.h> 83#include <ufs/ufs/extattr.h> 84#include <ufs/ufs/quota.h> 85#include <ufs/ufs/inode.h> 86#include <ufs/ufs/ufsmount.h> 87#include <ufs/ffs/fs.h> 88#include <ufs/ffs/softdep.h> 89#include <ufs/ffs/ffs_extern.h> 90#include <ufs/ufs/ufs_extern.h> 91 92#include <vm/vm.h> 93#include <vm/vm_extern.h> 94#include <vm/vm_object.h> 95 96#include <geom/geom.h> 97 98#include <ddb/ddb.h> 99 100#define KTR_SUJ 0 /* Define to KTR_SPARE. */ 101 102#ifndef SOFTUPDATES 103 104int 105softdep_flushfiles(oldmnt, flags, td) 106 struct mount *oldmnt; 107 int flags; 108 struct thread *td; 109{ 110 111 panic("softdep_flushfiles called"); 112} 113 114int 115softdep_mount(devvp, mp, fs, cred) 116 struct vnode *devvp; 117 struct mount *mp; 118 struct fs *fs; 119 struct ucred *cred; 120{ 121 122 return (0); 123} 124 125void 126softdep_initialize() 127{ 128 129 return; 130} 131 132void 133softdep_uninitialize() 134{ 135 136 return; 137} 138 139void 140softdep_unmount(mp) 141 struct mount *mp; 142{ 143 144 panic("softdep_unmount called"); 145} 146 147void 148softdep_setup_sbupdate(ump, fs, bp) 149 struct ufsmount *ump; 150 struct fs *fs; 151 struct buf *bp; 152{ 153 154 panic("softdep_setup_sbupdate called"); 155} 156 157void 158softdep_setup_inomapdep(bp, ip, newinum, mode) 159 struct buf *bp; 160 struct inode *ip; 161 ino_t newinum; 162 int mode; 163{ 164 165 panic("softdep_setup_inomapdep called"); 166} 167 168void 169softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags) 170 struct buf *bp; 171 struct mount *mp; 172 ufs2_daddr_t newblkno; 173 int frags; 174 int oldfrags; 175{ 176 177 panic("softdep_setup_blkmapdep called"); 178} 179 180void 181softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 182 struct inode *ip; 183 ufs_lbn_t lbn; 184 ufs2_daddr_t newblkno; 185 ufs2_daddr_t oldblkno; 186 long newsize; 187 long oldsize; 188 struct buf *bp; 189{ 190 191 panic("softdep_setup_allocdirect called"); 192} 193 194void 195softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 196 struct inode *ip; 197 ufs_lbn_t lbn; 198 ufs2_daddr_t newblkno; 199 ufs2_daddr_t oldblkno; 200 long newsize; 201 long oldsize; 202 struct buf *bp; 203{ 204 205 panic("softdep_setup_allocext called"); 206} 207 208void 209softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 210 struct inode *ip; 211 ufs_lbn_t lbn; 212 struct buf *bp; 213 int ptrno; 214 ufs2_daddr_t newblkno; 215 ufs2_daddr_t oldblkno; 216 struct buf *nbp; 217{ 218 219 panic("softdep_setup_allocindir_page called"); 220} 221 222void 223softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 224 struct buf *nbp; 225 struct inode *ip; 226 struct buf *bp; 227 int ptrno; 228 ufs2_daddr_t newblkno; 229{ 230 231 panic("softdep_setup_allocindir_meta called"); 232} 233 234void 235softdep_journal_freeblocks(ip, cred, length, flags) 236 struct inode *ip; 237 struct ucred *cred; 238 off_t length; 239 int flags; 240{ 241 242 panic("softdep_journal_freeblocks called"); 243} 244 245void 246softdep_journal_fsync(ip) 247 struct inode *ip; 248{ 249 250 panic("softdep_journal_fsync called"); 251} 252 253void 254softdep_setup_freeblocks(ip, length, flags) 255 struct inode *ip; 256 off_t length; 257 int flags; 258{ 259 260 panic("softdep_setup_freeblocks called"); 261} 262 263void 264softdep_freefile(pvp, ino, mode) 265 struct vnode *pvp; 266 ino_t ino; 267 int mode; 268{ 269 270 panic("softdep_freefile called"); 271} 272 273int 274softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 275 struct buf *bp; 276 struct inode *dp; 277 off_t diroffset; 278 ino_t newinum; 279 struct buf *newdirbp; 280 int isnewblk; 281{ 282 283 panic("softdep_setup_directory_add called"); 284} 285 286void 287softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize) 288 struct buf *bp; 289 struct inode *dp; 290 caddr_t base; 291 caddr_t oldloc; 292 caddr_t newloc; 293 int entrysize; 294{ 295 296 panic("softdep_change_directoryentry_offset called"); 297} 298 299void 300softdep_setup_remove(bp, dp, ip, isrmdir) 301 struct buf *bp; 302 struct inode *dp; 303 struct inode *ip; 304 int isrmdir; 305{ 306 307 panic("softdep_setup_remove called"); 308} 309 310void 311softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 312 struct buf *bp; 313 struct inode *dp; 314 struct inode *ip; 315 ino_t newinum; 316 int isrmdir; 317{ 318 319 panic("softdep_setup_directory_change called"); 320} 321 322void 323softdep_setup_blkfree(mp, bp, blkno, frags, wkhd) 324 struct mount *mp; 325 struct buf *bp; 326 ufs2_daddr_t blkno; 327 int frags; 328 struct workhead *wkhd; 329{ 330 331 panic("%s called", __FUNCTION__); 332} 333 334void 335softdep_setup_inofree(mp, bp, ino, wkhd) 336 struct mount *mp; 337 struct buf *bp; 338 ino_t ino; 339 struct workhead *wkhd; 340{ 341 342 panic("%s called", __FUNCTION__); 343} 344 345void 346softdep_setup_unlink(dp, ip) 347 struct inode *dp; 348 struct inode *ip; 349{ 350 351 panic("%s called", __FUNCTION__); 352} 353 354void 355softdep_setup_link(dp, ip) 356 struct inode *dp; 357 struct inode *ip; 358{ 359 360 panic("%s called", __FUNCTION__); 361} 362 363void 364softdep_revert_link(dp, ip) 365 struct inode *dp; 366 struct inode *ip; 367{ 368 369 panic("%s called", __FUNCTION__); 370} 371 372void 373softdep_setup_rmdir(dp, ip) 374 struct inode *dp; 375 struct inode *ip; 376{ 377 378 panic("%s called", __FUNCTION__); 379} 380 381void 382softdep_revert_rmdir(dp, ip) 383 struct inode *dp; 384 struct inode *ip; 385{ 386 387 panic("%s called", __FUNCTION__); 388} 389 390void 391softdep_setup_create(dp, ip) 392 struct inode *dp; 393 struct inode *ip; 394{ 395 396 panic("%s called", __FUNCTION__); 397} 398 399void 400softdep_revert_create(dp, ip) 401 struct inode *dp; 402 struct inode *ip; 403{ 404 405 panic("%s called", __FUNCTION__); 406} 407 408void 409softdep_setup_mkdir(dp, ip) 410 struct inode *dp; 411 struct inode *ip; 412{ 413 414 panic("%s called", __FUNCTION__); 415} 416 417void 418softdep_revert_mkdir(dp, ip) 419 struct inode *dp; 420 struct inode *ip; 421{ 422 423 panic("%s called", __FUNCTION__); 424} 425 426void 427softdep_setup_dotdot_link(dp, ip) 428 struct inode *dp; 429 struct inode *ip; 430{ 431 432 panic("%s called", __FUNCTION__); 433} 434 435int 436softdep_prealloc(vp, waitok) 437 struct vnode *vp; 438 int waitok; 439{ 440 441 panic("%s called", __FUNCTION__); 442} 443 444int 445softdep_journal_lookup(mp, vpp) 446 struct mount *mp; 447 struct vnode **vpp; 448{ 449 450 return (ENOENT); 451} 452 453void 454softdep_change_linkcnt(ip) 455 struct inode *ip; 456{ 457 458 panic("softdep_change_linkcnt called"); 459} 460 461void 462softdep_load_inodeblock(ip) 463 struct inode *ip; 464{ 465 466 panic("softdep_load_inodeblock called"); 467} 468 469void 470softdep_update_inodeblock(ip, bp, waitfor) 471 struct inode *ip; 472 struct buf *bp; 473 int waitfor; 474{ 475 476 panic("softdep_update_inodeblock called"); 477} 478 479int 480softdep_fsync(vp) 481 struct vnode *vp; /* the "in_core" copy of the inode */ 482{ 483 484 return (0); 485} 486 487void 488softdep_fsync_mountdev(vp) 489 struct vnode *vp; 490{ 491 492 return; 493} 494 495int 496softdep_flushworklist(oldmnt, countp, td) 497 struct mount *oldmnt; 498 int *countp; 499 struct thread *td; 500{ 501 502 *countp = 0; 503 return (0); 504} 505 506int 507softdep_sync_metadata(struct vnode *vp) 508{ 509 510 panic("softdep_sync_metadata called"); 511} 512 513int 514softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor) 515{ 516 517 panic("softdep_sync_buf called"); 518} 519 520int 521softdep_slowdown(vp) 522 struct vnode *vp; 523{ 524 525 panic("softdep_slowdown called"); 526} 527 528int 529softdep_request_cleanup(fs, vp, cred, resource) 530 struct fs *fs; 531 struct vnode *vp; 532 struct ucred *cred; 533 int resource; 534{ 535 536 return (0); 537} 538 539int 540softdep_check_suspend(struct mount *mp, 541 struct vnode *devvp, 542 int softdep_depcnt, 543 int softdep_accdepcnt, 544 int secondary_writes, 545 int secondary_accwrites) 546{ 547 struct bufobj *bo; 548 int error; 549 550 (void) softdep_depcnt, 551 (void) softdep_accdepcnt; 552 553 bo = &devvp->v_bufobj; 554 ASSERT_BO_WLOCKED(bo); 555 556 MNT_ILOCK(mp); 557 while (mp->mnt_secondary_writes != 0) { 558 BO_UNLOCK(bo); 559 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp), 560 (PUSER - 1) | PDROP, "secwr", 0); 561 BO_LOCK(bo); 562 MNT_ILOCK(mp); 563 } 564 565 /* 566 * Reasons for needing more work before suspend: 567 * - Dirty buffers on devvp. 568 * - Secondary writes occurred after start of vnode sync loop 569 */ 570 error = 0; 571 if (bo->bo_numoutput > 0 || 572 bo->bo_dirty.bv_cnt > 0 || 573 secondary_writes != 0 || 574 mp->mnt_secondary_writes != 0 || 575 secondary_accwrites != mp->mnt_secondary_accwrites) 576 error = EAGAIN; 577 BO_UNLOCK(bo); 578 return (error); 579} 580 581void 582softdep_get_depcounts(struct mount *mp, 583 int *softdepactivep, 584 int *softdepactiveaccp) 585{ 586 (void) mp; 587 *softdepactivep = 0; 588 *softdepactiveaccp = 0; 589} 590 591void 592softdep_buf_append(bp, wkhd) 593 struct buf *bp; 594 struct workhead *wkhd; 595{ 596 597 panic("softdep_buf_appendwork called"); 598} 599 600void 601softdep_inode_append(ip, cred, wkhd) 602 struct inode *ip; 603 struct ucred *cred; 604 struct workhead *wkhd; 605{ 606 607 panic("softdep_inode_appendwork called"); 608} 609 610void 611softdep_freework(wkhd) 612 struct workhead *wkhd; 613{ 614 615 panic("softdep_freework called"); 616} 617 618#else 619 620FEATURE(softupdates, "FFS soft-updates support"); 621 622static SYSCTL_NODE(_debug, OID_AUTO, softdep, CTLFLAG_RW, 0, 623 "soft updates stats"); 624static SYSCTL_NODE(_debug_softdep, OID_AUTO, total, CTLFLAG_RW, 0, 625 "total dependencies allocated"); 626static SYSCTL_NODE(_debug_softdep, OID_AUTO, highuse, CTLFLAG_RW, 0, 627 "high use dependencies allocated"); 628static SYSCTL_NODE(_debug_softdep, OID_AUTO, current, CTLFLAG_RW, 0, 629 "current dependencies allocated"); 630static SYSCTL_NODE(_debug_softdep, OID_AUTO, write, CTLFLAG_RW, 0, 631 "current dependencies written"); 632 633unsigned long dep_current[D_LAST + 1]; 634unsigned long dep_highuse[D_LAST + 1]; 635unsigned long dep_total[D_LAST + 1]; 636unsigned long dep_write[D_LAST + 1]; 637 638#define SOFTDEP_TYPE(type, str, long) \ 639 static MALLOC_DEFINE(M_ ## type, #str, long); \ 640 SYSCTL_ULONG(_debug_softdep_total, OID_AUTO, str, CTLFLAG_RD, \ 641 &dep_total[D_ ## type], 0, ""); \ 642 SYSCTL_ULONG(_debug_softdep_current, OID_AUTO, str, CTLFLAG_RD, \ 643 &dep_current[D_ ## type], 0, ""); \ 644 SYSCTL_ULONG(_debug_softdep_highuse, OID_AUTO, str, CTLFLAG_RD, \ 645 &dep_highuse[D_ ## type], 0, ""); \ 646 SYSCTL_ULONG(_debug_softdep_write, OID_AUTO, str, CTLFLAG_RD, \ 647 &dep_write[D_ ## type], 0, ""); 648 649SOFTDEP_TYPE(PAGEDEP, pagedep, "File page dependencies"); 650SOFTDEP_TYPE(INODEDEP, inodedep, "Inode dependencies"); 651SOFTDEP_TYPE(BMSAFEMAP, bmsafemap, 652 "Block or frag allocated from cyl group map"); 653SOFTDEP_TYPE(NEWBLK, newblk, "New block or frag allocation dependency"); 654SOFTDEP_TYPE(ALLOCDIRECT, allocdirect, "Block or frag dependency for an inode"); 655SOFTDEP_TYPE(INDIRDEP, indirdep, "Indirect block dependencies"); 656SOFTDEP_TYPE(ALLOCINDIR, allocindir, "Block dependency for an indirect block"); 657SOFTDEP_TYPE(FREEFRAG, freefrag, "Previously used frag for an inode"); 658SOFTDEP_TYPE(FREEBLKS, freeblks, "Blocks freed from an inode"); 659SOFTDEP_TYPE(FREEFILE, freefile, "Inode deallocated"); 660SOFTDEP_TYPE(DIRADD, diradd, "New directory entry"); 661SOFTDEP_TYPE(MKDIR, mkdir, "New directory"); 662SOFTDEP_TYPE(DIRREM, dirrem, "Directory entry deleted"); 663SOFTDEP_TYPE(NEWDIRBLK, newdirblk, "Unclaimed new directory block"); 664SOFTDEP_TYPE(FREEWORK, freework, "free an inode block"); 665SOFTDEP_TYPE(FREEDEP, freedep, "track a block free"); 666SOFTDEP_TYPE(JADDREF, jaddref, "Journal inode ref add"); 667SOFTDEP_TYPE(JREMREF, jremref, "Journal inode ref remove"); 668SOFTDEP_TYPE(JMVREF, jmvref, "Journal inode ref move"); 669SOFTDEP_TYPE(JNEWBLK, jnewblk, "Journal new block"); 670SOFTDEP_TYPE(JFREEBLK, jfreeblk, "Journal free block"); 671SOFTDEP_TYPE(JFREEFRAG, jfreefrag, "Journal free frag"); 672SOFTDEP_TYPE(JSEG, jseg, "Journal segment"); 673SOFTDEP_TYPE(JSEGDEP, jsegdep, "Journal segment complete"); 674SOFTDEP_TYPE(SBDEP, sbdep, "Superblock write dependency"); 675SOFTDEP_TYPE(JTRUNC, jtrunc, "Journal inode truncation"); 676SOFTDEP_TYPE(JFSYNC, jfsync, "Journal fsync complete"); 677 678static MALLOC_DEFINE(M_SENTINEL, "sentinel", "Worklist sentinel"); 679 680static MALLOC_DEFINE(M_SAVEDINO, "savedino", "Saved inodes"); 681static MALLOC_DEFINE(M_JBLOCKS, "jblocks", "Journal block locations"); 682static MALLOC_DEFINE(M_MOUNTDATA, "softdep", "Softdep per-mount data"); 683 684#define M_SOFTDEP_FLAGS (M_WAITOK) 685 686/* 687 * translate from workitem type to memory type 688 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 689 */ 690static struct malloc_type *memtype[] = { 691 NULL, 692 M_PAGEDEP, 693 M_INODEDEP, 694 M_BMSAFEMAP, 695 M_NEWBLK, 696 M_ALLOCDIRECT, 697 M_INDIRDEP, 698 M_ALLOCINDIR, 699 M_FREEFRAG, 700 M_FREEBLKS, 701 M_FREEFILE, 702 M_DIRADD, 703 M_MKDIR, 704 M_DIRREM, 705 M_NEWDIRBLK, 706 M_FREEWORK, 707 M_FREEDEP, 708 M_JADDREF, 709 M_JREMREF, 710 M_JMVREF, 711 M_JNEWBLK, 712 M_JFREEBLK, 713 M_JFREEFRAG, 714 M_JSEG, 715 M_JSEGDEP, 716 M_SBDEP, 717 M_JTRUNC, 718 M_JFSYNC, 719 M_SENTINEL 720}; 721 722#define DtoM(type) (memtype[type]) 723 724/* 725 * Names of malloc types. 726 */ 727#define TYPENAME(type) \ 728 ((unsigned)(type) <= D_LAST && (unsigned)(type) >= D_FIRST ? \ 729 memtype[type]->ks_shortdesc : "???") 730/* 731 * End system adaptation definitions. 732 */ 733 734#define DOTDOT_OFFSET offsetof(struct dirtemplate, dotdot_ino) 735#define DOT_OFFSET offsetof(struct dirtemplate, dot_ino) 736 737/* 738 * Internal function prototypes. 739 */ 740static void check_clear_deps(struct mount *); 741static void softdep_error(char *, int); 742static int softdep_process_worklist(struct mount *, int); 743static int softdep_waitidle(struct mount *, int); 744static void drain_output(struct vnode *); 745static struct buf *getdirtybuf(struct buf *, struct rwlock *, int); 746static int check_inodedep_free(struct inodedep *); 747static void clear_remove(struct mount *); 748static void clear_inodedeps(struct mount *); 749static void unlinked_inodedep(struct mount *, struct inodedep *); 750static void clear_unlinked_inodedep(struct inodedep *); 751static struct inodedep *first_unlinked_inodedep(struct ufsmount *); 752static int flush_pagedep_deps(struct vnode *, struct mount *, 753 struct diraddhd *); 754static int free_pagedep(struct pagedep *); 755static int flush_newblk_dep(struct vnode *, struct mount *, ufs_lbn_t); 756static int flush_inodedep_deps(struct vnode *, struct mount *, ino_t); 757static int flush_deplist(struct allocdirectlst *, int, int *); 758static int sync_cgs(struct mount *, int); 759static int handle_written_filepage(struct pagedep *, struct buf *, int); 760static int handle_written_sbdep(struct sbdep *, struct buf *); 761static void initiate_write_sbdep(struct sbdep *); 762static void diradd_inode_written(struct diradd *, struct inodedep *); 763static int handle_written_indirdep(struct indirdep *, struct buf *, 764 struct buf**, int); 765static int handle_written_inodeblock(struct inodedep *, struct buf *, int); 766static int jnewblk_rollforward(struct jnewblk *, struct fs *, struct cg *, 767 uint8_t *); 768static int handle_written_bmsafemap(struct bmsafemap *, struct buf *, int); 769static void handle_written_jaddref(struct jaddref *); 770static void handle_written_jremref(struct jremref *); 771static void handle_written_jseg(struct jseg *, struct buf *); 772static void handle_written_jnewblk(struct jnewblk *); 773static void handle_written_jblkdep(struct jblkdep *); 774static void handle_written_jfreefrag(struct jfreefrag *); 775static void complete_jseg(struct jseg *); 776static void complete_jsegs(struct jseg *); 777static void jseg_write(struct ufsmount *ump, struct jseg *, uint8_t *); 778static void jaddref_write(struct jaddref *, struct jseg *, uint8_t *); 779static void jremref_write(struct jremref *, struct jseg *, uint8_t *); 780static void jmvref_write(struct jmvref *, struct jseg *, uint8_t *); 781static void jtrunc_write(struct jtrunc *, struct jseg *, uint8_t *); 782static void jfsync_write(struct jfsync *, struct jseg *, uint8_t *data); 783static void jnewblk_write(struct jnewblk *, struct jseg *, uint8_t *); 784static void jfreeblk_write(struct jfreeblk *, struct jseg *, uint8_t *); 785static void jfreefrag_write(struct jfreefrag *, struct jseg *, uint8_t *); 786static inline void inoref_write(struct inoref *, struct jseg *, 787 struct jrefrec *); 788static void handle_allocdirect_partdone(struct allocdirect *, 789 struct workhead *); 790static struct jnewblk *cancel_newblk(struct newblk *, struct worklist *, 791 struct workhead *); 792static void indirdep_complete(struct indirdep *); 793static int indirblk_lookup(struct mount *, ufs2_daddr_t); 794static void indirblk_insert(struct freework *); 795static void indirblk_remove(struct freework *); 796static void handle_allocindir_partdone(struct allocindir *); 797static void initiate_write_filepage(struct pagedep *, struct buf *); 798static void initiate_write_indirdep(struct indirdep*, struct buf *); 799static void handle_written_mkdir(struct mkdir *, int); 800static int jnewblk_rollback(struct jnewblk *, struct fs *, struct cg *, 801 uint8_t *); 802static void initiate_write_bmsafemap(struct bmsafemap *, struct buf *); 803static void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *); 804static void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *); 805static void handle_workitem_freefile(struct freefile *); 806static int handle_workitem_remove(struct dirrem *, int); 807static struct dirrem *newdirrem(struct buf *, struct inode *, 808 struct inode *, int, struct dirrem **); 809static struct indirdep *indirdep_lookup(struct mount *, struct inode *, 810 struct buf *); 811static void cancel_indirdep(struct indirdep *, struct buf *, 812 struct freeblks *); 813static void free_indirdep(struct indirdep *); 814static void free_diradd(struct diradd *, struct workhead *); 815static void merge_diradd(struct inodedep *, struct diradd *); 816static void complete_diradd(struct diradd *); 817static struct diradd *diradd_lookup(struct pagedep *, int); 818static struct jremref *cancel_diradd_dotdot(struct inode *, struct dirrem *, 819 struct jremref *); 820static struct jremref *cancel_mkdir_dotdot(struct inode *, struct dirrem *, 821 struct jremref *); 822static void cancel_diradd(struct diradd *, struct dirrem *, struct jremref *, 823 struct jremref *, struct jremref *); 824static void dirrem_journal(struct dirrem *, struct jremref *, struct jremref *, 825 struct jremref *); 826static void cancel_allocindir(struct allocindir *, struct buf *bp, 827 struct freeblks *, int); 828static int setup_trunc_indir(struct freeblks *, struct inode *, 829 ufs_lbn_t, ufs_lbn_t, ufs2_daddr_t); 830static void complete_trunc_indir(struct freework *); 831static void trunc_indirdep(struct indirdep *, struct freeblks *, struct buf *, 832 int); 833static void complete_mkdir(struct mkdir *); 834static void free_newdirblk(struct newdirblk *); 835static void free_jremref(struct jremref *); 836static void free_jaddref(struct jaddref *); 837static void free_jsegdep(struct jsegdep *); 838static void free_jsegs(struct jblocks *); 839static void rele_jseg(struct jseg *); 840static void free_jseg(struct jseg *, struct jblocks *); 841static void free_jnewblk(struct jnewblk *); 842static void free_jblkdep(struct jblkdep *); 843static void free_jfreefrag(struct jfreefrag *); 844static void free_freedep(struct freedep *); 845static void journal_jremref(struct dirrem *, struct jremref *, 846 struct inodedep *); 847static void cancel_jnewblk(struct jnewblk *, struct workhead *); 848static int cancel_jaddref(struct jaddref *, struct inodedep *, 849 struct workhead *); 850static void cancel_jfreefrag(struct jfreefrag *); 851static inline void setup_freedirect(struct freeblks *, struct inode *, 852 int, int); 853static inline void setup_freeext(struct freeblks *, struct inode *, int, int); 854static inline void setup_freeindir(struct freeblks *, struct inode *, int, 855 ufs_lbn_t, int); 856static inline struct freeblks *newfreeblks(struct mount *, struct inode *); 857static void freeblks_free(struct ufsmount *, struct freeblks *, int); 858static void indir_trunc(struct freework *, ufs2_daddr_t, ufs_lbn_t); 859static ufs2_daddr_t blkcount(struct fs *, ufs2_daddr_t, off_t); 860static int trunc_check_buf(struct buf *, int *, ufs_lbn_t, int, int); 861static void trunc_dependencies(struct inode *, struct freeblks *, ufs_lbn_t, 862 int, int); 863static void trunc_pages(struct inode *, off_t, ufs2_daddr_t, int); 864static int cancel_pagedep(struct pagedep *, struct freeblks *, int); 865static int deallocate_dependencies(struct buf *, struct freeblks *, int); 866static void newblk_freefrag(struct newblk*); 867static void free_newblk(struct newblk *); 868static void cancel_allocdirect(struct allocdirectlst *, 869 struct allocdirect *, struct freeblks *); 870static int check_inode_unwritten(struct inodedep *); 871static int free_inodedep(struct inodedep *); 872static void freework_freeblock(struct freework *, u_long); 873static void freework_enqueue(struct freework *); 874static int handle_workitem_freeblocks(struct freeblks *, int); 875static int handle_complete_freeblocks(struct freeblks *, int); 876static void handle_workitem_indirblk(struct freework *); 877static void handle_written_freework(struct freework *); 878static void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *); 879static struct worklist *jnewblk_merge(struct worklist *, struct worklist *, 880 struct workhead *); 881static struct freefrag *setup_allocindir_phase2(struct buf *, struct inode *, 882 struct inodedep *, struct allocindir *, ufs_lbn_t); 883static struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t, 884 ufs2_daddr_t, ufs_lbn_t); 885static void handle_workitem_freefrag(struct freefrag *); 886static struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long, 887 ufs_lbn_t, u_long); 888static void allocdirect_merge(struct allocdirectlst *, 889 struct allocdirect *, struct allocdirect *); 890static struct freefrag *allocindir_merge(struct allocindir *, 891 struct allocindir *); 892static int bmsafemap_find(struct bmsafemap_hashhead *, int, 893 struct bmsafemap **); 894static struct bmsafemap *bmsafemap_lookup(struct mount *, struct buf *, 895 int cg, struct bmsafemap *); 896static int newblk_find(struct newblk_hashhead *, ufs2_daddr_t, int, 897 struct newblk **); 898static int newblk_lookup(struct mount *, ufs2_daddr_t, int, struct newblk **); 899static int inodedep_find(struct inodedep_hashhead *, ino_t, 900 struct inodedep **); 901static int inodedep_lookup(struct mount *, ino_t, int, struct inodedep **); 902static int pagedep_lookup(struct mount *, struct buf *bp, ino_t, ufs_lbn_t, 903 int, struct pagedep **); 904static int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t, 905 struct pagedep **); 906static void pause_timer(void *); 907static int request_cleanup(struct mount *, int); 908static int softdep_request_cleanup_flush(struct mount *, struct ufsmount *); 909static void schedule_cleanup(struct mount *); 910static void softdep_ast_cleanup_proc(struct thread *); 911static struct ufsmount *softdep_bp_to_mp(struct buf *bp); 912static int process_worklist_item(struct mount *, int, int); 913static void process_removes(struct vnode *); 914static void process_truncates(struct vnode *); 915static void jwork_move(struct workhead *, struct workhead *); 916static void jwork_insert(struct workhead *, struct jsegdep *); 917static void add_to_worklist(struct worklist *, int); 918static void wake_worklist(struct worklist *); 919static void wait_worklist(struct worklist *, char *); 920static void remove_from_worklist(struct worklist *); 921static void softdep_flush(void *); 922static void softdep_flushjournal(struct mount *); 923static int softdep_speedup(struct ufsmount *); 924static void worklist_speedup(struct mount *); 925static int journal_mount(struct mount *, struct fs *, struct ucred *); 926static void journal_unmount(struct ufsmount *); 927static int journal_space(struct ufsmount *, int); 928static void journal_suspend(struct ufsmount *); 929static int journal_unsuspend(struct ufsmount *ump); 930static void softdep_prelink(struct vnode *, struct vnode *); 931static void add_to_journal(struct worklist *); 932static void remove_from_journal(struct worklist *); 933static bool softdep_excess_items(struct ufsmount *, int); 934static void softdep_process_journal(struct mount *, struct worklist *, int); 935static struct jremref *newjremref(struct dirrem *, struct inode *, 936 struct inode *ip, off_t, nlink_t); 937static struct jaddref *newjaddref(struct inode *, ino_t, off_t, int16_t, 938 uint16_t); 939static inline void newinoref(struct inoref *, ino_t, ino_t, off_t, nlink_t, 940 uint16_t); 941static inline struct jsegdep *inoref_jseg(struct inoref *); 942static struct jmvref *newjmvref(struct inode *, ino_t, off_t, off_t); 943static struct jfreeblk *newjfreeblk(struct freeblks *, ufs_lbn_t, 944 ufs2_daddr_t, int); 945static void adjust_newfreework(struct freeblks *, int); 946static struct jtrunc *newjtrunc(struct freeblks *, off_t, int); 947static void move_newblock_dep(struct jaddref *, struct inodedep *); 948static void cancel_jfreeblk(struct freeblks *, ufs2_daddr_t); 949static struct jfreefrag *newjfreefrag(struct freefrag *, struct inode *, 950 ufs2_daddr_t, long, ufs_lbn_t); 951static struct freework *newfreework(struct ufsmount *, struct freeblks *, 952 struct freework *, ufs_lbn_t, ufs2_daddr_t, int, int, int); 953static int jwait(struct worklist *, int); 954static struct inodedep *inodedep_lookup_ip(struct inode *); 955static int bmsafemap_backgroundwrite(struct bmsafemap *, struct buf *); 956static struct freefile *handle_bufwait(struct inodedep *, struct workhead *); 957static void handle_jwork(struct workhead *); 958static struct mkdir *setup_newdir(struct diradd *, ino_t, ino_t, struct buf *, 959 struct mkdir **); 960static struct jblocks *jblocks_create(void); 961static ufs2_daddr_t jblocks_alloc(struct jblocks *, int, int *); 962static void jblocks_free(struct jblocks *, struct mount *, int); 963static void jblocks_destroy(struct jblocks *); 964static void jblocks_add(struct jblocks *, ufs2_daddr_t, int); 965 966/* 967 * Exported softdep operations. 968 */ 969static void softdep_disk_io_initiation(struct buf *); 970static void softdep_disk_write_complete(struct buf *); 971static void softdep_deallocate_dependencies(struct buf *); 972static int softdep_count_dependencies(struct buf *bp, int); 973 974/* 975 * Global lock over all of soft updates. 976 */ 977static struct mtx lk; 978MTX_SYSINIT(softdep_lock, &lk, "global softdep", MTX_DEF); 979 980#define ACQUIRE_GBLLOCK(lk) mtx_lock(lk) 981#define FREE_GBLLOCK(lk) mtx_unlock(lk) 982#define GBLLOCK_OWNED(lk) mtx_assert((lk), MA_OWNED) 983 984/* 985 * Per-filesystem soft-updates locking. 986 */ 987#define LOCK_PTR(ump) (&(ump)->um_softdep->sd_fslock) 988#define TRY_ACQUIRE_LOCK(ump) rw_try_wlock(&(ump)->um_softdep->sd_fslock) 989#define ACQUIRE_LOCK(ump) rw_wlock(&(ump)->um_softdep->sd_fslock) 990#define FREE_LOCK(ump) rw_wunlock(&(ump)->um_softdep->sd_fslock) 991#define LOCK_OWNED(ump) rw_assert(&(ump)->um_softdep->sd_fslock, \ 992 RA_WLOCKED) 993 994#define BUF_AREC(bp) lockallowrecurse(&(bp)->b_lock) 995#define BUF_NOREC(bp) lockdisablerecurse(&(bp)->b_lock) 996 997/* 998 * Worklist queue management. 999 * These routines require that the lock be held. 1000 */ 1001#ifndef /* NOT */ DEBUG 1002#define WORKLIST_INSERT(head, item) do { \ 1003 (item)->wk_state |= ONWORKLIST; \ 1004 LIST_INSERT_HEAD(head, item, wk_list); \ 1005} while (0) 1006#define WORKLIST_REMOVE(item) do { \ 1007 (item)->wk_state &= ~ONWORKLIST; \ 1008 LIST_REMOVE(item, wk_list); \ 1009} while (0) 1010#define WORKLIST_INSERT_UNLOCKED WORKLIST_INSERT 1011#define WORKLIST_REMOVE_UNLOCKED WORKLIST_REMOVE 1012 1013#else /* DEBUG */ 1014static void worklist_insert(struct workhead *, struct worklist *, int); 1015static void worklist_remove(struct worklist *, int); 1016 1017#define WORKLIST_INSERT(head, item) worklist_insert(head, item, 1) 1018#define WORKLIST_INSERT_UNLOCKED(head, item) worklist_insert(head, item, 0) 1019#define WORKLIST_REMOVE(item) worklist_remove(item, 1) 1020#define WORKLIST_REMOVE_UNLOCKED(item) worklist_remove(item, 0) 1021 1022static void 1023worklist_insert(head, item, locked) 1024 struct workhead *head; 1025 struct worklist *item; 1026 int locked; 1027{ 1028 1029 if (locked) 1030 LOCK_OWNED(VFSTOUFS(item->wk_mp)); 1031 if (item->wk_state & ONWORKLIST) 1032 panic("worklist_insert: %p %s(0x%X) already on list", 1033 item, TYPENAME(item->wk_type), item->wk_state); 1034 item->wk_state |= ONWORKLIST; 1035 LIST_INSERT_HEAD(head, item, wk_list); 1036} 1037 1038static void 1039worklist_remove(item, locked) 1040 struct worklist *item; 1041 int locked; 1042{ 1043 1044 if (locked) 1045 LOCK_OWNED(VFSTOUFS(item->wk_mp)); 1046 if ((item->wk_state & ONWORKLIST) == 0) 1047 panic("worklist_remove: %p %s(0x%X) not on list", 1048 item, TYPENAME(item->wk_type), item->wk_state); 1049 item->wk_state &= ~ONWORKLIST; 1050 LIST_REMOVE(item, wk_list); 1051} 1052#endif /* DEBUG */ 1053 1054/* 1055 * Merge two jsegdeps keeping only the oldest one as newer references 1056 * can't be discarded until after older references. 1057 */ 1058static inline struct jsegdep * 1059jsegdep_merge(struct jsegdep *one, struct jsegdep *two) 1060{ 1061 struct jsegdep *swp; 1062 1063 if (two == NULL) 1064 return (one); 1065 1066 if (one->jd_seg->js_seq > two->jd_seg->js_seq) { 1067 swp = one; 1068 one = two; 1069 two = swp; 1070 } 1071 WORKLIST_REMOVE(&two->jd_list); 1072 free_jsegdep(two); 1073 1074 return (one); 1075} 1076 1077/* 1078 * If two freedeps are compatible free one to reduce list size. 1079 */ 1080static inline struct freedep * 1081freedep_merge(struct freedep *one, struct freedep *two) 1082{ 1083 if (two == NULL) 1084 return (one); 1085 1086 if (one->fd_freework == two->fd_freework) { 1087 WORKLIST_REMOVE(&two->fd_list); 1088 free_freedep(two); 1089 } 1090 return (one); 1091} 1092 1093/* 1094 * Move journal work from one list to another. Duplicate freedeps and 1095 * jsegdeps are coalesced to keep the lists as small as possible. 1096 */ 1097static void 1098jwork_move(dst, src) 1099 struct workhead *dst; 1100 struct workhead *src; 1101{ 1102 struct freedep *freedep; 1103 struct jsegdep *jsegdep; 1104 struct worklist *wkn; 1105 struct worklist *wk; 1106 1107 KASSERT(dst != src, 1108 ("jwork_move: dst == src")); 1109 freedep = NULL; 1110 jsegdep = NULL; 1111 LIST_FOREACH_SAFE(wk, dst, wk_list, wkn) { 1112 if (wk->wk_type == D_JSEGDEP) 1113 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep); 1114 else if (wk->wk_type == D_FREEDEP) 1115 freedep = freedep_merge(WK_FREEDEP(wk), freedep); 1116 } 1117 1118 while ((wk = LIST_FIRST(src)) != NULL) { 1119 WORKLIST_REMOVE(wk); 1120 WORKLIST_INSERT(dst, wk); 1121 if (wk->wk_type == D_JSEGDEP) { 1122 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep); 1123 continue; 1124 } 1125 if (wk->wk_type == D_FREEDEP) 1126 freedep = freedep_merge(WK_FREEDEP(wk), freedep); 1127 } 1128} 1129 1130static void 1131jwork_insert(dst, jsegdep) 1132 struct workhead *dst; 1133 struct jsegdep *jsegdep; 1134{ 1135 struct jsegdep *jsegdepn; 1136 struct worklist *wk; 1137 1138 LIST_FOREACH(wk, dst, wk_list) 1139 if (wk->wk_type == D_JSEGDEP) 1140 break; 1141 if (wk == NULL) { 1142 WORKLIST_INSERT(dst, &jsegdep->jd_list); 1143 return; 1144 } 1145 jsegdepn = WK_JSEGDEP(wk); 1146 if (jsegdep->jd_seg->js_seq < jsegdepn->jd_seg->js_seq) { 1147 WORKLIST_REMOVE(wk); 1148 free_jsegdep(jsegdepn); 1149 WORKLIST_INSERT(dst, &jsegdep->jd_list); 1150 } else 1151 free_jsegdep(jsegdep); 1152} 1153 1154/* 1155 * Routines for tracking and managing workitems. 1156 */ 1157static void workitem_free(struct worklist *, int); 1158static void workitem_alloc(struct worklist *, int, struct mount *); 1159static void workitem_reassign(struct worklist *, int); 1160 1161#define WORKITEM_FREE(item, type) \ 1162 workitem_free((struct worklist *)(item), (type)) 1163#define WORKITEM_REASSIGN(item, type) \ 1164 workitem_reassign((struct worklist *)(item), (type)) 1165 1166static void 1167workitem_free(item, type) 1168 struct worklist *item; 1169 int type; 1170{ 1171 struct ufsmount *ump; 1172 1173#ifdef DEBUG 1174 if (item->wk_state & ONWORKLIST) 1175 panic("workitem_free: %s(0x%X) still on list", 1176 TYPENAME(item->wk_type), item->wk_state); 1177 if (item->wk_type != type && type != D_NEWBLK) 1178 panic("workitem_free: type mismatch %s != %s", 1179 TYPENAME(item->wk_type), TYPENAME(type)); 1180#endif 1181 if (item->wk_state & IOWAITING) 1182 wakeup(item); 1183 ump = VFSTOUFS(item->wk_mp); 1184 LOCK_OWNED(ump); 1185 KASSERT(ump->softdep_deps > 0, 1186 ("workitem_free: %s: softdep_deps going negative", 1187 ump->um_fs->fs_fsmnt)); 1188 if (--ump->softdep_deps == 0 && ump->softdep_req) 1189 wakeup(&ump->softdep_deps); 1190 KASSERT(dep_current[item->wk_type] > 0, 1191 ("workitem_free: %s: dep_current[%s] going negative", 1192 ump->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1193 KASSERT(ump->softdep_curdeps[item->wk_type] > 0, 1194 ("workitem_free: %s: softdep_curdeps[%s] going negative", 1195 ump->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1196 atomic_subtract_long(&dep_current[item->wk_type], 1); 1197 ump->softdep_curdeps[item->wk_type] -= 1; 1198 free(item, DtoM(type)); 1199} 1200 1201static void 1202workitem_alloc(item, type, mp) 1203 struct worklist *item; 1204 int type; 1205 struct mount *mp; 1206{ 1207 struct ufsmount *ump; 1208 1209 item->wk_type = type; 1210 item->wk_mp = mp; 1211 item->wk_state = 0; 1212 1213 ump = VFSTOUFS(mp); 1214 ACQUIRE_GBLLOCK(&lk); 1215 dep_current[type]++; 1216 if (dep_current[type] > dep_highuse[type]) 1217 dep_highuse[type] = dep_current[type]; 1218 dep_total[type]++; 1219 FREE_GBLLOCK(&lk); 1220 ACQUIRE_LOCK(ump); 1221 ump->softdep_curdeps[type] += 1; 1222 ump->softdep_deps++; 1223 ump->softdep_accdeps++; 1224 FREE_LOCK(ump); 1225} 1226 1227static void 1228workitem_reassign(item, newtype) 1229 struct worklist *item; 1230 int newtype; 1231{ 1232 struct ufsmount *ump; 1233 1234 ump = VFSTOUFS(item->wk_mp); 1235 LOCK_OWNED(ump); 1236 KASSERT(ump->softdep_curdeps[item->wk_type] > 0, 1237 ("workitem_reassign: %s: softdep_curdeps[%s] going negative", 1238 VFSTOUFS(item->wk_mp)->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1239 ump->softdep_curdeps[item->wk_type] -= 1; 1240 ump->softdep_curdeps[newtype] += 1; 1241 KASSERT(dep_current[item->wk_type] > 0, 1242 ("workitem_reassign: %s: dep_current[%s] going negative", 1243 VFSTOUFS(item->wk_mp)->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1244 ACQUIRE_GBLLOCK(&lk); 1245 dep_current[newtype]++; 1246 dep_current[item->wk_type]--; 1247 if (dep_current[newtype] > dep_highuse[newtype]) 1248 dep_highuse[newtype] = dep_current[newtype]; 1249 dep_total[newtype]++; 1250 FREE_GBLLOCK(&lk); 1251 item->wk_type = newtype; 1252} 1253 1254/* 1255 * Workitem queue management 1256 */ 1257static int max_softdeps; /* maximum number of structs before slowdown */ 1258static int tickdelay = 2; /* number of ticks to pause during slowdown */ 1259static int proc_waiting; /* tracks whether we have a timeout posted */ 1260static int *stat_countp; /* statistic to count in proc_waiting timeout */ 1261static struct callout softdep_callout; 1262static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 1263static int req_clear_remove; /* syncer process flush some freeblks */ 1264static int softdep_flushcache = 0; /* Should we do BIO_FLUSH? */ 1265 1266/* 1267 * runtime statistics 1268 */ 1269static int stat_flush_threads; /* number of softdep flushing threads */ 1270static int stat_worklist_push; /* number of worklist cleanups */ 1271static int stat_blk_limit_push; /* number of times block limit neared */ 1272static int stat_ino_limit_push; /* number of times inode limit neared */ 1273static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 1274static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 1275static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 1276static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 1277static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 1278static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 1279static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 1280static int stat_jaddref; /* bufs redirtied as ino bitmap can not write */ 1281static int stat_jnewblk; /* bufs redirtied as blk bitmap can not write */ 1282static int stat_journal_min; /* Times hit journal min threshold */ 1283static int stat_journal_low; /* Times hit journal low threshold */ 1284static int stat_journal_wait; /* Times blocked in jwait(). */ 1285static int stat_jwait_filepage; /* Times blocked in jwait() for filepage. */ 1286static int stat_jwait_freeblks; /* Times blocked in jwait() for freeblks. */ 1287static int stat_jwait_inode; /* Times blocked in jwait() for inodes. */ 1288static int stat_jwait_newblk; /* Times blocked in jwait() for newblks. */ 1289static int stat_cleanup_high_delay; /* Maximum cleanup delay (in ticks) */ 1290static int stat_cleanup_blkrequests; /* Number of block cleanup requests */ 1291static int stat_cleanup_inorequests; /* Number of inode cleanup requests */ 1292static int stat_cleanup_retries; /* Number of cleanups that needed to flush */ 1293static int stat_cleanup_failures; /* Number of cleanup requests that failed */ 1294static int stat_emptyjblocks; /* Number of potentially empty journal blocks */ 1295 1296SYSCTL_INT(_debug_softdep, OID_AUTO, max_softdeps, CTLFLAG_RW, 1297 &max_softdeps, 0, ""); 1298SYSCTL_INT(_debug_softdep, OID_AUTO, tickdelay, CTLFLAG_RW, 1299 &tickdelay, 0, ""); 1300SYSCTL_INT(_debug_softdep, OID_AUTO, flush_threads, CTLFLAG_RD, 1301 &stat_flush_threads, 0, ""); 1302SYSCTL_INT(_debug_softdep, OID_AUTO, worklist_push, CTLFLAG_RW, 1303 &stat_worklist_push, 0,""); 1304SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_push, CTLFLAG_RW, 1305 &stat_blk_limit_push, 0,""); 1306SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_push, CTLFLAG_RW, 1307 &stat_ino_limit_push, 0,""); 1308SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_hit, CTLFLAG_RW, 1309 &stat_blk_limit_hit, 0, ""); 1310SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_hit, CTLFLAG_RW, 1311 &stat_ino_limit_hit, 0, ""); 1312SYSCTL_INT(_debug_softdep, OID_AUTO, sync_limit_hit, CTLFLAG_RW, 1313 &stat_sync_limit_hit, 0, ""); 1314SYSCTL_INT(_debug_softdep, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, 1315 &stat_indir_blk_ptrs, 0, ""); 1316SYSCTL_INT(_debug_softdep, OID_AUTO, inode_bitmap, CTLFLAG_RW, 1317 &stat_inode_bitmap, 0, ""); 1318SYSCTL_INT(_debug_softdep, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, 1319 &stat_direct_blk_ptrs, 0, ""); 1320SYSCTL_INT(_debug_softdep, OID_AUTO, dir_entry, CTLFLAG_RW, 1321 &stat_dir_entry, 0, ""); 1322SYSCTL_INT(_debug_softdep, OID_AUTO, jaddref_rollback, CTLFLAG_RW, 1323 &stat_jaddref, 0, ""); 1324SYSCTL_INT(_debug_softdep, OID_AUTO, jnewblk_rollback, CTLFLAG_RW, 1325 &stat_jnewblk, 0, ""); 1326SYSCTL_INT(_debug_softdep, OID_AUTO, journal_low, CTLFLAG_RW, 1327 &stat_journal_low, 0, ""); 1328SYSCTL_INT(_debug_softdep, OID_AUTO, journal_min, CTLFLAG_RW, 1329 &stat_journal_min, 0, ""); 1330SYSCTL_INT(_debug_softdep, OID_AUTO, journal_wait, CTLFLAG_RW, 1331 &stat_journal_wait, 0, ""); 1332SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_filepage, CTLFLAG_RW, 1333 &stat_jwait_filepage, 0, ""); 1334SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_freeblks, CTLFLAG_RW, 1335 &stat_jwait_freeblks, 0, ""); 1336SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_inode, CTLFLAG_RW, 1337 &stat_jwait_inode, 0, ""); 1338SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_newblk, CTLFLAG_RW, 1339 &stat_jwait_newblk, 0, ""); 1340SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_blkrequests, CTLFLAG_RW, 1341 &stat_cleanup_blkrequests, 0, ""); 1342SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_inorequests, CTLFLAG_RW, 1343 &stat_cleanup_inorequests, 0, ""); 1344SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_high_delay, CTLFLAG_RW, 1345 &stat_cleanup_high_delay, 0, ""); 1346SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_retries, CTLFLAG_RW, 1347 &stat_cleanup_retries, 0, ""); 1348SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_failures, CTLFLAG_RW, 1349 &stat_cleanup_failures, 0, ""); 1350SYSCTL_INT(_debug_softdep, OID_AUTO, flushcache, CTLFLAG_RW, 1351 &softdep_flushcache, 0, ""); 1352SYSCTL_INT(_debug_softdep, OID_AUTO, emptyjblocks, CTLFLAG_RD, 1353 &stat_emptyjblocks, 0, ""); 1354 1355SYSCTL_DECL(_vfs_ffs); 1356 1357/* Whether to recompute the summary at mount time */ 1358static int compute_summary_at_mount = 0; 1359SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW, 1360 &compute_summary_at_mount, 0, "Recompute summary at mount"); 1361static int print_threads = 0; 1362SYSCTL_INT(_debug_softdep, OID_AUTO, print_threads, CTLFLAG_RW, 1363 &print_threads, 0, "Notify flusher thread start/stop"); 1364 1365/* List of all filesystems mounted with soft updates */ 1366static TAILQ_HEAD(, mount_softdeps) softdepmounts; 1367 1368/* 1369 * This function cleans the worklist for a filesystem. 1370 * Each filesystem running with soft dependencies gets its own 1371 * thread to run in this function. The thread is started up in 1372 * softdep_mount and shutdown in softdep_unmount. They show up 1373 * as part of the kernel "bufdaemon" process whose process 1374 * entry is available in bufdaemonproc. 1375 */ 1376static int searchfailed; 1377extern struct proc *bufdaemonproc; 1378static void 1379softdep_flush(addr) 1380 void *addr; 1381{ 1382 struct mount *mp; 1383 struct thread *td; 1384 struct ufsmount *ump; 1385 1386 td = curthread; 1387 td->td_pflags |= TDP_NORUNNINGBUF; 1388 mp = (struct mount *)addr; 1389 ump = VFSTOUFS(mp); 1390 atomic_add_int(&stat_flush_threads, 1); 1391 ACQUIRE_LOCK(ump); 1392 ump->softdep_flags &= ~FLUSH_STARTING; 1393 wakeup(&ump->softdep_flushtd); 1394 FREE_LOCK(ump); 1395 if (print_threads) { 1396 if (stat_flush_threads == 1) 1397 printf("Running %s at pid %d\n", bufdaemonproc->p_comm, 1398 bufdaemonproc->p_pid); 1399 printf("Start thread %s\n", td->td_name); 1400 } 1401 for (;;) { 1402 while (softdep_process_worklist(mp, 0) > 0 || 1403 (MOUNTEDSUJ(mp) && 1404 VFSTOUFS(mp)->softdep_jblocks->jb_suspended)) 1405 kthread_suspend_check(); 1406 ACQUIRE_LOCK(ump); 1407 if ((ump->softdep_flags & (FLUSH_CLEANUP | FLUSH_EXIT)) == 0) 1408 msleep(&ump->softdep_flushtd, LOCK_PTR(ump), PVM, 1409 "sdflush", hz / 2); 1410 ump->softdep_flags &= ~FLUSH_CLEANUP; 1411 /* 1412 * Check to see if we are done and need to exit. 1413 */ 1414 if ((ump->softdep_flags & FLUSH_EXIT) == 0) { 1415 FREE_LOCK(ump); 1416 continue; 1417 } 1418 ump->softdep_flags &= ~FLUSH_EXIT; 1419 FREE_LOCK(ump); 1420 wakeup(&ump->softdep_flags); 1421 if (print_threads) 1422 printf("Stop thread %s: searchfailed %d, did cleanups %d\n", td->td_name, searchfailed, ump->um_softdep->sd_cleanups); 1423 atomic_subtract_int(&stat_flush_threads, 1); 1424 kthread_exit(); 1425 panic("kthread_exit failed\n"); 1426 } 1427} 1428 1429static void 1430worklist_speedup(mp) 1431 struct mount *mp; 1432{ 1433 struct ufsmount *ump; 1434 1435 ump = VFSTOUFS(mp); 1436 LOCK_OWNED(ump); 1437 if ((ump->softdep_flags & (FLUSH_CLEANUP | FLUSH_EXIT)) == 0) 1438 ump->softdep_flags |= FLUSH_CLEANUP; 1439 wakeup(&ump->softdep_flushtd); 1440} 1441 1442static int 1443softdep_speedup(ump) 1444 struct ufsmount *ump; 1445{ 1446 struct ufsmount *altump; 1447 struct mount_softdeps *sdp; 1448 1449 LOCK_OWNED(ump); 1450 worklist_speedup(ump->um_mountp); 1451 bd_speedup(); 1452 /* 1453 * If we have global shortages, then we need other 1454 * filesystems to help with the cleanup. Here we wakeup a 1455 * flusher thread for a filesystem that is over its fair 1456 * share of resources. 1457 */ 1458 if (req_clear_inodedeps || req_clear_remove) { 1459 ACQUIRE_GBLLOCK(&lk); 1460 TAILQ_FOREACH(sdp, &softdepmounts, sd_next) { 1461 if ((altump = sdp->sd_ump) == ump) 1462 continue; 1463 if (((req_clear_inodedeps && 1464 altump->softdep_curdeps[D_INODEDEP] > 1465 max_softdeps / stat_flush_threads) || 1466 (req_clear_remove && 1467 altump->softdep_curdeps[D_DIRREM] > 1468 (max_softdeps / 2) / stat_flush_threads)) && 1469 TRY_ACQUIRE_LOCK(altump)) 1470 break; 1471 } 1472 if (sdp == NULL) { 1473 searchfailed++; 1474 FREE_GBLLOCK(&lk); 1475 } else { 1476 /* 1477 * Move to the end of the list so we pick a 1478 * different one on out next try. 1479 */ 1480 TAILQ_REMOVE(&softdepmounts, sdp, sd_next); 1481 TAILQ_INSERT_TAIL(&softdepmounts, sdp, sd_next); 1482 FREE_GBLLOCK(&lk); 1483 if ((altump->softdep_flags & 1484 (FLUSH_CLEANUP | FLUSH_EXIT)) == 0) 1485 altump->softdep_flags |= FLUSH_CLEANUP; 1486 altump->um_softdep->sd_cleanups++; 1487 wakeup(&altump->softdep_flushtd); 1488 FREE_LOCK(altump); 1489 } 1490 } 1491 return (speedup_syncer()); 1492} 1493 1494/* 1495 * Add an item to the end of the work queue. 1496 * This routine requires that the lock be held. 1497 * This is the only routine that adds items to the list. 1498 * The following routine is the only one that removes items 1499 * and does so in order from first to last. 1500 */ 1501 1502#define WK_HEAD 0x0001 /* Add to HEAD. */ 1503#define WK_NODELAY 0x0002 /* Process immediately. */ 1504 1505static void 1506add_to_worklist(wk, flags) 1507 struct worklist *wk; 1508 int flags; 1509{ 1510 struct ufsmount *ump; 1511 1512 ump = VFSTOUFS(wk->wk_mp); 1513 LOCK_OWNED(ump); 1514 if (wk->wk_state & ONWORKLIST) 1515 panic("add_to_worklist: %s(0x%X) already on list", 1516 TYPENAME(wk->wk_type), wk->wk_state); 1517 wk->wk_state |= ONWORKLIST; 1518 if (ump->softdep_on_worklist == 0) { 1519 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); 1520 ump->softdep_worklist_tail = wk; 1521 } else if (flags & WK_HEAD) { 1522 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); 1523 } else { 1524 LIST_INSERT_AFTER(ump->softdep_worklist_tail, wk, wk_list); 1525 ump->softdep_worklist_tail = wk; 1526 } 1527 ump->softdep_on_worklist += 1; 1528 if (flags & WK_NODELAY) 1529 worklist_speedup(wk->wk_mp); 1530} 1531 1532/* 1533 * Remove the item to be processed. If we are removing the last 1534 * item on the list, we need to recalculate the tail pointer. 1535 */ 1536static void 1537remove_from_worklist(wk) 1538 struct worklist *wk; 1539{ 1540 struct ufsmount *ump; 1541 1542 ump = VFSTOUFS(wk->wk_mp); 1543 if (ump->softdep_worklist_tail == wk) 1544 ump->softdep_worklist_tail = 1545 (struct worklist *)wk->wk_list.le_prev; 1546 WORKLIST_REMOVE(wk); 1547 ump->softdep_on_worklist -= 1; 1548} 1549 1550static void 1551wake_worklist(wk) 1552 struct worklist *wk; 1553{ 1554 if (wk->wk_state & IOWAITING) { 1555 wk->wk_state &= ~IOWAITING; 1556 wakeup(wk); 1557 } 1558} 1559 1560static void 1561wait_worklist(wk, wmesg) 1562 struct worklist *wk; 1563 char *wmesg; 1564{ 1565 struct ufsmount *ump; 1566 1567 ump = VFSTOUFS(wk->wk_mp); 1568 wk->wk_state |= IOWAITING; 1569 msleep(wk, LOCK_PTR(ump), PVM, wmesg, 0); 1570} 1571 1572/* 1573 * Process that runs once per second to handle items in the background queue. 1574 * 1575 * Note that we ensure that everything is done in the order in which they 1576 * appear in the queue. The code below depends on this property to ensure 1577 * that blocks of a file are freed before the inode itself is freed. This 1578 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 1579 * until all the old ones have been purged from the dependency lists. 1580 */ 1581static int 1582softdep_process_worklist(mp, full) 1583 struct mount *mp; 1584 int full; 1585{ 1586 int cnt, matchcnt; 1587 struct ufsmount *ump; 1588 long starttime; 1589 1590 KASSERT(mp != NULL, ("softdep_process_worklist: NULL mp")); 1591 if (MOUNTEDSOFTDEP(mp) == 0) 1592 return (0); 1593 matchcnt = 0; 1594 ump = VFSTOUFS(mp); 1595 ACQUIRE_LOCK(ump); 1596 starttime = time_second; 1597 softdep_process_journal(mp, NULL, full ? MNT_WAIT : 0); 1598 check_clear_deps(mp); 1599 while (ump->softdep_on_worklist > 0) { 1600 if ((cnt = process_worklist_item(mp, 10, LK_NOWAIT)) == 0) 1601 break; 1602 else 1603 matchcnt += cnt; 1604 check_clear_deps(mp); 1605 /* 1606 * We do not generally want to stop for buffer space, but if 1607 * we are really being a buffer hog, we will stop and wait. 1608 */ 1609 if (should_yield()) { 1610 FREE_LOCK(ump); 1611 kern_yield(PRI_USER); 1612 bwillwrite(); 1613 ACQUIRE_LOCK(ump); 1614 } 1615 /* 1616 * Never allow processing to run for more than one 1617 * second. This gives the syncer thread the opportunity 1618 * to pause if appropriate. 1619 */ 1620 if (!full && starttime != time_second) 1621 break; 1622 } 1623 if (full == 0) 1624 journal_unsuspend(ump); 1625 FREE_LOCK(ump); 1626 return (matchcnt); 1627} 1628 1629/* 1630 * Process all removes associated with a vnode if we are running out of 1631 * journal space. Any other process which attempts to flush these will 1632 * be unable as we have the vnodes locked. 1633 */ 1634static void 1635process_removes(vp) 1636 struct vnode *vp; 1637{ 1638 struct inodedep *inodedep; 1639 struct dirrem *dirrem; 1640 struct ufsmount *ump; 1641 struct mount *mp; 1642 ino_t inum; 1643 1644 mp = vp->v_mount; 1645 ump = VFSTOUFS(mp); 1646 LOCK_OWNED(ump); 1647 inum = VTOI(vp)->i_number; 1648 for (;;) { 1649top: 1650 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0) 1651 return; 1652 LIST_FOREACH(dirrem, &inodedep->id_dirremhd, dm_inonext) { 1653 /* 1654 * If another thread is trying to lock this vnode 1655 * it will fail but we must wait for it to do so 1656 * before we can proceed. 1657 */ 1658 if (dirrem->dm_state & INPROGRESS) { 1659 wait_worklist(&dirrem->dm_list, "pwrwait"); 1660 goto top; 1661 } 1662 if ((dirrem->dm_state & (COMPLETE | ONWORKLIST)) == 1663 (COMPLETE | ONWORKLIST)) 1664 break; 1665 } 1666 if (dirrem == NULL) 1667 return; 1668 remove_from_worklist(&dirrem->dm_list); 1669 FREE_LOCK(ump); 1670 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT)) 1671 panic("process_removes: suspended filesystem"); 1672 handle_workitem_remove(dirrem, 0); 1673 vn_finished_secondary_write(mp); 1674 ACQUIRE_LOCK(ump); 1675 } 1676} 1677 1678/* 1679 * Process all truncations associated with a vnode if we are running out 1680 * of journal space. This is called when the vnode lock is already held 1681 * and no other process can clear the truncation. This function returns 1682 * a value greater than zero if it did any work. 1683 */ 1684static void 1685process_truncates(vp) 1686 struct vnode *vp; 1687{ 1688 struct inodedep *inodedep; 1689 struct freeblks *freeblks; 1690 struct ufsmount *ump; 1691 struct mount *mp; 1692 ino_t inum; 1693 int cgwait; 1694 1695 mp = vp->v_mount; 1696 ump = VFSTOUFS(mp); 1697 LOCK_OWNED(ump); 1698 inum = VTOI(vp)->i_number; 1699 for (;;) { 1700 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0) 1701 return; 1702 cgwait = 0; 1703 TAILQ_FOREACH(freeblks, &inodedep->id_freeblklst, fb_next) { 1704 /* Journal entries not yet written. */ 1705 if (!LIST_EMPTY(&freeblks->fb_jblkdephd)) { 1706 jwait(&LIST_FIRST( 1707 &freeblks->fb_jblkdephd)->jb_list, 1708 MNT_WAIT); 1709 break; 1710 } 1711 /* Another thread is executing this item. */ 1712 if (freeblks->fb_state & INPROGRESS) { 1713 wait_worklist(&freeblks->fb_list, "ptrwait"); 1714 break; 1715 } 1716 /* Freeblks is waiting on a inode write. */ 1717 if ((freeblks->fb_state & COMPLETE) == 0) { 1718 FREE_LOCK(ump); 1719 ffs_update(vp, 1); 1720 ACQUIRE_LOCK(ump); 1721 break; 1722 } 1723 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST)) == 1724 (ALLCOMPLETE | ONWORKLIST)) { 1725 remove_from_worklist(&freeblks->fb_list); 1726 freeblks->fb_state |= INPROGRESS; 1727 FREE_LOCK(ump); 1728 if (vn_start_secondary_write(NULL, &mp, 1729 V_NOWAIT)) 1730 panic("process_truncates: " 1731 "suspended filesystem"); 1732 handle_workitem_freeblocks(freeblks, 0); 1733 vn_finished_secondary_write(mp); 1734 ACQUIRE_LOCK(ump); 1735 break; 1736 } 1737 if (freeblks->fb_cgwait) 1738 cgwait++; 1739 } 1740 if (cgwait) { 1741 FREE_LOCK(ump); 1742 sync_cgs(mp, MNT_WAIT); 1743 ffs_sync_snap(mp, MNT_WAIT); 1744 ACQUIRE_LOCK(ump); 1745 continue; 1746 } 1747 if (freeblks == NULL) 1748 break; 1749 } 1750 return; 1751} 1752 1753/* 1754 * Process one item on the worklist. 1755 */ 1756static int 1757process_worklist_item(mp, target, flags) 1758 struct mount *mp; 1759 int target; 1760 int flags; 1761{ 1762 struct worklist sentinel; 1763 struct worklist *wk; 1764 struct ufsmount *ump; 1765 int matchcnt; 1766 int error; 1767 1768 KASSERT(mp != NULL, ("process_worklist_item: NULL mp")); 1769 /* 1770 * If we are being called because of a process doing a 1771 * copy-on-write, then it is not safe to write as we may 1772 * recurse into the copy-on-write routine. 1773 */ 1774 if (curthread->td_pflags & TDP_COWINPROGRESS) 1775 return (-1); 1776 PHOLD(curproc); /* Don't let the stack go away. */ 1777 ump = VFSTOUFS(mp); 1778 LOCK_OWNED(ump); 1779 matchcnt = 0; 1780 sentinel.wk_mp = NULL; 1781 sentinel.wk_type = D_SENTINEL; 1782 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, &sentinel, wk_list); 1783 for (wk = LIST_NEXT(&sentinel, wk_list); wk != NULL; 1784 wk = LIST_NEXT(&sentinel, wk_list)) { 1785 if (wk->wk_type == D_SENTINEL) { 1786 LIST_REMOVE(&sentinel, wk_list); 1787 LIST_INSERT_AFTER(wk, &sentinel, wk_list); 1788 continue; 1789 } 1790 if (wk->wk_state & INPROGRESS) 1791 panic("process_worklist_item: %p already in progress.", 1792 wk); 1793 wk->wk_state |= INPROGRESS; 1794 remove_from_worklist(wk); 1795 FREE_LOCK(ump); 1796 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT)) 1797 panic("process_worklist_item: suspended filesystem"); 1798 switch (wk->wk_type) { 1799 case D_DIRREM: 1800 /* removal of a directory entry */ 1801 error = handle_workitem_remove(WK_DIRREM(wk), flags); 1802 break; 1803 1804 case D_FREEBLKS: 1805 /* releasing blocks and/or fragments from a file */ 1806 error = handle_workitem_freeblocks(WK_FREEBLKS(wk), 1807 flags); 1808 break; 1809 1810 case D_FREEFRAG: 1811 /* releasing a fragment when replaced as a file grows */ 1812 handle_workitem_freefrag(WK_FREEFRAG(wk)); 1813 error = 0; 1814 break; 1815 1816 case D_FREEFILE: 1817 /* releasing an inode when its link count drops to 0 */ 1818 handle_workitem_freefile(WK_FREEFILE(wk)); 1819 error = 0; 1820 break; 1821 1822 default: 1823 panic("%s_process_worklist: Unknown type %s", 1824 "softdep", TYPENAME(wk->wk_type)); 1825 /* NOTREACHED */ 1826 } 1827 vn_finished_secondary_write(mp); 1828 ACQUIRE_LOCK(ump); 1829 if (error == 0) { 1830 if (++matchcnt == target) 1831 break; 1832 continue; 1833 } 1834 /* 1835 * We have to retry the worklist item later. Wake up any 1836 * waiters who may be able to complete it immediately and 1837 * add the item back to the head so we don't try to execute 1838 * it again. 1839 */ 1840 wk->wk_state &= ~INPROGRESS; 1841 wake_worklist(wk); 1842 add_to_worklist(wk, WK_HEAD); 1843 } 1844 /* Sentinal could've become the tail from remove_from_worklist. */ 1845 if (ump->softdep_worklist_tail == &sentinel) 1846 ump->softdep_worklist_tail = 1847 (struct worklist *)sentinel.wk_list.le_prev; 1848 LIST_REMOVE(&sentinel, wk_list); 1849 PRELE(curproc); 1850 return (matchcnt); 1851} 1852 1853/* 1854 * Move dependencies from one buffer to another. 1855 */ 1856int 1857softdep_move_dependencies(oldbp, newbp) 1858 struct buf *oldbp; 1859 struct buf *newbp; 1860{ 1861 struct worklist *wk, *wktail; 1862 struct ufsmount *ump; 1863 int dirty; 1864 1865 if ((wk = LIST_FIRST(&oldbp->b_dep)) == NULL) 1866 return (0); 1867 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0, 1868 ("softdep_move_dependencies called on non-softdep filesystem")); 1869 dirty = 0; 1870 wktail = NULL; 1871 ump = VFSTOUFS(wk->wk_mp); 1872 ACQUIRE_LOCK(ump); 1873 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 1874 LIST_REMOVE(wk, wk_list); 1875 if (wk->wk_type == D_BMSAFEMAP && 1876 bmsafemap_backgroundwrite(WK_BMSAFEMAP(wk), newbp)) 1877 dirty = 1; 1878 if (wktail == NULL) 1879 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 1880 else 1881 LIST_INSERT_AFTER(wktail, wk, wk_list); 1882 wktail = wk; 1883 } 1884 FREE_LOCK(ump); 1885 1886 return (dirty); 1887} 1888 1889/* 1890 * Purge the work list of all items associated with a particular mount point. 1891 */ 1892int 1893softdep_flushworklist(oldmnt, countp, td) 1894 struct mount *oldmnt; 1895 int *countp; 1896 struct thread *td; 1897{ 1898 struct vnode *devvp; 1899 struct ufsmount *ump; 1900 int count, error; 1901 1902 /* 1903 * Alternately flush the block device associated with the mount 1904 * point and process any dependencies that the flushing 1905 * creates. We continue until no more worklist dependencies 1906 * are found. 1907 */ 1908 *countp = 0; 1909 error = 0; 1910 ump = VFSTOUFS(oldmnt); 1911 devvp = ump->um_devvp; 1912 while ((count = softdep_process_worklist(oldmnt, 1)) > 0) { 1913 *countp += count; 1914 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1915 error = VOP_FSYNC(devvp, MNT_WAIT, td); 1916 VOP_UNLOCK(devvp, 0); 1917 if (error != 0) 1918 break; 1919 } 1920 return (error); 1921} 1922 1923#define SU_WAITIDLE_RETRIES 20 1924static int 1925softdep_waitidle(struct mount *mp, int flags __unused) 1926{ 1927 struct ufsmount *ump; 1928 struct vnode *devvp; 1929 struct thread *td; 1930 int error, i; 1931 1932 ump = VFSTOUFS(mp); 1933 devvp = ump->um_devvp; 1934 td = curthread; 1935 error = 0; 1936 ACQUIRE_LOCK(ump); 1937 for (i = 0; i < SU_WAITIDLE_RETRIES && ump->softdep_deps != 0; i++) { 1938 ump->softdep_req = 1; 1939 KASSERT((flags & FORCECLOSE) == 0 || 1940 ump->softdep_on_worklist == 0, 1941 ("softdep_waitidle: work added after flush")); 1942 msleep(&ump->softdep_deps, LOCK_PTR(ump), PVM | PDROP, 1943 "softdeps", 10 * hz); 1944 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1945 error = VOP_FSYNC(devvp, MNT_WAIT, td); 1946 VOP_UNLOCK(devvp, 0); 1947 ACQUIRE_LOCK(ump); 1948 if (error != 0) 1949 break; 1950 } 1951 ump->softdep_req = 0; 1952 if (i == SU_WAITIDLE_RETRIES && error == 0 && ump->softdep_deps != 0) { 1953 error = EBUSY; 1954 printf("softdep_waitidle: Failed to flush worklist for %p\n", 1955 mp); 1956 } 1957 FREE_LOCK(ump); 1958 return (error); 1959} 1960 1961/* 1962 * Flush all vnodes and worklist items associated with a specified mount point. 1963 */ 1964int 1965softdep_flushfiles(oldmnt, flags, td) 1966 struct mount *oldmnt; 1967 int flags; 1968 struct thread *td; 1969{ 1970#ifdef QUOTA 1971 struct ufsmount *ump; 1972 int i; 1973#endif 1974 int error, early, depcount, loopcnt, retry_flush_count, retry; 1975 int morework; 1976 1977 KASSERT(MOUNTEDSOFTDEP(oldmnt) != 0, 1978 ("softdep_flushfiles called on non-softdep filesystem")); 1979 loopcnt = 10; 1980 retry_flush_count = 3; 1981retry_flush: 1982 error = 0; 1983 1984 /* 1985 * Alternately flush the vnodes associated with the mount 1986 * point and process any dependencies that the flushing 1987 * creates. In theory, this loop can happen at most twice, 1988 * but we give it a few extra just to be sure. 1989 */ 1990 for (; loopcnt > 0; loopcnt--) { 1991 /* 1992 * Do another flush in case any vnodes were brought in 1993 * as part of the cleanup operations. 1994 */ 1995 early = retry_flush_count == 1 || (oldmnt->mnt_kern_flag & 1996 MNTK_UNMOUNT) == 0 ? 0 : EARLYFLUSH; 1997 if ((error = ffs_flushfiles(oldmnt, flags | early, td)) != 0) 1998 break; 1999 if ((error = softdep_flushworklist(oldmnt, &depcount, td)) != 0 || 2000 depcount == 0) 2001 break; 2002 } 2003 /* 2004 * If we are unmounting then it is an error to fail. If we 2005 * are simply trying to downgrade to read-only, then filesystem 2006 * activity can keep us busy forever, so we just fail with EBUSY. 2007 */ 2008 if (loopcnt == 0) { 2009 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 2010 panic("softdep_flushfiles: looping"); 2011 error = EBUSY; 2012 } 2013 if (!error) 2014 error = softdep_waitidle(oldmnt, flags); 2015 if (!error) { 2016 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) { 2017 retry = 0; 2018 MNT_ILOCK(oldmnt); 2019 morework = oldmnt->mnt_nvnodelistsize > 0; 2020#ifdef QUOTA 2021 ump = VFSTOUFS(oldmnt); 2022 UFS_LOCK(ump); 2023 for (i = 0; i < MAXQUOTAS; i++) { 2024 if (ump->um_quotas[i] != NULLVP) 2025 morework = 1; 2026 } 2027 UFS_UNLOCK(ump); 2028#endif 2029 if (morework) { 2030 if (--retry_flush_count > 0) { 2031 retry = 1; 2032 loopcnt = 3; 2033 } else 2034 error = EBUSY; 2035 } 2036 MNT_IUNLOCK(oldmnt); 2037 if (retry) 2038 goto retry_flush; 2039 } 2040 } 2041 return (error); 2042} 2043 2044/* 2045 * Structure hashing. 2046 * 2047 * There are four types of structures that can be looked up: 2048 * 1) pagedep structures identified by mount point, inode number, 2049 * and logical block. 2050 * 2) inodedep structures identified by mount point and inode number. 2051 * 3) newblk structures identified by mount point and 2052 * physical block number. 2053 * 4) bmsafemap structures identified by mount point and 2054 * cylinder group number. 2055 * 2056 * The "pagedep" and "inodedep" dependency structures are hashed 2057 * separately from the file blocks and inodes to which they correspond. 2058 * This separation helps when the in-memory copy of an inode or 2059 * file block must be replaced. It also obviates the need to access 2060 * an inode or file page when simply updating (or de-allocating) 2061 * dependency structures. Lookup of newblk structures is needed to 2062 * find newly allocated blocks when trying to associate them with 2063 * their allocdirect or allocindir structure. 2064 * 2065 * The lookup routines optionally create and hash a new instance when 2066 * an existing entry is not found. The bmsafemap lookup routine always 2067 * allocates a new structure if an existing one is not found. 2068 */ 2069#define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 2070 2071/* 2072 * Structures and routines associated with pagedep caching. 2073 */ 2074#define PAGEDEP_HASH(ump, inum, lbn) \ 2075 (&(ump)->pagedep_hashtbl[((inum) + (lbn)) & (ump)->pagedep_hash_size]) 2076 2077static int 2078pagedep_find(pagedephd, ino, lbn, pagedeppp) 2079 struct pagedep_hashhead *pagedephd; 2080 ino_t ino; 2081 ufs_lbn_t lbn; 2082 struct pagedep **pagedeppp; 2083{ 2084 struct pagedep *pagedep; 2085 2086 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 2087 if (ino == pagedep->pd_ino && lbn == pagedep->pd_lbn) { 2088 *pagedeppp = pagedep; 2089 return (1); 2090 } 2091 } 2092 *pagedeppp = NULL; 2093 return (0); 2094} 2095/* 2096 * Look up a pagedep. Return 1 if found, 0 otherwise. 2097 * If not found, allocate if DEPALLOC flag is passed. 2098 * Found or allocated entry is returned in pagedeppp. 2099 */ 2100static int 2101pagedep_lookup(mp, bp, ino, lbn, flags, pagedeppp) 2102 struct mount *mp; 2103 struct buf *bp; 2104 ino_t ino; 2105 ufs_lbn_t lbn; 2106 int flags; 2107 struct pagedep **pagedeppp; 2108{ 2109 struct pagedep *pagedep; 2110 struct pagedep_hashhead *pagedephd; 2111 struct worklist *wk; 2112 struct ufsmount *ump; 2113 int ret; 2114 int i; 2115 2116 ump = VFSTOUFS(mp); 2117 LOCK_OWNED(ump); 2118 if (bp) { 2119 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 2120 if (wk->wk_type == D_PAGEDEP) { 2121 *pagedeppp = WK_PAGEDEP(wk); 2122 return (1); 2123 } 2124 } 2125 } 2126 pagedephd = PAGEDEP_HASH(ump, ino, lbn); 2127 ret = pagedep_find(pagedephd, ino, lbn, pagedeppp); 2128 if (ret) { 2129 if (((*pagedeppp)->pd_state & ONWORKLIST) == 0 && bp) 2130 WORKLIST_INSERT(&bp->b_dep, &(*pagedeppp)->pd_list); 2131 return (1); 2132 } 2133 if ((flags & DEPALLOC) == 0) 2134 return (0); 2135 FREE_LOCK(ump); 2136 pagedep = malloc(sizeof(struct pagedep), 2137 M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO); 2138 workitem_alloc(&pagedep->pd_list, D_PAGEDEP, mp); 2139 ACQUIRE_LOCK(ump); 2140 ret = pagedep_find(pagedephd, ino, lbn, pagedeppp); 2141 if (*pagedeppp) { 2142 /* 2143 * This should never happen since we only create pagedeps 2144 * with the vnode lock held. Could be an assert. 2145 */ 2146 WORKITEM_FREE(pagedep, D_PAGEDEP); 2147 return (ret); 2148 } 2149 pagedep->pd_ino = ino; 2150 pagedep->pd_lbn = lbn; 2151 LIST_INIT(&pagedep->pd_dirremhd); 2152 LIST_INIT(&pagedep->pd_pendinghd); 2153 for (i = 0; i < DAHASHSZ; i++) 2154 LIST_INIT(&pagedep->pd_diraddhd[i]); 2155 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 2156 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2157 *pagedeppp = pagedep; 2158 return (0); 2159} 2160 2161/* 2162 * Structures and routines associated with inodedep caching. 2163 */ 2164#define INODEDEP_HASH(ump, inum) \ 2165 (&(ump)->inodedep_hashtbl[(inum) & (ump)->inodedep_hash_size]) 2166 2167static int 2168inodedep_find(inodedephd, inum, inodedeppp) 2169 struct inodedep_hashhead *inodedephd; 2170 ino_t inum; 2171 struct inodedep **inodedeppp; 2172{ 2173 struct inodedep *inodedep; 2174 2175 LIST_FOREACH(inodedep, inodedephd, id_hash) 2176 if (inum == inodedep->id_ino) 2177 break; 2178 if (inodedep) { 2179 *inodedeppp = inodedep; 2180 return (1); 2181 } 2182 *inodedeppp = NULL; 2183 2184 return (0); 2185} 2186/* 2187 * Look up an inodedep. Return 1 if found, 0 if not found. 2188 * If not found, allocate if DEPALLOC flag is passed. 2189 * Found or allocated entry is returned in inodedeppp. 2190 */ 2191static int 2192inodedep_lookup(mp, inum, flags, inodedeppp) 2193 struct mount *mp; 2194 ino_t inum; 2195 int flags; 2196 struct inodedep **inodedeppp; 2197{ 2198 struct inodedep *inodedep; 2199 struct inodedep_hashhead *inodedephd; 2200 struct ufsmount *ump; 2201 struct fs *fs; 2202 2203 ump = VFSTOUFS(mp); 2204 LOCK_OWNED(ump); 2205 fs = ump->um_fs; 2206 inodedephd = INODEDEP_HASH(ump, inum); 2207 2208 if (inodedep_find(inodedephd, inum, inodedeppp)) 2209 return (1); 2210 if ((flags & DEPALLOC) == 0) 2211 return (0); 2212 /* 2213 * If the system is over its limit and our filesystem is 2214 * responsible for more than our share of that usage and 2215 * we are not in a rush, request some inodedep cleanup. 2216 */ 2217 if (softdep_excess_items(ump, D_INODEDEP)) 2218 schedule_cleanup(mp); 2219 else 2220 FREE_LOCK(ump); 2221 inodedep = malloc(sizeof(struct inodedep), 2222 M_INODEDEP, M_SOFTDEP_FLAGS); 2223 workitem_alloc(&inodedep->id_list, D_INODEDEP, mp); 2224 ACQUIRE_LOCK(ump); 2225 if (inodedep_find(inodedephd, inum, inodedeppp)) { 2226 WORKITEM_FREE(inodedep, D_INODEDEP); 2227 return (1); 2228 } 2229 inodedep->id_fs = fs; 2230 inodedep->id_ino = inum; 2231 inodedep->id_state = ALLCOMPLETE; 2232 inodedep->id_nlinkdelta = 0; 2233 inodedep->id_savedino1 = NULL; 2234 inodedep->id_savedsize = -1; 2235 inodedep->id_savedextsize = -1; 2236 inodedep->id_savednlink = -1; 2237 inodedep->id_bmsafemap = NULL; 2238 inodedep->id_mkdiradd = NULL; 2239 LIST_INIT(&inodedep->id_dirremhd); 2240 LIST_INIT(&inodedep->id_pendinghd); 2241 LIST_INIT(&inodedep->id_inowait); 2242 LIST_INIT(&inodedep->id_bufwait); 2243 TAILQ_INIT(&inodedep->id_inoreflst); 2244 TAILQ_INIT(&inodedep->id_inoupdt); 2245 TAILQ_INIT(&inodedep->id_newinoupdt); 2246 TAILQ_INIT(&inodedep->id_extupdt); 2247 TAILQ_INIT(&inodedep->id_newextupdt); 2248 TAILQ_INIT(&inodedep->id_freeblklst); 2249 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 2250 *inodedeppp = inodedep; 2251 return (0); 2252} 2253 2254/* 2255 * Structures and routines associated with newblk caching. 2256 */ 2257#define NEWBLK_HASH(ump, inum) \ 2258 (&(ump)->newblk_hashtbl[(inum) & (ump)->newblk_hash_size]) 2259 2260static int 2261newblk_find(newblkhd, newblkno, flags, newblkpp) 2262 struct newblk_hashhead *newblkhd; 2263 ufs2_daddr_t newblkno; 2264 int flags; 2265 struct newblk **newblkpp; 2266{ 2267 struct newblk *newblk; 2268 2269 LIST_FOREACH(newblk, newblkhd, nb_hash) { 2270 if (newblkno != newblk->nb_newblkno) 2271 continue; 2272 /* 2273 * If we're creating a new dependency don't match those that 2274 * have already been converted to allocdirects. This is for 2275 * a frag extend. 2276 */ 2277 if ((flags & DEPALLOC) && newblk->nb_list.wk_type != D_NEWBLK) 2278 continue; 2279 break; 2280 } 2281 if (newblk) { 2282 *newblkpp = newblk; 2283 return (1); 2284 } 2285 *newblkpp = NULL; 2286 return (0); 2287} 2288 2289/* 2290 * Look up a newblk. Return 1 if found, 0 if not found. 2291 * If not found, allocate if DEPALLOC flag is passed. 2292 * Found or allocated entry is returned in newblkpp. 2293 */ 2294static int 2295newblk_lookup(mp, newblkno, flags, newblkpp) 2296 struct mount *mp; 2297 ufs2_daddr_t newblkno; 2298 int flags; 2299 struct newblk **newblkpp; 2300{ 2301 struct newblk *newblk; 2302 struct newblk_hashhead *newblkhd; 2303 struct ufsmount *ump; 2304 2305 ump = VFSTOUFS(mp); 2306 LOCK_OWNED(ump); 2307 newblkhd = NEWBLK_HASH(ump, newblkno); 2308 if (newblk_find(newblkhd, newblkno, flags, newblkpp)) 2309 return (1); 2310 if ((flags & DEPALLOC) == 0) 2311 return (0); 2312 if (softdep_excess_items(ump, D_NEWBLK) || 2313 softdep_excess_items(ump, D_ALLOCDIRECT) || 2314 softdep_excess_items(ump, D_ALLOCINDIR)) 2315 schedule_cleanup(mp); 2316 else 2317 FREE_LOCK(ump); 2318 newblk = malloc(sizeof(union allblk), M_NEWBLK, 2319 M_SOFTDEP_FLAGS | M_ZERO); 2320 workitem_alloc(&newblk->nb_list, D_NEWBLK, mp); 2321 ACQUIRE_LOCK(ump); 2322 if (newblk_find(newblkhd, newblkno, flags, newblkpp)) { 2323 WORKITEM_FREE(newblk, D_NEWBLK); 2324 return (1); 2325 } 2326 newblk->nb_freefrag = NULL; 2327 LIST_INIT(&newblk->nb_indirdeps); 2328 LIST_INIT(&newblk->nb_newdirblk); 2329 LIST_INIT(&newblk->nb_jwork); 2330 newblk->nb_state = ATTACHED; 2331 newblk->nb_newblkno = newblkno; 2332 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 2333 *newblkpp = newblk; 2334 return (0); 2335} 2336 2337/* 2338 * Structures and routines associated with freed indirect block caching. 2339 */ 2340#define INDIR_HASH(ump, blkno) \ 2341 (&(ump)->indir_hashtbl[(blkno) & (ump)->indir_hash_size]) 2342 2343/* 2344 * Lookup an indirect block in the indir hash table. The freework is 2345 * removed and potentially freed. The caller must do a blocking journal 2346 * write before writing to the blkno. 2347 */ 2348static int 2349indirblk_lookup(mp, blkno) 2350 struct mount *mp; 2351 ufs2_daddr_t blkno; 2352{ 2353 struct freework *freework; 2354 struct indir_hashhead *wkhd; 2355 struct ufsmount *ump; 2356 2357 ump = VFSTOUFS(mp); 2358 wkhd = INDIR_HASH(ump, blkno); 2359 TAILQ_FOREACH(freework, wkhd, fw_next) { 2360 if (freework->fw_blkno != blkno) 2361 continue; 2362 indirblk_remove(freework); 2363 return (1); 2364 } 2365 return (0); 2366} 2367 2368/* 2369 * Insert an indirect block represented by freework into the indirblk 2370 * hash table so that it may prevent the block from being re-used prior 2371 * to the journal being written. 2372 */ 2373static void 2374indirblk_insert(freework) 2375 struct freework *freework; 2376{ 2377 struct jblocks *jblocks; 2378 struct jseg *jseg; 2379 struct ufsmount *ump; 2380 2381 ump = VFSTOUFS(freework->fw_list.wk_mp); 2382 jblocks = ump->softdep_jblocks; 2383 jseg = TAILQ_LAST(&jblocks->jb_segs, jseglst); 2384 if (jseg == NULL) 2385 return; 2386 2387 LIST_INSERT_HEAD(&jseg->js_indirs, freework, fw_segs); 2388 TAILQ_INSERT_HEAD(INDIR_HASH(ump, freework->fw_blkno), freework, 2389 fw_next); 2390 freework->fw_state &= ~DEPCOMPLETE; 2391} 2392 2393static void 2394indirblk_remove(freework) 2395 struct freework *freework; 2396{ 2397 struct ufsmount *ump; 2398 2399 ump = VFSTOUFS(freework->fw_list.wk_mp); 2400 LIST_REMOVE(freework, fw_segs); 2401 TAILQ_REMOVE(INDIR_HASH(ump, freework->fw_blkno), freework, fw_next); 2402 freework->fw_state |= DEPCOMPLETE; 2403 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE) 2404 WORKITEM_FREE(freework, D_FREEWORK); 2405} 2406 2407/* 2408 * Executed during filesystem system initialization before 2409 * mounting any filesystems. 2410 */ 2411void 2412softdep_initialize() 2413{ 2414 2415 TAILQ_INIT(&softdepmounts); 2416#ifdef __LP64__ 2417 max_softdeps = desiredvnodes * 4; 2418#else 2419 max_softdeps = desiredvnodes * 2; 2420#endif 2421 2422 /* initialise bioops hack */ 2423 bioops.io_start = softdep_disk_io_initiation; 2424 bioops.io_complete = softdep_disk_write_complete; 2425 bioops.io_deallocate = softdep_deallocate_dependencies; 2426 bioops.io_countdeps = softdep_count_dependencies; 2427 softdep_ast_cleanup = softdep_ast_cleanup_proc; 2428 2429 /* Initialize the callout with an mtx. */ 2430 callout_init_mtx(&softdep_callout, &lk, 0); 2431} 2432 2433/* 2434 * Executed after all filesystems have been unmounted during 2435 * filesystem module unload. 2436 */ 2437void 2438softdep_uninitialize() 2439{ 2440 2441 /* clear bioops hack */ 2442 bioops.io_start = NULL; 2443 bioops.io_complete = NULL; 2444 bioops.io_deallocate = NULL; 2445 bioops.io_countdeps = NULL; 2446 softdep_ast_cleanup = NULL; 2447 2448 callout_drain(&softdep_callout); 2449} 2450 2451/* 2452 * Called at mount time to notify the dependency code that a 2453 * filesystem wishes to use it. 2454 */ 2455int 2456softdep_mount(devvp, mp, fs, cred) 2457 struct vnode *devvp; 2458 struct mount *mp; 2459 struct fs *fs; 2460 struct ucred *cred; 2461{ 2462 struct csum_total cstotal; 2463 struct mount_softdeps *sdp; 2464 struct ufsmount *ump; 2465 struct cg *cgp; 2466 struct buf *bp; 2467 u_int cyl, i; 2468 int error; 2469 2470 sdp = malloc(sizeof(struct mount_softdeps), M_MOUNTDATA, 2471 M_WAITOK | M_ZERO); 2472 MNT_ILOCK(mp); 2473 mp->mnt_flag = (mp->mnt_flag & ~MNT_ASYNC) | MNT_SOFTDEP; 2474 if ((mp->mnt_kern_flag & MNTK_SOFTDEP) == 0) { 2475 mp->mnt_kern_flag = (mp->mnt_kern_flag & ~MNTK_ASYNC) | 2476 MNTK_SOFTDEP | MNTK_NOASYNC; 2477 } 2478 ump = VFSTOUFS(mp); 2479 ump->um_softdep = sdp; 2480 MNT_IUNLOCK(mp); 2481 rw_init(LOCK_PTR(ump), "per-fs softdep"); 2482 sdp->sd_ump = ump; 2483 LIST_INIT(&ump->softdep_workitem_pending); 2484 LIST_INIT(&ump->softdep_journal_pending); 2485 TAILQ_INIT(&ump->softdep_unlinked); 2486 LIST_INIT(&ump->softdep_dirtycg); 2487 ump->softdep_worklist_tail = NULL; 2488 ump->softdep_on_worklist = 0; 2489 ump->softdep_deps = 0; 2490 LIST_INIT(&ump->softdep_mkdirlisthd); 2491 ump->pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 2492 &ump->pagedep_hash_size); 2493 ump->pagedep_nextclean = 0; 2494 ump->inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, 2495 &ump->inodedep_hash_size); 2496 ump->inodedep_nextclean = 0; 2497 ump->newblk_hashtbl = hashinit(max_softdeps / 2, M_NEWBLK, 2498 &ump->newblk_hash_size); 2499 ump->bmsafemap_hashtbl = hashinit(1024, M_BMSAFEMAP, 2500 &ump->bmsafemap_hash_size); 2501 i = 1 << (ffs(desiredvnodes / 10) - 1); 2502 ump->indir_hashtbl = malloc(i * sizeof(struct indir_hashhead), 2503 M_FREEWORK, M_WAITOK); 2504 ump->indir_hash_size = i - 1; 2505 for (i = 0; i <= ump->indir_hash_size; i++) 2506 TAILQ_INIT(&ump->indir_hashtbl[i]); 2507 ACQUIRE_GBLLOCK(&lk); 2508 TAILQ_INSERT_TAIL(&softdepmounts, sdp, sd_next); 2509 FREE_GBLLOCK(&lk); 2510 if ((fs->fs_flags & FS_SUJ) && 2511 (error = journal_mount(mp, fs, cred)) != 0) { 2512 printf("Failed to start journal: %d\n", error); 2513 softdep_unmount(mp); 2514 return (error); 2515 } 2516 /* 2517 * Start our flushing thread in the bufdaemon process. 2518 */ 2519 ACQUIRE_LOCK(ump); 2520 ump->softdep_flags |= FLUSH_STARTING; 2521 FREE_LOCK(ump); 2522 kproc_kthread_add(&softdep_flush, mp, &bufdaemonproc, 2523 &ump->softdep_flushtd, 0, 0, "softdepflush", "%s worker", 2524 mp->mnt_stat.f_mntonname); 2525 ACQUIRE_LOCK(ump); 2526 while ((ump->softdep_flags & FLUSH_STARTING) != 0) { 2527 msleep(&ump->softdep_flushtd, LOCK_PTR(ump), PVM, "sdstart", 2528 hz / 2); 2529 } 2530 FREE_LOCK(ump); 2531 /* 2532 * When doing soft updates, the counters in the 2533 * superblock may have gotten out of sync. Recomputation 2534 * can take a long time and can be deferred for background 2535 * fsck. However, the old behavior of scanning the cylinder 2536 * groups and recalculating them at mount time is available 2537 * by setting vfs.ffs.compute_summary_at_mount to one. 2538 */ 2539 if (compute_summary_at_mount == 0 || fs->fs_clean != 0) 2540 return (0); 2541 bzero(&cstotal, sizeof cstotal); 2542 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 2543 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 2544 fs->fs_cgsize, cred, &bp)) != 0) { 2545 brelse(bp); 2546 softdep_unmount(mp); 2547 return (error); 2548 } 2549 cgp = (struct cg *)bp->b_data; 2550 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 2551 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 2552 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 2553 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 2554 fs->fs_cs(fs, cyl) = cgp->cg_cs; 2555 brelse(bp); 2556 } 2557#ifdef DEBUG 2558 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 2559 printf("%s: superblock summary recomputed\n", fs->fs_fsmnt); 2560#endif 2561 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 2562 return (0); 2563} 2564 2565void 2566softdep_unmount(mp) 2567 struct mount *mp; 2568{ 2569 struct ufsmount *ump; 2570#ifdef INVARIANTS 2571 int i; 2572#endif 2573 2574 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 2575 ("softdep_unmount called on non-softdep filesystem")); 2576 ump = VFSTOUFS(mp); 2577 MNT_ILOCK(mp); 2578 mp->mnt_flag &= ~MNT_SOFTDEP; 2579 if (MOUNTEDSUJ(mp) == 0) { 2580 MNT_IUNLOCK(mp); 2581 } else { 2582 mp->mnt_flag &= ~MNT_SUJ; 2583 MNT_IUNLOCK(mp); 2584 journal_unmount(ump); 2585 } 2586 /* 2587 * Shut down our flushing thread. Check for NULL is if 2588 * softdep_mount errors out before the thread has been created. 2589 */ 2590 if (ump->softdep_flushtd != NULL) { 2591 ACQUIRE_LOCK(ump); 2592 ump->softdep_flags |= FLUSH_EXIT; 2593 wakeup(&ump->softdep_flushtd); 2594 msleep(&ump->softdep_flags, LOCK_PTR(ump), PVM | PDROP, 2595 "sdwait", 0); 2596 KASSERT((ump->softdep_flags & FLUSH_EXIT) == 0, 2597 ("Thread shutdown failed")); 2598 } 2599 /* 2600 * Free up our resources. 2601 */ 2602 ACQUIRE_GBLLOCK(&lk); 2603 TAILQ_REMOVE(&softdepmounts, ump->um_softdep, sd_next); 2604 FREE_GBLLOCK(&lk); 2605 rw_destroy(LOCK_PTR(ump)); 2606 hashdestroy(ump->pagedep_hashtbl, M_PAGEDEP, ump->pagedep_hash_size); 2607 hashdestroy(ump->inodedep_hashtbl, M_INODEDEP, ump->inodedep_hash_size); 2608 hashdestroy(ump->newblk_hashtbl, M_NEWBLK, ump->newblk_hash_size); 2609 hashdestroy(ump->bmsafemap_hashtbl, M_BMSAFEMAP, 2610 ump->bmsafemap_hash_size); 2611 free(ump->indir_hashtbl, M_FREEWORK); 2612#ifdef INVARIANTS 2613 for (i = 0; i <= D_LAST; i++) 2614 KASSERT(ump->softdep_curdeps[i] == 0, 2615 ("Unmount %s: Dep type %s != 0 (%ld)", ump->um_fs->fs_fsmnt, 2616 TYPENAME(i), ump->softdep_curdeps[i])); 2617#endif 2618 free(ump->um_softdep, M_MOUNTDATA); 2619} 2620 2621static struct jblocks * 2622jblocks_create(void) 2623{ 2624 struct jblocks *jblocks; 2625 2626 jblocks = malloc(sizeof(*jblocks), M_JBLOCKS, M_WAITOK | M_ZERO); 2627 TAILQ_INIT(&jblocks->jb_segs); 2628 jblocks->jb_avail = 10; 2629 jblocks->jb_extent = malloc(sizeof(struct jextent) * jblocks->jb_avail, 2630 M_JBLOCKS, M_WAITOK | M_ZERO); 2631 2632 return (jblocks); 2633} 2634 2635static ufs2_daddr_t 2636jblocks_alloc(jblocks, bytes, actual) 2637 struct jblocks *jblocks; 2638 int bytes; 2639 int *actual; 2640{ 2641 ufs2_daddr_t daddr; 2642 struct jextent *jext; 2643 int freecnt; 2644 int blocks; 2645 2646 blocks = bytes / DEV_BSIZE; 2647 jext = &jblocks->jb_extent[jblocks->jb_head]; 2648 freecnt = jext->je_blocks - jblocks->jb_off; 2649 if (freecnt == 0) { 2650 jblocks->jb_off = 0; 2651 if (++jblocks->jb_head > jblocks->jb_used) 2652 jblocks->jb_head = 0; 2653 jext = &jblocks->jb_extent[jblocks->jb_head]; 2654 freecnt = jext->je_blocks; 2655 } 2656 if (freecnt > blocks) 2657 freecnt = blocks; 2658 *actual = freecnt * DEV_BSIZE; 2659 daddr = jext->je_daddr + jblocks->jb_off; 2660 jblocks->jb_off += freecnt; 2661 jblocks->jb_free -= freecnt; 2662 2663 return (daddr); 2664} 2665 2666static void 2667jblocks_free(jblocks, mp, bytes) 2668 struct jblocks *jblocks; 2669 struct mount *mp; 2670 int bytes; 2671{ 2672 2673 LOCK_OWNED(VFSTOUFS(mp)); 2674 jblocks->jb_free += bytes / DEV_BSIZE; 2675 if (jblocks->jb_suspended) 2676 worklist_speedup(mp); 2677 wakeup(jblocks); 2678} 2679 2680static void 2681jblocks_destroy(jblocks) 2682 struct jblocks *jblocks; 2683{ 2684 2685 if (jblocks->jb_extent) 2686 free(jblocks->jb_extent, M_JBLOCKS); 2687 free(jblocks, M_JBLOCKS); 2688} 2689 2690static void 2691jblocks_add(jblocks, daddr, blocks) 2692 struct jblocks *jblocks; 2693 ufs2_daddr_t daddr; 2694 int blocks; 2695{ 2696 struct jextent *jext; 2697 2698 jblocks->jb_blocks += blocks; 2699 jblocks->jb_free += blocks; 2700 jext = &jblocks->jb_extent[jblocks->jb_used]; 2701 /* Adding the first block. */ 2702 if (jext->je_daddr == 0) { 2703 jext->je_daddr = daddr; 2704 jext->je_blocks = blocks; 2705 return; 2706 } 2707 /* Extending the last extent. */ 2708 if (jext->je_daddr + jext->je_blocks == daddr) { 2709 jext->je_blocks += blocks; 2710 return; 2711 } 2712 /* Adding a new extent. */ 2713 if (++jblocks->jb_used == jblocks->jb_avail) { 2714 jblocks->jb_avail *= 2; 2715 jext = malloc(sizeof(struct jextent) * jblocks->jb_avail, 2716 M_JBLOCKS, M_WAITOK | M_ZERO); 2717 memcpy(jext, jblocks->jb_extent, 2718 sizeof(struct jextent) * jblocks->jb_used); 2719 free(jblocks->jb_extent, M_JBLOCKS); 2720 jblocks->jb_extent = jext; 2721 } 2722 jext = &jblocks->jb_extent[jblocks->jb_used]; 2723 jext->je_daddr = daddr; 2724 jext->je_blocks = blocks; 2725 return; 2726} 2727 2728int 2729softdep_journal_lookup(mp, vpp) 2730 struct mount *mp; 2731 struct vnode **vpp; 2732{ 2733 struct componentname cnp; 2734 struct vnode *dvp; 2735 ino_t sujournal; 2736 int error; 2737 2738 error = VFS_VGET(mp, UFS_ROOTINO, LK_EXCLUSIVE, &dvp); 2739 if (error) 2740 return (error); 2741 bzero(&cnp, sizeof(cnp)); 2742 cnp.cn_nameiop = LOOKUP; 2743 cnp.cn_flags = ISLASTCN; 2744 cnp.cn_thread = curthread; 2745 cnp.cn_cred = curthread->td_ucred; 2746 cnp.cn_pnbuf = SUJ_FILE; 2747 cnp.cn_nameptr = SUJ_FILE; 2748 cnp.cn_namelen = strlen(SUJ_FILE); 2749 error = ufs_lookup_ino(dvp, NULL, &cnp, &sujournal); 2750 vput(dvp); 2751 if (error != 0) 2752 return (error); 2753 error = VFS_VGET(mp, sujournal, LK_EXCLUSIVE, vpp); 2754 return (error); 2755} 2756 2757/* 2758 * Open and verify the journal file. 2759 */ 2760static int 2761journal_mount(mp, fs, cred) 2762 struct mount *mp; 2763 struct fs *fs; 2764 struct ucred *cred; 2765{ 2766 struct jblocks *jblocks; 2767 struct ufsmount *ump; 2768 struct vnode *vp; 2769 struct inode *ip; 2770 ufs2_daddr_t blkno; 2771 int bcount; 2772 int error; 2773 int i; 2774 2775 ump = VFSTOUFS(mp); 2776 ump->softdep_journal_tail = NULL; 2777 ump->softdep_on_journal = 0; 2778 ump->softdep_accdeps = 0; 2779 ump->softdep_req = 0; 2780 ump->softdep_jblocks = NULL; 2781 error = softdep_journal_lookup(mp, &vp); 2782 if (error != 0) { 2783 printf("Failed to find journal. Use tunefs to create one\n"); 2784 return (error); 2785 } 2786 ip = VTOI(vp); 2787 if (ip->i_size < SUJ_MIN) { 2788 error = ENOSPC; 2789 goto out; 2790 } 2791 bcount = lblkno(fs, ip->i_size); /* Only use whole blocks. */ 2792 jblocks = jblocks_create(); 2793 for (i = 0; i < bcount; i++) { 2794 error = ufs_bmaparray(vp, i, &blkno, NULL, NULL, NULL); 2795 if (error) 2796 break; 2797 jblocks_add(jblocks, blkno, fsbtodb(fs, fs->fs_frag)); 2798 } 2799 if (error) { 2800 jblocks_destroy(jblocks); 2801 goto out; 2802 } 2803 jblocks->jb_low = jblocks->jb_free / 3; /* Reserve 33%. */ 2804 jblocks->jb_min = jblocks->jb_free / 10; /* Suspend at 10%. */ 2805 ump->softdep_jblocks = jblocks; 2806out: 2807 if (error == 0) { 2808 MNT_ILOCK(mp); 2809 mp->mnt_flag |= MNT_SUJ; 2810 mp->mnt_flag &= ~MNT_SOFTDEP; 2811 MNT_IUNLOCK(mp); 2812 /* 2813 * Only validate the journal contents if the 2814 * filesystem is clean, otherwise we write the logs 2815 * but they'll never be used. If the filesystem was 2816 * still dirty when we mounted it the journal is 2817 * invalid and a new journal can only be valid if it 2818 * starts from a clean mount. 2819 */ 2820 if (fs->fs_clean) { 2821 DIP_SET(ip, i_modrev, fs->fs_mtime); 2822 ip->i_flags |= IN_MODIFIED; 2823 ffs_update(vp, 1); 2824 } 2825 } 2826 vput(vp); 2827 return (error); 2828} 2829 2830static void 2831journal_unmount(ump) 2832 struct ufsmount *ump; 2833{ 2834 2835 if (ump->softdep_jblocks) 2836 jblocks_destroy(ump->softdep_jblocks); 2837 ump->softdep_jblocks = NULL; 2838} 2839 2840/* 2841 * Called when a journal record is ready to be written. Space is allocated 2842 * and the journal entry is created when the journal is flushed to stable 2843 * store. 2844 */ 2845static void 2846add_to_journal(wk) 2847 struct worklist *wk; 2848{ 2849 struct ufsmount *ump; 2850 2851 ump = VFSTOUFS(wk->wk_mp); 2852 LOCK_OWNED(ump); 2853 if (wk->wk_state & ONWORKLIST) 2854 panic("add_to_journal: %s(0x%X) already on list", 2855 TYPENAME(wk->wk_type), wk->wk_state); 2856 wk->wk_state |= ONWORKLIST | DEPCOMPLETE; 2857 if (LIST_EMPTY(&ump->softdep_journal_pending)) { 2858 ump->softdep_jblocks->jb_age = ticks; 2859 LIST_INSERT_HEAD(&ump->softdep_journal_pending, wk, wk_list); 2860 } else 2861 LIST_INSERT_AFTER(ump->softdep_journal_tail, wk, wk_list); 2862 ump->softdep_journal_tail = wk; 2863 ump->softdep_on_journal += 1; 2864} 2865 2866/* 2867 * Remove an arbitrary item for the journal worklist maintain the tail 2868 * pointer. This happens when a new operation obviates the need to 2869 * journal an old operation. 2870 */ 2871static void 2872remove_from_journal(wk) 2873 struct worklist *wk; 2874{ 2875 struct ufsmount *ump; 2876 2877 ump = VFSTOUFS(wk->wk_mp); 2878 LOCK_OWNED(ump); 2879#ifdef SUJ_DEBUG 2880 { 2881 struct worklist *wkn; 2882 2883 LIST_FOREACH(wkn, &ump->softdep_journal_pending, wk_list) 2884 if (wkn == wk) 2885 break; 2886 if (wkn == NULL) 2887 panic("remove_from_journal: %p is not in journal", wk); 2888 } 2889#endif 2890 /* 2891 * We emulate a TAILQ to save space in most structures which do not 2892 * require TAILQ semantics. Here we must update the tail position 2893 * when removing the tail which is not the final entry. This works 2894 * only if the worklist linkage are at the beginning of the structure. 2895 */ 2896 if (ump->softdep_journal_tail == wk) 2897 ump->softdep_journal_tail = 2898 (struct worklist *)wk->wk_list.le_prev; 2899 WORKLIST_REMOVE(wk); 2900 ump->softdep_on_journal -= 1; 2901} 2902 2903/* 2904 * Check for journal space as well as dependency limits so the prelink 2905 * code can throttle both journaled and non-journaled filesystems. 2906 * Threshold is 0 for low and 1 for min. 2907 */ 2908static int 2909journal_space(ump, thresh) 2910 struct ufsmount *ump; 2911 int thresh; 2912{ 2913 struct jblocks *jblocks; 2914 int limit, avail; 2915 2916 jblocks = ump->softdep_jblocks; 2917 if (jblocks == NULL) 2918 return (1); 2919 /* 2920 * We use a tighter restriction here to prevent request_cleanup() 2921 * running in threads from running into locks we currently hold. 2922 * We have to be over the limit and our filesystem has to be 2923 * responsible for more than our share of that usage. 2924 */ 2925 limit = (max_softdeps / 10) * 9; 2926 if (dep_current[D_INODEDEP] > limit && 2927 ump->softdep_curdeps[D_INODEDEP] > limit / stat_flush_threads) 2928 return (0); 2929 if (thresh) 2930 thresh = jblocks->jb_min; 2931 else 2932 thresh = jblocks->jb_low; 2933 avail = (ump->softdep_on_journal * JREC_SIZE) / DEV_BSIZE; 2934 avail = jblocks->jb_free - avail; 2935 2936 return (avail > thresh); 2937} 2938 2939static void 2940journal_suspend(ump) 2941 struct ufsmount *ump; 2942{ 2943 struct jblocks *jblocks; 2944 struct mount *mp; 2945 2946 mp = UFSTOVFS(ump); 2947 jblocks = ump->softdep_jblocks; 2948 MNT_ILOCK(mp); 2949 if ((mp->mnt_kern_flag & MNTK_SUSPEND) == 0) { 2950 stat_journal_min++; 2951 mp->mnt_kern_flag |= MNTK_SUSPEND; 2952 mp->mnt_susp_owner = ump->softdep_flushtd; 2953 } 2954 jblocks->jb_suspended = 1; 2955 MNT_IUNLOCK(mp); 2956} 2957 2958static int 2959journal_unsuspend(struct ufsmount *ump) 2960{ 2961 struct jblocks *jblocks; 2962 struct mount *mp; 2963 2964 mp = UFSTOVFS(ump); 2965 jblocks = ump->softdep_jblocks; 2966 2967 if (jblocks != NULL && jblocks->jb_suspended && 2968 journal_space(ump, jblocks->jb_min)) { 2969 jblocks->jb_suspended = 0; 2970 FREE_LOCK(ump); 2971 mp->mnt_susp_owner = curthread; 2972 vfs_write_resume(mp, 0); 2973 ACQUIRE_LOCK(ump); 2974 return (1); 2975 } 2976 return (0); 2977} 2978 2979/* 2980 * Called before any allocation function to be certain that there is 2981 * sufficient space in the journal prior to creating any new records. 2982 * Since in the case of block allocation we may have multiple locked 2983 * buffers at the time of the actual allocation we can not block 2984 * when the journal records are created. Doing so would create a deadlock 2985 * if any of these buffers needed to be flushed to reclaim space. Instead 2986 * we require a sufficiently large amount of available space such that 2987 * each thread in the system could have passed this allocation check and 2988 * still have sufficient free space. With 20% of a minimum journal size 2989 * of 1MB we have 6553 records available. 2990 */ 2991int 2992softdep_prealloc(vp, waitok) 2993 struct vnode *vp; 2994 int waitok; 2995{ 2996 struct ufsmount *ump; 2997 2998 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 2999 ("softdep_prealloc called on non-softdep filesystem")); 3000 /* 3001 * Nothing to do if we are not running journaled soft updates. 3002 * If we currently hold the snapshot lock, we must avoid 3003 * handling other resources that could cause deadlock. Do not 3004 * touch quotas vnode since it is typically recursed with 3005 * other vnode locks held. 3006 */ 3007 if (DOINGSUJ(vp) == 0 || IS_SNAPSHOT(VTOI(vp)) || 3008 (vp->v_vflag & VV_SYSTEM) != 0) 3009 return (0); 3010 ump = VFSTOUFS(vp->v_mount); 3011 ACQUIRE_LOCK(ump); 3012 if (journal_space(ump, 0)) { 3013 FREE_LOCK(ump); 3014 return (0); 3015 } 3016 stat_journal_low++; 3017 FREE_LOCK(ump); 3018 if (waitok == MNT_NOWAIT) 3019 return (ENOSPC); 3020 /* 3021 * Attempt to sync this vnode once to flush any journal 3022 * work attached to it. 3023 */ 3024 if ((curthread->td_pflags & TDP_COWINPROGRESS) == 0) 3025 ffs_syncvnode(vp, waitok, 0); 3026 ACQUIRE_LOCK(ump); 3027 process_removes(vp); 3028 process_truncates(vp); 3029 if (journal_space(ump, 0) == 0) { 3030 softdep_speedup(ump); 3031 if (journal_space(ump, 1) == 0) 3032 journal_suspend(ump); 3033 } 3034 FREE_LOCK(ump); 3035 3036 return (0); 3037} 3038 3039/* 3040 * Before adjusting a link count on a vnode verify that we have sufficient 3041 * journal space. If not, process operations that depend on the currently 3042 * locked pair of vnodes to try to flush space as the syncer, buf daemon, 3043 * and softdep flush threads can not acquire these locks to reclaim space. 3044 */ 3045static void 3046softdep_prelink(dvp, vp) 3047 struct vnode *dvp; 3048 struct vnode *vp; 3049{ 3050 struct ufsmount *ump; 3051 3052 ump = VFSTOUFS(dvp->v_mount); 3053 LOCK_OWNED(ump); 3054 /* 3055 * Nothing to do if we have sufficient journal space. 3056 * If we currently hold the snapshot lock, we must avoid 3057 * handling other resources that could cause deadlock. 3058 */ 3059 if (journal_space(ump, 0) || (vp && IS_SNAPSHOT(VTOI(vp)))) 3060 return; 3061 stat_journal_low++; 3062 FREE_LOCK(ump); 3063 if (vp) 3064 ffs_syncvnode(vp, MNT_NOWAIT, 0); 3065 ffs_syncvnode(dvp, MNT_WAIT, 0); 3066 ACQUIRE_LOCK(ump); 3067 /* Process vp before dvp as it may create .. removes. */ 3068 if (vp) { 3069 process_removes(vp); 3070 process_truncates(vp); 3071 } 3072 process_removes(dvp); 3073 process_truncates(dvp); 3074 softdep_speedup(ump); 3075 process_worklist_item(UFSTOVFS(ump), 2, LK_NOWAIT); 3076 if (journal_space(ump, 0) == 0) { 3077 softdep_speedup(ump); 3078 if (journal_space(ump, 1) == 0) 3079 journal_suspend(ump); 3080 } 3081} 3082 3083static void 3084jseg_write(ump, jseg, data) 3085 struct ufsmount *ump; 3086 struct jseg *jseg; 3087 uint8_t *data; 3088{ 3089 struct jsegrec *rec; 3090 3091 rec = (struct jsegrec *)data; 3092 rec->jsr_seq = jseg->js_seq; 3093 rec->jsr_oldest = jseg->js_oldseq; 3094 rec->jsr_cnt = jseg->js_cnt; 3095 rec->jsr_blocks = jseg->js_size / ump->um_devvp->v_bufobj.bo_bsize; 3096 rec->jsr_crc = 0; 3097 rec->jsr_time = ump->um_fs->fs_mtime; 3098} 3099 3100static inline void 3101inoref_write(inoref, jseg, rec) 3102 struct inoref *inoref; 3103 struct jseg *jseg; 3104 struct jrefrec *rec; 3105{ 3106 3107 inoref->if_jsegdep->jd_seg = jseg; 3108 rec->jr_ino = inoref->if_ino; 3109 rec->jr_parent = inoref->if_parent; 3110 rec->jr_nlink = inoref->if_nlink; 3111 rec->jr_mode = inoref->if_mode; 3112 rec->jr_diroff = inoref->if_diroff; 3113} 3114 3115static void 3116jaddref_write(jaddref, jseg, data) 3117 struct jaddref *jaddref; 3118 struct jseg *jseg; 3119 uint8_t *data; 3120{ 3121 struct jrefrec *rec; 3122 3123 rec = (struct jrefrec *)data; 3124 rec->jr_op = JOP_ADDREF; 3125 inoref_write(&jaddref->ja_ref, jseg, rec); 3126} 3127 3128static void 3129jremref_write(jremref, jseg, data) 3130 struct jremref *jremref; 3131 struct jseg *jseg; 3132 uint8_t *data; 3133{ 3134 struct jrefrec *rec; 3135 3136 rec = (struct jrefrec *)data; 3137 rec->jr_op = JOP_REMREF; 3138 inoref_write(&jremref->jr_ref, jseg, rec); 3139} 3140 3141static void 3142jmvref_write(jmvref, jseg, data) 3143 struct jmvref *jmvref; 3144 struct jseg *jseg; 3145 uint8_t *data; 3146{ 3147 struct jmvrec *rec; 3148 3149 rec = (struct jmvrec *)data; 3150 rec->jm_op = JOP_MVREF; 3151 rec->jm_ino = jmvref->jm_ino; 3152 rec->jm_parent = jmvref->jm_parent; 3153 rec->jm_oldoff = jmvref->jm_oldoff; 3154 rec->jm_newoff = jmvref->jm_newoff; 3155} 3156 3157static void 3158jnewblk_write(jnewblk, jseg, data) 3159 struct jnewblk *jnewblk; 3160 struct jseg *jseg; 3161 uint8_t *data; 3162{ 3163 struct jblkrec *rec; 3164 3165 jnewblk->jn_jsegdep->jd_seg = jseg; 3166 rec = (struct jblkrec *)data; 3167 rec->jb_op = JOP_NEWBLK; 3168 rec->jb_ino = jnewblk->jn_ino; 3169 rec->jb_blkno = jnewblk->jn_blkno; 3170 rec->jb_lbn = jnewblk->jn_lbn; 3171 rec->jb_frags = jnewblk->jn_frags; 3172 rec->jb_oldfrags = jnewblk->jn_oldfrags; 3173} 3174 3175static void 3176jfreeblk_write(jfreeblk, jseg, data) 3177 struct jfreeblk *jfreeblk; 3178 struct jseg *jseg; 3179 uint8_t *data; 3180{ 3181 struct jblkrec *rec; 3182 3183 jfreeblk->jf_dep.jb_jsegdep->jd_seg = jseg; 3184 rec = (struct jblkrec *)data; 3185 rec->jb_op = JOP_FREEBLK; 3186 rec->jb_ino = jfreeblk->jf_ino; 3187 rec->jb_blkno = jfreeblk->jf_blkno; 3188 rec->jb_lbn = jfreeblk->jf_lbn; 3189 rec->jb_frags = jfreeblk->jf_frags; 3190 rec->jb_oldfrags = 0; 3191} 3192 3193static void 3194jfreefrag_write(jfreefrag, jseg, data) 3195 struct jfreefrag *jfreefrag; 3196 struct jseg *jseg; 3197 uint8_t *data; 3198{ 3199 struct jblkrec *rec; 3200 3201 jfreefrag->fr_jsegdep->jd_seg = jseg; 3202 rec = (struct jblkrec *)data; 3203 rec->jb_op = JOP_FREEBLK; 3204 rec->jb_ino = jfreefrag->fr_ino; 3205 rec->jb_blkno = jfreefrag->fr_blkno; 3206 rec->jb_lbn = jfreefrag->fr_lbn; 3207 rec->jb_frags = jfreefrag->fr_frags; 3208 rec->jb_oldfrags = 0; 3209} 3210 3211static void 3212jtrunc_write(jtrunc, jseg, data) 3213 struct jtrunc *jtrunc; 3214 struct jseg *jseg; 3215 uint8_t *data; 3216{ 3217 struct jtrncrec *rec; 3218 3219 jtrunc->jt_dep.jb_jsegdep->jd_seg = jseg; 3220 rec = (struct jtrncrec *)data; 3221 rec->jt_op = JOP_TRUNC; 3222 rec->jt_ino = jtrunc->jt_ino; 3223 rec->jt_size = jtrunc->jt_size; 3224 rec->jt_extsize = jtrunc->jt_extsize; 3225} 3226 3227static void 3228jfsync_write(jfsync, jseg, data) 3229 struct jfsync *jfsync; 3230 struct jseg *jseg; 3231 uint8_t *data; 3232{ 3233 struct jtrncrec *rec; 3234 3235 rec = (struct jtrncrec *)data; 3236 rec->jt_op = JOP_SYNC; 3237 rec->jt_ino = jfsync->jfs_ino; 3238 rec->jt_size = jfsync->jfs_size; 3239 rec->jt_extsize = jfsync->jfs_extsize; 3240} 3241 3242static void 3243softdep_flushjournal(mp) 3244 struct mount *mp; 3245{ 3246 struct jblocks *jblocks; 3247 struct ufsmount *ump; 3248 3249 if (MOUNTEDSUJ(mp) == 0) 3250 return; 3251 ump = VFSTOUFS(mp); 3252 jblocks = ump->softdep_jblocks; 3253 ACQUIRE_LOCK(ump); 3254 while (ump->softdep_on_journal) { 3255 jblocks->jb_needseg = 1; 3256 softdep_process_journal(mp, NULL, MNT_WAIT); 3257 } 3258 FREE_LOCK(ump); 3259} 3260 3261static void softdep_synchronize_completed(struct bio *); 3262static void softdep_synchronize(struct bio *, struct ufsmount *, void *); 3263 3264static void 3265softdep_synchronize_completed(bp) 3266 struct bio *bp; 3267{ 3268 struct jseg *oldest; 3269 struct jseg *jseg; 3270 struct ufsmount *ump; 3271 3272 /* 3273 * caller1 marks the last segment written before we issued the 3274 * synchronize cache. 3275 */ 3276 jseg = bp->bio_caller1; 3277 if (jseg == NULL) { 3278 g_destroy_bio(bp); 3279 return; 3280 } 3281 ump = VFSTOUFS(jseg->js_list.wk_mp); 3282 ACQUIRE_LOCK(ump); 3283 oldest = NULL; 3284 /* 3285 * Mark all the journal entries waiting on the synchronize cache 3286 * as completed so they may continue on. 3287 */ 3288 while (jseg != NULL && (jseg->js_state & COMPLETE) == 0) { 3289 jseg->js_state |= COMPLETE; 3290 oldest = jseg; 3291 jseg = TAILQ_PREV(jseg, jseglst, js_next); 3292 } 3293 /* 3294 * Restart deferred journal entry processing from the oldest 3295 * completed jseg. 3296 */ 3297 if (oldest) 3298 complete_jsegs(oldest); 3299 3300 FREE_LOCK(ump); 3301 g_destroy_bio(bp); 3302} 3303 3304/* 3305 * Send BIO_FLUSH/SYNCHRONIZE CACHE to the device to enforce write ordering 3306 * barriers. The journal must be written prior to any blocks that depend 3307 * on it and the journal can not be released until the blocks have be 3308 * written. This code handles both barriers simultaneously. 3309 */ 3310static void 3311softdep_synchronize(bp, ump, caller1) 3312 struct bio *bp; 3313 struct ufsmount *ump; 3314 void *caller1; 3315{ 3316 3317 bp->bio_cmd = BIO_FLUSH; 3318 bp->bio_flags |= BIO_ORDERED; 3319 bp->bio_data = NULL; 3320 bp->bio_offset = ump->um_cp->provider->mediasize; 3321 bp->bio_length = 0; 3322 bp->bio_done = softdep_synchronize_completed; 3323 bp->bio_caller1 = caller1; 3324 g_io_request(bp, 3325 (struct g_consumer *)ump->um_devvp->v_bufobj.bo_private); 3326} 3327 3328/* 3329 * Flush some journal records to disk. 3330 */ 3331static void 3332softdep_process_journal(mp, needwk, flags) 3333 struct mount *mp; 3334 struct worklist *needwk; 3335 int flags; 3336{ 3337 struct jblocks *jblocks; 3338 struct ufsmount *ump; 3339 struct worklist *wk; 3340 struct jseg *jseg; 3341 struct buf *bp; 3342 struct bio *bio; 3343 uint8_t *data; 3344 struct fs *fs; 3345 int shouldflush; 3346 int segwritten; 3347 int jrecmin; /* Minimum records per block. */ 3348 int jrecmax; /* Maximum records per block. */ 3349 int size; 3350 int cnt; 3351 int off; 3352 int devbsize; 3353 3354 if (MOUNTEDSUJ(mp) == 0) 3355 return; 3356 shouldflush = softdep_flushcache; 3357 bio = NULL; 3358 jseg = NULL; 3359 ump = VFSTOUFS(mp); 3360 LOCK_OWNED(ump); 3361 fs = ump->um_fs; 3362 jblocks = ump->softdep_jblocks; 3363 devbsize = ump->um_devvp->v_bufobj.bo_bsize; 3364 /* 3365 * We write anywhere between a disk block and fs block. The upper 3366 * bound is picked to prevent buffer cache fragmentation and limit 3367 * processing time per I/O. 3368 */ 3369 jrecmin = (devbsize / JREC_SIZE) - 1; /* -1 for seg header */ 3370 jrecmax = (fs->fs_bsize / devbsize) * jrecmin; 3371 segwritten = 0; 3372 for (;;) { 3373 cnt = ump->softdep_on_journal; 3374 /* 3375 * Criteria for writing a segment: 3376 * 1) We have a full block. 3377 * 2) We're called from jwait() and haven't found the 3378 * journal item yet. 3379 * 3) Always write if needseg is set. 3380 * 4) If we are called from process_worklist and have 3381 * not yet written anything we write a partial block 3382 * to enforce a 1 second maximum latency on journal 3383 * entries. 3384 */ 3385 if (cnt < (jrecmax - 1) && needwk == NULL && 3386 jblocks->jb_needseg == 0 && (segwritten || cnt == 0)) 3387 break; 3388 cnt++; 3389 /* 3390 * Verify some free journal space. softdep_prealloc() should 3391 * guarantee that we don't run out so this is indicative of 3392 * a problem with the flow control. Try to recover 3393 * gracefully in any event. 3394 */ 3395 while (jblocks->jb_free == 0) { 3396 if (flags != MNT_WAIT) 3397 break; 3398 printf("softdep: Out of journal space!\n"); 3399 softdep_speedup(ump); 3400 msleep(jblocks, LOCK_PTR(ump), PRIBIO, "jblocks", hz); 3401 } 3402 FREE_LOCK(ump); 3403 jseg = malloc(sizeof(*jseg), M_JSEG, M_SOFTDEP_FLAGS); 3404 workitem_alloc(&jseg->js_list, D_JSEG, mp); 3405 LIST_INIT(&jseg->js_entries); 3406 LIST_INIT(&jseg->js_indirs); 3407 jseg->js_state = ATTACHED; 3408 if (shouldflush == 0) 3409 jseg->js_state |= COMPLETE; 3410 else if (bio == NULL) 3411 bio = g_alloc_bio(); 3412 jseg->js_jblocks = jblocks; 3413 bp = geteblk(fs->fs_bsize, 0); 3414 ACQUIRE_LOCK(ump); 3415 /* 3416 * If there was a race while we were allocating the block 3417 * and jseg the entry we care about was likely written. 3418 * We bail out in both the WAIT and NOWAIT case and assume 3419 * the caller will loop if the entry it cares about is 3420 * not written. 3421 */ 3422 cnt = ump->softdep_on_journal; 3423 if (cnt + jblocks->jb_needseg == 0 || jblocks->jb_free == 0) { 3424 bp->b_flags |= B_INVAL | B_NOCACHE; 3425 WORKITEM_FREE(jseg, D_JSEG); 3426 FREE_LOCK(ump); 3427 brelse(bp); 3428 ACQUIRE_LOCK(ump); 3429 break; 3430 } 3431 /* 3432 * Calculate the disk block size required for the available 3433 * records rounded to the min size. 3434 */ 3435 if (cnt == 0) 3436 size = devbsize; 3437 else if (cnt < jrecmax) 3438 size = howmany(cnt, jrecmin) * devbsize; 3439 else 3440 size = fs->fs_bsize; 3441 /* 3442 * Allocate a disk block for this journal data and account 3443 * for truncation of the requested size if enough contiguous 3444 * space was not available. 3445 */ 3446 bp->b_blkno = jblocks_alloc(jblocks, size, &size); 3447 bp->b_lblkno = bp->b_blkno; 3448 bp->b_offset = bp->b_blkno * DEV_BSIZE; 3449 bp->b_bcount = size; 3450 bp->b_flags &= ~B_INVAL; 3451 bp->b_flags |= B_VALIDSUSPWRT | B_NOCOPY; 3452 /* 3453 * Initialize our jseg with cnt records. Assign the next 3454 * sequence number to it and link it in-order. 3455 */ 3456 cnt = MIN(cnt, (size / devbsize) * jrecmin); 3457 jseg->js_buf = bp; 3458 jseg->js_cnt = cnt; 3459 jseg->js_refs = cnt + 1; /* Self ref. */ 3460 jseg->js_size = size; 3461 jseg->js_seq = jblocks->jb_nextseq++; 3462 if (jblocks->jb_oldestseg == NULL) 3463 jblocks->jb_oldestseg = jseg; 3464 jseg->js_oldseq = jblocks->jb_oldestseg->js_seq; 3465 TAILQ_INSERT_TAIL(&jblocks->jb_segs, jseg, js_next); 3466 if (jblocks->jb_writeseg == NULL) 3467 jblocks->jb_writeseg = jseg; 3468 /* 3469 * Start filling in records from the pending list. 3470 */ 3471 data = bp->b_data; 3472 off = 0; 3473 3474 /* 3475 * Always put a header on the first block. 3476 * XXX As with below, there might not be a chance to get 3477 * into the loop. Ensure that something valid is written. 3478 */ 3479 jseg_write(ump, jseg, data); 3480 off += JREC_SIZE; 3481 data = bp->b_data + off; 3482 3483 /* 3484 * XXX Something is wrong here. There's no work to do, 3485 * but we need to perform and I/O and allow it to complete 3486 * anyways. 3487 */ 3488 if (LIST_EMPTY(&ump->softdep_journal_pending)) 3489 stat_emptyjblocks++; 3490 3491 while ((wk = LIST_FIRST(&ump->softdep_journal_pending)) 3492 != NULL) { 3493 if (cnt == 0) 3494 break; 3495 /* Place a segment header on every device block. */ 3496 if ((off % devbsize) == 0) { 3497 jseg_write(ump, jseg, data); 3498 off += JREC_SIZE; 3499 data = bp->b_data + off; 3500 } 3501 if (wk == needwk) 3502 needwk = NULL; 3503 remove_from_journal(wk); 3504 wk->wk_state |= INPROGRESS; 3505 WORKLIST_INSERT(&jseg->js_entries, wk); 3506 switch (wk->wk_type) { 3507 case D_JADDREF: 3508 jaddref_write(WK_JADDREF(wk), jseg, data); 3509 break; 3510 case D_JREMREF: 3511 jremref_write(WK_JREMREF(wk), jseg, data); 3512 break; 3513 case D_JMVREF: 3514 jmvref_write(WK_JMVREF(wk), jseg, data); 3515 break; 3516 case D_JNEWBLK: 3517 jnewblk_write(WK_JNEWBLK(wk), jseg, data); 3518 break; 3519 case D_JFREEBLK: 3520 jfreeblk_write(WK_JFREEBLK(wk), jseg, data); 3521 break; 3522 case D_JFREEFRAG: 3523 jfreefrag_write(WK_JFREEFRAG(wk), jseg, data); 3524 break; 3525 case D_JTRUNC: 3526 jtrunc_write(WK_JTRUNC(wk), jseg, data); 3527 break; 3528 case D_JFSYNC: 3529 jfsync_write(WK_JFSYNC(wk), jseg, data); 3530 break; 3531 default: 3532 panic("process_journal: Unknown type %s", 3533 TYPENAME(wk->wk_type)); 3534 /* NOTREACHED */ 3535 } 3536 off += JREC_SIZE; 3537 data = bp->b_data + off; 3538 cnt--; 3539 } 3540 3541 /* Clear any remaining space so we don't leak kernel data */ 3542 if (size > off) 3543 bzero(data, size - off); 3544 3545 /* 3546 * Write this one buffer and continue. 3547 */ 3548 segwritten = 1; 3549 jblocks->jb_needseg = 0; 3550 WORKLIST_INSERT(&bp->b_dep, &jseg->js_list); 3551 FREE_LOCK(ump); 3552 pbgetvp(ump->um_devvp, bp); 3553 /* 3554 * We only do the blocking wait once we find the journal 3555 * entry we're looking for. 3556 */ 3557 if (needwk == NULL && flags == MNT_WAIT) 3558 bwrite(bp); 3559 else 3560 bawrite(bp); 3561 ACQUIRE_LOCK(ump); 3562 } 3563 /* 3564 * If we wrote a segment issue a synchronize cache so the journal 3565 * is reflected on disk before the data is written. Since reclaiming 3566 * journal space also requires writing a journal record this 3567 * process also enforces a barrier before reclamation. 3568 */ 3569 if (segwritten && shouldflush) { 3570 softdep_synchronize(bio, ump, 3571 TAILQ_LAST(&jblocks->jb_segs, jseglst)); 3572 } else if (bio) 3573 g_destroy_bio(bio); 3574 /* 3575 * If we've suspended the filesystem because we ran out of journal 3576 * space either try to sync it here to make some progress or 3577 * unsuspend it if we already have. 3578 */ 3579 if (flags == 0 && jblocks->jb_suspended) { 3580 if (journal_unsuspend(ump)) 3581 return; 3582 FREE_LOCK(ump); 3583 VFS_SYNC(mp, MNT_NOWAIT); 3584 ffs_sbupdate(ump, MNT_WAIT, 0); 3585 ACQUIRE_LOCK(ump); 3586 } 3587} 3588 3589/* 3590 * Complete a jseg, allowing all dependencies awaiting journal writes 3591 * to proceed. Each journal dependency also attaches a jsegdep to dependent 3592 * structures so that the journal segment can be freed to reclaim space. 3593 */ 3594static void 3595complete_jseg(jseg) 3596 struct jseg *jseg; 3597{ 3598 struct worklist *wk; 3599 struct jmvref *jmvref; 3600#ifdef INVARIANTS 3601 int i = 0; 3602#endif 3603 3604 while ((wk = LIST_FIRST(&jseg->js_entries)) != NULL) { 3605 WORKLIST_REMOVE(wk); 3606 wk->wk_state &= ~INPROGRESS; 3607 wk->wk_state |= COMPLETE; 3608 KASSERT(i++ < jseg->js_cnt, 3609 ("handle_written_jseg: overflow %d >= %d", 3610 i - 1, jseg->js_cnt)); 3611 switch (wk->wk_type) { 3612 case D_JADDREF: 3613 handle_written_jaddref(WK_JADDREF(wk)); 3614 break; 3615 case D_JREMREF: 3616 handle_written_jremref(WK_JREMREF(wk)); 3617 break; 3618 case D_JMVREF: 3619 rele_jseg(jseg); /* No jsegdep. */ 3620 jmvref = WK_JMVREF(wk); 3621 LIST_REMOVE(jmvref, jm_deps); 3622 if ((jmvref->jm_pagedep->pd_state & ONWORKLIST) == 0) 3623 free_pagedep(jmvref->jm_pagedep); 3624 WORKITEM_FREE(jmvref, D_JMVREF); 3625 break; 3626 case D_JNEWBLK: 3627 handle_written_jnewblk(WK_JNEWBLK(wk)); 3628 break; 3629 case D_JFREEBLK: 3630 handle_written_jblkdep(&WK_JFREEBLK(wk)->jf_dep); 3631 break; 3632 case D_JTRUNC: 3633 handle_written_jblkdep(&WK_JTRUNC(wk)->jt_dep); 3634 break; 3635 case D_JFSYNC: 3636 rele_jseg(jseg); /* No jsegdep. */ 3637 WORKITEM_FREE(wk, D_JFSYNC); 3638 break; 3639 case D_JFREEFRAG: 3640 handle_written_jfreefrag(WK_JFREEFRAG(wk)); 3641 break; 3642 default: 3643 panic("handle_written_jseg: Unknown type %s", 3644 TYPENAME(wk->wk_type)); 3645 /* NOTREACHED */ 3646 } 3647 } 3648 /* Release the self reference so the structure may be freed. */ 3649 rele_jseg(jseg); 3650} 3651 3652/* 3653 * Determine which jsegs are ready for completion processing. Waits for 3654 * synchronize cache to complete as well as forcing in-order completion 3655 * of journal entries. 3656 */ 3657static void 3658complete_jsegs(jseg) 3659 struct jseg *jseg; 3660{ 3661 struct jblocks *jblocks; 3662 struct jseg *jsegn; 3663 3664 jblocks = jseg->js_jblocks; 3665 /* 3666 * Don't allow out of order completions. If this isn't the first 3667 * block wait for it to write before we're done. 3668 */ 3669 if (jseg != jblocks->jb_writeseg) 3670 return; 3671 /* Iterate through available jsegs processing their entries. */ 3672 while (jseg && (jseg->js_state & ALLCOMPLETE) == ALLCOMPLETE) { 3673 jblocks->jb_oldestwrseq = jseg->js_oldseq; 3674 jsegn = TAILQ_NEXT(jseg, js_next); 3675 complete_jseg(jseg); 3676 jseg = jsegn; 3677 } 3678 jblocks->jb_writeseg = jseg; 3679 /* 3680 * Attempt to free jsegs now that oldestwrseq may have advanced. 3681 */ 3682 free_jsegs(jblocks); 3683} 3684 3685/* 3686 * Mark a jseg as DEPCOMPLETE and throw away the buffer. Attempt to handle 3687 * the final completions. 3688 */ 3689static void 3690handle_written_jseg(jseg, bp) 3691 struct jseg *jseg; 3692 struct buf *bp; 3693{ 3694 3695 if (jseg->js_refs == 0) 3696 panic("handle_written_jseg: No self-reference on %p", jseg); 3697 jseg->js_state |= DEPCOMPLETE; 3698 /* 3699 * We'll never need this buffer again, set flags so it will be 3700 * discarded. 3701 */ 3702 bp->b_flags |= B_INVAL | B_NOCACHE; 3703 pbrelvp(bp); 3704 complete_jsegs(jseg); 3705} 3706 3707static inline struct jsegdep * 3708inoref_jseg(inoref) 3709 struct inoref *inoref; 3710{ 3711 struct jsegdep *jsegdep; 3712 3713 jsegdep = inoref->if_jsegdep; 3714 inoref->if_jsegdep = NULL; 3715 3716 return (jsegdep); 3717} 3718 3719/* 3720 * Called once a jremref has made it to stable store. The jremref is marked 3721 * complete and we attempt to free it. Any pagedeps writes sleeping waiting 3722 * for the jremref to complete will be awoken by free_jremref. 3723 */ 3724static void 3725handle_written_jremref(jremref) 3726 struct jremref *jremref; 3727{ 3728 struct inodedep *inodedep; 3729 struct jsegdep *jsegdep; 3730 struct dirrem *dirrem; 3731 3732 /* Grab the jsegdep. */ 3733 jsegdep = inoref_jseg(&jremref->jr_ref); 3734 /* 3735 * Remove us from the inoref list. 3736 */ 3737 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 3738 0, &inodedep) == 0) 3739 panic("handle_written_jremref: Lost inodedep"); 3740 TAILQ_REMOVE(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps); 3741 /* 3742 * Complete the dirrem. 3743 */ 3744 dirrem = jremref->jr_dirrem; 3745 jremref->jr_dirrem = NULL; 3746 LIST_REMOVE(jremref, jr_deps); 3747 jsegdep->jd_state |= jremref->jr_state & MKDIR_PARENT; 3748 jwork_insert(&dirrem->dm_jwork, jsegdep); 3749 if (LIST_EMPTY(&dirrem->dm_jremrefhd) && 3750 (dirrem->dm_state & COMPLETE) != 0) 3751 add_to_worklist(&dirrem->dm_list, 0); 3752 free_jremref(jremref); 3753} 3754 3755/* 3756 * Called once a jaddref has made it to stable store. The dependency is 3757 * marked complete and any dependent structures are added to the inode 3758 * bufwait list to be completed as soon as it is written. If a bitmap write 3759 * depends on this entry we move the inode into the inodedephd of the 3760 * bmsafemap dependency and attempt to remove the jaddref from the bmsafemap. 3761 */ 3762static void 3763handle_written_jaddref(jaddref) 3764 struct jaddref *jaddref; 3765{ 3766 struct jsegdep *jsegdep; 3767 struct inodedep *inodedep; 3768 struct diradd *diradd; 3769 struct mkdir *mkdir; 3770 3771 /* Grab the jsegdep. */ 3772 jsegdep = inoref_jseg(&jaddref->ja_ref); 3773 mkdir = NULL; 3774 diradd = NULL; 3775 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino, 3776 0, &inodedep) == 0) 3777 panic("handle_written_jaddref: Lost inodedep."); 3778 if (jaddref->ja_diradd == NULL) 3779 panic("handle_written_jaddref: No dependency"); 3780 if (jaddref->ja_diradd->da_list.wk_type == D_DIRADD) { 3781 diradd = jaddref->ja_diradd; 3782 WORKLIST_INSERT(&inodedep->id_bufwait, &diradd->da_list); 3783 } else if (jaddref->ja_state & MKDIR_PARENT) { 3784 mkdir = jaddref->ja_mkdir; 3785 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir->md_list); 3786 } else if (jaddref->ja_state & MKDIR_BODY) 3787 mkdir = jaddref->ja_mkdir; 3788 else 3789 panic("handle_written_jaddref: Unknown dependency %p", 3790 jaddref->ja_diradd); 3791 jaddref->ja_diradd = NULL; /* also clears ja_mkdir */ 3792 /* 3793 * Remove us from the inode list. 3794 */ 3795 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, if_deps); 3796 /* 3797 * The mkdir may be waiting on the jaddref to clear before freeing. 3798 */ 3799 if (mkdir) { 3800 KASSERT(mkdir->md_list.wk_type == D_MKDIR, 3801 ("handle_written_jaddref: Incorrect type for mkdir %s", 3802 TYPENAME(mkdir->md_list.wk_type))); 3803 mkdir->md_jaddref = NULL; 3804 diradd = mkdir->md_diradd; 3805 mkdir->md_state |= DEPCOMPLETE; 3806 complete_mkdir(mkdir); 3807 } 3808 jwork_insert(&diradd->da_jwork, jsegdep); 3809 if (jaddref->ja_state & NEWBLOCK) { 3810 inodedep->id_state |= ONDEPLIST; 3811 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_inodedephd, 3812 inodedep, id_deps); 3813 } 3814 free_jaddref(jaddref); 3815} 3816 3817/* 3818 * Called once a jnewblk journal is written. The allocdirect or allocindir 3819 * is placed in the bmsafemap to await notification of a written bitmap. If 3820 * the operation was canceled we add the segdep to the appropriate 3821 * dependency to free the journal space once the canceling operation 3822 * completes. 3823 */ 3824static void 3825handle_written_jnewblk(jnewblk) 3826 struct jnewblk *jnewblk; 3827{ 3828 struct bmsafemap *bmsafemap; 3829 struct freefrag *freefrag; 3830 struct freework *freework; 3831 struct jsegdep *jsegdep; 3832 struct newblk *newblk; 3833 3834 /* Grab the jsegdep. */ 3835 jsegdep = jnewblk->jn_jsegdep; 3836 jnewblk->jn_jsegdep = NULL; 3837 if (jnewblk->jn_dep == NULL) 3838 panic("handle_written_jnewblk: No dependency for the segdep."); 3839 switch (jnewblk->jn_dep->wk_type) { 3840 case D_NEWBLK: 3841 case D_ALLOCDIRECT: 3842 case D_ALLOCINDIR: 3843 /* 3844 * Add the written block to the bmsafemap so it can 3845 * be notified when the bitmap is on disk. 3846 */ 3847 newblk = WK_NEWBLK(jnewblk->jn_dep); 3848 newblk->nb_jnewblk = NULL; 3849 if ((newblk->nb_state & GOINGAWAY) == 0) { 3850 bmsafemap = newblk->nb_bmsafemap; 3851 newblk->nb_state |= ONDEPLIST; 3852 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, 3853 nb_deps); 3854 } 3855 jwork_insert(&newblk->nb_jwork, jsegdep); 3856 break; 3857 case D_FREEFRAG: 3858 /* 3859 * A newblock being removed by a freefrag when replaced by 3860 * frag extension. 3861 */ 3862 freefrag = WK_FREEFRAG(jnewblk->jn_dep); 3863 freefrag->ff_jdep = NULL; 3864 jwork_insert(&freefrag->ff_jwork, jsegdep); 3865 break; 3866 case D_FREEWORK: 3867 /* 3868 * A direct block was removed by truncate. 3869 */ 3870 freework = WK_FREEWORK(jnewblk->jn_dep); 3871 freework->fw_jnewblk = NULL; 3872 jwork_insert(&freework->fw_freeblks->fb_jwork, jsegdep); 3873 break; 3874 default: 3875 panic("handle_written_jnewblk: Unknown type %d.", 3876 jnewblk->jn_dep->wk_type); 3877 } 3878 jnewblk->jn_dep = NULL; 3879 free_jnewblk(jnewblk); 3880} 3881 3882/* 3883 * Cancel a jfreefrag that won't be needed, probably due to colliding with 3884 * an in-flight allocation that has not yet been committed. Divorce us 3885 * from the freefrag and mark it DEPCOMPLETE so that it may be added 3886 * to the worklist. 3887 */ 3888static void 3889cancel_jfreefrag(jfreefrag) 3890 struct jfreefrag *jfreefrag; 3891{ 3892 struct freefrag *freefrag; 3893 3894 if (jfreefrag->fr_jsegdep) { 3895 free_jsegdep(jfreefrag->fr_jsegdep); 3896 jfreefrag->fr_jsegdep = NULL; 3897 } 3898 freefrag = jfreefrag->fr_freefrag; 3899 jfreefrag->fr_freefrag = NULL; 3900 free_jfreefrag(jfreefrag); 3901 freefrag->ff_state |= DEPCOMPLETE; 3902 CTR1(KTR_SUJ, "cancel_jfreefrag: blkno %jd", freefrag->ff_blkno); 3903} 3904 3905/* 3906 * Free a jfreefrag when the parent freefrag is rendered obsolete. 3907 */ 3908static void 3909free_jfreefrag(jfreefrag) 3910 struct jfreefrag *jfreefrag; 3911{ 3912 3913 if (jfreefrag->fr_state & INPROGRESS) 3914 WORKLIST_REMOVE(&jfreefrag->fr_list); 3915 else if (jfreefrag->fr_state & ONWORKLIST) 3916 remove_from_journal(&jfreefrag->fr_list); 3917 if (jfreefrag->fr_freefrag != NULL) 3918 panic("free_jfreefrag: Still attached to a freefrag."); 3919 WORKITEM_FREE(jfreefrag, D_JFREEFRAG); 3920} 3921 3922/* 3923 * Called when the journal write for a jfreefrag completes. The parent 3924 * freefrag is added to the worklist if this completes its dependencies. 3925 */ 3926static void 3927handle_written_jfreefrag(jfreefrag) 3928 struct jfreefrag *jfreefrag; 3929{ 3930 struct jsegdep *jsegdep; 3931 struct freefrag *freefrag; 3932 3933 /* Grab the jsegdep. */ 3934 jsegdep = jfreefrag->fr_jsegdep; 3935 jfreefrag->fr_jsegdep = NULL; 3936 freefrag = jfreefrag->fr_freefrag; 3937 if (freefrag == NULL) 3938 panic("handle_written_jfreefrag: No freefrag."); 3939 freefrag->ff_state |= DEPCOMPLETE; 3940 freefrag->ff_jdep = NULL; 3941 jwork_insert(&freefrag->ff_jwork, jsegdep); 3942 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE) 3943 add_to_worklist(&freefrag->ff_list, 0); 3944 jfreefrag->fr_freefrag = NULL; 3945 free_jfreefrag(jfreefrag); 3946} 3947 3948/* 3949 * Called when the journal write for a jfreeblk completes. The jfreeblk 3950 * is removed from the freeblks list of pending journal writes and the 3951 * jsegdep is moved to the freeblks jwork to be completed when all blocks 3952 * have been reclaimed. 3953 */ 3954static void 3955handle_written_jblkdep(jblkdep) 3956 struct jblkdep *jblkdep; 3957{ 3958 struct freeblks *freeblks; 3959 struct jsegdep *jsegdep; 3960 3961 /* Grab the jsegdep. */ 3962 jsegdep = jblkdep->jb_jsegdep; 3963 jblkdep->jb_jsegdep = NULL; 3964 freeblks = jblkdep->jb_freeblks; 3965 LIST_REMOVE(jblkdep, jb_deps); 3966 jwork_insert(&freeblks->fb_jwork, jsegdep); 3967 /* 3968 * If the freeblks is all journaled, we can add it to the worklist. 3969 */ 3970 if (LIST_EMPTY(&freeblks->fb_jblkdephd) && 3971 (freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 3972 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 3973 3974 free_jblkdep(jblkdep); 3975} 3976 3977static struct jsegdep * 3978newjsegdep(struct worklist *wk) 3979{ 3980 struct jsegdep *jsegdep; 3981 3982 jsegdep = malloc(sizeof(*jsegdep), M_JSEGDEP, M_SOFTDEP_FLAGS); 3983 workitem_alloc(&jsegdep->jd_list, D_JSEGDEP, wk->wk_mp); 3984 jsegdep->jd_seg = NULL; 3985 3986 return (jsegdep); 3987} 3988 3989static struct jmvref * 3990newjmvref(dp, ino, oldoff, newoff) 3991 struct inode *dp; 3992 ino_t ino; 3993 off_t oldoff; 3994 off_t newoff; 3995{ 3996 struct jmvref *jmvref; 3997 3998 jmvref = malloc(sizeof(*jmvref), M_JMVREF, M_SOFTDEP_FLAGS); 3999 workitem_alloc(&jmvref->jm_list, D_JMVREF, ITOVFS(dp)); 4000 jmvref->jm_list.wk_state = ATTACHED | DEPCOMPLETE; 4001 jmvref->jm_parent = dp->i_number; 4002 jmvref->jm_ino = ino; 4003 jmvref->jm_oldoff = oldoff; 4004 jmvref->jm_newoff = newoff; 4005 4006 return (jmvref); 4007} 4008 4009/* 4010 * Allocate a new jremref that tracks the removal of ip from dp with the 4011 * directory entry offset of diroff. Mark the entry as ATTACHED and 4012 * DEPCOMPLETE as we have all the information required for the journal write 4013 * and the directory has already been removed from the buffer. The caller 4014 * is responsible for linking the jremref into the pagedep and adding it 4015 * to the journal to write. The MKDIR_PARENT flag is set if we're doing 4016 * a DOTDOT addition so handle_workitem_remove() can properly assign 4017 * the jsegdep when we're done. 4018 */ 4019static struct jremref * 4020newjremref(struct dirrem *dirrem, struct inode *dp, struct inode *ip, 4021 off_t diroff, nlink_t nlink) 4022{ 4023 struct jremref *jremref; 4024 4025 jremref = malloc(sizeof(*jremref), M_JREMREF, M_SOFTDEP_FLAGS); 4026 workitem_alloc(&jremref->jr_list, D_JREMREF, ITOVFS(dp)); 4027 jremref->jr_state = ATTACHED; 4028 newinoref(&jremref->jr_ref, ip->i_number, dp->i_number, diroff, 4029 nlink, ip->i_mode); 4030 jremref->jr_dirrem = dirrem; 4031 4032 return (jremref); 4033} 4034 4035static inline void 4036newinoref(struct inoref *inoref, ino_t ino, ino_t parent, off_t diroff, 4037 nlink_t nlink, uint16_t mode) 4038{ 4039 4040 inoref->if_jsegdep = newjsegdep(&inoref->if_list); 4041 inoref->if_diroff = diroff; 4042 inoref->if_ino = ino; 4043 inoref->if_parent = parent; 4044 inoref->if_nlink = nlink; 4045 inoref->if_mode = mode; 4046} 4047 4048/* 4049 * Allocate a new jaddref to track the addition of ino to dp at diroff. The 4050 * directory offset may not be known until later. The caller is responsible 4051 * adding the entry to the journal when this information is available. nlink 4052 * should be the link count prior to the addition and mode is only required 4053 * to have the correct FMT. 4054 */ 4055static struct jaddref * 4056newjaddref(struct inode *dp, ino_t ino, off_t diroff, int16_t nlink, 4057 uint16_t mode) 4058{ 4059 struct jaddref *jaddref; 4060 4061 jaddref = malloc(sizeof(*jaddref), M_JADDREF, M_SOFTDEP_FLAGS); 4062 workitem_alloc(&jaddref->ja_list, D_JADDREF, ITOVFS(dp)); 4063 jaddref->ja_state = ATTACHED; 4064 jaddref->ja_mkdir = NULL; 4065 newinoref(&jaddref->ja_ref, ino, dp->i_number, diroff, nlink, mode); 4066 4067 return (jaddref); 4068} 4069 4070/* 4071 * Create a new free dependency for a freework. The caller is responsible 4072 * for adjusting the reference count when it has the lock held. The freedep 4073 * will track an outstanding bitmap write that will ultimately clear the 4074 * freework to continue. 4075 */ 4076static struct freedep * 4077newfreedep(struct freework *freework) 4078{ 4079 struct freedep *freedep; 4080 4081 freedep = malloc(sizeof(*freedep), M_FREEDEP, M_SOFTDEP_FLAGS); 4082 workitem_alloc(&freedep->fd_list, D_FREEDEP, freework->fw_list.wk_mp); 4083 freedep->fd_freework = freework; 4084 4085 return (freedep); 4086} 4087 4088/* 4089 * Free a freedep structure once the buffer it is linked to is written. If 4090 * this is the last reference to the freework schedule it for completion. 4091 */ 4092static void 4093free_freedep(freedep) 4094 struct freedep *freedep; 4095{ 4096 struct freework *freework; 4097 4098 freework = freedep->fd_freework; 4099 freework->fw_freeblks->fb_cgwait--; 4100 if (--freework->fw_ref == 0) 4101 freework_enqueue(freework); 4102 WORKITEM_FREE(freedep, D_FREEDEP); 4103} 4104 4105/* 4106 * Allocate a new freework structure that may be a level in an indirect 4107 * when parent is not NULL or a top level block when it is. The top level 4108 * freework structures are allocated without the per-filesystem lock held 4109 * and before the freeblks is visible outside of softdep_setup_freeblocks(). 4110 */ 4111static struct freework * 4112newfreework(ump, freeblks, parent, lbn, nb, frags, off, journal) 4113 struct ufsmount *ump; 4114 struct freeblks *freeblks; 4115 struct freework *parent; 4116 ufs_lbn_t lbn; 4117 ufs2_daddr_t nb; 4118 int frags; 4119 int off; 4120 int journal; 4121{ 4122 struct freework *freework; 4123 4124 freework = malloc(sizeof(*freework), M_FREEWORK, M_SOFTDEP_FLAGS); 4125 workitem_alloc(&freework->fw_list, D_FREEWORK, freeblks->fb_list.wk_mp); 4126 freework->fw_state = ATTACHED; 4127 freework->fw_jnewblk = NULL; 4128 freework->fw_freeblks = freeblks; 4129 freework->fw_parent = parent; 4130 freework->fw_lbn = lbn; 4131 freework->fw_blkno = nb; 4132 freework->fw_frags = frags; 4133 freework->fw_indir = NULL; 4134 freework->fw_ref = (MOUNTEDSUJ(UFSTOVFS(ump)) == 0 || 4135 lbn >= -UFS_NXADDR) ? 0 : NINDIR(ump->um_fs) + 1; 4136 freework->fw_start = freework->fw_off = off; 4137 if (journal) 4138 newjfreeblk(freeblks, lbn, nb, frags); 4139 if (parent == NULL) { 4140 ACQUIRE_LOCK(ump); 4141 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list); 4142 freeblks->fb_ref++; 4143 FREE_LOCK(ump); 4144 } 4145 4146 return (freework); 4147} 4148 4149/* 4150 * Eliminate a jfreeblk for a block that does not need journaling. 4151 */ 4152static void 4153cancel_jfreeblk(freeblks, blkno) 4154 struct freeblks *freeblks; 4155 ufs2_daddr_t blkno; 4156{ 4157 struct jfreeblk *jfreeblk; 4158 struct jblkdep *jblkdep; 4159 4160 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) { 4161 if (jblkdep->jb_list.wk_type != D_JFREEBLK) 4162 continue; 4163 jfreeblk = WK_JFREEBLK(&jblkdep->jb_list); 4164 if (jfreeblk->jf_blkno == blkno) 4165 break; 4166 } 4167 if (jblkdep == NULL) 4168 return; 4169 CTR1(KTR_SUJ, "cancel_jfreeblk: blkno %jd", blkno); 4170 free_jsegdep(jblkdep->jb_jsegdep); 4171 LIST_REMOVE(jblkdep, jb_deps); 4172 WORKITEM_FREE(jfreeblk, D_JFREEBLK); 4173} 4174 4175/* 4176 * Allocate a new jfreeblk to journal top level block pointer when truncating 4177 * a file. The caller must add this to the worklist when the per-filesystem 4178 * lock is held. 4179 */ 4180static struct jfreeblk * 4181newjfreeblk(freeblks, lbn, blkno, frags) 4182 struct freeblks *freeblks; 4183 ufs_lbn_t lbn; 4184 ufs2_daddr_t blkno; 4185 int frags; 4186{ 4187 struct jfreeblk *jfreeblk; 4188 4189 jfreeblk = malloc(sizeof(*jfreeblk), M_JFREEBLK, M_SOFTDEP_FLAGS); 4190 workitem_alloc(&jfreeblk->jf_dep.jb_list, D_JFREEBLK, 4191 freeblks->fb_list.wk_mp); 4192 jfreeblk->jf_dep.jb_jsegdep = newjsegdep(&jfreeblk->jf_dep.jb_list); 4193 jfreeblk->jf_dep.jb_freeblks = freeblks; 4194 jfreeblk->jf_ino = freeblks->fb_inum; 4195 jfreeblk->jf_lbn = lbn; 4196 jfreeblk->jf_blkno = blkno; 4197 jfreeblk->jf_frags = frags; 4198 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jfreeblk->jf_dep, jb_deps); 4199 4200 return (jfreeblk); 4201} 4202 4203/* 4204 * The journal is only prepared to handle full-size block numbers, so we 4205 * have to adjust the record to reflect the change to a full-size block. 4206 * For example, suppose we have a block made up of fragments 8-15 and 4207 * want to free its last two fragments. We are given a request that says: 4208 * FREEBLK ino=5, blkno=14, lbn=0, frags=2, oldfrags=0 4209 * where frags are the number of fragments to free and oldfrags are the 4210 * number of fragments to keep. To block align it, we have to change it to 4211 * have a valid full-size blkno, so it becomes: 4212 * FREEBLK ino=5, blkno=8, lbn=0, frags=2, oldfrags=6 4213 */ 4214static void 4215adjust_newfreework(freeblks, frag_offset) 4216 struct freeblks *freeblks; 4217 int frag_offset; 4218{ 4219 struct jfreeblk *jfreeblk; 4220 4221 KASSERT((LIST_FIRST(&freeblks->fb_jblkdephd) != NULL && 4222 LIST_FIRST(&freeblks->fb_jblkdephd)->jb_list.wk_type == D_JFREEBLK), 4223 ("adjust_newfreework: Missing freeblks dependency")); 4224 4225 jfreeblk = WK_JFREEBLK(LIST_FIRST(&freeblks->fb_jblkdephd)); 4226 jfreeblk->jf_blkno -= frag_offset; 4227 jfreeblk->jf_frags += frag_offset; 4228} 4229 4230/* 4231 * Allocate a new jtrunc to track a partial truncation. 4232 */ 4233static struct jtrunc * 4234newjtrunc(freeblks, size, extsize) 4235 struct freeblks *freeblks; 4236 off_t size; 4237 int extsize; 4238{ 4239 struct jtrunc *jtrunc; 4240 4241 jtrunc = malloc(sizeof(*jtrunc), M_JTRUNC, M_SOFTDEP_FLAGS); 4242 workitem_alloc(&jtrunc->jt_dep.jb_list, D_JTRUNC, 4243 freeblks->fb_list.wk_mp); 4244 jtrunc->jt_dep.jb_jsegdep = newjsegdep(&jtrunc->jt_dep.jb_list); 4245 jtrunc->jt_dep.jb_freeblks = freeblks; 4246 jtrunc->jt_ino = freeblks->fb_inum; 4247 jtrunc->jt_size = size; 4248 jtrunc->jt_extsize = extsize; 4249 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jtrunc->jt_dep, jb_deps); 4250 4251 return (jtrunc); 4252} 4253 4254/* 4255 * If we're canceling a new bitmap we have to search for another ref 4256 * to move into the bmsafemap dep. This might be better expressed 4257 * with another structure. 4258 */ 4259static void 4260move_newblock_dep(jaddref, inodedep) 4261 struct jaddref *jaddref; 4262 struct inodedep *inodedep; 4263{ 4264 struct inoref *inoref; 4265 struct jaddref *jaddrefn; 4266 4267 jaddrefn = NULL; 4268 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref; 4269 inoref = TAILQ_NEXT(inoref, if_deps)) { 4270 if ((jaddref->ja_state & NEWBLOCK) && 4271 inoref->if_list.wk_type == D_JADDREF) { 4272 jaddrefn = (struct jaddref *)inoref; 4273 break; 4274 } 4275 } 4276 if (jaddrefn == NULL) 4277 return; 4278 jaddrefn->ja_state &= ~(ATTACHED | UNDONE); 4279 jaddrefn->ja_state |= jaddref->ja_state & 4280 (ATTACHED | UNDONE | NEWBLOCK); 4281 jaddref->ja_state &= ~(ATTACHED | UNDONE | NEWBLOCK); 4282 jaddref->ja_state |= ATTACHED; 4283 LIST_REMOVE(jaddref, ja_bmdeps); 4284 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_jaddrefhd, jaddrefn, 4285 ja_bmdeps); 4286} 4287 4288/* 4289 * Cancel a jaddref either before it has been written or while it is being 4290 * written. This happens when a link is removed before the add reaches 4291 * the disk. The jaddref dependency is kept linked into the bmsafemap 4292 * and inode to prevent the link count or bitmap from reaching the disk 4293 * until handle_workitem_remove() re-adjusts the counts and bitmaps as 4294 * required. 4295 * 4296 * Returns 1 if the canceled addref requires journaling of the remove and 4297 * 0 otherwise. 4298 */ 4299static int 4300cancel_jaddref(jaddref, inodedep, wkhd) 4301 struct jaddref *jaddref; 4302 struct inodedep *inodedep; 4303 struct workhead *wkhd; 4304{ 4305 struct inoref *inoref; 4306 struct jsegdep *jsegdep; 4307 int needsj; 4308 4309 KASSERT((jaddref->ja_state & COMPLETE) == 0, 4310 ("cancel_jaddref: Canceling complete jaddref")); 4311 if (jaddref->ja_state & (INPROGRESS | COMPLETE)) 4312 needsj = 1; 4313 else 4314 needsj = 0; 4315 if (inodedep == NULL) 4316 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino, 4317 0, &inodedep) == 0) 4318 panic("cancel_jaddref: Lost inodedep"); 4319 /* 4320 * We must adjust the nlink of any reference operation that follows 4321 * us so that it is consistent with the in-memory reference. This 4322 * ensures that inode nlink rollbacks always have the correct link. 4323 */ 4324 if (needsj == 0) { 4325 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref; 4326 inoref = TAILQ_NEXT(inoref, if_deps)) { 4327 if (inoref->if_state & GOINGAWAY) 4328 break; 4329 inoref->if_nlink--; 4330 } 4331 } 4332 jsegdep = inoref_jseg(&jaddref->ja_ref); 4333 if (jaddref->ja_state & NEWBLOCK) 4334 move_newblock_dep(jaddref, inodedep); 4335 wake_worklist(&jaddref->ja_list); 4336 jaddref->ja_mkdir = NULL; 4337 if (jaddref->ja_state & INPROGRESS) { 4338 jaddref->ja_state &= ~INPROGRESS; 4339 WORKLIST_REMOVE(&jaddref->ja_list); 4340 jwork_insert(wkhd, jsegdep); 4341 } else { 4342 free_jsegdep(jsegdep); 4343 if (jaddref->ja_state & DEPCOMPLETE) 4344 remove_from_journal(&jaddref->ja_list); 4345 } 4346 jaddref->ja_state |= (GOINGAWAY | DEPCOMPLETE); 4347 /* 4348 * Leave NEWBLOCK jaddrefs on the inodedep so handle_workitem_remove 4349 * can arrange for them to be freed with the bitmap. Otherwise we 4350 * no longer need this addref attached to the inoreflst and it 4351 * will incorrectly adjust nlink if we leave it. 4352 */ 4353 if ((jaddref->ja_state & NEWBLOCK) == 0) { 4354 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, 4355 if_deps); 4356 jaddref->ja_state |= COMPLETE; 4357 free_jaddref(jaddref); 4358 return (needsj); 4359 } 4360 /* 4361 * Leave the head of the list for jsegdeps for fast merging. 4362 */ 4363 if (LIST_FIRST(wkhd) != NULL) { 4364 jaddref->ja_state |= ONWORKLIST; 4365 LIST_INSERT_AFTER(LIST_FIRST(wkhd), &jaddref->ja_list, wk_list); 4366 } else 4367 WORKLIST_INSERT(wkhd, &jaddref->ja_list); 4368 4369 return (needsj); 4370} 4371 4372/* 4373 * Attempt to free a jaddref structure when some work completes. This 4374 * should only succeed once the entry is written and all dependencies have 4375 * been notified. 4376 */ 4377static void 4378free_jaddref(jaddref) 4379 struct jaddref *jaddref; 4380{ 4381 4382 if ((jaddref->ja_state & ALLCOMPLETE) != ALLCOMPLETE) 4383 return; 4384 if (jaddref->ja_ref.if_jsegdep) 4385 panic("free_jaddref: segdep attached to jaddref %p(0x%X)\n", 4386 jaddref, jaddref->ja_state); 4387 if (jaddref->ja_state & NEWBLOCK) 4388 LIST_REMOVE(jaddref, ja_bmdeps); 4389 if (jaddref->ja_state & (INPROGRESS | ONWORKLIST)) 4390 panic("free_jaddref: Bad state %p(0x%X)", 4391 jaddref, jaddref->ja_state); 4392 if (jaddref->ja_mkdir != NULL) 4393 panic("free_jaddref: Work pending, 0x%X\n", jaddref->ja_state); 4394 WORKITEM_FREE(jaddref, D_JADDREF); 4395} 4396 4397/* 4398 * Free a jremref structure once it has been written or discarded. 4399 */ 4400static void 4401free_jremref(jremref) 4402 struct jremref *jremref; 4403{ 4404 4405 if (jremref->jr_ref.if_jsegdep) 4406 free_jsegdep(jremref->jr_ref.if_jsegdep); 4407 if (jremref->jr_state & INPROGRESS) 4408 panic("free_jremref: IO still pending"); 4409 WORKITEM_FREE(jremref, D_JREMREF); 4410} 4411 4412/* 4413 * Free a jnewblk structure. 4414 */ 4415static void 4416free_jnewblk(jnewblk) 4417 struct jnewblk *jnewblk; 4418{ 4419 4420 if ((jnewblk->jn_state & ALLCOMPLETE) != ALLCOMPLETE) 4421 return; 4422 LIST_REMOVE(jnewblk, jn_deps); 4423 if (jnewblk->jn_dep != NULL) 4424 panic("free_jnewblk: Dependency still attached."); 4425 WORKITEM_FREE(jnewblk, D_JNEWBLK); 4426} 4427 4428/* 4429 * Cancel a jnewblk which has been been made redundant by frag extension. 4430 */ 4431static void 4432cancel_jnewblk(jnewblk, wkhd) 4433 struct jnewblk *jnewblk; 4434 struct workhead *wkhd; 4435{ 4436 struct jsegdep *jsegdep; 4437 4438 CTR1(KTR_SUJ, "cancel_jnewblk: blkno %jd", jnewblk->jn_blkno); 4439 jsegdep = jnewblk->jn_jsegdep; 4440 if (jnewblk->jn_jsegdep == NULL || jnewblk->jn_dep == NULL) 4441 panic("cancel_jnewblk: Invalid state"); 4442 jnewblk->jn_jsegdep = NULL; 4443 jnewblk->jn_dep = NULL; 4444 jnewblk->jn_state |= GOINGAWAY; 4445 if (jnewblk->jn_state & INPROGRESS) { 4446 jnewblk->jn_state &= ~INPROGRESS; 4447 WORKLIST_REMOVE(&jnewblk->jn_list); 4448 jwork_insert(wkhd, jsegdep); 4449 } else { 4450 free_jsegdep(jsegdep); 4451 remove_from_journal(&jnewblk->jn_list); 4452 } 4453 wake_worklist(&jnewblk->jn_list); 4454 WORKLIST_INSERT(wkhd, &jnewblk->jn_list); 4455} 4456 4457static void 4458free_jblkdep(jblkdep) 4459 struct jblkdep *jblkdep; 4460{ 4461 4462 if (jblkdep->jb_list.wk_type == D_JFREEBLK) 4463 WORKITEM_FREE(jblkdep, D_JFREEBLK); 4464 else if (jblkdep->jb_list.wk_type == D_JTRUNC) 4465 WORKITEM_FREE(jblkdep, D_JTRUNC); 4466 else 4467 panic("free_jblkdep: Unexpected type %s", 4468 TYPENAME(jblkdep->jb_list.wk_type)); 4469} 4470 4471/* 4472 * Free a single jseg once it is no longer referenced in memory or on 4473 * disk. Reclaim journal blocks and dependencies waiting for the segment 4474 * to disappear. 4475 */ 4476static void 4477free_jseg(jseg, jblocks) 4478 struct jseg *jseg; 4479 struct jblocks *jblocks; 4480{ 4481 struct freework *freework; 4482 4483 /* 4484 * Free freework structures that were lingering to indicate freed 4485 * indirect blocks that forced journal write ordering on reallocate. 4486 */ 4487 while ((freework = LIST_FIRST(&jseg->js_indirs)) != NULL) 4488 indirblk_remove(freework); 4489 if (jblocks->jb_oldestseg == jseg) 4490 jblocks->jb_oldestseg = TAILQ_NEXT(jseg, js_next); 4491 TAILQ_REMOVE(&jblocks->jb_segs, jseg, js_next); 4492 jblocks_free(jblocks, jseg->js_list.wk_mp, jseg->js_size); 4493 KASSERT(LIST_EMPTY(&jseg->js_entries), 4494 ("free_jseg: Freed jseg has valid entries.")); 4495 WORKITEM_FREE(jseg, D_JSEG); 4496} 4497 4498/* 4499 * Free all jsegs that meet the criteria for being reclaimed and update 4500 * oldestseg. 4501 */ 4502static void 4503free_jsegs(jblocks) 4504 struct jblocks *jblocks; 4505{ 4506 struct jseg *jseg; 4507 4508 /* 4509 * Free only those jsegs which have none allocated before them to 4510 * preserve the journal space ordering. 4511 */ 4512 while ((jseg = TAILQ_FIRST(&jblocks->jb_segs)) != NULL) { 4513 /* 4514 * Only reclaim space when nothing depends on this journal 4515 * set and another set has written that it is no longer 4516 * valid. 4517 */ 4518 if (jseg->js_refs != 0) { 4519 jblocks->jb_oldestseg = jseg; 4520 return; 4521 } 4522 if ((jseg->js_state & ALLCOMPLETE) != ALLCOMPLETE) 4523 break; 4524 if (jseg->js_seq > jblocks->jb_oldestwrseq) 4525 break; 4526 /* 4527 * We can free jsegs that didn't write entries when 4528 * oldestwrseq == js_seq. 4529 */ 4530 if (jseg->js_seq == jblocks->jb_oldestwrseq && 4531 jseg->js_cnt != 0) 4532 break; 4533 free_jseg(jseg, jblocks); 4534 } 4535 /* 4536 * If we exited the loop above we still must discover the 4537 * oldest valid segment. 4538 */ 4539 if (jseg) 4540 for (jseg = jblocks->jb_oldestseg; jseg != NULL; 4541 jseg = TAILQ_NEXT(jseg, js_next)) 4542 if (jseg->js_refs != 0) 4543 break; 4544 jblocks->jb_oldestseg = jseg; 4545 /* 4546 * The journal has no valid records but some jsegs may still be 4547 * waiting on oldestwrseq to advance. We force a small record 4548 * out to permit these lingering records to be reclaimed. 4549 */ 4550 if (jblocks->jb_oldestseg == NULL && !TAILQ_EMPTY(&jblocks->jb_segs)) 4551 jblocks->jb_needseg = 1; 4552} 4553 4554/* 4555 * Release one reference to a jseg and free it if the count reaches 0. This 4556 * should eventually reclaim journal space as well. 4557 */ 4558static void 4559rele_jseg(jseg) 4560 struct jseg *jseg; 4561{ 4562 4563 KASSERT(jseg->js_refs > 0, 4564 ("free_jseg: Invalid refcnt %d", jseg->js_refs)); 4565 if (--jseg->js_refs != 0) 4566 return; 4567 free_jsegs(jseg->js_jblocks); 4568} 4569 4570/* 4571 * Release a jsegdep and decrement the jseg count. 4572 */ 4573static void 4574free_jsegdep(jsegdep) 4575 struct jsegdep *jsegdep; 4576{ 4577 4578 if (jsegdep->jd_seg) 4579 rele_jseg(jsegdep->jd_seg); 4580 WORKITEM_FREE(jsegdep, D_JSEGDEP); 4581} 4582 4583/* 4584 * Wait for a journal item to make it to disk. Initiate journal processing 4585 * if required. 4586 */ 4587static int 4588jwait(wk, waitfor) 4589 struct worklist *wk; 4590 int waitfor; 4591{ 4592 4593 LOCK_OWNED(VFSTOUFS(wk->wk_mp)); 4594 /* 4595 * Blocking journal waits cause slow synchronous behavior. Record 4596 * stats on the frequency of these blocking operations. 4597 */ 4598 if (waitfor == MNT_WAIT) { 4599 stat_journal_wait++; 4600 switch (wk->wk_type) { 4601 case D_JREMREF: 4602 case D_JMVREF: 4603 stat_jwait_filepage++; 4604 break; 4605 case D_JTRUNC: 4606 case D_JFREEBLK: 4607 stat_jwait_freeblks++; 4608 break; 4609 case D_JNEWBLK: 4610 stat_jwait_newblk++; 4611 break; 4612 case D_JADDREF: 4613 stat_jwait_inode++; 4614 break; 4615 default: 4616 break; 4617 } 4618 } 4619 /* 4620 * If IO has not started we process the journal. We can't mark the 4621 * worklist item as IOWAITING because we drop the lock while 4622 * processing the journal and the worklist entry may be freed after 4623 * this point. The caller may call back in and re-issue the request. 4624 */ 4625 if ((wk->wk_state & INPROGRESS) == 0) { 4626 softdep_process_journal(wk->wk_mp, wk, waitfor); 4627 if (waitfor != MNT_WAIT) 4628 return (EBUSY); 4629 return (0); 4630 } 4631 if (waitfor != MNT_WAIT) 4632 return (EBUSY); 4633 wait_worklist(wk, "jwait"); 4634 return (0); 4635} 4636 4637/* 4638 * Lookup an inodedep based on an inode pointer and set the nlinkdelta as 4639 * appropriate. This is a convenience function to reduce duplicate code 4640 * for the setup and revert functions below. 4641 */ 4642static struct inodedep * 4643inodedep_lookup_ip(ip) 4644 struct inode *ip; 4645{ 4646 struct inodedep *inodedep; 4647 4648 KASSERT(ip->i_nlink >= ip->i_effnlink, 4649 ("inodedep_lookup_ip: bad delta")); 4650 (void) inodedep_lookup(ITOVFS(ip), ip->i_number, DEPALLOC, 4651 &inodedep); 4652 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 4653 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked")); 4654 4655 return (inodedep); 4656} 4657 4658/* 4659 * Called prior to creating a new inode and linking it to a directory. The 4660 * jaddref structure must already be allocated by softdep_setup_inomapdep 4661 * and it is discovered here so we can initialize the mode and update 4662 * nlinkdelta. 4663 */ 4664void 4665softdep_setup_create(dp, ip) 4666 struct inode *dp; 4667 struct inode *ip; 4668{ 4669 struct inodedep *inodedep; 4670 struct jaddref *jaddref; 4671 struct vnode *dvp; 4672 4673 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4674 ("softdep_setup_create called on non-softdep filesystem")); 4675 KASSERT(ip->i_nlink == 1, 4676 ("softdep_setup_create: Invalid link count.")); 4677 dvp = ITOV(dp); 4678 ACQUIRE_LOCK(ITOUMP(dp)); 4679 inodedep = inodedep_lookup_ip(ip); 4680 if (DOINGSUJ(dvp)) { 4681 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4682 inoreflst); 4683 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 4684 ("softdep_setup_create: No addref structure present.")); 4685 } 4686 softdep_prelink(dvp, NULL); 4687 FREE_LOCK(ITOUMP(dp)); 4688} 4689 4690/* 4691 * Create a jaddref structure to track the addition of a DOTDOT link when 4692 * we are reparenting an inode as part of a rename. This jaddref will be 4693 * found by softdep_setup_directory_change. Adjusts nlinkdelta for 4694 * non-journaling softdep. 4695 */ 4696void 4697softdep_setup_dotdot_link(dp, ip) 4698 struct inode *dp; 4699 struct inode *ip; 4700{ 4701 struct inodedep *inodedep; 4702 struct jaddref *jaddref; 4703 struct vnode *dvp; 4704 4705 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4706 ("softdep_setup_dotdot_link called on non-softdep filesystem")); 4707 dvp = ITOV(dp); 4708 jaddref = NULL; 4709 /* 4710 * We don't set MKDIR_PARENT as this is not tied to a mkdir and 4711 * is used as a normal link would be. 4712 */ 4713 if (DOINGSUJ(dvp)) 4714 jaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET, 4715 dp->i_effnlink - 1, dp->i_mode); 4716 ACQUIRE_LOCK(ITOUMP(dp)); 4717 inodedep = inodedep_lookup_ip(dp); 4718 if (jaddref) 4719 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 4720 if_deps); 4721 softdep_prelink(dvp, ITOV(ip)); 4722 FREE_LOCK(ITOUMP(dp)); 4723} 4724 4725/* 4726 * Create a jaddref structure to track a new link to an inode. The directory 4727 * offset is not known until softdep_setup_directory_add or 4728 * softdep_setup_directory_change. Adjusts nlinkdelta for non-journaling 4729 * softdep. 4730 */ 4731void 4732softdep_setup_link(dp, ip) 4733 struct inode *dp; 4734 struct inode *ip; 4735{ 4736 struct inodedep *inodedep; 4737 struct jaddref *jaddref; 4738 struct vnode *dvp; 4739 4740 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4741 ("softdep_setup_link called on non-softdep filesystem")); 4742 dvp = ITOV(dp); 4743 jaddref = NULL; 4744 if (DOINGSUJ(dvp)) 4745 jaddref = newjaddref(dp, ip->i_number, 0, ip->i_effnlink - 1, 4746 ip->i_mode); 4747 ACQUIRE_LOCK(ITOUMP(dp)); 4748 inodedep = inodedep_lookup_ip(ip); 4749 if (jaddref) 4750 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 4751 if_deps); 4752 softdep_prelink(dvp, ITOV(ip)); 4753 FREE_LOCK(ITOUMP(dp)); 4754} 4755 4756/* 4757 * Called to create the jaddref structures to track . and .. references as 4758 * well as lookup and further initialize the incomplete jaddref created 4759 * by softdep_setup_inomapdep when the inode was allocated. Adjusts 4760 * nlinkdelta for non-journaling softdep. 4761 */ 4762void 4763softdep_setup_mkdir(dp, ip) 4764 struct inode *dp; 4765 struct inode *ip; 4766{ 4767 struct inodedep *inodedep; 4768 struct jaddref *dotdotaddref; 4769 struct jaddref *dotaddref; 4770 struct jaddref *jaddref; 4771 struct vnode *dvp; 4772 4773 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4774 ("softdep_setup_mkdir called on non-softdep filesystem")); 4775 dvp = ITOV(dp); 4776 dotaddref = dotdotaddref = NULL; 4777 if (DOINGSUJ(dvp)) { 4778 dotaddref = newjaddref(ip, ip->i_number, DOT_OFFSET, 1, 4779 ip->i_mode); 4780 dotaddref->ja_state |= MKDIR_BODY; 4781 dotdotaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET, 4782 dp->i_effnlink - 1, dp->i_mode); 4783 dotdotaddref->ja_state |= MKDIR_PARENT; 4784 } 4785 ACQUIRE_LOCK(ITOUMP(dp)); 4786 inodedep = inodedep_lookup_ip(ip); 4787 if (DOINGSUJ(dvp)) { 4788 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4789 inoreflst); 4790 KASSERT(jaddref != NULL, 4791 ("softdep_setup_mkdir: No addref structure present.")); 4792 KASSERT(jaddref->ja_parent == dp->i_number, 4793 ("softdep_setup_mkdir: bad parent %ju", 4794 (uintmax_t)jaddref->ja_parent)); 4795 TAILQ_INSERT_BEFORE(&jaddref->ja_ref, &dotaddref->ja_ref, 4796 if_deps); 4797 } 4798 inodedep = inodedep_lookup_ip(dp); 4799 if (DOINGSUJ(dvp)) 4800 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, 4801 &dotdotaddref->ja_ref, if_deps); 4802 softdep_prelink(ITOV(dp), NULL); 4803 FREE_LOCK(ITOUMP(dp)); 4804} 4805 4806/* 4807 * Called to track nlinkdelta of the inode and parent directories prior to 4808 * unlinking a directory. 4809 */ 4810void 4811softdep_setup_rmdir(dp, ip) 4812 struct inode *dp; 4813 struct inode *ip; 4814{ 4815 struct vnode *dvp; 4816 4817 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4818 ("softdep_setup_rmdir called on non-softdep filesystem")); 4819 dvp = ITOV(dp); 4820 ACQUIRE_LOCK(ITOUMP(dp)); 4821 (void) inodedep_lookup_ip(ip); 4822 (void) inodedep_lookup_ip(dp); 4823 softdep_prelink(dvp, ITOV(ip)); 4824 FREE_LOCK(ITOUMP(dp)); 4825} 4826 4827/* 4828 * Called to track nlinkdelta of the inode and parent directories prior to 4829 * unlink. 4830 */ 4831void 4832softdep_setup_unlink(dp, ip) 4833 struct inode *dp; 4834 struct inode *ip; 4835{ 4836 struct vnode *dvp; 4837 4838 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4839 ("softdep_setup_unlink called on non-softdep filesystem")); 4840 dvp = ITOV(dp); 4841 ACQUIRE_LOCK(ITOUMP(dp)); 4842 (void) inodedep_lookup_ip(ip); 4843 (void) inodedep_lookup_ip(dp); 4844 softdep_prelink(dvp, ITOV(ip)); 4845 FREE_LOCK(ITOUMP(dp)); 4846} 4847 4848/* 4849 * Called to release the journal structures created by a failed non-directory 4850 * creation. Adjusts nlinkdelta for non-journaling softdep. 4851 */ 4852void 4853softdep_revert_create(dp, ip) 4854 struct inode *dp; 4855 struct inode *ip; 4856{ 4857 struct inodedep *inodedep; 4858 struct jaddref *jaddref; 4859 struct vnode *dvp; 4860 4861 KASSERT(MOUNTEDSOFTDEP(ITOVFS((dp))) != 0, 4862 ("softdep_revert_create called on non-softdep filesystem")); 4863 dvp = ITOV(dp); 4864 ACQUIRE_LOCK(ITOUMP(dp)); 4865 inodedep = inodedep_lookup_ip(ip); 4866 if (DOINGSUJ(dvp)) { 4867 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4868 inoreflst); 4869 KASSERT(jaddref->ja_parent == dp->i_number, 4870 ("softdep_revert_create: addref parent mismatch")); 4871 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4872 } 4873 FREE_LOCK(ITOUMP(dp)); 4874} 4875 4876/* 4877 * Called to release the journal structures created by a failed link 4878 * addition. Adjusts nlinkdelta for non-journaling softdep. 4879 */ 4880void 4881softdep_revert_link(dp, ip) 4882 struct inode *dp; 4883 struct inode *ip; 4884{ 4885 struct inodedep *inodedep; 4886 struct jaddref *jaddref; 4887 struct vnode *dvp; 4888 4889 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4890 ("softdep_revert_link called on non-softdep filesystem")); 4891 dvp = ITOV(dp); 4892 ACQUIRE_LOCK(ITOUMP(dp)); 4893 inodedep = inodedep_lookup_ip(ip); 4894 if (DOINGSUJ(dvp)) { 4895 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4896 inoreflst); 4897 KASSERT(jaddref->ja_parent == dp->i_number, 4898 ("softdep_revert_link: addref parent mismatch")); 4899 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4900 } 4901 FREE_LOCK(ITOUMP(dp)); 4902} 4903 4904/* 4905 * Called to release the journal structures created by a failed mkdir 4906 * attempt. Adjusts nlinkdelta for non-journaling softdep. 4907 */ 4908void 4909softdep_revert_mkdir(dp, ip) 4910 struct inode *dp; 4911 struct inode *ip; 4912{ 4913 struct inodedep *inodedep; 4914 struct jaddref *jaddref; 4915 struct jaddref *dotaddref; 4916 struct vnode *dvp; 4917 4918 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4919 ("softdep_revert_mkdir called on non-softdep filesystem")); 4920 dvp = ITOV(dp); 4921 4922 ACQUIRE_LOCK(ITOUMP(dp)); 4923 inodedep = inodedep_lookup_ip(dp); 4924 if (DOINGSUJ(dvp)) { 4925 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4926 inoreflst); 4927 KASSERT(jaddref->ja_parent == ip->i_number, 4928 ("softdep_revert_mkdir: dotdot addref parent mismatch")); 4929 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4930 } 4931 inodedep = inodedep_lookup_ip(ip); 4932 if (DOINGSUJ(dvp)) { 4933 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4934 inoreflst); 4935 KASSERT(jaddref->ja_parent == dp->i_number, 4936 ("softdep_revert_mkdir: addref parent mismatch")); 4937 dotaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref, 4938 inoreflst, if_deps); 4939 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4940 KASSERT(dotaddref->ja_parent == ip->i_number, 4941 ("softdep_revert_mkdir: dot addref parent mismatch")); 4942 cancel_jaddref(dotaddref, inodedep, &inodedep->id_inowait); 4943 } 4944 FREE_LOCK(ITOUMP(dp)); 4945} 4946 4947/* 4948 * Called to correct nlinkdelta after a failed rmdir. 4949 */ 4950void 4951softdep_revert_rmdir(dp, ip) 4952 struct inode *dp; 4953 struct inode *ip; 4954{ 4955 4956 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4957 ("softdep_revert_rmdir called on non-softdep filesystem")); 4958 ACQUIRE_LOCK(ITOUMP(dp)); 4959 (void) inodedep_lookup_ip(ip); 4960 (void) inodedep_lookup_ip(dp); 4961 FREE_LOCK(ITOUMP(dp)); 4962} 4963 4964/* 4965 * Protecting the freemaps (or bitmaps). 4966 * 4967 * To eliminate the need to execute fsck before mounting a filesystem 4968 * after a power failure, one must (conservatively) guarantee that the 4969 * on-disk copy of the bitmaps never indicate that a live inode or block is 4970 * free. So, when a block or inode is allocated, the bitmap should be 4971 * updated (on disk) before any new pointers. When a block or inode is 4972 * freed, the bitmap should not be updated until all pointers have been 4973 * reset. The latter dependency is handled by the delayed de-allocation 4974 * approach described below for block and inode de-allocation. The former 4975 * dependency is handled by calling the following procedure when a block or 4976 * inode is allocated. When an inode is allocated an "inodedep" is created 4977 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 4978 * Each "inodedep" is also inserted into the hash indexing structure so 4979 * that any additional link additions can be made dependent on the inode 4980 * allocation. 4981 * 4982 * The ufs filesystem maintains a number of free block counts (e.g., per 4983 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 4984 * in addition to the bitmaps. These counts are used to improve efficiency 4985 * during allocation and therefore must be consistent with the bitmaps. 4986 * There is no convenient way to guarantee post-crash consistency of these 4987 * counts with simple update ordering, for two main reasons: (1) The counts 4988 * and bitmaps for a single cylinder group block are not in the same disk 4989 * sector. If a disk write is interrupted (e.g., by power failure), one may 4990 * be written and the other not. (2) Some of the counts are located in the 4991 * superblock rather than the cylinder group block. So, we focus our soft 4992 * updates implementation on protecting the bitmaps. When mounting a 4993 * filesystem, we recompute the auxiliary counts from the bitmaps. 4994 */ 4995 4996/* 4997 * Called just after updating the cylinder group block to allocate an inode. 4998 */ 4999void 5000softdep_setup_inomapdep(bp, ip, newinum, mode) 5001 struct buf *bp; /* buffer for cylgroup block with inode map */ 5002 struct inode *ip; /* inode related to allocation */ 5003 ino_t newinum; /* new inode number being allocated */ 5004 int mode; 5005{ 5006 struct inodedep *inodedep; 5007 struct bmsafemap *bmsafemap; 5008 struct jaddref *jaddref; 5009 struct mount *mp; 5010 struct fs *fs; 5011 5012 mp = ITOVFS(ip); 5013 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5014 ("softdep_setup_inomapdep called on non-softdep filesystem")); 5015 fs = VFSTOUFS(mp)->um_fs; 5016 jaddref = NULL; 5017 5018 /* 5019 * Allocate the journal reference add structure so that the bitmap 5020 * can be dependent on it. 5021 */ 5022 if (MOUNTEDSUJ(mp)) { 5023 jaddref = newjaddref(ip, newinum, 0, 0, mode); 5024 jaddref->ja_state |= NEWBLOCK; 5025 } 5026 5027 /* 5028 * Create a dependency for the newly allocated inode. 5029 * Panic if it already exists as something is seriously wrong. 5030 * Otherwise add it to the dependency list for the buffer holding 5031 * the cylinder group map from which it was allocated. 5032 * 5033 * We have to preallocate a bmsafemap entry in case it is needed 5034 * in bmsafemap_lookup since once we allocate the inodedep, we 5035 * have to finish initializing it before we can FREE_LOCK(). 5036 * By preallocating, we avoid FREE_LOCK() while doing a malloc 5037 * in bmsafemap_lookup. We cannot call bmsafemap_lookup before 5038 * creating the inodedep as it can be freed during the time 5039 * that we FREE_LOCK() while allocating the inodedep. We must 5040 * call workitem_alloc() before entering the locked section as 5041 * it also acquires the lock and we must avoid trying doing so 5042 * recursively. 5043 */ 5044 bmsafemap = malloc(sizeof(struct bmsafemap), 5045 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 5046 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); 5047 ACQUIRE_LOCK(ITOUMP(ip)); 5048 if ((inodedep_lookup(mp, newinum, DEPALLOC, &inodedep))) 5049 panic("softdep_setup_inomapdep: dependency %p for new" 5050 "inode already exists", inodedep); 5051 bmsafemap = bmsafemap_lookup(mp, bp, ino_to_cg(fs, newinum), bmsafemap); 5052 if (jaddref) { 5053 LIST_INSERT_HEAD(&bmsafemap->sm_jaddrefhd, jaddref, ja_bmdeps); 5054 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 5055 if_deps); 5056 } else { 5057 inodedep->id_state |= ONDEPLIST; 5058 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 5059 } 5060 inodedep->id_bmsafemap = bmsafemap; 5061 inodedep->id_state &= ~DEPCOMPLETE; 5062 FREE_LOCK(ITOUMP(ip)); 5063} 5064 5065/* 5066 * Called just after updating the cylinder group block to 5067 * allocate block or fragment. 5068 */ 5069void 5070softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags) 5071 struct buf *bp; /* buffer for cylgroup block with block map */ 5072 struct mount *mp; /* filesystem doing allocation */ 5073 ufs2_daddr_t newblkno; /* number of newly allocated block */ 5074 int frags; /* Number of fragments. */ 5075 int oldfrags; /* Previous number of fragments for extend. */ 5076{ 5077 struct newblk *newblk; 5078 struct bmsafemap *bmsafemap; 5079 struct jnewblk *jnewblk; 5080 struct ufsmount *ump; 5081 struct fs *fs; 5082 5083 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5084 ("softdep_setup_blkmapdep called on non-softdep filesystem")); 5085 ump = VFSTOUFS(mp); 5086 fs = ump->um_fs; 5087 jnewblk = NULL; 5088 /* 5089 * Create a dependency for the newly allocated block. 5090 * Add it to the dependency list for the buffer holding 5091 * the cylinder group map from which it was allocated. 5092 */ 5093 if (MOUNTEDSUJ(mp)) { 5094 jnewblk = malloc(sizeof(*jnewblk), M_JNEWBLK, M_SOFTDEP_FLAGS); 5095 workitem_alloc(&jnewblk->jn_list, D_JNEWBLK, mp); 5096 jnewblk->jn_jsegdep = newjsegdep(&jnewblk->jn_list); 5097 jnewblk->jn_state = ATTACHED; 5098 jnewblk->jn_blkno = newblkno; 5099 jnewblk->jn_frags = frags; 5100 jnewblk->jn_oldfrags = oldfrags; 5101#ifdef SUJ_DEBUG 5102 { 5103 struct cg *cgp; 5104 uint8_t *blksfree; 5105 long bno; 5106 int i; 5107 5108 cgp = (struct cg *)bp->b_data; 5109 blksfree = cg_blksfree(cgp); 5110 bno = dtogd(fs, jnewblk->jn_blkno); 5111 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; 5112 i++) { 5113 if (isset(blksfree, bno + i)) 5114 panic("softdep_setup_blkmapdep: " 5115 "free fragment %d from %d-%d " 5116 "state 0x%X dep %p", i, 5117 jnewblk->jn_oldfrags, 5118 jnewblk->jn_frags, 5119 jnewblk->jn_state, 5120 jnewblk->jn_dep); 5121 } 5122 } 5123#endif 5124 } 5125 5126 CTR3(KTR_SUJ, 5127 "softdep_setup_blkmapdep: blkno %jd frags %d oldfrags %d", 5128 newblkno, frags, oldfrags); 5129 ACQUIRE_LOCK(ump); 5130 if (newblk_lookup(mp, newblkno, DEPALLOC, &newblk) != 0) 5131 panic("softdep_setup_blkmapdep: found block"); 5132 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp, 5133 dtog(fs, newblkno), NULL); 5134 if (jnewblk) { 5135 jnewblk->jn_dep = (struct worklist *)newblk; 5136 LIST_INSERT_HEAD(&bmsafemap->sm_jnewblkhd, jnewblk, jn_deps); 5137 } else { 5138 newblk->nb_state |= ONDEPLIST; 5139 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 5140 } 5141 newblk->nb_bmsafemap = bmsafemap; 5142 newblk->nb_jnewblk = jnewblk; 5143 FREE_LOCK(ump); 5144} 5145 5146#define BMSAFEMAP_HASH(ump, cg) \ 5147 (&(ump)->bmsafemap_hashtbl[(cg) & (ump)->bmsafemap_hash_size]) 5148 5149static int 5150bmsafemap_find(bmsafemaphd, cg, bmsafemapp) 5151 struct bmsafemap_hashhead *bmsafemaphd; 5152 int cg; 5153 struct bmsafemap **bmsafemapp; 5154{ 5155 struct bmsafemap *bmsafemap; 5156 5157 LIST_FOREACH(bmsafemap, bmsafemaphd, sm_hash) 5158 if (bmsafemap->sm_cg == cg) 5159 break; 5160 if (bmsafemap) { 5161 *bmsafemapp = bmsafemap; 5162 return (1); 5163 } 5164 *bmsafemapp = NULL; 5165 5166 return (0); 5167} 5168 5169/* 5170 * Find the bmsafemap associated with a cylinder group buffer. 5171 * If none exists, create one. The buffer must be locked when 5172 * this routine is called and this routine must be called with 5173 * the softdep lock held. To avoid giving up the lock while 5174 * allocating a new bmsafemap, a preallocated bmsafemap may be 5175 * provided. If it is provided but not needed, it is freed. 5176 */ 5177static struct bmsafemap * 5178bmsafemap_lookup(mp, bp, cg, newbmsafemap) 5179 struct mount *mp; 5180 struct buf *bp; 5181 int cg; 5182 struct bmsafemap *newbmsafemap; 5183{ 5184 struct bmsafemap_hashhead *bmsafemaphd; 5185 struct bmsafemap *bmsafemap, *collision; 5186 struct worklist *wk; 5187 struct ufsmount *ump; 5188 5189 ump = VFSTOUFS(mp); 5190 LOCK_OWNED(ump); 5191 KASSERT(bp != NULL, ("bmsafemap_lookup: missing buffer")); 5192 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5193 if (wk->wk_type == D_BMSAFEMAP) { 5194 if (newbmsafemap) 5195 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP); 5196 return (WK_BMSAFEMAP(wk)); 5197 } 5198 } 5199 bmsafemaphd = BMSAFEMAP_HASH(ump, cg); 5200 if (bmsafemap_find(bmsafemaphd, cg, &bmsafemap) == 1) { 5201 if (newbmsafemap) 5202 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP); 5203 return (bmsafemap); 5204 } 5205 if (newbmsafemap) { 5206 bmsafemap = newbmsafemap; 5207 } else { 5208 FREE_LOCK(ump); 5209 bmsafemap = malloc(sizeof(struct bmsafemap), 5210 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 5211 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); 5212 ACQUIRE_LOCK(ump); 5213 } 5214 bmsafemap->sm_buf = bp; 5215 LIST_INIT(&bmsafemap->sm_inodedephd); 5216 LIST_INIT(&bmsafemap->sm_inodedepwr); 5217 LIST_INIT(&bmsafemap->sm_newblkhd); 5218 LIST_INIT(&bmsafemap->sm_newblkwr); 5219 LIST_INIT(&bmsafemap->sm_jaddrefhd); 5220 LIST_INIT(&bmsafemap->sm_jnewblkhd); 5221 LIST_INIT(&bmsafemap->sm_freehd); 5222 LIST_INIT(&bmsafemap->sm_freewr); 5223 if (bmsafemap_find(bmsafemaphd, cg, &collision) == 1) { 5224 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 5225 return (collision); 5226 } 5227 bmsafemap->sm_cg = cg; 5228 LIST_INSERT_HEAD(bmsafemaphd, bmsafemap, sm_hash); 5229 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next); 5230 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 5231 return (bmsafemap); 5232} 5233 5234/* 5235 * Direct block allocation dependencies. 5236 * 5237 * When a new block is allocated, the corresponding disk locations must be 5238 * initialized (with zeros or new data) before the on-disk inode points to 5239 * them. Also, the freemap from which the block was allocated must be 5240 * updated (on disk) before the inode's pointer. These two dependencies are 5241 * independent of each other and are needed for all file blocks and indirect 5242 * blocks that are pointed to directly by the inode. Just before the 5243 * "in-core" version of the inode is updated with a newly allocated block 5244 * number, a procedure (below) is called to setup allocation dependency 5245 * structures. These structures are removed when the corresponding 5246 * dependencies are satisfied or when the block allocation becomes obsolete 5247 * (i.e., the file is deleted, the block is de-allocated, or the block is a 5248 * fragment that gets upgraded). All of these cases are handled in 5249 * procedures described later. 5250 * 5251 * When a file extension causes a fragment to be upgraded, either to a larger 5252 * fragment or to a full block, the on-disk location may change (if the 5253 * previous fragment could not simply be extended). In this case, the old 5254 * fragment must be de-allocated, but not until after the inode's pointer has 5255 * been updated. In most cases, this is handled by later procedures, which 5256 * will construct a "freefrag" structure to be added to the workitem queue 5257 * when the inode update is complete (or obsolete). The main exception to 5258 * this is when an allocation occurs while a pending allocation dependency 5259 * (for the same block pointer) remains. This case is handled in the main 5260 * allocation dependency setup procedure by immediately freeing the 5261 * unreferenced fragments. 5262 */ 5263void 5264softdep_setup_allocdirect(ip, off, newblkno, oldblkno, newsize, oldsize, bp) 5265 struct inode *ip; /* inode to which block is being added */ 5266 ufs_lbn_t off; /* block pointer within inode */ 5267 ufs2_daddr_t newblkno; /* disk block number being added */ 5268 ufs2_daddr_t oldblkno; /* previous block number, 0 unless frag */ 5269 long newsize; /* size of new block */ 5270 long oldsize; /* size of new block */ 5271 struct buf *bp; /* bp for allocated block */ 5272{ 5273 struct allocdirect *adp, *oldadp; 5274 struct allocdirectlst *adphead; 5275 struct freefrag *freefrag; 5276 struct inodedep *inodedep; 5277 struct pagedep *pagedep; 5278 struct jnewblk *jnewblk; 5279 struct newblk *newblk; 5280 struct mount *mp; 5281 ufs_lbn_t lbn; 5282 5283 lbn = bp->b_lblkno; 5284 mp = ITOVFS(ip); 5285 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5286 ("softdep_setup_allocdirect called on non-softdep filesystem")); 5287 if (oldblkno && oldblkno != newblkno) 5288 /* 5289 * The usual case is that a smaller fragment that 5290 * was just allocated has been replaced with a bigger 5291 * fragment or a full-size block. If it is marked as 5292 * B_DELWRI, the current contents have not been written 5293 * to disk. It is possible that the block was written 5294 * earlier, but very uncommon. If the block has never 5295 * been written, there is no need to send a BIO_DELETE 5296 * for it when it is freed. The gain from avoiding the 5297 * TRIMs for the common case of unwritten blocks far 5298 * exceeds the cost of the write amplification for the 5299 * uncommon case of failing to send a TRIM for a block 5300 * that had been written. 5301 */ 5302 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn, 5303 (bp->b_flags & B_DELWRI) != 0 ? NOTRIM_KEY : SINGLETON_KEY); 5304 else 5305 freefrag = NULL; 5306 5307 CTR6(KTR_SUJ, 5308 "softdep_setup_allocdirect: ino %d blkno %jd oldblkno %jd " 5309 "off %jd newsize %ld oldsize %d", 5310 ip->i_number, newblkno, oldblkno, off, newsize, oldsize); 5311 ACQUIRE_LOCK(ITOUMP(ip)); 5312 if (off >= UFS_NDADDR) { 5313 if (lbn > 0) 5314 panic("softdep_setup_allocdirect: bad lbn %jd, off %jd", 5315 lbn, off); 5316 /* allocating an indirect block */ 5317 if (oldblkno != 0) 5318 panic("softdep_setup_allocdirect: non-zero indir"); 5319 } else { 5320 if (off != lbn) 5321 panic("softdep_setup_allocdirect: lbn %jd != off %jd", 5322 lbn, off); 5323 /* 5324 * Allocating a direct block. 5325 * 5326 * If we are allocating a directory block, then we must 5327 * allocate an associated pagedep to track additions and 5328 * deletions. 5329 */ 5330 if ((ip->i_mode & IFMT) == IFDIR) 5331 pagedep_lookup(mp, bp, ip->i_number, off, DEPALLOC, 5332 &pagedep); 5333 } 5334 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0) 5335 panic("softdep_setup_allocdirect: lost block"); 5336 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5337 ("softdep_setup_allocdirect: newblk already initialized")); 5338 /* 5339 * Convert the newblk to an allocdirect. 5340 */ 5341 WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT); 5342 adp = (struct allocdirect *)newblk; 5343 newblk->nb_freefrag = freefrag; 5344 adp->ad_offset = off; 5345 adp->ad_oldblkno = oldblkno; 5346 adp->ad_newsize = newsize; 5347 adp->ad_oldsize = oldsize; 5348 5349 /* 5350 * Finish initializing the journal. 5351 */ 5352 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5353 jnewblk->jn_ino = ip->i_number; 5354 jnewblk->jn_lbn = lbn; 5355 add_to_journal(&jnewblk->jn_list); 5356 } 5357 if (freefrag && freefrag->ff_jdep != NULL && 5358 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5359 add_to_journal(freefrag->ff_jdep); 5360 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 5361 adp->ad_inodedep = inodedep; 5362 5363 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list); 5364 /* 5365 * The list of allocdirects must be kept in sorted and ascending 5366 * order so that the rollback routines can quickly determine the 5367 * first uncommitted block (the size of the file stored on disk 5368 * ends at the end of the lowest committed fragment, or if there 5369 * are no fragments, at the end of the highest committed block). 5370 * Since files generally grow, the typical case is that the new 5371 * block is to be added at the end of the list. We speed this 5372 * special case by checking against the last allocdirect in the 5373 * list before laboriously traversing the list looking for the 5374 * insertion point. 5375 */ 5376 adphead = &inodedep->id_newinoupdt; 5377 oldadp = TAILQ_LAST(adphead, allocdirectlst); 5378 if (oldadp == NULL || oldadp->ad_offset <= off) { 5379 /* insert at end of list */ 5380 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 5381 if (oldadp != NULL && oldadp->ad_offset == off) 5382 allocdirect_merge(adphead, adp, oldadp); 5383 FREE_LOCK(ITOUMP(ip)); 5384 return; 5385 } 5386 TAILQ_FOREACH(oldadp, adphead, ad_next) { 5387 if (oldadp->ad_offset >= off) 5388 break; 5389 } 5390 if (oldadp == NULL) 5391 panic("softdep_setup_allocdirect: lost entry"); 5392 /* insert in middle of list */ 5393 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 5394 if (oldadp->ad_offset == off) 5395 allocdirect_merge(adphead, adp, oldadp); 5396 5397 FREE_LOCK(ITOUMP(ip)); 5398} 5399 5400/* 5401 * Merge a newer and older journal record to be stored either in a 5402 * newblock or freefrag. This handles aggregating journal records for 5403 * fragment allocation into a second record as well as replacing a 5404 * journal free with an aborted journal allocation. A segment for the 5405 * oldest record will be placed on wkhd if it has been written. If not 5406 * the segment for the newer record will suffice. 5407 */ 5408static struct worklist * 5409jnewblk_merge(new, old, wkhd) 5410 struct worklist *new; 5411 struct worklist *old; 5412 struct workhead *wkhd; 5413{ 5414 struct jnewblk *njnewblk; 5415 struct jnewblk *jnewblk; 5416 5417 /* Handle NULLs to simplify callers. */ 5418 if (new == NULL) 5419 return (old); 5420 if (old == NULL) 5421 return (new); 5422 /* Replace a jfreefrag with a jnewblk. */ 5423 if (new->wk_type == D_JFREEFRAG) { 5424 if (WK_JNEWBLK(old)->jn_blkno != WK_JFREEFRAG(new)->fr_blkno) 5425 panic("jnewblk_merge: blkno mismatch: %p, %p", 5426 old, new); 5427 cancel_jfreefrag(WK_JFREEFRAG(new)); 5428 return (old); 5429 } 5430 if (old->wk_type != D_JNEWBLK || new->wk_type != D_JNEWBLK) 5431 panic("jnewblk_merge: Bad type: old %d new %d\n", 5432 old->wk_type, new->wk_type); 5433 /* 5434 * Handle merging of two jnewblk records that describe 5435 * different sets of fragments in the same block. 5436 */ 5437 jnewblk = WK_JNEWBLK(old); 5438 njnewblk = WK_JNEWBLK(new); 5439 if (jnewblk->jn_blkno != njnewblk->jn_blkno) 5440 panic("jnewblk_merge: Merging disparate blocks."); 5441 /* 5442 * The record may be rolled back in the cg. 5443 */ 5444 if (jnewblk->jn_state & UNDONE) { 5445 jnewblk->jn_state &= ~UNDONE; 5446 njnewblk->jn_state |= UNDONE; 5447 njnewblk->jn_state &= ~ATTACHED; 5448 } 5449 /* 5450 * We modify the newer addref and free the older so that if neither 5451 * has been written the most up-to-date copy will be on disk. If 5452 * both have been written but rolled back we only temporarily need 5453 * one of them to fix the bits when the cg write completes. 5454 */ 5455 jnewblk->jn_state |= ATTACHED | COMPLETE; 5456 njnewblk->jn_oldfrags = jnewblk->jn_oldfrags; 5457 cancel_jnewblk(jnewblk, wkhd); 5458 WORKLIST_REMOVE(&jnewblk->jn_list); 5459 free_jnewblk(jnewblk); 5460 return (new); 5461} 5462 5463/* 5464 * Replace an old allocdirect dependency with a newer one. 5465 */ 5466static void 5467allocdirect_merge(adphead, newadp, oldadp) 5468 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 5469 struct allocdirect *newadp; /* allocdirect being added */ 5470 struct allocdirect *oldadp; /* existing allocdirect being checked */ 5471{ 5472 struct worklist *wk; 5473 struct freefrag *freefrag; 5474 5475 freefrag = NULL; 5476 LOCK_OWNED(VFSTOUFS(newadp->ad_list.wk_mp)); 5477 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 5478 newadp->ad_oldsize != oldadp->ad_newsize || 5479 newadp->ad_offset >= UFS_NDADDR) 5480 panic("%s %jd != new %jd || old size %ld != new %ld", 5481 "allocdirect_merge: old blkno", 5482 (intmax_t)newadp->ad_oldblkno, 5483 (intmax_t)oldadp->ad_newblkno, 5484 newadp->ad_oldsize, oldadp->ad_newsize); 5485 newadp->ad_oldblkno = oldadp->ad_oldblkno; 5486 newadp->ad_oldsize = oldadp->ad_oldsize; 5487 /* 5488 * If the old dependency had a fragment to free or had never 5489 * previously had a block allocated, then the new dependency 5490 * can immediately post its freefrag and adopt the old freefrag. 5491 * This action is done by swapping the freefrag dependencies. 5492 * The new dependency gains the old one's freefrag, and the 5493 * old one gets the new one and then immediately puts it on 5494 * the worklist when it is freed by free_newblk. It is 5495 * not possible to do this swap when the old dependency had a 5496 * non-zero size but no previous fragment to free. This condition 5497 * arises when the new block is an extension of the old block. 5498 * Here, the first part of the fragment allocated to the new 5499 * dependency is part of the block currently claimed on disk by 5500 * the old dependency, so cannot legitimately be freed until the 5501 * conditions for the new dependency are fulfilled. 5502 */ 5503 freefrag = newadp->ad_freefrag; 5504 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 5505 newadp->ad_freefrag = oldadp->ad_freefrag; 5506 oldadp->ad_freefrag = freefrag; 5507 } 5508 /* 5509 * If we are tracking a new directory-block allocation, 5510 * move it from the old allocdirect to the new allocdirect. 5511 */ 5512 if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) { 5513 WORKLIST_REMOVE(wk); 5514 if (!LIST_EMPTY(&oldadp->ad_newdirblk)) 5515 panic("allocdirect_merge: extra newdirblk"); 5516 WORKLIST_INSERT(&newadp->ad_newdirblk, wk); 5517 } 5518 TAILQ_REMOVE(adphead, oldadp, ad_next); 5519 /* 5520 * We need to move any journal dependencies over to the freefrag 5521 * that releases this block if it exists. Otherwise we are 5522 * extending an existing block and we'll wait until that is 5523 * complete to release the journal space and extend the 5524 * new journal to cover this old space as well. 5525 */ 5526 if (freefrag == NULL) { 5527 if (oldadp->ad_newblkno != newadp->ad_newblkno) 5528 panic("allocdirect_merge: %jd != %jd", 5529 oldadp->ad_newblkno, newadp->ad_newblkno); 5530 newadp->ad_block.nb_jnewblk = (struct jnewblk *) 5531 jnewblk_merge(&newadp->ad_block.nb_jnewblk->jn_list, 5532 &oldadp->ad_block.nb_jnewblk->jn_list, 5533 &newadp->ad_block.nb_jwork); 5534 oldadp->ad_block.nb_jnewblk = NULL; 5535 cancel_newblk(&oldadp->ad_block, NULL, 5536 &newadp->ad_block.nb_jwork); 5537 } else { 5538 wk = (struct worklist *) cancel_newblk(&oldadp->ad_block, 5539 &freefrag->ff_list, &freefrag->ff_jwork); 5540 freefrag->ff_jdep = jnewblk_merge(freefrag->ff_jdep, wk, 5541 &freefrag->ff_jwork); 5542 } 5543 free_newblk(&oldadp->ad_block); 5544} 5545 5546/* 5547 * Allocate a jfreefrag structure to journal a single block free. 5548 */ 5549static struct jfreefrag * 5550newjfreefrag(freefrag, ip, blkno, size, lbn) 5551 struct freefrag *freefrag; 5552 struct inode *ip; 5553 ufs2_daddr_t blkno; 5554 long size; 5555 ufs_lbn_t lbn; 5556{ 5557 struct jfreefrag *jfreefrag; 5558 struct fs *fs; 5559 5560 fs = ITOFS(ip); 5561 jfreefrag = malloc(sizeof(struct jfreefrag), M_JFREEFRAG, 5562 M_SOFTDEP_FLAGS); 5563 workitem_alloc(&jfreefrag->fr_list, D_JFREEFRAG, ITOVFS(ip)); 5564 jfreefrag->fr_jsegdep = newjsegdep(&jfreefrag->fr_list); 5565 jfreefrag->fr_state = ATTACHED | DEPCOMPLETE; 5566 jfreefrag->fr_ino = ip->i_number; 5567 jfreefrag->fr_lbn = lbn; 5568 jfreefrag->fr_blkno = blkno; 5569 jfreefrag->fr_frags = numfrags(fs, size); 5570 jfreefrag->fr_freefrag = freefrag; 5571 5572 return (jfreefrag); 5573} 5574 5575/* 5576 * Allocate a new freefrag structure. 5577 */ 5578static struct freefrag * 5579newfreefrag(ip, blkno, size, lbn, key) 5580 struct inode *ip; 5581 ufs2_daddr_t blkno; 5582 long size; 5583 ufs_lbn_t lbn; 5584 u_long key; 5585{ 5586 struct freefrag *freefrag; 5587 struct ufsmount *ump; 5588 struct fs *fs; 5589 5590 CTR4(KTR_SUJ, "newfreefrag: ino %d blkno %jd size %ld lbn %jd", 5591 ip->i_number, blkno, size, lbn); 5592 ump = ITOUMP(ip); 5593 fs = ump->um_fs; 5594 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 5595 panic("newfreefrag: frag size"); 5596 freefrag = malloc(sizeof(struct freefrag), 5597 M_FREEFRAG, M_SOFTDEP_FLAGS); 5598 workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ump)); 5599 freefrag->ff_state = ATTACHED; 5600 LIST_INIT(&freefrag->ff_jwork); 5601 freefrag->ff_inum = ip->i_number; 5602 freefrag->ff_vtype = ITOV(ip)->v_type; 5603 freefrag->ff_blkno = blkno; 5604 freefrag->ff_fragsize = size; 5605 freefrag->ff_key = key; 5606 5607 if (MOUNTEDSUJ(UFSTOVFS(ump))) { 5608 freefrag->ff_jdep = (struct worklist *) 5609 newjfreefrag(freefrag, ip, blkno, size, lbn); 5610 } else { 5611 freefrag->ff_state |= DEPCOMPLETE; 5612 freefrag->ff_jdep = NULL; 5613 } 5614 5615 return (freefrag); 5616} 5617 5618/* 5619 * This workitem de-allocates fragments that were replaced during 5620 * file block allocation. 5621 */ 5622static void 5623handle_workitem_freefrag(freefrag) 5624 struct freefrag *freefrag; 5625{ 5626 struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp); 5627 struct workhead wkhd; 5628 5629 CTR3(KTR_SUJ, 5630 "handle_workitem_freefrag: ino %d blkno %jd size %ld", 5631 freefrag->ff_inum, freefrag->ff_blkno, freefrag->ff_fragsize); 5632 /* 5633 * It would be illegal to add new completion items to the 5634 * freefrag after it was schedule to be done so it must be 5635 * safe to modify the list head here. 5636 */ 5637 LIST_INIT(&wkhd); 5638 ACQUIRE_LOCK(ump); 5639 LIST_SWAP(&freefrag->ff_jwork, &wkhd, worklist, wk_list); 5640 /* 5641 * If the journal has not been written we must cancel it here. 5642 */ 5643 if (freefrag->ff_jdep) { 5644 if (freefrag->ff_jdep->wk_type != D_JNEWBLK) 5645 panic("handle_workitem_freefrag: Unexpected type %d\n", 5646 freefrag->ff_jdep->wk_type); 5647 cancel_jnewblk(WK_JNEWBLK(freefrag->ff_jdep), &wkhd); 5648 } 5649 FREE_LOCK(ump); 5650 ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno, 5651 freefrag->ff_fragsize, freefrag->ff_inum, freefrag->ff_vtype, 5652 &wkhd, freefrag->ff_key); 5653 ACQUIRE_LOCK(ump); 5654 WORKITEM_FREE(freefrag, D_FREEFRAG); 5655 FREE_LOCK(ump); 5656} 5657 5658/* 5659 * Set up a dependency structure for an external attributes data block. 5660 * This routine follows much of the structure of softdep_setup_allocdirect. 5661 * See the description of softdep_setup_allocdirect above for details. 5662 */ 5663void 5664softdep_setup_allocext(ip, off, newblkno, oldblkno, newsize, oldsize, bp) 5665 struct inode *ip; 5666 ufs_lbn_t off; 5667 ufs2_daddr_t newblkno; 5668 ufs2_daddr_t oldblkno; 5669 long newsize; 5670 long oldsize; 5671 struct buf *bp; 5672{ 5673 struct allocdirect *adp, *oldadp; 5674 struct allocdirectlst *adphead; 5675 struct freefrag *freefrag; 5676 struct inodedep *inodedep; 5677 struct jnewblk *jnewblk; 5678 struct newblk *newblk; 5679 struct mount *mp; 5680 struct ufsmount *ump; 5681 ufs_lbn_t lbn; 5682 5683 mp = ITOVFS(ip); 5684 ump = VFSTOUFS(mp); 5685 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5686 ("softdep_setup_allocext called on non-softdep filesystem")); 5687 KASSERT(off < UFS_NXADDR, 5688 ("softdep_setup_allocext: lbn %lld > UFS_NXADDR", (long long)off)); 5689 5690 lbn = bp->b_lblkno; 5691 if (oldblkno && oldblkno != newblkno) 5692 /* 5693 * The usual case is that a smaller fragment that 5694 * was just allocated has been replaced with a bigger 5695 * fragment or a full-size block. If it is marked as 5696 * B_DELWRI, the current contents have not been written 5697 * to disk. It is possible that the block was written 5698 * earlier, but very uncommon. If the block has never 5699 * been written, there is no need to send a BIO_DELETE 5700 * for it when it is freed. The gain from avoiding the 5701 * TRIMs for the common case of unwritten blocks far 5702 * exceeds the cost of the write amplification for the 5703 * uncommon case of failing to send a TRIM for a block 5704 * that had been written. 5705 */ 5706 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn, 5707 (bp->b_flags & B_DELWRI) != 0 ? NOTRIM_KEY : SINGLETON_KEY); 5708 else 5709 freefrag = NULL; 5710 5711 ACQUIRE_LOCK(ump); 5712 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0) 5713 panic("softdep_setup_allocext: lost block"); 5714 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5715 ("softdep_setup_allocext: newblk already initialized")); 5716 /* 5717 * Convert the newblk to an allocdirect. 5718 */ 5719 WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT); 5720 adp = (struct allocdirect *)newblk; 5721 newblk->nb_freefrag = freefrag; 5722 adp->ad_offset = off; 5723 adp->ad_oldblkno = oldblkno; 5724 adp->ad_newsize = newsize; 5725 adp->ad_oldsize = oldsize; 5726 adp->ad_state |= EXTDATA; 5727 5728 /* 5729 * Finish initializing the journal. 5730 */ 5731 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5732 jnewblk->jn_ino = ip->i_number; 5733 jnewblk->jn_lbn = lbn; 5734 add_to_journal(&jnewblk->jn_list); 5735 } 5736 if (freefrag && freefrag->ff_jdep != NULL && 5737 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5738 add_to_journal(freefrag->ff_jdep); 5739 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 5740 adp->ad_inodedep = inodedep; 5741 5742 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list); 5743 /* 5744 * The list of allocdirects must be kept in sorted and ascending 5745 * order so that the rollback routines can quickly determine the 5746 * first uncommitted block (the size of the file stored on disk 5747 * ends at the end of the lowest committed fragment, or if there 5748 * are no fragments, at the end of the highest committed block). 5749 * Since files generally grow, the typical case is that the new 5750 * block is to be added at the end of the list. We speed this 5751 * special case by checking against the last allocdirect in the 5752 * list before laboriously traversing the list looking for the 5753 * insertion point. 5754 */ 5755 adphead = &inodedep->id_newextupdt; 5756 oldadp = TAILQ_LAST(adphead, allocdirectlst); 5757 if (oldadp == NULL || oldadp->ad_offset <= off) { 5758 /* insert at end of list */ 5759 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 5760 if (oldadp != NULL && oldadp->ad_offset == off) 5761 allocdirect_merge(adphead, adp, oldadp); 5762 FREE_LOCK(ump); 5763 return; 5764 } 5765 TAILQ_FOREACH(oldadp, adphead, ad_next) { 5766 if (oldadp->ad_offset >= off) 5767 break; 5768 } 5769 if (oldadp == NULL) 5770 panic("softdep_setup_allocext: lost entry"); 5771 /* insert in middle of list */ 5772 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 5773 if (oldadp->ad_offset == off) 5774 allocdirect_merge(adphead, adp, oldadp); 5775 FREE_LOCK(ump); 5776} 5777 5778/* 5779 * Indirect block allocation dependencies. 5780 * 5781 * The same dependencies that exist for a direct block also exist when 5782 * a new block is allocated and pointed to by an entry in a block of 5783 * indirect pointers. The undo/redo states described above are also 5784 * used here. Because an indirect block contains many pointers that 5785 * may have dependencies, a second copy of the entire in-memory indirect 5786 * block is kept. The buffer cache copy is always completely up-to-date. 5787 * The second copy, which is used only as a source for disk writes, 5788 * contains only the safe pointers (i.e., those that have no remaining 5789 * update dependencies). The second copy is freed when all pointers 5790 * are safe. The cache is not allowed to replace indirect blocks with 5791 * pending update dependencies. If a buffer containing an indirect 5792 * block with dependencies is written, these routines will mark it 5793 * dirty again. It can only be successfully written once all the 5794 * dependencies are removed. The ffs_fsync routine in conjunction with 5795 * softdep_sync_metadata work together to get all the dependencies 5796 * removed so that a file can be successfully written to disk. Three 5797 * procedures are used when setting up indirect block pointer 5798 * dependencies. The division is necessary because of the organization 5799 * of the "balloc" routine and because of the distinction between file 5800 * pages and file metadata blocks. 5801 */ 5802 5803/* 5804 * Allocate a new allocindir structure. 5805 */ 5806static struct allocindir * 5807newallocindir(ip, ptrno, newblkno, oldblkno, lbn) 5808 struct inode *ip; /* inode for file being extended */ 5809 int ptrno; /* offset of pointer in indirect block */ 5810 ufs2_daddr_t newblkno; /* disk block number being added */ 5811 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 5812 ufs_lbn_t lbn; 5813{ 5814 struct newblk *newblk; 5815 struct allocindir *aip; 5816 struct freefrag *freefrag; 5817 struct jnewblk *jnewblk; 5818 5819 if (oldblkno) 5820 freefrag = newfreefrag(ip, oldblkno, ITOFS(ip)->fs_bsize, lbn, 5821 SINGLETON_KEY); 5822 else 5823 freefrag = NULL; 5824 ACQUIRE_LOCK(ITOUMP(ip)); 5825 if (newblk_lookup(ITOVFS(ip), newblkno, 0, &newblk) == 0) 5826 panic("new_allocindir: lost block"); 5827 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5828 ("newallocindir: newblk already initialized")); 5829 WORKITEM_REASSIGN(newblk, D_ALLOCINDIR); 5830 newblk->nb_freefrag = freefrag; 5831 aip = (struct allocindir *)newblk; 5832 aip->ai_offset = ptrno; 5833 aip->ai_oldblkno = oldblkno; 5834 aip->ai_lbn = lbn; 5835 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5836 jnewblk->jn_ino = ip->i_number; 5837 jnewblk->jn_lbn = lbn; 5838 add_to_journal(&jnewblk->jn_list); 5839 } 5840 if (freefrag && freefrag->ff_jdep != NULL && 5841 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5842 add_to_journal(freefrag->ff_jdep); 5843 return (aip); 5844} 5845 5846/* 5847 * Called just before setting an indirect block pointer 5848 * to a newly allocated file page. 5849 */ 5850void 5851softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 5852 struct inode *ip; /* inode for file being extended */ 5853 ufs_lbn_t lbn; /* allocated block number within file */ 5854 struct buf *bp; /* buffer with indirect blk referencing page */ 5855 int ptrno; /* offset of pointer in indirect block */ 5856 ufs2_daddr_t newblkno; /* disk block number being added */ 5857 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 5858 struct buf *nbp; /* buffer holding allocated page */ 5859{ 5860 struct inodedep *inodedep; 5861 struct freefrag *freefrag; 5862 struct allocindir *aip; 5863 struct pagedep *pagedep; 5864 struct mount *mp; 5865 struct ufsmount *ump; 5866 5867 mp = ITOVFS(ip); 5868 ump = VFSTOUFS(mp); 5869 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5870 ("softdep_setup_allocindir_page called on non-softdep filesystem")); 5871 KASSERT(lbn == nbp->b_lblkno, 5872 ("softdep_setup_allocindir_page: lbn %jd != lblkno %jd", 5873 lbn, bp->b_lblkno)); 5874 CTR4(KTR_SUJ, 5875 "softdep_setup_allocindir_page: ino %d blkno %jd oldblkno %jd " 5876 "lbn %jd", ip->i_number, newblkno, oldblkno, lbn); 5877 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page"); 5878 aip = newallocindir(ip, ptrno, newblkno, oldblkno, lbn); 5879 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 5880 /* 5881 * If we are allocating a directory page, then we must 5882 * allocate an associated pagedep to track additions and 5883 * deletions. 5884 */ 5885 if ((ip->i_mode & IFMT) == IFDIR) 5886 pagedep_lookup(mp, nbp, ip->i_number, lbn, DEPALLOC, &pagedep); 5887 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list); 5888 freefrag = setup_allocindir_phase2(bp, ip, inodedep, aip, lbn); 5889 FREE_LOCK(ump); 5890 if (freefrag) 5891 handle_workitem_freefrag(freefrag); 5892} 5893 5894/* 5895 * Called just before setting an indirect block pointer to a 5896 * newly allocated indirect block. 5897 */ 5898void 5899softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 5900 struct buf *nbp; /* newly allocated indirect block */ 5901 struct inode *ip; /* inode for file being extended */ 5902 struct buf *bp; /* indirect block referencing allocated block */ 5903 int ptrno; /* offset of pointer in indirect block */ 5904 ufs2_daddr_t newblkno; /* disk block number being added */ 5905{ 5906 struct inodedep *inodedep; 5907 struct allocindir *aip; 5908 struct ufsmount *ump; 5909 ufs_lbn_t lbn; 5910 5911 ump = ITOUMP(ip); 5912 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 5913 ("softdep_setup_allocindir_meta called on non-softdep filesystem")); 5914 CTR3(KTR_SUJ, 5915 "softdep_setup_allocindir_meta: ino %d blkno %jd ptrno %d", 5916 ip->i_number, newblkno, ptrno); 5917 lbn = nbp->b_lblkno; 5918 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta"); 5919 aip = newallocindir(ip, ptrno, newblkno, 0, lbn); 5920 inodedep_lookup(UFSTOVFS(ump), ip->i_number, DEPALLOC, &inodedep); 5921 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list); 5922 if (setup_allocindir_phase2(bp, ip, inodedep, aip, lbn)) 5923 panic("softdep_setup_allocindir_meta: Block already existed"); 5924 FREE_LOCK(ump); 5925} 5926 5927static void 5928indirdep_complete(indirdep) 5929 struct indirdep *indirdep; 5930{ 5931 struct allocindir *aip; 5932 5933 LIST_REMOVE(indirdep, ir_next); 5934 indirdep->ir_state |= DEPCOMPLETE; 5935 5936 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) { 5937 LIST_REMOVE(aip, ai_next); 5938 free_newblk(&aip->ai_block); 5939 } 5940 /* 5941 * If this indirdep is not attached to a buf it was simply waiting 5942 * on completion to clear completehd. free_indirdep() asserts 5943 * that nothing is dangling. 5944 */ 5945 if ((indirdep->ir_state & ONWORKLIST) == 0) 5946 free_indirdep(indirdep); 5947} 5948 5949static struct indirdep * 5950indirdep_lookup(mp, ip, bp) 5951 struct mount *mp; 5952 struct inode *ip; 5953 struct buf *bp; 5954{ 5955 struct indirdep *indirdep, *newindirdep; 5956 struct newblk *newblk; 5957 struct ufsmount *ump; 5958 struct worklist *wk; 5959 struct fs *fs; 5960 ufs2_daddr_t blkno; 5961 5962 ump = VFSTOUFS(mp); 5963 LOCK_OWNED(ump); 5964 indirdep = NULL; 5965 newindirdep = NULL; 5966 fs = ump->um_fs; 5967 for (;;) { 5968 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5969 if (wk->wk_type != D_INDIRDEP) 5970 continue; 5971 indirdep = WK_INDIRDEP(wk); 5972 break; 5973 } 5974 /* Found on the buffer worklist, no new structure to free. */ 5975 if (indirdep != NULL && newindirdep == NULL) 5976 return (indirdep); 5977 if (indirdep != NULL && newindirdep != NULL) 5978 panic("indirdep_lookup: simultaneous create"); 5979 /* None found on the buffer and a new structure is ready. */ 5980 if (indirdep == NULL && newindirdep != NULL) 5981 break; 5982 /* None found and no new structure available. */ 5983 FREE_LOCK(ump); 5984 newindirdep = malloc(sizeof(struct indirdep), 5985 M_INDIRDEP, M_SOFTDEP_FLAGS); 5986 workitem_alloc(&newindirdep->ir_list, D_INDIRDEP, mp); 5987 newindirdep->ir_state = ATTACHED; 5988 if (I_IS_UFS1(ip)) 5989 newindirdep->ir_state |= UFS1FMT; 5990 TAILQ_INIT(&newindirdep->ir_trunc); 5991 newindirdep->ir_saveddata = NULL; 5992 LIST_INIT(&newindirdep->ir_deplisthd); 5993 LIST_INIT(&newindirdep->ir_donehd); 5994 LIST_INIT(&newindirdep->ir_writehd); 5995 LIST_INIT(&newindirdep->ir_completehd); 5996 if (bp->b_blkno == bp->b_lblkno) { 5997 ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp, 5998 NULL, NULL); 5999 bp->b_blkno = blkno; 6000 } 6001 newindirdep->ir_freeblks = NULL; 6002 newindirdep->ir_savebp = 6003 getblk(ump->um_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0); 6004 newindirdep->ir_bp = bp; 6005 BUF_KERNPROC(newindirdep->ir_savebp); 6006 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 6007 ACQUIRE_LOCK(ump); 6008 } 6009 indirdep = newindirdep; 6010 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 6011 /* 6012 * If the block is not yet allocated we don't set DEPCOMPLETE so 6013 * that we don't free dependencies until the pointers are valid. 6014 * This could search b_dep for D_ALLOCDIRECT/D_ALLOCINDIR rather 6015 * than using the hash. 6016 */ 6017 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk)) 6018 LIST_INSERT_HEAD(&newblk->nb_indirdeps, indirdep, ir_next); 6019 else 6020 indirdep->ir_state |= DEPCOMPLETE; 6021 return (indirdep); 6022} 6023 6024/* 6025 * Called to finish the allocation of the "aip" allocated 6026 * by one of the two routines above. 6027 */ 6028static struct freefrag * 6029setup_allocindir_phase2(bp, ip, inodedep, aip, lbn) 6030 struct buf *bp; /* in-memory copy of the indirect block */ 6031 struct inode *ip; /* inode for file being extended */ 6032 struct inodedep *inodedep; /* Inodedep for ip */ 6033 struct allocindir *aip; /* allocindir allocated by the above routines */ 6034 ufs_lbn_t lbn; /* Logical block number for this block. */ 6035{ 6036 struct fs *fs; 6037 struct indirdep *indirdep; 6038 struct allocindir *oldaip; 6039 struct freefrag *freefrag; 6040 struct mount *mp; 6041 struct ufsmount *ump; 6042 6043 mp = ITOVFS(ip); 6044 ump = VFSTOUFS(mp); 6045 LOCK_OWNED(ump); 6046 fs = ump->um_fs; 6047 if (bp->b_lblkno >= 0) 6048 panic("setup_allocindir_phase2: not indir blk"); 6049 KASSERT(aip->ai_offset >= 0 && aip->ai_offset < NINDIR(fs), 6050 ("setup_allocindir_phase2: Bad offset %d", aip->ai_offset)); 6051 indirdep = indirdep_lookup(mp, ip, bp); 6052 KASSERT(indirdep->ir_savebp != NULL, 6053 ("setup_allocindir_phase2 NULL ir_savebp")); 6054 aip->ai_indirdep = indirdep; 6055 /* 6056 * Check for an unwritten dependency for this indirect offset. If 6057 * there is, merge the old dependency into the new one. This happens 6058 * as a result of reallocblk only. 6059 */ 6060 freefrag = NULL; 6061 if (aip->ai_oldblkno != 0) { 6062 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) { 6063 if (oldaip->ai_offset == aip->ai_offset) { 6064 freefrag = allocindir_merge(aip, oldaip); 6065 goto done; 6066 } 6067 } 6068 LIST_FOREACH(oldaip, &indirdep->ir_donehd, ai_next) { 6069 if (oldaip->ai_offset == aip->ai_offset) { 6070 freefrag = allocindir_merge(aip, oldaip); 6071 goto done; 6072 } 6073 } 6074 } 6075done: 6076 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 6077 return (freefrag); 6078} 6079 6080/* 6081 * Merge two allocindirs which refer to the same block. Move newblock 6082 * dependencies and setup the freefrags appropriately. 6083 */ 6084static struct freefrag * 6085allocindir_merge(aip, oldaip) 6086 struct allocindir *aip; 6087 struct allocindir *oldaip; 6088{ 6089 struct freefrag *freefrag; 6090 struct worklist *wk; 6091 6092 if (oldaip->ai_newblkno != aip->ai_oldblkno) 6093 panic("allocindir_merge: blkno"); 6094 aip->ai_oldblkno = oldaip->ai_oldblkno; 6095 freefrag = aip->ai_freefrag; 6096 aip->ai_freefrag = oldaip->ai_freefrag; 6097 oldaip->ai_freefrag = NULL; 6098 KASSERT(freefrag != NULL, ("setup_allocindir_phase2: No freefrag")); 6099 /* 6100 * If we are tracking a new directory-block allocation, 6101 * move it from the old allocindir to the new allocindir. 6102 */ 6103 if ((wk = LIST_FIRST(&oldaip->ai_newdirblk)) != NULL) { 6104 WORKLIST_REMOVE(wk); 6105 if (!LIST_EMPTY(&oldaip->ai_newdirblk)) 6106 panic("allocindir_merge: extra newdirblk"); 6107 WORKLIST_INSERT(&aip->ai_newdirblk, wk); 6108 } 6109 /* 6110 * We can skip journaling for this freefrag and just complete 6111 * any pending journal work for the allocindir that is being 6112 * removed after the freefrag completes. 6113 */ 6114 if (freefrag->ff_jdep) 6115 cancel_jfreefrag(WK_JFREEFRAG(freefrag->ff_jdep)); 6116 LIST_REMOVE(oldaip, ai_next); 6117 freefrag->ff_jdep = (struct worklist *)cancel_newblk(&oldaip->ai_block, 6118 &freefrag->ff_list, &freefrag->ff_jwork); 6119 free_newblk(&oldaip->ai_block); 6120 6121 return (freefrag); 6122} 6123 6124static inline void 6125setup_freedirect(freeblks, ip, i, needj) 6126 struct freeblks *freeblks; 6127 struct inode *ip; 6128 int i; 6129 int needj; 6130{ 6131 struct ufsmount *ump; 6132 ufs2_daddr_t blkno; 6133 int frags; 6134 6135 blkno = DIP(ip, i_db[i]); 6136 if (blkno == 0) 6137 return; 6138 DIP_SET(ip, i_db[i], 0); 6139 ump = ITOUMP(ip); 6140 frags = sblksize(ump->um_fs, ip->i_size, i); 6141 frags = numfrags(ump->um_fs, frags); 6142 newfreework(ump, freeblks, NULL, i, blkno, frags, 0, needj); 6143} 6144 6145static inline void 6146setup_freeext(freeblks, ip, i, needj) 6147 struct freeblks *freeblks; 6148 struct inode *ip; 6149 int i; 6150 int needj; 6151{ 6152 struct ufsmount *ump; 6153 ufs2_daddr_t blkno; 6154 int frags; 6155 6156 blkno = ip->i_din2->di_extb[i]; 6157 if (blkno == 0) 6158 return; 6159 ip->i_din2->di_extb[i] = 0; 6160 ump = ITOUMP(ip); 6161 frags = sblksize(ump->um_fs, ip->i_din2->di_extsize, i); 6162 frags = numfrags(ump->um_fs, frags); 6163 newfreework(ump, freeblks, NULL, -1 - i, blkno, frags, 0, needj); 6164} 6165 6166static inline void 6167setup_freeindir(freeblks, ip, i, lbn, needj) 6168 struct freeblks *freeblks; 6169 struct inode *ip; 6170 int i; 6171 ufs_lbn_t lbn; 6172 int needj; 6173{ 6174 struct ufsmount *ump; 6175 ufs2_daddr_t blkno; 6176 6177 blkno = DIP(ip, i_ib[i]); 6178 if (blkno == 0) 6179 return; 6180 DIP_SET(ip, i_ib[i], 0); 6181 ump = ITOUMP(ip); 6182 newfreework(ump, freeblks, NULL, lbn, blkno, ump->um_fs->fs_frag, 6183 0, needj); 6184} 6185 6186static inline struct freeblks * 6187newfreeblks(mp, ip) 6188 struct mount *mp; 6189 struct inode *ip; 6190{ 6191 struct freeblks *freeblks; 6192 6193 freeblks = malloc(sizeof(struct freeblks), 6194 M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO); 6195 workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp); 6196 LIST_INIT(&freeblks->fb_jblkdephd); 6197 LIST_INIT(&freeblks->fb_jwork); 6198 freeblks->fb_ref = 0; 6199 freeblks->fb_cgwait = 0; 6200 freeblks->fb_state = ATTACHED; 6201 freeblks->fb_uid = ip->i_uid; 6202 freeblks->fb_inum = ip->i_number; 6203 freeblks->fb_vtype = ITOV(ip)->v_type; 6204 freeblks->fb_modrev = DIP(ip, i_modrev); 6205 freeblks->fb_devvp = ITODEVVP(ip); 6206 freeblks->fb_chkcnt = 0; 6207 freeblks->fb_len = 0; 6208 6209 return (freeblks); 6210} 6211 6212static void 6213trunc_indirdep(indirdep, freeblks, bp, off) 6214 struct indirdep *indirdep; 6215 struct freeblks *freeblks; 6216 struct buf *bp; 6217 int off; 6218{ 6219 struct allocindir *aip, *aipn; 6220 6221 /* 6222 * The first set of allocindirs won't be in savedbp. 6223 */ 6224 LIST_FOREACH_SAFE(aip, &indirdep->ir_deplisthd, ai_next, aipn) 6225 if (aip->ai_offset > off) 6226 cancel_allocindir(aip, bp, freeblks, 1); 6227 LIST_FOREACH_SAFE(aip, &indirdep->ir_donehd, ai_next, aipn) 6228 if (aip->ai_offset > off) 6229 cancel_allocindir(aip, bp, freeblks, 1); 6230 /* 6231 * These will exist in savedbp. 6232 */ 6233 LIST_FOREACH_SAFE(aip, &indirdep->ir_writehd, ai_next, aipn) 6234 if (aip->ai_offset > off) 6235 cancel_allocindir(aip, NULL, freeblks, 0); 6236 LIST_FOREACH_SAFE(aip, &indirdep->ir_completehd, ai_next, aipn) 6237 if (aip->ai_offset > off) 6238 cancel_allocindir(aip, NULL, freeblks, 0); 6239} 6240 6241/* 6242 * Follow the chain of indirects down to lastlbn creating a freework 6243 * structure for each. This will be used to start indir_trunc() at 6244 * the right offset and create the journal records for the parrtial 6245 * truncation. A second step will handle the truncated dependencies. 6246 */ 6247static int 6248setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno) 6249 struct freeblks *freeblks; 6250 struct inode *ip; 6251 ufs_lbn_t lbn; 6252 ufs_lbn_t lastlbn; 6253 ufs2_daddr_t blkno; 6254{ 6255 struct indirdep *indirdep; 6256 struct indirdep *indirn; 6257 struct freework *freework; 6258 struct newblk *newblk; 6259 struct mount *mp; 6260 struct ufsmount *ump; 6261 struct buf *bp; 6262 uint8_t *start; 6263 uint8_t *end; 6264 ufs_lbn_t lbnadd; 6265 int level; 6266 int error; 6267 int off; 6268 6269 6270 freework = NULL; 6271 if (blkno == 0) 6272 return (0); 6273 mp = freeblks->fb_list.wk_mp; 6274 ump = VFSTOUFS(mp); 6275 bp = getblk(ITOV(ip), lbn, mp->mnt_stat.f_iosize, 0, 0, 0); 6276 if ((bp->b_flags & B_CACHE) == 0) { 6277 bp->b_blkno = blkptrtodb(VFSTOUFS(mp), blkno); 6278 bp->b_iocmd = BIO_READ; 6279 bp->b_flags &= ~B_INVAL; 6280 bp->b_ioflags &= ~BIO_ERROR; 6281 vfs_busy_pages(bp, 0); 6282 bp->b_iooffset = dbtob(bp->b_blkno); 6283 bstrategy(bp); 6284#ifdef RACCT 6285 if (racct_enable) { 6286 PROC_LOCK(curproc); 6287 racct_add_buf(curproc, bp, 0); 6288 PROC_UNLOCK(curproc); 6289 } 6290#endif /* RACCT */ 6291 curthread->td_ru.ru_inblock++; 6292 error = bufwait(bp); 6293 if (error) { 6294 brelse(bp); 6295 return (error); 6296 } 6297 } 6298 level = lbn_level(lbn); 6299 lbnadd = lbn_offset(ump->um_fs, level); 6300 /* 6301 * Compute the offset of the last block we want to keep. Store 6302 * in the freework the first block we want to completely free. 6303 */ 6304 off = (lastlbn - -(lbn + level)) / lbnadd; 6305 if (off + 1 == NINDIR(ump->um_fs)) 6306 goto nowork; 6307 freework = newfreework(ump, freeblks, NULL, lbn, blkno, 0, off + 1, 0); 6308 /* 6309 * Link the freework into the indirdep. This will prevent any new 6310 * allocations from proceeding until we are finished with the 6311 * truncate and the block is written. 6312 */ 6313 ACQUIRE_LOCK(ump); 6314 indirdep = indirdep_lookup(mp, ip, bp); 6315 if (indirdep->ir_freeblks) 6316 panic("setup_trunc_indir: indirdep already truncated."); 6317 TAILQ_INSERT_TAIL(&indirdep->ir_trunc, freework, fw_next); 6318 freework->fw_indir = indirdep; 6319 /* 6320 * Cancel any allocindirs that will not make it to disk. 6321 * We have to do this for all copies of the indirdep that 6322 * live on this newblk. 6323 */ 6324 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 6325 if (newblk_lookup(mp, dbtofsb(ump->um_fs, bp->b_blkno), 0, 6326 &newblk) == 0) 6327 panic("setup_trunc_indir: lost block"); 6328 LIST_FOREACH(indirn, &newblk->nb_indirdeps, ir_next) 6329 trunc_indirdep(indirn, freeblks, bp, off); 6330 } else 6331 trunc_indirdep(indirdep, freeblks, bp, off); 6332 FREE_LOCK(ump); 6333 /* 6334 * Creation is protected by the buf lock. The saveddata is only 6335 * needed if a full truncation follows a partial truncation but it 6336 * is difficult to allocate in that case so we fetch it anyway. 6337 */ 6338 if (indirdep->ir_saveddata == NULL) 6339 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP, 6340 M_SOFTDEP_FLAGS); 6341nowork: 6342 /* Fetch the blkno of the child and the zero start offset. */ 6343 if (I_IS_UFS1(ip)) { 6344 blkno = ((ufs1_daddr_t *)bp->b_data)[off]; 6345 start = (uint8_t *)&((ufs1_daddr_t *)bp->b_data)[off+1]; 6346 } else { 6347 blkno = ((ufs2_daddr_t *)bp->b_data)[off]; 6348 start = (uint8_t *)&((ufs2_daddr_t *)bp->b_data)[off+1]; 6349 } 6350 if (freework) { 6351 /* Zero the truncated pointers. */ 6352 end = bp->b_data + bp->b_bcount; 6353 bzero(start, end - start); 6354 bdwrite(bp); 6355 } else 6356 bqrelse(bp); 6357 if (level == 0) 6358 return (0); 6359 lbn++; /* adjust level */ 6360 lbn -= (off * lbnadd); 6361 return setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno); 6362} 6363 6364/* 6365 * Complete the partial truncation of an indirect block setup by 6366 * setup_trunc_indir(). This zeros the truncated pointers in the saved 6367 * copy and writes them to disk before the freeblks is allowed to complete. 6368 */ 6369static void 6370complete_trunc_indir(freework) 6371 struct freework *freework; 6372{ 6373 struct freework *fwn; 6374 struct indirdep *indirdep; 6375 struct ufsmount *ump; 6376 struct buf *bp; 6377 uintptr_t start; 6378 int count; 6379 6380 ump = VFSTOUFS(freework->fw_list.wk_mp); 6381 LOCK_OWNED(ump); 6382 indirdep = freework->fw_indir; 6383 for (;;) { 6384 bp = indirdep->ir_bp; 6385 /* See if the block was discarded. */ 6386 if (bp == NULL) 6387 break; 6388 /* Inline part of getdirtybuf(). We dont want bremfree. */ 6389 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0) 6390 break; 6391 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 6392 LOCK_PTR(ump)) == 0) 6393 BUF_UNLOCK(bp); 6394 ACQUIRE_LOCK(ump); 6395 } 6396 freework->fw_state |= DEPCOMPLETE; 6397 TAILQ_REMOVE(&indirdep->ir_trunc, freework, fw_next); 6398 /* 6399 * Zero the pointers in the saved copy. 6400 */ 6401 if (indirdep->ir_state & UFS1FMT) 6402 start = sizeof(ufs1_daddr_t); 6403 else 6404 start = sizeof(ufs2_daddr_t); 6405 start *= freework->fw_start; 6406 count = indirdep->ir_savebp->b_bcount - start; 6407 start += (uintptr_t)indirdep->ir_savebp->b_data; 6408 bzero((char *)start, count); 6409 /* 6410 * We need to start the next truncation in the list if it has not 6411 * been started yet. 6412 */ 6413 fwn = TAILQ_FIRST(&indirdep->ir_trunc); 6414 if (fwn != NULL) { 6415 if (fwn->fw_freeblks == indirdep->ir_freeblks) 6416 TAILQ_REMOVE(&indirdep->ir_trunc, fwn, fw_next); 6417 if ((fwn->fw_state & ONWORKLIST) == 0) 6418 freework_enqueue(fwn); 6419 } 6420 /* 6421 * If bp is NULL the block was fully truncated, restore 6422 * the saved block list otherwise free it if it is no 6423 * longer needed. 6424 */ 6425 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 6426 if (bp == NULL) 6427 bcopy(indirdep->ir_saveddata, 6428 indirdep->ir_savebp->b_data, 6429 indirdep->ir_savebp->b_bcount); 6430 free(indirdep->ir_saveddata, M_INDIRDEP); 6431 indirdep->ir_saveddata = NULL; 6432 } 6433 /* 6434 * When bp is NULL there is a full truncation pending. We 6435 * must wait for this full truncation to be journaled before 6436 * we can release this freework because the disk pointers will 6437 * never be written as zero. 6438 */ 6439 if (bp == NULL) { 6440 if (LIST_EMPTY(&indirdep->ir_freeblks->fb_jblkdephd)) 6441 handle_written_freework(freework); 6442 else 6443 WORKLIST_INSERT(&indirdep->ir_freeblks->fb_freeworkhd, 6444 &freework->fw_list); 6445 } else { 6446 /* Complete when the real copy is written. */ 6447 WORKLIST_INSERT(&bp->b_dep, &freework->fw_list); 6448 BUF_UNLOCK(bp); 6449 } 6450} 6451 6452/* 6453 * Calculate the number of blocks we are going to release where datablocks 6454 * is the current total and length is the new file size. 6455 */ 6456static ufs2_daddr_t 6457blkcount(fs, datablocks, length) 6458 struct fs *fs; 6459 ufs2_daddr_t datablocks; 6460 off_t length; 6461{ 6462 off_t totblks, numblks; 6463 6464 totblks = 0; 6465 numblks = howmany(length, fs->fs_bsize); 6466 if (numblks <= UFS_NDADDR) { 6467 totblks = howmany(length, fs->fs_fsize); 6468 goto out; 6469 } 6470 totblks = blkstofrags(fs, numblks); 6471 numblks -= UFS_NDADDR; 6472 /* 6473 * Count all single, then double, then triple indirects required. 6474 * Subtracting one indirects worth of blocks for each pass 6475 * acknowledges one of each pointed to by the inode. 6476 */ 6477 for (;;) { 6478 totblks += blkstofrags(fs, howmany(numblks, NINDIR(fs))); 6479 numblks -= NINDIR(fs); 6480 if (numblks <= 0) 6481 break; 6482 numblks = howmany(numblks, NINDIR(fs)); 6483 } 6484out: 6485 totblks = fsbtodb(fs, totblks); 6486 /* 6487 * Handle sparse files. We can't reclaim more blocks than the inode 6488 * references. We will correct it later in handle_complete_freeblks() 6489 * when we know the real count. 6490 */ 6491 if (totblks > datablocks) 6492 return (0); 6493 return (datablocks - totblks); 6494} 6495 6496/* 6497 * Handle freeblocks for journaled softupdate filesystems. 6498 * 6499 * Contrary to normal softupdates, we must preserve the block pointers in 6500 * indirects until their subordinates are free. This is to avoid journaling 6501 * every block that is freed which may consume more space than the journal 6502 * itself. The recovery program will see the free block journals at the 6503 * base of the truncated area and traverse them to reclaim space. The 6504 * pointers in the inode may be cleared immediately after the journal 6505 * records are written because each direct and indirect pointer in the 6506 * inode is recorded in a journal. This permits full truncation to proceed 6507 * asynchronously. The write order is journal -> inode -> cgs -> indirects. 6508 * 6509 * The algorithm is as follows: 6510 * 1) Traverse the in-memory state and create journal entries to release 6511 * the relevant blocks and full indirect trees. 6512 * 2) Traverse the indirect block chain adding partial truncation freework 6513 * records to indirects in the path to lastlbn. The freework will 6514 * prevent new allocation dependencies from being satisfied in this 6515 * indirect until the truncation completes. 6516 * 3) Read and lock the inode block, performing an update with the new size 6517 * and pointers. This prevents truncated data from becoming valid on 6518 * disk through step 4. 6519 * 4) Reap unsatisfied dependencies that are beyond the truncated area, 6520 * eliminate journal work for those records that do not require it. 6521 * 5) Schedule the journal records to be written followed by the inode block. 6522 * 6) Allocate any necessary frags for the end of file. 6523 * 7) Zero any partially truncated blocks. 6524 * 6525 * From this truncation proceeds asynchronously using the freework and 6526 * indir_trunc machinery. The file will not be extended again into a 6527 * partially truncated indirect block until all work is completed but 6528 * the normal dependency mechanism ensures that it is rolled back/forward 6529 * as appropriate. Further truncation may occur without delay and is 6530 * serialized in indir_trunc(). 6531 */ 6532void 6533softdep_journal_freeblocks(ip, cred, length, flags) 6534 struct inode *ip; /* The inode whose length is to be reduced */ 6535 struct ucred *cred; 6536 off_t length; /* The new length for the file */ 6537 int flags; /* IO_EXT and/or IO_NORMAL */ 6538{ 6539 struct freeblks *freeblks, *fbn; 6540 struct worklist *wk, *wkn; 6541 struct inodedep *inodedep; 6542 struct jblkdep *jblkdep; 6543 struct allocdirect *adp, *adpn; 6544 struct ufsmount *ump; 6545 struct fs *fs; 6546 struct buf *bp; 6547 struct vnode *vp; 6548 struct mount *mp; 6549 ufs2_daddr_t extblocks, datablocks; 6550 ufs_lbn_t tmpval, lbn, lastlbn; 6551 int frags, lastoff, iboff, allocblock, needj, error, i; 6552 6553 ump = ITOUMP(ip); 6554 mp = UFSTOVFS(ump); 6555 fs = ump->um_fs; 6556 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 6557 ("softdep_journal_freeblocks called on non-softdep filesystem")); 6558 vp = ITOV(ip); 6559 needj = 1; 6560 iboff = -1; 6561 allocblock = 0; 6562 extblocks = 0; 6563 datablocks = 0; 6564 frags = 0; 6565 freeblks = newfreeblks(mp, ip); 6566 ACQUIRE_LOCK(ump); 6567 /* 6568 * If we're truncating a removed file that will never be written 6569 * we don't need to journal the block frees. The canceled journals 6570 * for the allocations will suffice. 6571 */ 6572 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6573 if ((inodedep->id_state & (UNLINKED | DEPCOMPLETE)) == UNLINKED && 6574 length == 0) 6575 needj = 0; 6576 CTR3(KTR_SUJ, "softdep_journal_freeblks: ip %d length %ld needj %d", 6577 ip->i_number, length, needj); 6578 FREE_LOCK(ump); 6579 /* 6580 * Calculate the lbn that we are truncating to. This results in -1 6581 * if we're truncating the 0 bytes. So it is the last lbn we want 6582 * to keep, not the first lbn we want to truncate. 6583 */ 6584 lastlbn = lblkno(fs, length + fs->fs_bsize - 1) - 1; 6585 lastoff = blkoff(fs, length); 6586 /* 6587 * Compute frags we are keeping in lastlbn. 0 means all. 6588 */ 6589 if (lastlbn >= 0 && lastlbn < UFS_NDADDR) { 6590 frags = fragroundup(fs, lastoff); 6591 /* adp offset of last valid allocdirect. */ 6592 iboff = lastlbn; 6593 } else if (lastlbn > 0) 6594 iboff = UFS_NDADDR; 6595 if (fs->fs_magic == FS_UFS2_MAGIC) 6596 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 6597 /* 6598 * Handle normal data blocks and indirects. This section saves 6599 * values used after the inode update to complete frag and indirect 6600 * truncation. 6601 */ 6602 if ((flags & IO_NORMAL) != 0) { 6603 /* 6604 * Handle truncation of whole direct and indirect blocks. 6605 */ 6606 for (i = iboff + 1; i < UFS_NDADDR; i++) 6607 setup_freedirect(freeblks, ip, i, needj); 6608 for (i = 0, tmpval = NINDIR(fs), lbn = UFS_NDADDR; 6609 i < UFS_NIADDR; 6610 i++, lbn += tmpval, tmpval *= NINDIR(fs)) { 6611 /* Release a whole indirect tree. */ 6612 if (lbn > lastlbn) { 6613 setup_freeindir(freeblks, ip, i, -lbn -i, 6614 needj); 6615 continue; 6616 } 6617 iboff = i + UFS_NDADDR; 6618 /* 6619 * Traverse partially truncated indirect tree. 6620 */ 6621 if (lbn <= lastlbn && lbn + tmpval - 1 > lastlbn) 6622 setup_trunc_indir(freeblks, ip, -lbn - i, 6623 lastlbn, DIP(ip, i_ib[i])); 6624 } 6625 /* 6626 * Handle partial truncation to a frag boundary. 6627 */ 6628 if (frags) { 6629 ufs2_daddr_t blkno; 6630 long oldfrags; 6631 6632 oldfrags = blksize(fs, ip, lastlbn); 6633 blkno = DIP(ip, i_db[lastlbn]); 6634 if (blkno && oldfrags != frags) { 6635 oldfrags -= frags; 6636 oldfrags = numfrags(fs, oldfrags); 6637 blkno += numfrags(fs, frags); 6638 newfreework(ump, freeblks, NULL, lastlbn, 6639 blkno, oldfrags, 0, needj); 6640 if (needj) 6641 adjust_newfreework(freeblks, 6642 numfrags(fs, frags)); 6643 } else if (blkno == 0) 6644 allocblock = 1; 6645 } 6646 /* 6647 * Add a journal record for partial truncate if we are 6648 * handling indirect blocks. Non-indirects need no extra 6649 * journaling. 6650 */ 6651 if (length != 0 && lastlbn >= UFS_NDADDR) { 6652 ip->i_flag |= IN_TRUNCATED; 6653 newjtrunc(freeblks, length, 0); 6654 } 6655 ip->i_size = length; 6656 DIP_SET(ip, i_size, ip->i_size); 6657 ip->i_flag |= IN_SIZEMOD | IN_CHANGE; 6658 datablocks = DIP(ip, i_blocks) - extblocks; 6659 if (length != 0) 6660 datablocks = blkcount(fs, datablocks, length); 6661 freeblks->fb_len = length; 6662 } 6663 if ((flags & IO_EXT) != 0) { 6664 for (i = 0; i < UFS_NXADDR; i++) 6665 setup_freeext(freeblks, ip, i, needj); 6666 ip->i_din2->di_extsize = 0; 6667 datablocks += extblocks; 6668 ip->i_flag |= IN_SIZEMOD | IN_CHANGE; 6669 } 6670#ifdef QUOTA 6671 /* Reference the quotas in case the block count is wrong in the end. */ 6672 quotaref(vp, freeblks->fb_quota); 6673 (void) chkdq(ip, -datablocks, NOCRED, FORCE); 6674#endif 6675 freeblks->fb_chkcnt = -datablocks; 6676 UFS_LOCK(ump); 6677 fs->fs_pendingblocks += datablocks; 6678 UFS_UNLOCK(ump); 6679 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks); 6680 /* 6681 * Handle truncation of incomplete alloc direct dependencies. We 6682 * hold the inode block locked to prevent incomplete dependencies 6683 * from reaching the disk while we are eliminating those that 6684 * have been truncated. This is a partially inlined ffs_update(). 6685 */ 6686 ufs_itimes(vp); 6687 ip->i_flag &= ~(IN_LAZYACCESS | IN_LAZYMOD | IN_MODIFIED); 6688 error = bread(ump->um_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 6689 (int)fs->fs_bsize, cred, &bp); 6690 if (error) { 6691 brelse(bp); 6692 softdep_error("softdep_journal_freeblocks", error); 6693 return; 6694 } 6695 if (bp->b_bufsize == fs->fs_bsize) 6696 bp->b_flags |= B_CLUSTEROK; 6697 softdep_update_inodeblock(ip, bp, 0); 6698 if (ump->um_fstype == UFS1) 6699 *((struct ufs1_dinode *)bp->b_data + 6700 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1; 6701 else 6702 *((struct ufs2_dinode *)bp->b_data + 6703 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2; 6704 ACQUIRE_LOCK(ump); 6705 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6706 if ((inodedep->id_state & IOSTARTED) != 0) 6707 panic("softdep_setup_freeblocks: inode busy"); 6708 /* 6709 * Add the freeblks structure to the list of operations that 6710 * must await the zero'ed inode being written to disk. If we 6711 * still have a bitmap dependency (needj), then the inode 6712 * has never been written to disk, so we can process the 6713 * freeblks below once we have deleted the dependencies. 6714 */ 6715 if (needj) 6716 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list); 6717 else 6718 freeblks->fb_state |= COMPLETE; 6719 if ((flags & IO_NORMAL) != 0) { 6720 TAILQ_FOREACH_SAFE(adp, &inodedep->id_inoupdt, ad_next, adpn) { 6721 if (adp->ad_offset > iboff) 6722 cancel_allocdirect(&inodedep->id_inoupdt, adp, 6723 freeblks); 6724 /* 6725 * Truncate the allocdirect. We could eliminate 6726 * or modify journal records as well. 6727 */ 6728 else if (adp->ad_offset == iboff && frags) 6729 adp->ad_newsize = frags; 6730 } 6731 } 6732 if ((flags & IO_EXT) != 0) 6733 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 6734 cancel_allocdirect(&inodedep->id_extupdt, adp, 6735 freeblks); 6736 /* 6737 * Scan the bufwait list for newblock dependencies that will never 6738 * make it to disk. 6739 */ 6740 LIST_FOREACH_SAFE(wk, &inodedep->id_bufwait, wk_list, wkn) { 6741 if (wk->wk_type != D_ALLOCDIRECT) 6742 continue; 6743 adp = WK_ALLOCDIRECT(wk); 6744 if (((flags & IO_NORMAL) != 0 && (adp->ad_offset > iboff)) || 6745 ((flags & IO_EXT) != 0 && (adp->ad_state & EXTDATA))) { 6746 cancel_jfreeblk(freeblks, adp->ad_newblkno); 6747 cancel_newblk(WK_NEWBLK(wk), NULL, &freeblks->fb_jwork); 6748 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk); 6749 } 6750 } 6751 /* 6752 * Add journal work. 6753 */ 6754 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) 6755 add_to_journal(&jblkdep->jb_list); 6756 FREE_LOCK(ump); 6757 bdwrite(bp); 6758 /* 6759 * Truncate dependency structures beyond length. 6760 */ 6761 trunc_dependencies(ip, freeblks, lastlbn, frags, flags); 6762 /* 6763 * This is only set when we need to allocate a fragment because 6764 * none existed at the end of a frag-sized file. It handles only 6765 * allocating a new, zero filled block. 6766 */ 6767 if (allocblock) { 6768 ip->i_size = length - lastoff; 6769 DIP_SET(ip, i_size, ip->i_size); 6770 error = UFS_BALLOC(vp, length - 1, 1, cred, BA_CLRBUF, &bp); 6771 if (error != 0) { 6772 softdep_error("softdep_journal_freeblks", error); 6773 return; 6774 } 6775 ip->i_size = length; 6776 DIP_SET(ip, i_size, length); 6777 ip->i_flag |= IN_SIZEMOD | IN_CHANGE | IN_UPDATE; 6778 allocbuf(bp, frags); 6779 ffs_update(vp, 0); 6780 bawrite(bp); 6781 } else if (lastoff != 0 && vp->v_type != VDIR) { 6782 int size; 6783 6784 /* 6785 * Zero the end of a truncated frag or block. 6786 */ 6787 size = sblksize(fs, length, lastlbn); 6788 error = bread(vp, lastlbn, size, cred, &bp); 6789 if (error) { 6790 softdep_error("softdep_journal_freeblks", error); 6791 return; 6792 } 6793 bzero((char *)bp->b_data + lastoff, size - lastoff); 6794 bawrite(bp); 6795 6796 } 6797 ACQUIRE_LOCK(ump); 6798 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6799 TAILQ_INSERT_TAIL(&inodedep->id_freeblklst, freeblks, fb_next); 6800 freeblks->fb_state |= DEPCOMPLETE | ONDEPLIST; 6801 /* 6802 * We zero earlier truncations so they don't erroneously 6803 * update i_blocks. 6804 */ 6805 if (freeblks->fb_len == 0 && (flags & IO_NORMAL) != 0) 6806 TAILQ_FOREACH(fbn, &inodedep->id_freeblklst, fb_next) 6807 fbn->fb_len = 0; 6808 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE && 6809 LIST_EMPTY(&freeblks->fb_jblkdephd)) 6810 freeblks->fb_state |= INPROGRESS; 6811 else 6812 freeblks = NULL; 6813 FREE_LOCK(ump); 6814 if (freeblks) 6815 handle_workitem_freeblocks(freeblks, 0); 6816 trunc_pages(ip, length, extblocks, flags); 6817 6818} 6819 6820/* 6821 * Flush a JOP_SYNC to the journal. 6822 */ 6823void 6824softdep_journal_fsync(ip) 6825 struct inode *ip; 6826{ 6827 struct jfsync *jfsync; 6828 struct ufsmount *ump; 6829 6830 ump = ITOUMP(ip); 6831 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 6832 ("softdep_journal_fsync called on non-softdep filesystem")); 6833 if ((ip->i_flag & IN_TRUNCATED) == 0) 6834 return; 6835 ip->i_flag &= ~IN_TRUNCATED; 6836 jfsync = malloc(sizeof(*jfsync), M_JFSYNC, M_SOFTDEP_FLAGS | M_ZERO); 6837 workitem_alloc(&jfsync->jfs_list, D_JFSYNC, UFSTOVFS(ump)); 6838 jfsync->jfs_size = ip->i_size; 6839 jfsync->jfs_ino = ip->i_number; 6840 ACQUIRE_LOCK(ump); 6841 add_to_journal(&jfsync->jfs_list); 6842 jwait(&jfsync->jfs_list, MNT_WAIT); 6843 FREE_LOCK(ump); 6844} 6845 6846/* 6847 * Block de-allocation dependencies. 6848 * 6849 * When blocks are de-allocated, the on-disk pointers must be nullified before 6850 * the blocks are made available for use by other files. (The true 6851 * requirement is that old pointers must be nullified before new on-disk 6852 * pointers are set. We chose this slightly more stringent requirement to 6853 * reduce complexity.) Our implementation handles this dependency by updating 6854 * the inode (or indirect block) appropriately but delaying the actual block 6855 * de-allocation (i.e., freemap and free space count manipulation) until 6856 * after the updated versions reach stable storage. After the disk is 6857 * updated, the blocks can be safely de-allocated whenever it is convenient. 6858 * This implementation handles only the common case of reducing a file's 6859 * length to zero. Other cases are handled by the conventional synchronous 6860 * write approach. 6861 * 6862 * The ffs implementation with which we worked double-checks 6863 * the state of the block pointers and file size as it reduces 6864 * a file's length. Some of this code is replicated here in our 6865 * soft updates implementation. The freeblks->fb_chkcnt field is 6866 * used to transfer a part of this information to the procedure 6867 * that eventually de-allocates the blocks. 6868 * 6869 * This routine should be called from the routine that shortens 6870 * a file's length, before the inode's size or block pointers 6871 * are modified. It will save the block pointer information for 6872 * later release and zero the inode so that the calling routine 6873 * can release it. 6874 */ 6875void 6876softdep_setup_freeblocks(ip, length, flags) 6877 struct inode *ip; /* The inode whose length is to be reduced */ 6878 off_t length; /* The new length for the file */ 6879 int flags; /* IO_EXT and/or IO_NORMAL */ 6880{ 6881 struct ufs1_dinode *dp1; 6882 struct ufs2_dinode *dp2; 6883 struct freeblks *freeblks; 6884 struct inodedep *inodedep; 6885 struct allocdirect *adp; 6886 struct ufsmount *ump; 6887 struct buf *bp; 6888 struct fs *fs; 6889 ufs2_daddr_t extblocks, datablocks; 6890 struct mount *mp; 6891 int i, delay, error; 6892 ufs_lbn_t tmpval; 6893 ufs_lbn_t lbn; 6894 6895 ump = ITOUMP(ip); 6896 mp = UFSTOVFS(ump); 6897 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 6898 ("softdep_setup_freeblocks called on non-softdep filesystem")); 6899 CTR2(KTR_SUJ, "softdep_setup_freeblks: ip %d length %ld", 6900 ip->i_number, length); 6901 KASSERT(length == 0, ("softdep_setup_freeblocks: non-zero length")); 6902 fs = ump->um_fs; 6903 if ((error = bread(ump->um_devvp, 6904 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 6905 (int)fs->fs_bsize, NOCRED, &bp)) != 0) { 6906 brelse(bp); 6907 softdep_error("softdep_setup_freeblocks", error); 6908 return; 6909 } 6910 freeblks = newfreeblks(mp, ip); 6911 extblocks = 0; 6912 datablocks = 0; 6913 if (fs->fs_magic == FS_UFS2_MAGIC) 6914 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 6915 if ((flags & IO_NORMAL) != 0) { 6916 for (i = 0; i < UFS_NDADDR; i++) 6917 setup_freedirect(freeblks, ip, i, 0); 6918 for (i = 0, tmpval = NINDIR(fs), lbn = UFS_NDADDR; 6919 i < UFS_NIADDR; 6920 i++, lbn += tmpval, tmpval *= NINDIR(fs)) 6921 setup_freeindir(freeblks, ip, i, -lbn -i, 0); 6922 ip->i_size = 0; 6923 DIP_SET(ip, i_size, 0); 6924 ip->i_flag |= IN_SIZEMOD | IN_CHANGE; 6925 datablocks = DIP(ip, i_blocks) - extblocks; 6926 } 6927 if ((flags & IO_EXT) != 0) { 6928 for (i = 0; i < UFS_NXADDR; i++) 6929 setup_freeext(freeblks, ip, i, 0); 6930 ip->i_din2->di_extsize = 0; 6931 datablocks += extblocks; 6932 ip->i_flag |= IN_SIZEMOD | IN_CHANGE; 6933 } 6934#ifdef QUOTA 6935 /* Reference the quotas in case the block count is wrong in the end. */ 6936 quotaref(ITOV(ip), freeblks->fb_quota); 6937 (void) chkdq(ip, -datablocks, NOCRED, FORCE); 6938#endif 6939 freeblks->fb_chkcnt = -datablocks; 6940 UFS_LOCK(ump); 6941 fs->fs_pendingblocks += datablocks; 6942 UFS_UNLOCK(ump); 6943 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks); 6944 /* 6945 * Push the zero'ed inode to its disk buffer so that we are free 6946 * to delete its dependencies below. Once the dependencies are gone 6947 * the buffer can be safely released. 6948 */ 6949 if (ump->um_fstype == UFS1) { 6950 dp1 = ((struct ufs1_dinode *)bp->b_data + 6951 ino_to_fsbo(fs, ip->i_number)); 6952 ip->i_din1->di_freelink = dp1->di_freelink; 6953 *dp1 = *ip->i_din1; 6954 } else { 6955 dp2 = ((struct ufs2_dinode *)bp->b_data + 6956 ino_to_fsbo(fs, ip->i_number)); 6957 ip->i_din2->di_freelink = dp2->di_freelink; 6958 *dp2 = *ip->i_din2; 6959 } 6960 /* 6961 * Find and eliminate any inode dependencies. 6962 */ 6963 ACQUIRE_LOCK(ump); 6964 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6965 if ((inodedep->id_state & IOSTARTED) != 0) 6966 panic("softdep_setup_freeblocks: inode busy"); 6967 /* 6968 * Add the freeblks structure to the list of operations that 6969 * must await the zero'ed inode being written to disk. If we 6970 * still have a bitmap dependency (delay == 0), then the inode 6971 * has never been written to disk, so we can process the 6972 * freeblks below once we have deleted the dependencies. 6973 */ 6974 delay = (inodedep->id_state & DEPCOMPLETE); 6975 if (delay) 6976 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list); 6977 else 6978 freeblks->fb_state |= COMPLETE; 6979 /* 6980 * Because the file length has been truncated to zero, any 6981 * pending block allocation dependency structures associated 6982 * with this inode are obsolete and can simply be de-allocated. 6983 * We must first merge the two dependency lists to get rid of 6984 * any duplicate freefrag structures, then purge the merged list. 6985 * If we still have a bitmap dependency, then the inode has never 6986 * been written to disk, so we can free any fragments without delay. 6987 */ 6988 if (flags & IO_NORMAL) { 6989 merge_inode_lists(&inodedep->id_newinoupdt, 6990 &inodedep->id_inoupdt); 6991 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 6992 cancel_allocdirect(&inodedep->id_inoupdt, adp, 6993 freeblks); 6994 } 6995 if (flags & IO_EXT) { 6996 merge_inode_lists(&inodedep->id_newextupdt, 6997 &inodedep->id_extupdt); 6998 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 6999 cancel_allocdirect(&inodedep->id_extupdt, adp, 7000 freeblks); 7001 } 7002 FREE_LOCK(ump); 7003 bdwrite(bp); 7004 trunc_dependencies(ip, freeblks, -1, 0, flags); 7005 ACQUIRE_LOCK(ump); 7006 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 7007 (void) free_inodedep(inodedep); 7008 freeblks->fb_state |= DEPCOMPLETE; 7009 /* 7010 * If the inode with zeroed block pointers is now on disk 7011 * we can start freeing blocks. 7012 */ 7013 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 7014 freeblks->fb_state |= INPROGRESS; 7015 else 7016 freeblks = NULL; 7017 FREE_LOCK(ump); 7018 if (freeblks) 7019 handle_workitem_freeblocks(freeblks, 0); 7020 trunc_pages(ip, length, extblocks, flags); 7021} 7022 7023/* 7024 * Eliminate pages from the page cache that back parts of this inode and 7025 * adjust the vnode pager's idea of our size. This prevents stale data 7026 * from hanging around in the page cache. 7027 */ 7028static void 7029trunc_pages(ip, length, extblocks, flags) 7030 struct inode *ip; 7031 off_t length; 7032 ufs2_daddr_t extblocks; 7033 int flags; 7034{ 7035 struct vnode *vp; 7036 struct fs *fs; 7037 ufs_lbn_t lbn; 7038 off_t end, extend; 7039 7040 vp = ITOV(ip); 7041 fs = ITOFS(ip); 7042 extend = OFF_TO_IDX(lblktosize(fs, -extblocks)); 7043 if ((flags & IO_EXT) != 0) 7044 vn_pages_remove(vp, extend, 0); 7045 if ((flags & IO_NORMAL) == 0) 7046 return; 7047 BO_LOCK(&vp->v_bufobj); 7048 drain_output(vp); 7049 BO_UNLOCK(&vp->v_bufobj); 7050 /* 7051 * The vnode pager eliminates file pages we eliminate indirects 7052 * below. 7053 */ 7054 vnode_pager_setsize(vp, length); 7055 /* 7056 * Calculate the end based on the last indirect we want to keep. If 7057 * the block extends into indirects we can just use the negative of 7058 * its lbn. Doubles and triples exist at lower numbers so we must 7059 * be careful not to remove those, if they exist. double and triple 7060 * indirect lbns do not overlap with others so it is not important 7061 * to verify how many levels are required. 7062 */ 7063 lbn = lblkno(fs, length); 7064 if (lbn >= UFS_NDADDR) { 7065 /* Calculate the virtual lbn of the triple indirect. */ 7066 lbn = -lbn - (UFS_NIADDR - 1); 7067 end = OFF_TO_IDX(lblktosize(fs, lbn)); 7068 } else 7069 end = extend; 7070 vn_pages_remove(vp, OFF_TO_IDX(OFF_MAX), end); 7071} 7072 7073/* 7074 * See if the buf bp is in the range eliminated by truncation. 7075 */ 7076static int 7077trunc_check_buf(bp, blkoffp, lastlbn, lastoff, flags) 7078 struct buf *bp; 7079 int *blkoffp; 7080 ufs_lbn_t lastlbn; 7081 int lastoff; 7082 int flags; 7083{ 7084 ufs_lbn_t lbn; 7085 7086 *blkoffp = 0; 7087 /* Only match ext/normal blocks as appropriate. */ 7088 if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) || 7089 ((flags & IO_NORMAL) == 0 && (bp->b_xflags & BX_ALTDATA) == 0)) 7090 return (0); 7091 /* ALTDATA is always a full truncation. */ 7092 if ((bp->b_xflags & BX_ALTDATA) != 0) 7093 return (1); 7094 /* -1 is full truncation. */ 7095 if (lastlbn == -1) 7096 return (1); 7097 /* 7098 * If this is a partial truncate we only want those 7099 * blocks and indirect blocks that cover the range 7100 * we're after. 7101 */ 7102 lbn = bp->b_lblkno; 7103 if (lbn < 0) 7104 lbn = -(lbn + lbn_level(lbn)); 7105 if (lbn < lastlbn) 7106 return (0); 7107 /* Here we only truncate lblkno if it's partial. */ 7108 if (lbn == lastlbn) { 7109 if (lastoff == 0) 7110 return (0); 7111 *blkoffp = lastoff; 7112 } 7113 return (1); 7114} 7115 7116/* 7117 * Eliminate any dependencies that exist in memory beyond lblkno:off 7118 */ 7119static void 7120trunc_dependencies(ip, freeblks, lastlbn, lastoff, flags) 7121 struct inode *ip; 7122 struct freeblks *freeblks; 7123 ufs_lbn_t lastlbn; 7124 int lastoff; 7125 int flags; 7126{ 7127 struct bufobj *bo; 7128 struct vnode *vp; 7129 struct buf *bp; 7130 int blkoff; 7131 7132 /* 7133 * We must wait for any I/O in progress to finish so that 7134 * all potential buffers on the dirty list will be visible. 7135 * Once they are all there, walk the list and get rid of 7136 * any dependencies. 7137 */ 7138 vp = ITOV(ip); 7139 bo = &vp->v_bufobj; 7140 BO_LOCK(bo); 7141 drain_output(vp); 7142 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) 7143 bp->b_vflags &= ~BV_SCANNED; 7144restart: 7145 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) { 7146 if (bp->b_vflags & BV_SCANNED) 7147 continue; 7148 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) { 7149 bp->b_vflags |= BV_SCANNED; 7150 continue; 7151 } 7152 KASSERT(bp->b_bufobj == bo, ("Wrong object in buffer")); 7153 if ((bp = getdirtybuf(bp, BO_LOCKPTR(bo), MNT_WAIT)) == NULL) 7154 goto restart; 7155 BO_UNLOCK(bo); 7156 if (deallocate_dependencies(bp, freeblks, blkoff)) 7157 bqrelse(bp); 7158 else 7159 brelse(bp); 7160 BO_LOCK(bo); 7161 goto restart; 7162 } 7163 /* 7164 * Now do the work of vtruncbuf while also matching indirect blocks. 7165 */ 7166 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) 7167 bp->b_vflags &= ~BV_SCANNED; 7168cleanrestart: 7169 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) { 7170 if (bp->b_vflags & BV_SCANNED) 7171 continue; 7172 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) { 7173 bp->b_vflags |= BV_SCANNED; 7174 continue; 7175 } 7176 if (BUF_LOCK(bp, 7177 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 7178 BO_LOCKPTR(bo)) == ENOLCK) { 7179 BO_LOCK(bo); 7180 goto cleanrestart; 7181 } 7182 BO_LOCK(bo); 7183 bp->b_vflags |= BV_SCANNED; 7184 BO_UNLOCK(bo); 7185 bremfree(bp); 7186 if (blkoff != 0) { 7187 allocbuf(bp, blkoff); 7188 bqrelse(bp); 7189 } else { 7190 bp->b_flags |= B_INVAL | B_NOCACHE | B_RELBUF; 7191 brelse(bp); 7192 } 7193 BO_LOCK(bo); 7194 goto cleanrestart; 7195 } 7196 drain_output(vp); 7197 BO_UNLOCK(bo); 7198} 7199 7200static int 7201cancel_pagedep(pagedep, freeblks, blkoff) 7202 struct pagedep *pagedep; 7203 struct freeblks *freeblks; 7204 int blkoff; 7205{ 7206 struct jremref *jremref; 7207 struct jmvref *jmvref; 7208 struct dirrem *dirrem, *tmp; 7209 int i; 7210 7211 /* 7212 * Copy any directory remove dependencies to the list 7213 * to be processed after the freeblks proceeds. If 7214 * directory entry never made it to disk they 7215 * can be dumped directly onto the work list. 7216 */ 7217 LIST_FOREACH_SAFE(dirrem, &pagedep->pd_dirremhd, dm_next, tmp) { 7218 /* Skip this directory removal if it is intended to remain. */ 7219 if (dirrem->dm_offset < blkoff) 7220 continue; 7221 /* 7222 * If there are any dirrems we wait for the journal write 7223 * to complete and then restart the buf scan as the lock 7224 * has been dropped. 7225 */ 7226 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) { 7227 jwait(&jremref->jr_list, MNT_WAIT); 7228 return (ERESTART); 7229 } 7230 LIST_REMOVE(dirrem, dm_next); 7231 dirrem->dm_dirinum = pagedep->pd_ino; 7232 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &dirrem->dm_list); 7233 } 7234 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) { 7235 jwait(&jmvref->jm_list, MNT_WAIT); 7236 return (ERESTART); 7237 } 7238 /* 7239 * When we're partially truncating a pagedep we just want to flush 7240 * journal entries and return. There can not be any adds in the 7241 * truncated portion of the directory and newblk must remain if 7242 * part of the block remains. 7243 */ 7244 if (blkoff != 0) { 7245 struct diradd *dap; 7246 7247 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 7248 if (dap->da_offset > blkoff) 7249 panic("cancel_pagedep: diradd %p off %d > %d", 7250 dap, dap->da_offset, blkoff); 7251 for (i = 0; i < DAHASHSZ; i++) 7252 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) 7253 if (dap->da_offset > blkoff) 7254 panic("cancel_pagedep: diradd %p off %d > %d", 7255 dap, dap->da_offset, blkoff); 7256 return (0); 7257 } 7258 /* 7259 * There should be no directory add dependencies present 7260 * as the directory could not be truncated until all 7261 * children were removed. 7262 */ 7263 KASSERT(LIST_FIRST(&pagedep->pd_pendinghd) == NULL, 7264 ("deallocate_dependencies: pendinghd != NULL")); 7265 for (i = 0; i < DAHASHSZ; i++) 7266 KASSERT(LIST_FIRST(&pagedep->pd_diraddhd[i]) == NULL, 7267 ("deallocate_dependencies: diraddhd != NULL")); 7268 if ((pagedep->pd_state & NEWBLOCK) != 0) 7269 free_newdirblk(pagedep->pd_newdirblk); 7270 if (free_pagedep(pagedep) == 0) 7271 panic("Failed to free pagedep %p", pagedep); 7272 return (0); 7273} 7274 7275/* 7276 * Reclaim any dependency structures from a buffer that is about to 7277 * be reallocated to a new vnode. The buffer must be locked, thus, 7278 * no I/O completion operations can occur while we are manipulating 7279 * its associated dependencies. The mutex is held so that other I/O's 7280 * associated with related dependencies do not occur. 7281 */ 7282static int 7283deallocate_dependencies(bp, freeblks, off) 7284 struct buf *bp; 7285 struct freeblks *freeblks; 7286 int off; 7287{ 7288 struct indirdep *indirdep; 7289 struct pagedep *pagedep; 7290 struct worklist *wk, *wkn; 7291 struct ufsmount *ump; 7292 7293 ump = softdep_bp_to_mp(bp); 7294 if (ump == NULL) 7295 goto done; 7296 ACQUIRE_LOCK(ump); 7297 LIST_FOREACH_SAFE(wk, &bp->b_dep, wk_list, wkn) { 7298 switch (wk->wk_type) { 7299 case D_INDIRDEP: 7300 indirdep = WK_INDIRDEP(wk); 7301 if (bp->b_lblkno >= 0 || 7302 bp->b_blkno != indirdep->ir_savebp->b_lblkno) 7303 panic("deallocate_dependencies: not indir"); 7304 cancel_indirdep(indirdep, bp, freeblks); 7305 continue; 7306 7307 case D_PAGEDEP: 7308 pagedep = WK_PAGEDEP(wk); 7309 if (cancel_pagedep(pagedep, freeblks, off)) { 7310 FREE_LOCK(ump); 7311 return (ERESTART); 7312 } 7313 continue; 7314 7315 case D_ALLOCINDIR: 7316 /* 7317 * Simply remove the allocindir, we'll find it via 7318 * the indirdep where we can clear pointers if 7319 * needed. 7320 */ 7321 WORKLIST_REMOVE(wk); 7322 continue; 7323 7324 case D_FREEWORK: 7325 /* 7326 * A truncation is waiting for the zero'd pointers 7327 * to be written. It can be freed when the freeblks 7328 * is journaled. 7329 */ 7330 WORKLIST_REMOVE(wk); 7331 wk->wk_state |= ONDEPLIST; 7332 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk); 7333 break; 7334 7335 case D_ALLOCDIRECT: 7336 if (off != 0) 7337 continue; 7338 /* FALLTHROUGH */ 7339 default: 7340 panic("deallocate_dependencies: Unexpected type %s", 7341 TYPENAME(wk->wk_type)); 7342 /* NOTREACHED */ 7343 } 7344 } 7345 FREE_LOCK(ump); 7346done: 7347 /* 7348 * Don't throw away this buf, we were partially truncating and 7349 * some deps may always remain. 7350 */ 7351 if (off) { 7352 allocbuf(bp, off); 7353 bp->b_vflags |= BV_SCANNED; 7354 return (EBUSY); 7355 } 7356 bp->b_flags |= B_INVAL | B_NOCACHE; 7357 7358 return (0); 7359} 7360 7361/* 7362 * An allocdirect is being canceled due to a truncate. We must make sure 7363 * the journal entry is released in concert with the blkfree that releases 7364 * the storage. Completed journal entries must not be released until the 7365 * space is no longer pointed to by the inode or in the bitmap. 7366 */ 7367static void 7368cancel_allocdirect(adphead, adp, freeblks) 7369 struct allocdirectlst *adphead; 7370 struct allocdirect *adp; 7371 struct freeblks *freeblks; 7372{ 7373 struct freework *freework; 7374 struct newblk *newblk; 7375 struct worklist *wk; 7376 7377 TAILQ_REMOVE(adphead, adp, ad_next); 7378 newblk = (struct newblk *)adp; 7379 freework = NULL; 7380 /* 7381 * Find the correct freework structure. 7382 */ 7383 LIST_FOREACH(wk, &freeblks->fb_freeworkhd, wk_list) { 7384 if (wk->wk_type != D_FREEWORK) 7385 continue; 7386 freework = WK_FREEWORK(wk); 7387 if (freework->fw_blkno == newblk->nb_newblkno) 7388 break; 7389 } 7390 if (freework == NULL) 7391 panic("cancel_allocdirect: Freework not found"); 7392 /* 7393 * If a newblk exists at all we still have the journal entry that 7394 * initiated the allocation so we do not need to journal the free. 7395 */ 7396 cancel_jfreeblk(freeblks, freework->fw_blkno); 7397 /* 7398 * If the journal hasn't been written the jnewblk must be passed 7399 * to the call to ffs_blkfree that reclaims the space. We accomplish 7400 * this by linking the journal dependency into the freework to be 7401 * freed when freework_freeblock() is called. If the journal has 7402 * been written we can simply reclaim the journal space when the 7403 * freeblks work is complete. 7404 */ 7405 freework->fw_jnewblk = cancel_newblk(newblk, &freework->fw_list, 7406 &freeblks->fb_jwork); 7407 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list); 7408} 7409 7410 7411/* 7412 * Cancel a new block allocation. May be an indirect or direct block. We 7413 * remove it from various lists and return any journal record that needs to 7414 * be resolved by the caller. 7415 * 7416 * A special consideration is made for indirects which were never pointed 7417 * at on disk and will never be found once this block is released. 7418 */ 7419static struct jnewblk * 7420cancel_newblk(newblk, wk, wkhd) 7421 struct newblk *newblk; 7422 struct worklist *wk; 7423 struct workhead *wkhd; 7424{ 7425 struct jnewblk *jnewblk; 7426 7427 CTR1(KTR_SUJ, "cancel_newblk: blkno %jd", newblk->nb_newblkno); 7428 7429 newblk->nb_state |= GOINGAWAY; 7430 /* 7431 * Previously we traversed the completedhd on each indirdep 7432 * attached to this newblk to cancel them and gather journal 7433 * work. Since we need only the oldest journal segment and 7434 * the lowest point on the tree will always have the oldest 7435 * journal segment we are free to release the segments 7436 * of any subordinates and may leave the indirdep list to 7437 * indirdep_complete() when this newblk is freed. 7438 */ 7439 if (newblk->nb_state & ONDEPLIST) { 7440 newblk->nb_state &= ~ONDEPLIST; 7441 LIST_REMOVE(newblk, nb_deps); 7442 } 7443 if (newblk->nb_state & ONWORKLIST) 7444 WORKLIST_REMOVE(&newblk->nb_list); 7445 /* 7446 * If the journal entry hasn't been written we save a pointer to 7447 * the dependency that frees it until it is written or the 7448 * superseding operation completes. 7449 */ 7450 jnewblk = newblk->nb_jnewblk; 7451 if (jnewblk != NULL && wk != NULL) { 7452 newblk->nb_jnewblk = NULL; 7453 jnewblk->jn_dep = wk; 7454 } 7455 if (!LIST_EMPTY(&newblk->nb_jwork)) 7456 jwork_move(wkhd, &newblk->nb_jwork); 7457 /* 7458 * When truncating we must free the newdirblk early to remove 7459 * the pagedep from the hash before returning. 7460 */ 7461 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL) 7462 free_newdirblk(WK_NEWDIRBLK(wk)); 7463 if (!LIST_EMPTY(&newblk->nb_newdirblk)) 7464 panic("cancel_newblk: extra newdirblk"); 7465 7466 return (jnewblk); 7467} 7468 7469/* 7470 * Schedule the freefrag associated with a newblk to be released once 7471 * the pointers are written and the previous block is no longer needed. 7472 */ 7473static void 7474newblk_freefrag(newblk) 7475 struct newblk *newblk; 7476{ 7477 struct freefrag *freefrag; 7478 7479 if (newblk->nb_freefrag == NULL) 7480 return; 7481 freefrag = newblk->nb_freefrag; 7482 newblk->nb_freefrag = NULL; 7483 freefrag->ff_state |= COMPLETE; 7484 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE) 7485 add_to_worklist(&freefrag->ff_list, 0); 7486} 7487 7488/* 7489 * Free a newblk. Generate a new freefrag work request if appropriate. 7490 * This must be called after the inode pointer and any direct block pointers 7491 * are valid or fully removed via truncate or frag extension. 7492 */ 7493static void 7494free_newblk(newblk) 7495 struct newblk *newblk; 7496{ 7497 struct indirdep *indirdep; 7498 struct worklist *wk; 7499 7500 KASSERT(newblk->nb_jnewblk == NULL, 7501 ("free_newblk: jnewblk %p still attached", newblk->nb_jnewblk)); 7502 KASSERT(newblk->nb_list.wk_type != D_NEWBLK, 7503 ("free_newblk: unclaimed newblk")); 7504 LOCK_OWNED(VFSTOUFS(newblk->nb_list.wk_mp)); 7505 newblk_freefrag(newblk); 7506 if (newblk->nb_state & ONDEPLIST) 7507 LIST_REMOVE(newblk, nb_deps); 7508 if (newblk->nb_state & ONWORKLIST) 7509 WORKLIST_REMOVE(&newblk->nb_list); 7510 LIST_REMOVE(newblk, nb_hash); 7511 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL) 7512 free_newdirblk(WK_NEWDIRBLK(wk)); 7513 if (!LIST_EMPTY(&newblk->nb_newdirblk)) 7514 panic("free_newblk: extra newdirblk"); 7515 while ((indirdep = LIST_FIRST(&newblk->nb_indirdeps)) != NULL) 7516 indirdep_complete(indirdep); 7517 handle_jwork(&newblk->nb_jwork); 7518 WORKITEM_FREE(newblk, D_NEWBLK); 7519} 7520 7521/* 7522 * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep. 7523 */ 7524static void 7525free_newdirblk(newdirblk) 7526 struct newdirblk *newdirblk; 7527{ 7528 struct pagedep *pagedep; 7529 struct diradd *dap; 7530 struct worklist *wk; 7531 7532 LOCK_OWNED(VFSTOUFS(newdirblk->db_list.wk_mp)); 7533 WORKLIST_REMOVE(&newdirblk->db_list); 7534 /* 7535 * If the pagedep is still linked onto the directory buffer 7536 * dependency chain, then some of the entries on the 7537 * pd_pendinghd list may not be committed to disk yet. In 7538 * this case, we will simply clear the NEWBLOCK flag and 7539 * let the pd_pendinghd list be processed when the pagedep 7540 * is next written. If the pagedep is no longer on the buffer 7541 * dependency chain, then all the entries on the pd_pending 7542 * list are committed to disk and we can free them here. 7543 */ 7544 pagedep = newdirblk->db_pagedep; 7545 pagedep->pd_state &= ~NEWBLOCK; 7546 if ((pagedep->pd_state & ONWORKLIST) == 0) { 7547 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 7548 free_diradd(dap, NULL); 7549 /* 7550 * If no dependencies remain, the pagedep will be freed. 7551 */ 7552 free_pagedep(pagedep); 7553 } 7554 /* Should only ever be one item in the list. */ 7555 while ((wk = LIST_FIRST(&newdirblk->db_mkdir)) != NULL) { 7556 WORKLIST_REMOVE(wk); 7557 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 7558 } 7559 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 7560} 7561 7562/* 7563 * Prepare an inode to be freed. The actual free operation is not 7564 * done until the zero'ed inode has been written to disk. 7565 */ 7566void 7567softdep_freefile(pvp, ino, mode) 7568 struct vnode *pvp; 7569 ino_t ino; 7570 int mode; 7571{ 7572 struct inode *ip = VTOI(pvp); 7573 struct inodedep *inodedep; 7574 struct freefile *freefile; 7575 struct freeblks *freeblks; 7576 struct ufsmount *ump; 7577 7578 ump = ITOUMP(ip); 7579 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 7580 ("softdep_freefile called on non-softdep filesystem")); 7581 /* 7582 * This sets up the inode de-allocation dependency. 7583 */ 7584 freefile = malloc(sizeof(struct freefile), 7585 M_FREEFILE, M_SOFTDEP_FLAGS); 7586 workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount); 7587 freefile->fx_mode = mode; 7588 freefile->fx_oldinum = ino; 7589 freefile->fx_devvp = ump->um_devvp; 7590 LIST_INIT(&freefile->fx_jwork); 7591 UFS_LOCK(ump); 7592 ump->um_fs->fs_pendinginodes += 1; 7593 UFS_UNLOCK(ump); 7594 7595 /* 7596 * If the inodedep does not exist, then the zero'ed inode has 7597 * been written to disk. If the allocated inode has never been 7598 * written to disk, then the on-disk inode is zero'ed. In either 7599 * case we can free the file immediately. If the journal was 7600 * canceled before being written the inode will never make it to 7601 * disk and we must send the canceled journal entrys to 7602 * ffs_freefile() to be cleared in conjunction with the bitmap. 7603 * Any blocks waiting on the inode to write can be safely freed 7604 * here as it will never been written. 7605 */ 7606 ACQUIRE_LOCK(ump); 7607 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep); 7608 if (inodedep) { 7609 /* 7610 * Clear out freeblks that no longer need to reference 7611 * this inode. 7612 */ 7613 while ((freeblks = 7614 TAILQ_FIRST(&inodedep->id_freeblklst)) != NULL) { 7615 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, 7616 fb_next); 7617 freeblks->fb_state &= ~ONDEPLIST; 7618 } 7619 /* 7620 * Remove this inode from the unlinked list. 7621 */ 7622 if (inodedep->id_state & UNLINKED) { 7623 /* 7624 * Save the journal work to be freed with the bitmap 7625 * before we clear UNLINKED. Otherwise it can be lost 7626 * if the inode block is written. 7627 */ 7628 handle_bufwait(inodedep, &freefile->fx_jwork); 7629 clear_unlinked_inodedep(inodedep); 7630 /* 7631 * Re-acquire inodedep as we've dropped the 7632 * per-filesystem lock in clear_unlinked_inodedep(). 7633 */ 7634 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep); 7635 } 7636 } 7637 if (inodedep == NULL || check_inode_unwritten(inodedep)) { 7638 FREE_LOCK(ump); 7639 handle_workitem_freefile(freefile); 7640 return; 7641 } 7642 if ((inodedep->id_state & DEPCOMPLETE) == 0) 7643 inodedep->id_state |= GOINGAWAY; 7644 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 7645 FREE_LOCK(ump); 7646 if (ip->i_number == ino) 7647 ip->i_flag |= IN_MODIFIED; 7648} 7649 7650/* 7651 * Check to see if an inode has never been written to disk. If 7652 * so free the inodedep and return success, otherwise return failure. 7653 * 7654 * If we still have a bitmap dependency, then the inode has never 7655 * been written to disk. Drop the dependency as it is no longer 7656 * necessary since the inode is being deallocated. We set the 7657 * ALLCOMPLETE flags since the bitmap now properly shows that the 7658 * inode is not allocated. Even if the inode is actively being 7659 * written, it has been rolled back to its zero'ed state, so we 7660 * are ensured that a zero inode is what is on the disk. For short 7661 * lived files, this change will usually result in removing all the 7662 * dependencies from the inode so that it can be freed immediately. 7663 */ 7664static int 7665check_inode_unwritten(inodedep) 7666 struct inodedep *inodedep; 7667{ 7668 7669 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp)); 7670 7671 if ((inodedep->id_state & (DEPCOMPLETE | UNLINKED)) != 0 || 7672 !LIST_EMPTY(&inodedep->id_dirremhd) || 7673 !LIST_EMPTY(&inodedep->id_pendinghd) || 7674 !LIST_EMPTY(&inodedep->id_bufwait) || 7675 !LIST_EMPTY(&inodedep->id_inowait) || 7676 !TAILQ_EMPTY(&inodedep->id_inoreflst) || 7677 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 7678 !TAILQ_EMPTY(&inodedep->id_newinoupdt) || 7679 !TAILQ_EMPTY(&inodedep->id_extupdt) || 7680 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 7681 !TAILQ_EMPTY(&inodedep->id_freeblklst) || 7682 inodedep->id_mkdiradd != NULL || 7683 inodedep->id_nlinkdelta != 0) 7684 return (0); 7685 /* 7686 * Another process might be in initiate_write_inodeblock_ufs[12] 7687 * trying to allocate memory without holding "Softdep Lock". 7688 */ 7689 if ((inodedep->id_state & IOSTARTED) != 0 && 7690 inodedep->id_savedino1 == NULL) 7691 return (0); 7692 7693 if (inodedep->id_state & ONDEPLIST) 7694 LIST_REMOVE(inodedep, id_deps); 7695 inodedep->id_state &= ~ONDEPLIST; 7696 inodedep->id_state |= ALLCOMPLETE; 7697 inodedep->id_bmsafemap = NULL; 7698 if (inodedep->id_state & ONWORKLIST) 7699 WORKLIST_REMOVE(&inodedep->id_list); 7700 if (inodedep->id_savedino1 != NULL) { 7701 free(inodedep->id_savedino1, M_SAVEDINO); 7702 inodedep->id_savedino1 = NULL; 7703 } 7704 if (free_inodedep(inodedep) == 0) 7705 panic("check_inode_unwritten: busy inode"); 7706 return (1); 7707} 7708 7709static int 7710check_inodedep_free(inodedep) 7711 struct inodedep *inodedep; 7712{ 7713 7714 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp)); 7715 if ((inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 7716 !LIST_EMPTY(&inodedep->id_dirremhd) || 7717 !LIST_EMPTY(&inodedep->id_pendinghd) || 7718 !LIST_EMPTY(&inodedep->id_bufwait) || 7719 !LIST_EMPTY(&inodedep->id_inowait) || 7720 !TAILQ_EMPTY(&inodedep->id_inoreflst) || 7721 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 7722 !TAILQ_EMPTY(&inodedep->id_newinoupdt) || 7723 !TAILQ_EMPTY(&inodedep->id_extupdt) || 7724 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 7725 !TAILQ_EMPTY(&inodedep->id_freeblklst) || 7726 inodedep->id_mkdiradd != NULL || 7727 inodedep->id_nlinkdelta != 0 || 7728 inodedep->id_savedino1 != NULL) 7729 return (0); 7730 return (1); 7731} 7732 7733/* 7734 * Try to free an inodedep structure. Return 1 if it could be freed. 7735 */ 7736static int 7737free_inodedep(inodedep) 7738 struct inodedep *inodedep; 7739{ 7740 7741 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp)); 7742 if ((inodedep->id_state & (ONWORKLIST | UNLINKED)) != 0 || 7743 !check_inodedep_free(inodedep)) 7744 return (0); 7745 if (inodedep->id_state & ONDEPLIST) 7746 LIST_REMOVE(inodedep, id_deps); 7747 LIST_REMOVE(inodedep, id_hash); 7748 WORKITEM_FREE(inodedep, D_INODEDEP); 7749 return (1); 7750} 7751 7752/* 7753 * Free the block referenced by a freework structure. The parent freeblks 7754 * structure is released and completed when the final cg bitmap reaches 7755 * the disk. This routine may be freeing a jnewblk which never made it to 7756 * disk in which case we do not have to wait as the operation is undone 7757 * in memory immediately. 7758 */ 7759static void 7760freework_freeblock(freework, key) 7761 struct freework *freework; 7762 u_long key; 7763{ 7764 struct freeblks *freeblks; 7765 struct jnewblk *jnewblk; 7766 struct ufsmount *ump; 7767 struct workhead wkhd; 7768 struct fs *fs; 7769 int bsize; 7770 int needj; 7771 7772 ump = VFSTOUFS(freework->fw_list.wk_mp); 7773 LOCK_OWNED(ump); 7774 /* 7775 * Handle partial truncate separately. 7776 */ 7777 if (freework->fw_indir) { 7778 complete_trunc_indir(freework); 7779 return; 7780 } 7781 freeblks = freework->fw_freeblks; 7782 fs = ump->um_fs; 7783 needj = MOUNTEDSUJ(freeblks->fb_list.wk_mp) != 0; 7784 bsize = lfragtosize(fs, freework->fw_frags); 7785 LIST_INIT(&wkhd); 7786 /* 7787 * DEPCOMPLETE is cleared in indirblk_insert() if the block lives 7788 * on the indirblk hashtable and prevents premature freeing. 7789 */ 7790 freework->fw_state |= DEPCOMPLETE; 7791 /* 7792 * SUJ needs to wait for the segment referencing freed indirect 7793 * blocks to expire so that we know the checker will not confuse 7794 * a re-allocated indirect block with its old contents. 7795 */ 7796 if (needj && freework->fw_lbn <= -UFS_NDADDR) 7797 indirblk_insert(freework); 7798 /* 7799 * If we are canceling an existing jnewblk pass it to the free 7800 * routine, otherwise pass the freeblk which will ultimately 7801 * release the freeblks. If we're not journaling, we can just 7802 * free the freeblks immediately. 7803 */ 7804 jnewblk = freework->fw_jnewblk; 7805 if (jnewblk != NULL) { 7806 cancel_jnewblk(jnewblk, &wkhd); 7807 needj = 0; 7808 } else if (needj) { 7809 freework->fw_state |= DELAYEDFREE; 7810 freeblks->fb_cgwait++; 7811 WORKLIST_INSERT(&wkhd, &freework->fw_list); 7812 } 7813 FREE_LOCK(ump); 7814 freeblks_free(ump, freeblks, btodb(bsize)); 7815 CTR4(KTR_SUJ, 7816 "freework_freeblock: ino %jd blkno %jd lbn %jd size %d", 7817 freeblks->fb_inum, freework->fw_blkno, freework->fw_lbn, bsize); 7818 ffs_blkfree(ump, fs, freeblks->fb_devvp, freework->fw_blkno, bsize, 7819 freeblks->fb_inum, freeblks->fb_vtype, &wkhd, key); 7820 ACQUIRE_LOCK(ump); 7821 /* 7822 * The jnewblk will be discarded and the bits in the map never 7823 * made it to disk. We can immediately free the freeblk. 7824 */ 7825 if (needj == 0) 7826 handle_written_freework(freework); 7827} 7828 7829/* 7830 * We enqueue freework items that need processing back on the freeblks and 7831 * add the freeblks to the worklist. This makes it easier to find all work 7832 * required to flush a truncation in process_truncates(). 7833 */ 7834static void 7835freework_enqueue(freework) 7836 struct freework *freework; 7837{ 7838 struct freeblks *freeblks; 7839 7840 freeblks = freework->fw_freeblks; 7841 if ((freework->fw_state & INPROGRESS) == 0) 7842 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list); 7843 if ((freeblks->fb_state & 7844 (ONWORKLIST | INPROGRESS | ALLCOMPLETE)) == ALLCOMPLETE && 7845 LIST_EMPTY(&freeblks->fb_jblkdephd)) 7846 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 7847} 7848 7849/* 7850 * Start, continue, or finish the process of freeing an indirect block tree. 7851 * The free operation may be paused at any point with fw_off containing the 7852 * offset to restart from. This enables us to implement some flow control 7853 * for large truncates which may fan out and generate a huge number of 7854 * dependencies. 7855 */ 7856static void 7857handle_workitem_indirblk(freework) 7858 struct freework *freework; 7859{ 7860 struct freeblks *freeblks; 7861 struct ufsmount *ump; 7862 struct fs *fs; 7863 7864 freeblks = freework->fw_freeblks; 7865 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7866 fs = ump->um_fs; 7867 if (freework->fw_state & DEPCOMPLETE) { 7868 handle_written_freework(freework); 7869 return; 7870 } 7871 if (freework->fw_off == NINDIR(fs)) { 7872 freework_freeblock(freework, SINGLETON_KEY); 7873 return; 7874 } 7875 freework->fw_state |= INPROGRESS; 7876 FREE_LOCK(ump); 7877 indir_trunc(freework, fsbtodb(fs, freework->fw_blkno), 7878 freework->fw_lbn); 7879 ACQUIRE_LOCK(ump); 7880} 7881 7882/* 7883 * Called when a freework structure attached to a cg buf is written. The 7884 * ref on either the parent or the freeblks structure is released and 7885 * the freeblks is added back to the worklist if there is more work to do. 7886 */ 7887static void 7888handle_written_freework(freework) 7889 struct freework *freework; 7890{ 7891 struct freeblks *freeblks; 7892 struct freework *parent; 7893 7894 freeblks = freework->fw_freeblks; 7895 parent = freework->fw_parent; 7896 if (freework->fw_state & DELAYEDFREE) 7897 freeblks->fb_cgwait--; 7898 freework->fw_state |= COMPLETE; 7899 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE) 7900 WORKITEM_FREE(freework, D_FREEWORK); 7901 if (parent) { 7902 if (--parent->fw_ref == 0) 7903 freework_enqueue(parent); 7904 return; 7905 } 7906 if (--freeblks->fb_ref != 0) 7907 return; 7908 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST | INPROGRESS)) == 7909 ALLCOMPLETE && LIST_EMPTY(&freeblks->fb_jblkdephd)) 7910 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 7911} 7912 7913/* 7914 * This workitem routine performs the block de-allocation. 7915 * The workitem is added to the pending list after the updated 7916 * inode block has been written to disk. As mentioned above, 7917 * checks regarding the number of blocks de-allocated (compared 7918 * to the number of blocks allocated for the file) are also 7919 * performed in this function. 7920 */ 7921static int 7922handle_workitem_freeblocks(freeblks, flags) 7923 struct freeblks *freeblks; 7924 int flags; 7925{ 7926 struct freework *freework; 7927 struct newblk *newblk; 7928 struct allocindir *aip; 7929 struct ufsmount *ump; 7930 struct worklist *wk; 7931 u_long key; 7932 7933 KASSERT(LIST_EMPTY(&freeblks->fb_jblkdephd), 7934 ("handle_workitem_freeblocks: Journal entries not written.")); 7935 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7936 key = ffs_blkrelease_start(ump, freeblks->fb_devvp, freeblks->fb_inum); 7937 ACQUIRE_LOCK(ump); 7938 while ((wk = LIST_FIRST(&freeblks->fb_freeworkhd)) != NULL) { 7939 WORKLIST_REMOVE(wk); 7940 switch (wk->wk_type) { 7941 case D_DIRREM: 7942 wk->wk_state |= COMPLETE; 7943 add_to_worklist(wk, 0); 7944 continue; 7945 7946 case D_ALLOCDIRECT: 7947 free_newblk(WK_NEWBLK(wk)); 7948 continue; 7949 7950 case D_ALLOCINDIR: 7951 aip = WK_ALLOCINDIR(wk); 7952 freework = NULL; 7953 if (aip->ai_state & DELAYEDFREE) { 7954 FREE_LOCK(ump); 7955 freework = newfreework(ump, freeblks, NULL, 7956 aip->ai_lbn, aip->ai_newblkno, 7957 ump->um_fs->fs_frag, 0, 0); 7958 ACQUIRE_LOCK(ump); 7959 } 7960 newblk = WK_NEWBLK(wk); 7961 if (newblk->nb_jnewblk) { 7962 freework->fw_jnewblk = newblk->nb_jnewblk; 7963 newblk->nb_jnewblk->jn_dep = &freework->fw_list; 7964 newblk->nb_jnewblk = NULL; 7965 } 7966 free_newblk(newblk); 7967 continue; 7968 7969 case D_FREEWORK: 7970 freework = WK_FREEWORK(wk); 7971 if (freework->fw_lbn <= -UFS_NDADDR) 7972 handle_workitem_indirblk(freework); 7973 else 7974 freework_freeblock(freework, key); 7975 continue; 7976 default: 7977 panic("handle_workitem_freeblocks: Unknown type %s", 7978 TYPENAME(wk->wk_type)); 7979 } 7980 } 7981 if (freeblks->fb_ref != 0) { 7982 freeblks->fb_state &= ~INPROGRESS; 7983 wake_worklist(&freeblks->fb_list); 7984 freeblks = NULL; 7985 } 7986 FREE_LOCK(ump); 7987 ffs_blkrelease_finish(ump, key); 7988 if (freeblks) 7989 return handle_complete_freeblocks(freeblks, flags); 7990 return (0); 7991} 7992 7993/* 7994 * Handle completion of block free via truncate. This allows fs_pending 7995 * to track the actual free block count more closely than if we only updated 7996 * it at the end. We must be careful to handle cases where the block count 7997 * on free was incorrect. 7998 */ 7999static void 8000freeblks_free(ump, freeblks, blocks) 8001 struct ufsmount *ump; 8002 struct freeblks *freeblks; 8003 int blocks; 8004{ 8005 struct fs *fs; 8006 ufs2_daddr_t remain; 8007 8008 UFS_LOCK(ump); 8009 remain = -freeblks->fb_chkcnt; 8010 freeblks->fb_chkcnt += blocks; 8011 if (remain > 0) { 8012 if (remain < blocks) 8013 blocks = remain; 8014 fs = ump->um_fs; 8015 fs->fs_pendingblocks -= blocks; 8016 } 8017 UFS_UNLOCK(ump); 8018} 8019 8020/* 8021 * Once all of the freework workitems are complete we can retire the 8022 * freeblocks dependency and any journal work awaiting completion. This 8023 * can not be called until all other dependencies are stable on disk. 8024 */ 8025static int 8026handle_complete_freeblocks(freeblks, flags) 8027 struct freeblks *freeblks; 8028 int flags; 8029{ 8030 struct inodedep *inodedep; 8031 struct inode *ip; 8032 struct vnode *vp; 8033 struct fs *fs; 8034 struct ufsmount *ump; 8035 ufs2_daddr_t spare; 8036 8037 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 8038 fs = ump->um_fs; 8039 flags = LK_EXCLUSIVE | flags; 8040 spare = freeblks->fb_chkcnt; 8041 8042 /* 8043 * If we did not release the expected number of blocks we may have 8044 * to adjust the inode block count here. Only do so if it wasn't 8045 * a truncation to zero and the modrev still matches. 8046 */ 8047 if (spare && freeblks->fb_len != 0) { 8048 if (ffs_vgetf(freeblks->fb_list.wk_mp, freeblks->fb_inum, 8049 flags, &vp, FFSV_FORCEINSMQ) != 0) 8050 return (EBUSY); 8051 ip = VTOI(vp); 8052 if (ip->i_mode == 0) { 8053 vgone(vp); 8054 } else if (DIP(ip, i_modrev) == freeblks->fb_modrev) { 8055 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - spare); 8056 ip->i_flag |= IN_CHANGE; 8057 /* 8058 * We must wait so this happens before the 8059 * journal is reclaimed. 8060 */ 8061 ffs_update(vp, 1); 8062 } 8063 vput(vp); 8064 } 8065 if (spare < 0) { 8066 UFS_LOCK(ump); 8067 fs->fs_pendingblocks += spare; 8068 UFS_UNLOCK(ump); 8069 } 8070#ifdef QUOTA 8071 /* Handle spare. */ 8072 if (spare) 8073 quotaadj(freeblks->fb_quota, ump, -spare); 8074 quotarele(freeblks->fb_quota); 8075#endif 8076 ACQUIRE_LOCK(ump); 8077 if (freeblks->fb_state & ONDEPLIST) { 8078 inodedep_lookup(freeblks->fb_list.wk_mp, freeblks->fb_inum, 8079 0, &inodedep); 8080 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, fb_next); 8081 freeblks->fb_state &= ~ONDEPLIST; 8082 if (TAILQ_EMPTY(&inodedep->id_freeblklst)) 8083 free_inodedep(inodedep); 8084 } 8085 /* 8086 * All of the freeblock deps must be complete prior to this call 8087 * so it's now safe to complete earlier outstanding journal entries. 8088 */ 8089 handle_jwork(&freeblks->fb_jwork); 8090 WORKITEM_FREE(freeblks, D_FREEBLKS); 8091 FREE_LOCK(ump); 8092 return (0); 8093} 8094 8095/* 8096 * Release blocks associated with the freeblks and stored in the indirect 8097 * block dbn. If level is greater than SINGLE, the block is an indirect block 8098 * and recursive calls to indirtrunc must be used to cleanse other indirect 8099 * blocks. 8100 * 8101 * This handles partial and complete truncation of blocks. Partial is noted 8102 * with goingaway == 0. In this case the freework is completed after the 8103 * zero'd indirects are written to disk. For full truncation the freework 8104 * is completed after the block is freed. 8105 */ 8106static void 8107indir_trunc(freework, dbn, lbn) 8108 struct freework *freework; 8109 ufs2_daddr_t dbn; 8110 ufs_lbn_t lbn; 8111{ 8112 struct freework *nfreework; 8113 struct workhead wkhd; 8114 struct freeblks *freeblks; 8115 struct buf *bp; 8116 struct fs *fs; 8117 struct indirdep *indirdep; 8118 struct mount *mp; 8119 struct ufsmount *ump; 8120 ufs1_daddr_t *bap1; 8121 ufs2_daddr_t nb, nnb, *bap2; 8122 ufs_lbn_t lbnadd, nlbn; 8123 u_long key; 8124 int nblocks, ufs1fmt, freedblocks; 8125 int goingaway, freedeps, needj, level, cnt, i; 8126 8127 freeblks = freework->fw_freeblks; 8128 mp = freeblks->fb_list.wk_mp; 8129 ump = VFSTOUFS(mp); 8130 fs = ump->um_fs; 8131 /* 8132 * Get buffer of block pointers to be freed. There are three cases: 8133 * 8134 * 1) Partial truncate caches the indirdep pointer in the freework 8135 * which provides us a back copy to the save bp which holds the 8136 * pointers we want to clear. When this completes the zero 8137 * pointers are written to the real copy. 8138 * 2) The indirect is being completely truncated, cancel_indirdep() 8139 * eliminated the real copy and placed the indirdep on the saved 8140 * copy. The indirdep and buf are discarded when this completes. 8141 * 3) The indirect was not in memory, we read a copy off of the disk 8142 * using the devvp and drop and invalidate the buffer when we're 8143 * done. 8144 */ 8145 goingaway = 1; 8146 indirdep = NULL; 8147 if (freework->fw_indir != NULL) { 8148 goingaway = 0; 8149 indirdep = freework->fw_indir; 8150 bp = indirdep->ir_savebp; 8151 if (bp == NULL || bp->b_blkno != dbn) 8152 panic("indir_trunc: Bad saved buf %p blkno %jd", 8153 bp, (intmax_t)dbn); 8154 } else if ((bp = incore(&freeblks->fb_devvp->v_bufobj, dbn)) != NULL) { 8155 /* 8156 * The lock prevents the buf dep list from changing and 8157 * indirects on devvp should only ever have one dependency. 8158 */ 8159 indirdep = WK_INDIRDEP(LIST_FIRST(&bp->b_dep)); 8160 if (indirdep == NULL || (indirdep->ir_state & GOINGAWAY) == 0) 8161 panic("indir_trunc: Bad indirdep %p from buf %p", 8162 indirdep, bp); 8163 } else if (bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 8164 NOCRED, &bp) != 0) { 8165 brelse(bp); 8166 return; 8167 } 8168 ACQUIRE_LOCK(ump); 8169 /* Protects against a race with complete_trunc_indir(). */ 8170 freework->fw_state &= ~INPROGRESS; 8171 /* 8172 * If we have an indirdep we need to enforce the truncation order 8173 * and discard it when it is complete. 8174 */ 8175 if (indirdep) { 8176 if (freework != TAILQ_FIRST(&indirdep->ir_trunc) && 8177 !TAILQ_EMPTY(&indirdep->ir_trunc)) { 8178 /* 8179 * Add the complete truncate to the list on the 8180 * indirdep to enforce in-order processing. 8181 */ 8182 if (freework->fw_indir == NULL) 8183 TAILQ_INSERT_TAIL(&indirdep->ir_trunc, 8184 freework, fw_next); 8185 FREE_LOCK(ump); 8186 return; 8187 } 8188 /* 8189 * If we're goingaway, free the indirdep. Otherwise it will 8190 * linger until the write completes. 8191 */ 8192 if (goingaway) 8193 free_indirdep(indirdep); 8194 } 8195 FREE_LOCK(ump); 8196 /* Initialize pointers depending on block size. */ 8197 if (ump->um_fstype == UFS1) { 8198 bap1 = (ufs1_daddr_t *)bp->b_data; 8199 nb = bap1[freework->fw_off]; 8200 ufs1fmt = 1; 8201 bap2 = NULL; 8202 } else { 8203 bap2 = (ufs2_daddr_t *)bp->b_data; 8204 nb = bap2[freework->fw_off]; 8205 ufs1fmt = 0; 8206 bap1 = NULL; 8207 } 8208 level = lbn_level(lbn); 8209 needj = MOUNTEDSUJ(UFSTOVFS(ump)) != 0; 8210 lbnadd = lbn_offset(fs, level); 8211 nblocks = btodb(fs->fs_bsize); 8212 nfreework = freework; 8213 freedeps = 0; 8214 cnt = 0; 8215 /* 8216 * Reclaim blocks. Traverses into nested indirect levels and 8217 * arranges for the current level to be freed when subordinates 8218 * are free when journaling. 8219 */ 8220 key = ffs_blkrelease_start(ump, freeblks->fb_devvp, freeblks->fb_inum); 8221 for (i = freework->fw_off; i < NINDIR(fs); i++, nb = nnb) { 8222 if (UFS_CHECK_BLKNO(mp, freeblks->fb_inum, nb, 8223 fs->fs_bsize) != 0) 8224 nb = 0; 8225 if (i != NINDIR(fs) - 1) { 8226 if (ufs1fmt) 8227 nnb = bap1[i+1]; 8228 else 8229 nnb = bap2[i+1]; 8230 } else 8231 nnb = 0; 8232 if (nb == 0) 8233 continue; 8234 cnt++; 8235 if (level != 0) { 8236 nlbn = (lbn + 1) - (i * lbnadd); 8237 if (needj != 0) { 8238 nfreework = newfreework(ump, freeblks, freework, 8239 nlbn, nb, fs->fs_frag, 0, 0); 8240 freedeps++; 8241 } 8242 indir_trunc(nfreework, fsbtodb(fs, nb), nlbn); 8243 } else { 8244 struct freedep *freedep; 8245 8246 /* 8247 * Attempt to aggregate freedep dependencies for 8248 * all blocks being released to the same CG. 8249 */ 8250 LIST_INIT(&wkhd); 8251 if (needj != 0 && 8252 (nnb == 0 || (dtog(fs, nb) != dtog(fs, nnb)))) { 8253 freedep = newfreedep(freework); 8254 WORKLIST_INSERT_UNLOCKED(&wkhd, 8255 &freedep->fd_list); 8256 freedeps++; 8257 } 8258 CTR3(KTR_SUJ, 8259 "indir_trunc: ino %jd blkno %jd size %d", 8260 freeblks->fb_inum, nb, fs->fs_bsize); 8261 ffs_blkfree(ump, fs, freeblks->fb_devvp, nb, 8262 fs->fs_bsize, freeblks->fb_inum, 8263 freeblks->fb_vtype, &wkhd, key); 8264 } 8265 } 8266 ffs_blkrelease_finish(ump, key); 8267 if (goingaway) { 8268 bp->b_flags |= B_INVAL | B_NOCACHE; 8269 brelse(bp); 8270 } 8271 freedblocks = 0; 8272 if (level == 0) 8273 freedblocks = (nblocks * cnt); 8274 if (needj == 0) 8275 freedblocks += nblocks; 8276 freeblks_free(ump, freeblks, freedblocks); 8277 /* 8278 * If we are journaling set up the ref counts and offset so this 8279 * indirect can be completed when its children are free. 8280 */ 8281 if (needj) { 8282 ACQUIRE_LOCK(ump); 8283 freework->fw_off = i; 8284 freework->fw_ref += freedeps; 8285 freework->fw_ref -= NINDIR(fs) + 1; 8286 if (level == 0) 8287 freeblks->fb_cgwait += freedeps; 8288 if (freework->fw_ref == 0) 8289 freework_freeblock(freework, SINGLETON_KEY); 8290 FREE_LOCK(ump); 8291 return; 8292 } 8293 /* 8294 * If we're not journaling we can free the indirect now. 8295 */ 8296 dbn = dbtofsb(fs, dbn); 8297 CTR3(KTR_SUJ, 8298 "indir_trunc 2: ino %jd blkno %jd size %d", 8299 freeblks->fb_inum, dbn, fs->fs_bsize); 8300 ffs_blkfree(ump, fs, freeblks->fb_devvp, dbn, fs->fs_bsize, 8301 freeblks->fb_inum, freeblks->fb_vtype, NULL, SINGLETON_KEY); 8302 /* Non SUJ softdep does single-threaded truncations. */ 8303 if (freework->fw_blkno == dbn) { 8304 freework->fw_state |= ALLCOMPLETE; 8305 ACQUIRE_LOCK(ump); 8306 handle_written_freework(freework); 8307 FREE_LOCK(ump); 8308 } 8309 return; 8310} 8311 8312/* 8313 * Cancel an allocindir when it is removed via truncation. When bp is not 8314 * NULL the indirect never appeared on disk and is scheduled to be freed 8315 * independently of the indir so we can more easily track journal work. 8316 */ 8317static void 8318cancel_allocindir(aip, bp, freeblks, trunc) 8319 struct allocindir *aip; 8320 struct buf *bp; 8321 struct freeblks *freeblks; 8322 int trunc; 8323{ 8324 struct indirdep *indirdep; 8325 struct freefrag *freefrag; 8326 struct newblk *newblk; 8327 8328 newblk = (struct newblk *)aip; 8329 LIST_REMOVE(aip, ai_next); 8330 /* 8331 * We must eliminate the pointer in bp if it must be freed on its 8332 * own due to partial truncate or pending journal work. 8333 */ 8334 if (bp && (trunc || newblk->nb_jnewblk)) { 8335 /* 8336 * Clear the pointer and mark the aip to be freed 8337 * directly if it never existed on disk. 8338 */ 8339 aip->ai_state |= DELAYEDFREE; 8340 indirdep = aip->ai_indirdep; 8341 if (indirdep->ir_state & UFS1FMT) 8342 ((ufs1_daddr_t *)bp->b_data)[aip->ai_offset] = 0; 8343 else 8344 ((ufs2_daddr_t *)bp->b_data)[aip->ai_offset] = 0; 8345 } 8346 /* 8347 * When truncating the previous pointer will be freed via 8348 * savedbp. Eliminate the freefrag which would dup free. 8349 */ 8350 if (trunc && (freefrag = newblk->nb_freefrag) != NULL) { 8351 newblk->nb_freefrag = NULL; 8352 if (freefrag->ff_jdep) 8353 cancel_jfreefrag( 8354 WK_JFREEFRAG(freefrag->ff_jdep)); 8355 jwork_move(&freeblks->fb_jwork, &freefrag->ff_jwork); 8356 WORKITEM_FREE(freefrag, D_FREEFRAG); 8357 } 8358 /* 8359 * If the journal hasn't been written the jnewblk must be passed 8360 * to the call to ffs_blkfree that reclaims the space. We accomplish 8361 * this by leaving the journal dependency on the newblk to be freed 8362 * when a freework is created in handle_workitem_freeblocks(). 8363 */ 8364 cancel_newblk(newblk, NULL, &freeblks->fb_jwork); 8365 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list); 8366} 8367 8368/* 8369 * Create the mkdir dependencies for . and .. in a new directory. Link them 8370 * in to a newdirblk so any subsequent additions are tracked properly. The 8371 * caller is responsible for adding the mkdir1 dependency to the journal 8372 * and updating id_mkdiradd. This function returns with the per-filesystem 8373 * lock held. 8374 */ 8375static struct mkdir * 8376setup_newdir(dap, newinum, dinum, newdirbp, mkdirp) 8377 struct diradd *dap; 8378 ino_t newinum; 8379 ino_t dinum; 8380 struct buf *newdirbp; 8381 struct mkdir **mkdirp; 8382{ 8383 struct newblk *newblk; 8384 struct pagedep *pagedep; 8385 struct inodedep *inodedep; 8386 struct newdirblk *newdirblk; 8387 struct mkdir *mkdir1, *mkdir2; 8388 struct worklist *wk; 8389 struct jaddref *jaddref; 8390 struct ufsmount *ump; 8391 struct mount *mp; 8392 8393 mp = dap->da_list.wk_mp; 8394 ump = VFSTOUFS(mp); 8395 newdirblk = malloc(sizeof(struct newdirblk), M_NEWDIRBLK, 8396 M_SOFTDEP_FLAGS); 8397 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 8398 LIST_INIT(&newdirblk->db_mkdir); 8399 mkdir1 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS); 8400 workitem_alloc(&mkdir1->md_list, D_MKDIR, mp); 8401 mkdir1->md_state = ATTACHED | MKDIR_BODY; 8402 mkdir1->md_diradd = dap; 8403 mkdir1->md_jaddref = NULL; 8404 mkdir2 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS); 8405 workitem_alloc(&mkdir2->md_list, D_MKDIR, mp); 8406 mkdir2->md_state = ATTACHED | MKDIR_PARENT; 8407 mkdir2->md_diradd = dap; 8408 mkdir2->md_jaddref = NULL; 8409 if (MOUNTEDSUJ(mp) == 0) { 8410 mkdir1->md_state |= DEPCOMPLETE; 8411 mkdir2->md_state |= DEPCOMPLETE; 8412 } 8413 /* 8414 * Dependency on "." and ".." being written to disk. 8415 */ 8416 mkdir1->md_buf = newdirbp; 8417 ACQUIRE_LOCK(VFSTOUFS(mp)); 8418 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir1, md_mkdirs); 8419 /* 8420 * We must link the pagedep, allocdirect, and newdirblk for 8421 * the initial file page so the pointer to the new directory 8422 * is not written until the directory contents are live and 8423 * any subsequent additions are not marked live until the 8424 * block is reachable via the inode. 8425 */ 8426 if (pagedep_lookup(mp, newdirbp, newinum, 0, 0, &pagedep) == 0) 8427 panic("setup_newdir: lost pagedep"); 8428 LIST_FOREACH(wk, &newdirbp->b_dep, wk_list) 8429 if (wk->wk_type == D_ALLOCDIRECT) 8430 break; 8431 if (wk == NULL) 8432 panic("setup_newdir: lost allocdirect"); 8433 if (pagedep->pd_state & NEWBLOCK) 8434 panic("setup_newdir: NEWBLOCK already set"); 8435 newblk = WK_NEWBLK(wk); 8436 pagedep->pd_state |= NEWBLOCK; 8437 pagedep->pd_newdirblk = newdirblk; 8438 newdirblk->db_pagedep = pagedep; 8439 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list); 8440 WORKLIST_INSERT(&newdirblk->db_mkdir, &mkdir1->md_list); 8441 /* 8442 * Look up the inodedep for the parent directory so that we 8443 * can link mkdir2 into the pending dotdot jaddref or 8444 * the inode write if there is none. If the inode is 8445 * ALLCOMPLETE and no jaddref is present all dependencies have 8446 * been satisfied and mkdir2 can be freed. 8447 */ 8448 inodedep_lookup(mp, dinum, 0, &inodedep); 8449 if (MOUNTEDSUJ(mp)) { 8450 if (inodedep == NULL) 8451 panic("setup_newdir: Lost parent."); 8452 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 8453 inoreflst); 8454 KASSERT(jaddref != NULL && jaddref->ja_parent == newinum && 8455 (jaddref->ja_state & MKDIR_PARENT), 8456 ("setup_newdir: bad dotdot jaddref %p", jaddref)); 8457 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir2, md_mkdirs); 8458 mkdir2->md_jaddref = jaddref; 8459 jaddref->ja_mkdir = mkdir2; 8460 } else if (inodedep == NULL || 8461 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 8462 dap->da_state &= ~MKDIR_PARENT; 8463 WORKITEM_FREE(mkdir2, D_MKDIR); 8464 mkdir2 = NULL; 8465 } else { 8466 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir2, md_mkdirs); 8467 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir2->md_list); 8468 } 8469 *mkdirp = mkdir2; 8470 8471 return (mkdir1); 8472} 8473 8474/* 8475 * Directory entry addition dependencies. 8476 * 8477 * When adding a new directory entry, the inode (with its incremented link 8478 * count) must be written to disk before the directory entry's pointer to it. 8479 * Also, if the inode is newly allocated, the corresponding freemap must be 8480 * updated (on disk) before the directory entry's pointer. These requirements 8481 * are met via undo/redo on the directory entry's pointer, which consists 8482 * simply of the inode number. 8483 * 8484 * As directory entries are added and deleted, the free space within a 8485 * directory block can become fragmented. The ufs filesystem will compact 8486 * a fragmented directory block to make space for a new entry. When this 8487 * occurs, the offsets of previously added entries change. Any "diradd" 8488 * dependency structures corresponding to these entries must be updated with 8489 * the new offsets. 8490 */ 8491 8492/* 8493 * This routine is called after the in-memory inode's link 8494 * count has been incremented, but before the directory entry's 8495 * pointer to the inode has been set. 8496 */ 8497int 8498softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 8499 struct buf *bp; /* buffer containing directory block */ 8500 struct inode *dp; /* inode for directory */ 8501 off_t diroffset; /* offset of new entry in directory */ 8502 ino_t newinum; /* inode referenced by new directory entry */ 8503 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 8504 int isnewblk; /* entry is in a newly allocated block */ 8505{ 8506 int offset; /* offset of new entry within directory block */ 8507 ufs_lbn_t lbn; /* block in directory containing new entry */ 8508 struct fs *fs; 8509 struct diradd *dap; 8510 struct newblk *newblk; 8511 struct pagedep *pagedep; 8512 struct inodedep *inodedep; 8513 struct newdirblk *newdirblk; 8514 struct mkdir *mkdir1, *mkdir2; 8515 struct jaddref *jaddref; 8516 struct ufsmount *ump; 8517 struct mount *mp; 8518 int isindir; 8519 8520 mp = ITOVFS(dp); 8521 ump = VFSTOUFS(mp); 8522 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 8523 ("softdep_setup_directory_add called on non-softdep filesystem")); 8524 /* 8525 * Whiteouts have no dependencies. 8526 */ 8527 if (newinum == UFS_WINO) { 8528 if (newdirbp != NULL) 8529 bdwrite(newdirbp); 8530 return (0); 8531 } 8532 jaddref = NULL; 8533 mkdir1 = mkdir2 = NULL; 8534 fs = ump->um_fs; 8535 lbn = lblkno(fs, diroffset); 8536 offset = blkoff(fs, diroffset); 8537 dap = malloc(sizeof(struct diradd), M_DIRADD, 8538 M_SOFTDEP_FLAGS|M_ZERO); 8539 workitem_alloc(&dap->da_list, D_DIRADD, mp); 8540 dap->da_offset = offset; 8541 dap->da_newinum = newinum; 8542 dap->da_state = ATTACHED; 8543 LIST_INIT(&dap->da_jwork); 8544 isindir = bp->b_lblkno >= UFS_NDADDR; 8545 newdirblk = NULL; 8546 if (isnewblk && 8547 (isindir ? blkoff(fs, diroffset) : fragoff(fs, diroffset)) == 0) { 8548 newdirblk = malloc(sizeof(struct newdirblk), 8549 M_NEWDIRBLK, M_SOFTDEP_FLAGS); 8550 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 8551 LIST_INIT(&newdirblk->db_mkdir); 8552 } 8553 /* 8554 * If we're creating a new directory setup the dependencies and set 8555 * the dap state to wait for them. Otherwise it's COMPLETE and 8556 * we can move on. 8557 */ 8558 if (newdirbp == NULL) { 8559 dap->da_state |= DEPCOMPLETE; 8560 ACQUIRE_LOCK(ump); 8561 } else { 8562 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 8563 mkdir1 = setup_newdir(dap, newinum, dp->i_number, newdirbp, 8564 &mkdir2); 8565 } 8566 /* 8567 * Link into parent directory pagedep to await its being written. 8568 */ 8569 pagedep_lookup(mp, bp, dp->i_number, lbn, DEPALLOC, &pagedep); 8570#ifdef DEBUG 8571 if (diradd_lookup(pagedep, offset) != NULL) 8572 panic("softdep_setup_directory_add: %p already at off %d\n", 8573 diradd_lookup(pagedep, offset), offset); 8574#endif 8575 dap->da_pagedep = pagedep; 8576 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 8577 da_pdlist); 8578 inodedep_lookup(mp, newinum, DEPALLOC, &inodedep); 8579 /* 8580 * If we're journaling, link the diradd into the jaddref so it 8581 * may be completed after the journal entry is written. Otherwise, 8582 * link the diradd into its inodedep. If the inode is not yet 8583 * written place it on the bufwait list, otherwise do the post-inode 8584 * write processing to put it on the id_pendinghd list. 8585 */ 8586 if (MOUNTEDSUJ(mp)) { 8587 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 8588 inoreflst); 8589 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 8590 ("softdep_setup_directory_add: bad jaddref %p", jaddref)); 8591 jaddref->ja_diroff = diroffset; 8592 jaddref->ja_diradd = dap; 8593 add_to_journal(&jaddref->ja_list); 8594 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 8595 diradd_inode_written(dap, inodedep); 8596 else 8597 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 8598 /* 8599 * Add the journal entries for . and .. links now that the primary 8600 * link is written. 8601 */ 8602 if (mkdir1 != NULL && MOUNTEDSUJ(mp)) { 8603 jaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref, 8604 inoreflst, if_deps); 8605 KASSERT(jaddref != NULL && 8606 jaddref->ja_ino == jaddref->ja_parent && 8607 (jaddref->ja_state & MKDIR_BODY), 8608 ("softdep_setup_directory_add: bad dot jaddref %p", 8609 jaddref)); 8610 mkdir1->md_jaddref = jaddref; 8611 jaddref->ja_mkdir = mkdir1; 8612 /* 8613 * It is important that the dotdot journal entry 8614 * is added prior to the dot entry since dot writes 8615 * both the dot and dotdot links. These both must 8616 * be added after the primary link for the journal 8617 * to remain consistent. 8618 */ 8619 add_to_journal(&mkdir2->md_jaddref->ja_list); 8620 add_to_journal(&jaddref->ja_list); 8621 } 8622 /* 8623 * If we are adding a new directory remember this diradd so that if 8624 * we rename it we can keep the dot and dotdot dependencies. If 8625 * we are adding a new name for an inode that has a mkdiradd we 8626 * must be in rename and we have to move the dot and dotdot 8627 * dependencies to this new name. The old name is being orphaned 8628 * soon. 8629 */ 8630 if (mkdir1 != NULL) { 8631 if (inodedep->id_mkdiradd != NULL) 8632 panic("softdep_setup_directory_add: Existing mkdir"); 8633 inodedep->id_mkdiradd = dap; 8634 } else if (inodedep->id_mkdiradd) 8635 merge_diradd(inodedep, dap); 8636 if (newdirblk != NULL) { 8637 /* 8638 * There is nothing to do if we are already tracking 8639 * this block. 8640 */ 8641 if ((pagedep->pd_state & NEWBLOCK) != 0) { 8642 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 8643 FREE_LOCK(ump); 8644 return (0); 8645 } 8646 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk) 8647 == 0) 8648 panic("softdep_setup_directory_add: lost entry"); 8649 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list); 8650 pagedep->pd_state |= NEWBLOCK; 8651 pagedep->pd_newdirblk = newdirblk; 8652 newdirblk->db_pagedep = pagedep; 8653 FREE_LOCK(ump); 8654 /* 8655 * If we extended into an indirect signal direnter to sync. 8656 */ 8657 if (isindir) 8658 return (1); 8659 return (0); 8660 } 8661 FREE_LOCK(ump); 8662 return (0); 8663} 8664 8665/* 8666 * This procedure is called to change the offset of a directory 8667 * entry when compacting a directory block which must be owned 8668 * exclusively by the caller. Note that the actual entry movement 8669 * must be done in this procedure to ensure that no I/O completions 8670 * occur while the move is in progress. 8671 */ 8672void 8673softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize) 8674 struct buf *bp; /* Buffer holding directory block. */ 8675 struct inode *dp; /* inode for directory */ 8676 caddr_t base; /* address of dp->i_offset */ 8677 caddr_t oldloc; /* address of old directory location */ 8678 caddr_t newloc; /* address of new directory location */ 8679 int entrysize; /* size of directory entry */ 8680{ 8681 int offset, oldoffset, newoffset; 8682 struct pagedep *pagedep; 8683 struct jmvref *jmvref; 8684 struct diradd *dap; 8685 struct direct *de; 8686 struct mount *mp; 8687 struct ufsmount *ump; 8688 ufs_lbn_t lbn; 8689 int flags; 8690 8691 mp = ITOVFS(dp); 8692 ump = VFSTOUFS(mp); 8693 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 8694 ("softdep_change_directoryentry_offset called on " 8695 "non-softdep filesystem")); 8696 de = (struct direct *)oldloc; 8697 jmvref = NULL; 8698 flags = 0; 8699 /* 8700 * Moves are always journaled as it would be too complex to 8701 * determine if any affected adds or removes are present in the 8702 * journal. 8703 */ 8704 if (MOUNTEDSUJ(mp)) { 8705 flags = DEPALLOC; 8706 jmvref = newjmvref(dp, de->d_ino, 8707 dp->i_offset + (oldloc - base), 8708 dp->i_offset + (newloc - base)); 8709 } 8710 lbn = lblkno(ump->um_fs, dp->i_offset); 8711 offset = blkoff(ump->um_fs, dp->i_offset); 8712 oldoffset = offset + (oldloc - base); 8713 newoffset = offset + (newloc - base); 8714 ACQUIRE_LOCK(ump); 8715 if (pagedep_lookup(mp, bp, dp->i_number, lbn, flags, &pagedep) == 0) 8716 goto done; 8717 dap = diradd_lookup(pagedep, oldoffset); 8718 if (dap) { 8719 dap->da_offset = newoffset; 8720 newoffset = DIRADDHASH(newoffset); 8721 oldoffset = DIRADDHASH(oldoffset); 8722 if ((dap->da_state & ALLCOMPLETE) != ALLCOMPLETE && 8723 newoffset != oldoffset) { 8724 LIST_REMOVE(dap, da_pdlist); 8725 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[newoffset], 8726 dap, da_pdlist); 8727 } 8728 } 8729done: 8730 if (jmvref) { 8731 jmvref->jm_pagedep = pagedep; 8732 LIST_INSERT_HEAD(&pagedep->pd_jmvrefhd, jmvref, jm_deps); 8733 add_to_journal(&jmvref->jm_list); 8734 } 8735 bcopy(oldloc, newloc, entrysize); 8736 FREE_LOCK(ump); 8737} 8738 8739/* 8740 * Move the mkdir dependencies and journal work from one diradd to another 8741 * when renaming a directory. The new name must depend on the mkdir deps 8742 * completing as the old name did. Directories can only have one valid link 8743 * at a time so one must be canonical. 8744 */ 8745static void 8746merge_diradd(inodedep, newdap) 8747 struct inodedep *inodedep; 8748 struct diradd *newdap; 8749{ 8750 struct diradd *olddap; 8751 struct mkdir *mkdir, *nextmd; 8752 struct ufsmount *ump; 8753 short state; 8754 8755 olddap = inodedep->id_mkdiradd; 8756 inodedep->id_mkdiradd = newdap; 8757 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8758 newdap->da_state &= ~DEPCOMPLETE; 8759 ump = VFSTOUFS(inodedep->id_list.wk_mp); 8760 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir; 8761 mkdir = nextmd) { 8762 nextmd = LIST_NEXT(mkdir, md_mkdirs); 8763 if (mkdir->md_diradd != olddap) 8764 continue; 8765 mkdir->md_diradd = newdap; 8766 state = mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY); 8767 newdap->da_state |= state; 8768 olddap->da_state &= ~state; 8769 if ((olddap->da_state & 8770 (MKDIR_PARENT | MKDIR_BODY)) == 0) 8771 break; 8772 } 8773 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 8774 panic("merge_diradd: unfound ref"); 8775 } 8776 /* 8777 * Any mkdir related journal items are not safe to be freed until 8778 * the new name is stable. 8779 */ 8780 jwork_move(&newdap->da_jwork, &olddap->da_jwork); 8781 olddap->da_state |= DEPCOMPLETE; 8782 complete_diradd(olddap); 8783} 8784 8785/* 8786 * Move the diradd to the pending list when all diradd dependencies are 8787 * complete. 8788 */ 8789static void 8790complete_diradd(dap) 8791 struct diradd *dap; 8792{ 8793 struct pagedep *pagedep; 8794 8795 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 8796 if (dap->da_state & DIRCHG) 8797 pagedep = dap->da_previous->dm_pagedep; 8798 else 8799 pagedep = dap->da_pagedep; 8800 LIST_REMOVE(dap, da_pdlist); 8801 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 8802 } 8803} 8804 8805/* 8806 * Cancel a diradd when a dirrem overlaps with it. We must cancel the journal 8807 * add entries and conditonally journal the remove. 8808 */ 8809static void 8810cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref) 8811 struct diradd *dap; 8812 struct dirrem *dirrem; 8813 struct jremref *jremref; 8814 struct jremref *dotremref; 8815 struct jremref *dotdotremref; 8816{ 8817 struct inodedep *inodedep; 8818 struct jaddref *jaddref; 8819 struct inoref *inoref; 8820 struct ufsmount *ump; 8821 struct mkdir *mkdir; 8822 8823 /* 8824 * If no remove references were allocated we're on a non-journaled 8825 * filesystem and can skip the cancel step. 8826 */ 8827 if (jremref == NULL) { 8828 free_diradd(dap, NULL); 8829 return; 8830 } 8831 /* 8832 * Cancel the primary name an free it if it does not require 8833 * journaling. 8834 */ 8835 if (inodedep_lookup(dap->da_list.wk_mp, dap->da_newinum, 8836 0, &inodedep) != 0) { 8837 /* Abort the addref that reference this diradd. */ 8838 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 8839 if (inoref->if_list.wk_type != D_JADDREF) 8840 continue; 8841 jaddref = (struct jaddref *)inoref; 8842 if (jaddref->ja_diradd != dap) 8843 continue; 8844 if (cancel_jaddref(jaddref, inodedep, 8845 &dirrem->dm_jwork) == 0) { 8846 free_jremref(jremref); 8847 jremref = NULL; 8848 } 8849 break; 8850 } 8851 } 8852 /* 8853 * Cancel subordinate names and free them if they do not require 8854 * journaling. 8855 */ 8856 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8857 ump = VFSTOUFS(dap->da_list.wk_mp); 8858 LIST_FOREACH(mkdir, &ump->softdep_mkdirlisthd, md_mkdirs) { 8859 if (mkdir->md_diradd != dap) 8860 continue; 8861 if ((jaddref = mkdir->md_jaddref) == NULL) 8862 continue; 8863 mkdir->md_jaddref = NULL; 8864 if (mkdir->md_state & MKDIR_PARENT) { 8865 if (cancel_jaddref(jaddref, NULL, 8866 &dirrem->dm_jwork) == 0) { 8867 free_jremref(dotdotremref); 8868 dotdotremref = NULL; 8869 } 8870 } else { 8871 if (cancel_jaddref(jaddref, inodedep, 8872 &dirrem->dm_jwork) == 0) { 8873 free_jremref(dotremref); 8874 dotremref = NULL; 8875 } 8876 } 8877 } 8878 } 8879 8880 if (jremref) 8881 journal_jremref(dirrem, jremref, inodedep); 8882 if (dotremref) 8883 journal_jremref(dirrem, dotremref, inodedep); 8884 if (dotdotremref) 8885 journal_jremref(dirrem, dotdotremref, NULL); 8886 jwork_move(&dirrem->dm_jwork, &dap->da_jwork); 8887 free_diradd(dap, &dirrem->dm_jwork); 8888} 8889 8890/* 8891 * Free a diradd dependency structure. 8892 */ 8893static void 8894free_diradd(dap, wkhd) 8895 struct diradd *dap; 8896 struct workhead *wkhd; 8897{ 8898 struct dirrem *dirrem; 8899 struct pagedep *pagedep; 8900 struct inodedep *inodedep; 8901 struct mkdir *mkdir, *nextmd; 8902 struct ufsmount *ump; 8903 8904 ump = VFSTOUFS(dap->da_list.wk_mp); 8905 LOCK_OWNED(ump); 8906 LIST_REMOVE(dap, da_pdlist); 8907 if (dap->da_state & ONWORKLIST) 8908 WORKLIST_REMOVE(&dap->da_list); 8909 if ((dap->da_state & DIRCHG) == 0) { 8910 pagedep = dap->da_pagedep; 8911 } else { 8912 dirrem = dap->da_previous; 8913 pagedep = dirrem->dm_pagedep; 8914 dirrem->dm_dirinum = pagedep->pd_ino; 8915 dirrem->dm_state |= COMPLETE; 8916 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 8917 add_to_worklist(&dirrem->dm_list, 0); 8918 } 8919 if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum, 8920 0, &inodedep) != 0) 8921 if (inodedep->id_mkdiradd == dap) 8922 inodedep->id_mkdiradd = NULL; 8923 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8924 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir; 8925 mkdir = nextmd) { 8926 nextmd = LIST_NEXT(mkdir, md_mkdirs); 8927 if (mkdir->md_diradd != dap) 8928 continue; 8929 dap->da_state &= 8930 ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)); 8931 LIST_REMOVE(mkdir, md_mkdirs); 8932 if (mkdir->md_state & ONWORKLIST) 8933 WORKLIST_REMOVE(&mkdir->md_list); 8934 if (mkdir->md_jaddref != NULL) 8935 panic("free_diradd: Unexpected jaddref"); 8936 WORKITEM_FREE(mkdir, D_MKDIR); 8937 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 8938 break; 8939 } 8940 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 8941 panic("free_diradd: unfound ref"); 8942 } 8943 if (inodedep) 8944 free_inodedep(inodedep); 8945 /* 8946 * Free any journal segments waiting for the directory write. 8947 */ 8948 handle_jwork(&dap->da_jwork); 8949 WORKITEM_FREE(dap, D_DIRADD); 8950} 8951 8952/* 8953 * Directory entry removal dependencies. 8954 * 8955 * When removing a directory entry, the entry's inode pointer must be 8956 * zero'ed on disk before the corresponding inode's link count is decremented 8957 * (possibly freeing the inode for re-use). This dependency is handled by 8958 * updating the directory entry but delaying the inode count reduction until 8959 * after the directory block has been written to disk. After this point, the 8960 * inode count can be decremented whenever it is convenient. 8961 */ 8962 8963/* 8964 * This routine should be called immediately after removing 8965 * a directory entry. The inode's link count should not be 8966 * decremented by the calling procedure -- the soft updates 8967 * code will do this task when it is safe. 8968 */ 8969void 8970softdep_setup_remove(bp, dp, ip, isrmdir) 8971 struct buf *bp; /* buffer containing directory block */ 8972 struct inode *dp; /* inode for the directory being modified */ 8973 struct inode *ip; /* inode for directory entry being removed */ 8974 int isrmdir; /* indicates if doing RMDIR */ 8975{ 8976 struct dirrem *dirrem, *prevdirrem; 8977 struct inodedep *inodedep; 8978 struct ufsmount *ump; 8979 int direct; 8980 8981 ump = ITOUMP(ip); 8982 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 8983 ("softdep_setup_remove called on non-softdep filesystem")); 8984 /* 8985 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. We want 8986 * newdirrem() to setup the full directory remove which requires 8987 * isrmdir > 1. 8988 */ 8989 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 8990 /* 8991 * Add the dirrem to the inodedep's pending remove list for quick 8992 * discovery later. 8993 */ 8994 if (inodedep_lookup(UFSTOVFS(ump), ip->i_number, 0, &inodedep) == 0) 8995 panic("softdep_setup_remove: Lost inodedep."); 8996 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked")); 8997 dirrem->dm_state |= ONDEPLIST; 8998 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 8999 9000 /* 9001 * If the COMPLETE flag is clear, then there were no active 9002 * entries and we want to roll back to a zeroed entry until 9003 * the new inode is committed to disk. If the COMPLETE flag is 9004 * set then we have deleted an entry that never made it to 9005 * disk. If the entry we deleted resulted from a name change, 9006 * then the old name still resides on disk. We cannot delete 9007 * its inode (returned to us in prevdirrem) until the zeroed 9008 * directory entry gets to disk. The new inode has never been 9009 * referenced on the disk, so can be deleted immediately. 9010 */ 9011 if ((dirrem->dm_state & COMPLETE) == 0) { 9012 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 9013 dm_next); 9014 FREE_LOCK(ump); 9015 } else { 9016 if (prevdirrem != NULL) 9017 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 9018 prevdirrem, dm_next); 9019 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 9020 direct = LIST_EMPTY(&dirrem->dm_jremrefhd); 9021 FREE_LOCK(ump); 9022 if (direct) 9023 handle_workitem_remove(dirrem, 0); 9024 } 9025} 9026 9027/* 9028 * Check for an entry matching 'offset' on both the pd_dirraddhd list and the 9029 * pd_pendinghd list of a pagedep. 9030 */ 9031static struct diradd * 9032diradd_lookup(pagedep, offset) 9033 struct pagedep *pagedep; 9034 int offset; 9035{ 9036 struct diradd *dap; 9037 9038 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 9039 if (dap->da_offset == offset) 9040 return (dap); 9041 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 9042 if (dap->da_offset == offset) 9043 return (dap); 9044 return (NULL); 9045} 9046 9047/* 9048 * Search for a .. diradd dependency in a directory that is being removed. 9049 * If the directory was renamed to a new parent we have a diradd rather 9050 * than a mkdir for the .. entry. We need to cancel it now before 9051 * it is found in truncate(). 9052 */ 9053static struct jremref * 9054cancel_diradd_dotdot(ip, dirrem, jremref) 9055 struct inode *ip; 9056 struct dirrem *dirrem; 9057 struct jremref *jremref; 9058{ 9059 struct pagedep *pagedep; 9060 struct diradd *dap; 9061 struct worklist *wk; 9062 9063 if (pagedep_lookup(ITOVFS(ip), NULL, ip->i_number, 0, 0, &pagedep) == 0) 9064 return (jremref); 9065 dap = diradd_lookup(pagedep, DOTDOT_OFFSET); 9066 if (dap == NULL) 9067 return (jremref); 9068 cancel_diradd(dap, dirrem, jremref, NULL, NULL); 9069 /* 9070 * Mark any journal work as belonging to the parent so it is freed 9071 * with the .. reference. 9072 */ 9073 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list) 9074 wk->wk_state |= MKDIR_PARENT; 9075 return (NULL); 9076} 9077 9078/* 9079 * Cancel the MKDIR_PARENT mkdir component of a diradd when we're going to 9080 * replace it with a dirrem/diradd pair as a result of re-parenting a 9081 * directory. This ensures that we don't simultaneously have a mkdir and 9082 * a diradd for the same .. entry. 9083 */ 9084static struct jremref * 9085cancel_mkdir_dotdot(ip, dirrem, jremref) 9086 struct inode *ip; 9087 struct dirrem *dirrem; 9088 struct jremref *jremref; 9089{ 9090 struct inodedep *inodedep; 9091 struct jaddref *jaddref; 9092 struct ufsmount *ump; 9093 struct mkdir *mkdir; 9094 struct diradd *dap; 9095 struct mount *mp; 9096 9097 mp = ITOVFS(ip); 9098 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 9099 return (jremref); 9100 dap = inodedep->id_mkdiradd; 9101 if (dap == NULL || (dap->da_state & MKDIR_PARENT) == 0) 9102 return (jremref); 9103 ump = VFSTOUFS(inodedep->id_list.wk_mp); 9104 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir; 9105 mkdir = LIST_NEXT(mkdir, md_mkdirs)) 9106 if (mkdir->md_diradd == dap && mkdir->md_state & MKDIR_PARENT) 9107 break; 9108 if (mkdir == NULL) 9109 panic("cancel_mkdir_dotdot: Unable to find mkdir\n"); 9110 if ((jaddref = mkdir->md_jaddref) != NULL) { 9111 mkdir->md_jaddref = NULL; 9112 jaddref->ja_state &= ~MKDIR_PARENT; 9113 if (inodedep_lookup(mp, jaddref->ja_ino, 0, &inodedep) == 0) 9114 panic("cancel_mkdir_dotdot: Lost parent inodedep"); 9115 if (cancel_jaddref(jaddref, inodedep, &dirrem->dm_jwork)) { 9116 journal_jremref(dirrem, jremref, inodedep); 9117 jremref = NULL; 9118 } 9119 } 9120 if (mkdir->md_state & ONWORKLIST) 9121 WORKLIST_REMOVE(&mkdir->md_list); 9122 mkdir->md_state |= ALLCOMPLETE; 9123 complete_mkdir(mkdir); 9124 return (jremref); 9125} 9126 9127static void 9128journal_jremref(dirrem, jremref, inodedep) 9129 struct dirrem *dirrem; 9130 struct jremref *jremref; 9131 struct inodedep *inodedep; 9132{ 9133 9134 if (inodedep == NULL) 9135 if (inodedep_lookup(jremref->jr_list.wk_mp, 9136 jremref->jr_ref.if_ino, 0, &inodedep) == 0) 9137 panic("journal_jremref: Lost inodedep"); 9138 LIST_INSERT_HEAD(&dirrem->dm_jremrefhd, jremref, jr_deps); 9139 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps); 9140 add_to_journal(&jremref->jr_list); 9141} 9142 9143static void 9144dirrem_journal(dirrem, jremref, dotremref, dotdotremref) 9145 struct dirrem *dirrem; 9146 struct jremref *jremref; 9147 struct jremref *dotremref; 9148 struct jremref *dotdotremref; 9149{ 9150 struct inodedep *inodedep; 9151 9152 9153 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 0, 9154 &inodedep) == 0) 9155 panic("dirrem_journal: Lost inodedep"); 9156 journal_jremref(dirrem, jremref, inodedep); 9157 if (dotremref) 9158 journal_jremref(dirrem, dotremref, inodedep); 9159 if (dotdotremref) 9160 journal_jremref(dirrem, dotdotremref, NULL); 9161} 9162 9163/* 9164 * Allocate a new dirrem if appropriate and return it along with 9165 * its associated pagedep. Called without a lock, returns with lock. 9166 */ 9167static struct dirrem * 9168newdirrem(bp, dp, ip, isrmdir, prevdirremp) 9169 struct buf *bp; /* buffer containing directory block */ 9170 struct inode *dp; /* inode for the directory being modified */ 9171 struct inode *ip; /* inode for directory entry being removed */ 9172 int isrmdir; /* indicates if doing RMDIR */ 9173 struct dirrem **prevdirremp; /* previously referenced inode, if any */ 9174{ 9175 int offset; 9176 ufs_lbn_t lbn; 9177 struct diradd *dap; 9178 struct dirrem *dirrem; 9179 struct pagedep *pagedep; 9180 struct jremref *jremref; 9181 struct jremref *dotremref; 9182 struct jremref *dotdotremref; 9183 struct vnode *dvp; 9184 struct ufsmount *ump; 9185 9186 /* 9187 * Whiteouts have no deletion dependencies. 9188 */ 9189 if (ip == NULL) 9190 panic("newdirrem: whiteout"); 9191 dvp = ITOV(dp); 9192 ump = ITOUMP(dp); 9193 9194 /* 9195 * If the system is over its limit and our filesystem is 9196 * responsible for more than our share of that usage and 9197 * we are not a snapshot, request some inodedep cleanup. 9198 * Limiting the number of dirrem structures will also limit 9199 * the number of freefile and freeblks structures. 9200 */ 9201 ACQUIRE_LOCK(ump); 9202 if (!IS_SNAPSHOT(ip) && softdep_excess_items(ump, D_DIRREM)) 9203 schedule_cleanup(UFSTOVFS(ump)); 9204 else 9205 FREE_LOCK(ump); 9206 dirrem = malloc(sizeof(struct dirrem), M_DIRREM, M_SOFTDEP_FLAGS | 9207 M_ZERO); 9208 workitem_alloc(&dirrem->dm_list, D_DIRREM, dvp->v_mount); 9209 LIST_INIT(&dirrem->dm_jremrefhd); 9210 LIST_INIT(&dirrem->dm_jwork); 9211 dirrem->dm_state = isrmdir ? RMDIR : 0; 9212 dirrem->dm_oldinum = ip->i_number; 9213 *prevdirremp = NULL; 9214 /* 9215 * Allocate remove reference structures to track journal write 9216 * dependencies. We will always have one for the link and 9217 * when doing directories we will always have one more for dot. 9218 * When renaming a directory we skip the dotdot link change so 9219 * this is not needed. 9220 */ 9221 jremref = dotremref = dotdotremref = NULL; 9222 if (DOINGSUJ(dvp)) { 9223 if (isrmdir) { 9224 jremref = newjremref(dirrem, dp, ip, dp->i_offset, 9225 ip->i_effnlink + 2); 9226 dotremref = newjremref(dirrem, ip, ip, DOT_OFFSET, 9227 ip->i_effnlink + 1); 9228 dotdotremref = newjremref(dirrem, ip, dp, DOTDOT_OFFSET, 9229 dp->i_effnlink + 1); 9230 dotdotremref->jr_state |= MKDIR_PARENT; 9231 } else 9232 jremref = newjremref(dirrem, dp, ip, dp->i_offset, 9233 ip->i_effnlink + 1); 9234 } 9235 ACQUIRE_LOCK(ump); 9236 lbn = lblkno(ump->um_fs, dp->i_offset); 9237 offset = blkoff(ump->um_fs, dp->i_offset); 9238 pagedep_lookup(UFSTOVFS(ump), bp, dp->i_number, lbn, DEPALLOC, 9239 &pagedep); 9240 dirrem->dm_pagedep = pagedep; 9241 dirrem->dm_offset = offset; 9242 /* 9243 * If we're renaming a .. link to a new directory, cancel any 9244 * existing MKDIR_PARENT mkdir. If it has already been canceled 9245 * the jremref is preserved for any potential diradd in this 9246 * location. This can not coincide with a rmdir. 9247 */ 9248 if (dp->i_offset == DOTDOT_OFFSET) { 9249 if (isrmdir) 9250 panic("newdirrem: .. directory change during remove?"); 9251 jremref = cancel_mkdir_dotdot(dp, dirrem, jremref); 9252 } 9253 /* 9254 * If we're removing a directory search for the .. dependency now and 9255 * cancel it. Any pending journal work will be added to the dirrem 9256 * to be completed when the workitem remove completes. 9257 */ 9258 if (isrmdir) 9259 dotdotremref = cancel_diradd_dotdot(ip, dirrem, dotdotremref); 9260 /* 9261 * Check for a diradd dependency for the same directory entry. 9262 * If present, then both dependencies become obsolete and can 9263 * be de-allocated. 9264 */ 9265 dap = diradd_lookup(pagedep, offset); 9266 if (dap == NULL) { 9267 /* 9268 * Link the jremref structures into the dirrem so they are 9269 * written prior to the pagedep. 9270 */ 9271 if (jremref) 9272 dirrem_journal(dirrem, jremref, dotremref, 9273 dotdotremref); 9274 return (dirrem); 9275 } 9276 /* 9277 * Must be ATTACHED at this point. 9278 */ 9279 if ((dap->da_state & ATTACHED) == 0) 9280 panic("newdirrem: not ATTACHED"); 9281 if (dap->da_newinum != ip->i_number) 9282 panic("newdirrem: inum %ju should be %ju", 9283 (uintmax_t)ip->i_number, (uintmax_t)dap->da_newinum); 9284 /* 9285 * If we are deleting a changed name that never made it to disk, 9286 * then return the dirrem describing the previous inode (which 9287 * represents the inode currently referenced from this entry on disk). 9288 */ 9289 if ((dap->da_state & DIRCHG) != 0) { 9290 *prevdirremp = dap->da_previous; 9291 dap->da_state &= ~DIRCHG; 9292 dap->da_pagedep = pagedep; 9293 } 9294 /* 9295 * We are deleting an entry that never made it to disk. 9296 * Mark it COMPLETE so we can delete its inode immediately. 9297 */ 9298 dirrem->dm_state |= COMPLETE; 9299 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref); 9300#ifdef SUJ_DEBUG 9301 if (isrmdir == 0) { 9302 struct worklist *wk; 9303 9304 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list) 9305 if (wk->wk_state & (MKDIR_BODY | MKDIR_PARENT)) 9306 panic("bad wk %p (0x%X)\n", wk, wk->wk_state); 9307 } 9308#endif 9309 9310 return (dirrem); 9311} 9312 9313/* 9314 * Directory entry change dependencies. 9315 * 9316 * Changing an existing directory entry requires that an add operation 9317 * be completed first followed by a deletion. The semantics for the addition 9318 * are identical to the description of adding a new entry above except 9319 * that the rollback is to the old inode number rather than zero. Once 9320 * the addition dependency is completed, the removal is done as described 9321 * in the removal routine above. 9322 */ 9323 9324/* 9325 * This routine should be called immediately after changing 9326 * a directory entry. The inode's link count should not be 9327 * decremented by the calling procedure -- the soft updates 9328 * code will perform this task when it is safe. 9329 */ 9330void 9331softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 9332 struct buf *bp; /* buffer containing directory block */ 9333 struct inode *dp; /* inode for the directory being modified */ 9334 struct inode *ip; /* inode for directory entry being removed */ 9335 ino_t newinum; /* new inode number for changed entry */ 9336 int isrmdir; /* indicates if doing RMDIR */ 9337{ 9338 int offset; 9339 struct diradd *dap = NULL; 9340 struct dirrem *dirrem, *prevdirrem; 9341 struct pagedep *pagedep; 9342 struct inodedep *inodedep; 9343 struct jaddref *jaddref; 9344 struct mount *mp; 9345 struct ufsmount *ump; 9346 9347 mp = ITOVFS(dp); 9348 ump = VFSTOUFS(mp); 9349 offset = blkoff(ump->um_fs, dp->i_offset); 9350 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 9351 ("softdep_setup_directory_change called on non-softdep filesystem")); 9352 9353 /* 9354 * Whiteouts do not need diradd dependencies. 9355 */ 9356 if (newinum != UFS_WINO) { 9357 dap = malloc(sizeof(struct diradd), 9358 M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO); 9359 workitem_alloc(&dap->da_list, D_DIRADD, mp); 9360 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 9361 dap->da_offset = offset; 9362 dap->da_newinum = newinum; 9363 LIST_INIT(&dap->da_jwork); 9364 } 9365 9366 /* 9367 * Allocate a new dirrem and ACQUIRE_LOCK. 9368 */ 9369 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 9370 pagedep = dirrem->dm_pagedep; 9371 /* 9372 * The possible values for isrmdir: 9373 * 0 - non-directory file rename 9374 * 1 - directory rename within same directory 9375 * inum - directory rename to new directory of given inode number 9376 * When renaming to a new directory, we are both deleting and 9377 * creating a new directory entry, so the link count on the new 9378 * directory should not change. Thus we do not need the followup 9379 * dirrem which is usually done in handle_workitem_remove. We set 9380 * the DIRCHG flag to tell handle_workitem_remove to skip the 9381 * followup dirrem. 9382 */ 9383 if (isrmdir > 1) 9384 dirrem->dm_state |= DIRCHG; 9385 9386 /* 9387 * Whiteouts have no additional dependencies, 9388 * so just put the dirrem on the correct list. 9389 */ 9390 if (newinum == UFS_WINO) { 9391 if ((dirrem->dm_state & COMPLETE) == 0) { 9392 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 9393 dm_next); 9394 } else { 9395 dirrem->dm_dirinum = pagedep->pd_ino; 9396 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 9397 add_to_worklist(&dirrem->dm_list, 0); 9398 } 9399 FREE_LOCK(ump); 9400 return; 9401 } 9402 /* 9403 * Add the dirrem to the inodedep's pending remove list for quick 9404 * discovery later. A valid nlinkdelta ensures that this lookup 9405 * will not fail. 9406 */ 9407 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 9408 panic("softdep_setup_directory_change: Lost inodedep."); 9409 dirrem->dm_state |= ONDEPLIST; 9410 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 9411 9412 /* 9413 * If the COMPLETE flag is clear, then there were no active 9414 * entries and we want to roll back to the previous inode until 9415 * the new inode is committed to disk. If the COMPLETE flag is 9416 * set, then we have deleted an entry that never made it to disk. 9417 * If the entry we deleted resulted from a name change, then the old 9418 * inode reference still resides on disk. Any rollback that we do 9419 * needs to be to that old inode (returned to us in prevdirrem). If 9420 * the entry we deleted resulted from a create, then there is 9421 * no entry on the disk, so we want to roll back to zero rather 9422 * than the uncommitted inode. In either of the COMPLETE cases we 9423 * want to immediately free the unwritten and unreferenced inode. 9424 */ 9425 if ((dirrem->dm_state & COMPLETE) == 0) { 9426 dap->da_previous = dirrem; 9427 } else { 9428 if (prevdirrem != NULL) { 9429 dap->da_previous = prevdirrem; 9430 } else { 9431 dap->da_state &= ~DIRCHG; 9432 dap->da_pagedep = pagedep; 9433 } 9434 dirrem->dm_dirinum = pagedep->pd_ino; 9435 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 9436 add_to_worklist(&dirrem->dm_list, 0); 9437 } 9438 /* 9439 * Lookup the jaddref for this journal entry. We must finish 9440 * initializing it and make the diradd write dependent on it. 9441 * If we're not journaling, put it on the id_bufwait list if the 9442 * inode is not yet written. If it is written, do the post-inode 9443 * write processing to put it on the id_pendinghd list. 9444 */ 9445 inodedep_lookup(mp, newinum, DEPALLOC, &inodedep); 9446 if (MOUNTEDSUJ(mp)) { 9447 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 9448 inoreflst); 9449 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 9450 ("softdep_setup_directory_change: bad jaddref %p", 9451 jaddref)); 9452 jaddref->ja_diroff = dp->i_offset; 9453 jaddref->ja_diradd = dap; 9454 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 9455 dap, da_pdlist); 9456 add_to_journal(&jaddref->ja_list); 9457 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 9458 dap->da_state |= COMPLETE; 9459 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 9460 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 9461 } else { 9462 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 9463 dap, da_pdlist); 9464 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 9465 } 9466 /* 9467 * If we're making a new name for a directory that has not been 9468 * committed when need to move the dot and dotdot references to 9469 * this new name. 9470 */ 9471 if (inodedep->id_mkdiradd && dp->i_offset != DOTDOT_OFFSET) 9472 merge_diradd(inodedep, dap); 9473 FREE_LOCK(ump); 9474} 9475 9476/* 9477 * Called whenever the link count on an inode is changed. 9478 * It creates an inode dependency so that the new reference(s) 9479 * to the inode cannot be committed to disk until the updated 9480 * inode has been written. 9481 */ 9482void 9483softdep_change_linkcnt(ip) 9484 struct inode *ip; /* the inode with the increased link count */ 9485{ 9486 struct inodedep *inodedep; 9487 struct ufsmount *ump; 9488 9489 ump = ITOUMP(ip); 9490 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 9491 ("softdep_change_linkcnt called on non-softdep filesystem")); 9492 ACQUIRE_LOCK(ump); 9493 inodedep_lookup(UFSTOVFS(ump), ip->i_number, DEPALLOC, &inodedep); 9494 if (ip->i_nlink < ip->i_effnlink) 9495 panic("softdep_change_linkcnt: bad delta"); 9496 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9497 FREE_LOCK(ump); 9498} 9499 9500/* 9501 * Attach a sbdep dependency to the superblock buf so that we can keep 9502 * track of the head of the linked list of referenced but unlinked inodes. 9503 */ 9504void 9505softdep_setup_sbupdate(ump, fs, bp) 9506 struct ufsmount *ump; 9507 struct fs *fs; 9508 struct buf *bp; 9509{ 9510 struct sbdep *sbdep; 9511 struct worklist *wk; 9512 9513 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 9514 ("softdep_setup_sbupdate called on non-softdep filesystem")); 9515 LIST_FOREACH(wk, &bp->b_dep, wk_list) 9516 if (wk->wk_type == D_SBDEP) 9517 break; 9518 if (wk != NULL) 9519 return; 9520 sbdep = malloc(sizeof(struct sbdep), M_SBDEP, M_SOFTDEP_FLAGS); 9521 workitem_alloc(&sbdep->sb_list, D_SBDEP, UFSTOVFS(ump)); 9522 sbdep->sb_fs = fs; 9523 sbdep->sb_ump = ump; 9524 ACQUIRE_LOCK(ump); 9525 WORKLIST_INSERT(&bp->b_dep, &sbdep->sb_list); 9526 FREE_LOCK(ump); 9527} 9528 9529/* 9530 * Return the first unlinked inodedep which is ready to be the head of the 9531 * list. The inodedep and all those after it must have valid next pointers. 9532 */ 9533static struct inodedep * 9534first_unlinked_inodedep(ump) 9535 struct ufsmount *ump; 9536{ 9537 struct inodedep *inodedep; 9538 struct inodedep *idp; 9539 9540 LOCK_OWNED(ump); 9541 for (inodedep = TAILQ_LAST(&ump->softdep_unlinked, inodedeplst); 9542 inodedep; inodedep = idp) { 9543 if ((inodedep->id_state & UNLINKNEXT) == 0) 9544 return (NULL); 9545 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9546 if (idp == NULL || (idp->id_state & UNLINKNEXT) == 0) 9547 break; 9548 if ((inodedep->id_state & UNLINKPREV) == 0) 9549 break; 9550 } 9551 return (inodedep); 9552} 9553 9554/* 9555 * Set the sujfree unlinked head pointer prior to writing a superblock. 9556 */ 9557static void 9558initiate_write_sbdep(sbdep) 9559 struct sbdep *sbdep; 9560{ 9561 struct inodedep *inodedep; 9562 struct fs *bpfs; 9563 struct fs *fs; 9564 9565 bpfs = sbdep->sb_fs; 9566 fs = sbdep->sb_ump->um_fs; 9567 inodedep = first_unlinked_inodedep(sbdep->sb_ump); 9568 if (inodedep) { 9569 fs->fs_sujfree = inodedep->id_ino; 9570 inodedep->id_state |= UNLINKPREV; 9571 } else 9572 fs->fs_sujfree = 0; 9573 bpfs->fs_sujfree = fs->fs_sujfree; 9574} 9575 9576/* 9577 * After a superblock is written determine whether it must be written again 9578 * due to a changing unlinked list head. 9579 */ 9580static int 9581handle_written_sbdep(sbdep, bp) 9582 struct sbdep *sbdep; 9583 struct buf *bp; 9584{ 9585 struct inodedep *inodedep; 9586 struct fs *fs; 9587 9588 LOCK_OWNED(sbdep->sb_ump); 9589 fs = sbdep->sb_fs; 9590 /* 9591 * If the superblock doesn't match the in-memory list start over. 9592 */ 9593 inodedep = first_unlinked_inodedep(sbdep->sb_ump); 9594 if ((inodedep && fs->fs_sujfree != inodedep->id_ino) || 9595 (inodedep == NULL && fs->fs_sujfree != 0)) { 9596 bdirty(bp); 9597 return (1); 9598 } 9599 WORKITEM_FREE(sbdep, D_SBDEP); 9600 if (fs->fs_sujfree == 0) 9601 return (0); 9602 /* 9603 * Now that we have a record of this inode in stable store allow it 9604 * to be written to free up pending work. Inodes may see a lot of 9605 * write activity after they are unlinked which we must not hold up. 9606 */ 9607 for (; inodedep != NULL; inodedep = TAILQ_NEXT(inodedep, id_unlinked)) { 9608 if ((inodedep->id_state & UNLINKLINKS) != UNLINKLINKS) 9609 panic("handle_written_sbdep: Bad inodedep %p (0x%X)", 9610 inodedep, inodedep->id_state); 9611 if (inodedep->id_state & UNLINKONLIST) 9612 break; 9613 inodedep->id_state |= DEPCOMPLETE | UNLINKONLIST; 9614 } 9615 9616 return (0); 9617} 9618 9619/* 9620 * Mark an inodedep as unlinked and insert it into the in-memory unlinked list. 9621 */ 9622static void 9623unlinked_inodedep(mp, inodedep) 9624 struct mount *mp; 9625 struct inodedep *inodedep; 9626{ 9627 struct ufsmount *ump; 9628 9629 ump = VFSTOUFS(mp); 9630 LOCK_OWNED(ump); 9631 if (MOUNTEDSUJ(mp) == 0) 9632 return; 9633 ump->um_fs->fs_fmod = 1; 9634 if (inodedep->id_state & UNLINKED) 9635 panic("unlinked_inodedep: %p already unlinked\n", inodedep); 9636 inodedep->id_state |= UNLINKED; 9637 TAILQ_INSERT_HEAD(&ump->softdep_unlinked, inodedep, id_unlinked); 9638} 9639 9640/* 9641 * Remove an inodedep from the unlinked inodedep list. This may require 9642 * disk writes if the inode has made it that far. 9643 */ 9644static void 9645clear_unlinked_inodedep(inodedep) 9646 struct inodedep *inodedep; 9647{ 9648 struct ufsmount *ump; 9649 struct inodedep *idp; 9650 struct inodedep *idn; 9651 struct fs *fs; 9652 struct buf *bp; 9653 ino_t ino; 9654 ino_t nino; 9655 ino_t pino; 9656 int error; 9657 9658 ump = VFSTOUFS(inodedep->id_list.wk_mp); 9659 fs = ump->um_fs; 9660 ino = inodedep->id_ino; 9661 error = 0; 9662 for (;;) { 9663 LOCK_OWNED(ump); 9664 KASSERT((inodedep->id_state & UNLINKED) != 0, 9665 ("clear_unlinked_inodedep: inodedep %p not unlinked", 9666 inodedep)); 9667 /* 9668 * If nothing has yet been written simply remove us from 9669 * the in memory list and return. This is the most common 9670 * case where handle_workitem_remove() loses the final 9671 * reference. 9672 */ 9673 if ((inodedep->id_state & UNLINKLINKS) == 0) 9674 break; 9675 /* 9676 * If we have a NEXT pointer and no PREV pointer we can simply 9677 * clear NEXT's PREV and remove ourselves from the list. Be 9678 * careful not to clear PREV if the superblock points at 9679 * next as well. 9680 */ 9681 idn = TAILQ_NEXT(inodedep, id_unlinked); 9682 if ((inodedep->id_state & UNLINKLINKS) == UNLINKNEXT) { 9683 if (idn && fs->fs_sujfree != idn->id_ino) 9684 idn->id_state &= ~UNLINKPREV; 9685 break; 9686 } 9687 /* 9688 * Here we have an inodedep which is actually linked into 9689 * the list. We must remove it by forcing a write to the 9690 * link before us, whether it be the superblock or an inode. 9691 * Unfortunately the list may change while we're waiting 9692 * on the buf lock for either resource so we must loop until 9693 * we lock the right one. If both the superblock and an 9694 * inode point to this inode we must clear the inode first 9695 * followed by the superblock. 9696 */ 9697 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9698 pino = 0; 9699 if (idp && (idp->id_state & UNLINKNEXT)) 9700 pino = idp->id_ino; 9701 FREE_LOCK(ump); 9702 if (pino == 0) { 9703 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc), 9704 (int)fs->fs_sbsize, 0, 0, 0); 9705 } else { 9706 error = bread(ump->um_devvp, 9707 fsbtodb(fs, ino_to_fsba(fs, pino)), 9708 (int)fs->fs_bsize, NOCRED, &bp); 9709 if (error) 9710 brelse(bp); 9711 } 9712 ACQUIRE_LOCK(ump); 9713 if (error) 9714 break; 9715 /* If the list has changed restart the loop. */ 9716 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9717 nino = 0; 9718 if (idp && (idp->id_state & UNLINKNEXT)) 9719 nino = idp->id_ino; 9720 if (nino != pino || 9721 (inodedep->id_state & UNLINKPREV) != UNLINKPREV) { 9722 FREE_LOCK(ump); 9723 brelse(bp); 9724 ACQUIRE_LOCK(ump); 9725 continue; 9726 } 9727 nino = 0; 9728 idn = TAILQ_NEXT(inodedep, id_unlinked); 9729 if (idn) 9730 nino = idn->id_ino; 9731 /* 9732 * Remove us from the in memory list. After this we cannot 9733 * access the inodedep. 9734 */ 9735 KASSERT((inodedep->id_state & UNLINKED) != 0, 9736 ("clear_unlinked_inodedep: inodedep %p not unlinked", 9737 inodedep)); 9738 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST); 9739 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked); 9740 FREE_LOCK(ump); 9741 /* 9742 * The predecessor's next pointer is manually updated here 9743 * so that the NEXT flag is never cleared for an element 9744 * that is in the list. 9745 */ 9746 if (pino == 0) { 9747 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize); 9748 ffs_oldfscompat_write((struct fs *)bp->b_data, ump); 9749 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data, 9750 bp); 9751 } else if (fs->fs_magic == FS_UFS1_MAGIC) 9752 ((struct ufs1_dinode *)bp->b_data + 9753 ino_to_fsbo(fs, pino))->di_freelink = nino; 9754 else 9755 ((struct ufs2_dinode *)bp->b_data + 9756 ino_to_fsbo(fs, pino))->di_freelink = nino; 9757 /* 9758 * If the bwrite fails we have no recourse to recover. The 9759 * filesystem is corrupted already. 9760 */ 9761 bwrite(bp); 9762 ACQUIRE_LOCK(ump); 9763 /* 9764 * If the superblock pointer still needs to be cleared force 9765 * a write here. 9766 */ 9767 if (fs->fs_sujfree == ino) { 9768 FREE_LOCK(ump); 9769 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc), 9770 (int)fs->fs_sbsize, 0, 0, 0); 9771 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize); 9772 ffs_oldfscompat_write((struct fs *)bp->b_data, ump); 9773 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data, 9774 bp); 9775 bwrite(bp); 9776 ACQUIRE_LOCK(ump); 9777 } 9778 9779 if (fs->fs_sujfree != ino) 9780 return; 9781 panic("clear_unlinked_inodedep: Failed to clear free head"); 9782 } 9783 if (inodedep->id_ino == fs->fs_sujfree) 9784 panic("clear_unlinked_inodedep: Freeing head of free list"); 9785 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST); 9786 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked); 9787 return; 9788} 9789 9790/* 9791 * This workitem decrements the inode's link count. 9792 * If the link count reaches zero, the file is removed. 9793 */ 9794static int 9795handle_workitem_remove(dirrem, flags) 9796 struct dirrem *dirrem; 9797 int flags; 9798{ 9799 struct inodedep *inodedep; 9800 struct workhead dotdotwk; 9801 struct worklist *wk; 9802 struct ufsmount *ump; 9803 struct mount *mp; 9804 struct vnode *vp; 9805 struct inode *ip; 9806 ino_t oldinum; 9807 9808 if (dirrem->dm_state & ONWORKLIST) 9809 panic("handle_workitem_remove: dirrem %p still on worklist", 9810 dirrem); 9811 oldinum = dirrem->dm_oldinum; 9812 mp = dirrem->dm_list.wk_mp; 9813 ump = VFSTOUFS(mp); 9814 flags |= LK_EXCLUSIVE; 9815 if (ffs_vgetf(mp, oldinum, flags, &vp, FFSV_FORCEINSMQ) != 0) 9816 return (EBUSY); 9817 ip = VTOI(vp); 9818 MPASS(ip->i_mode != 0); 9819 ACQUIRE_LOCK(ump); 9820 if ((inodedep_lookup(mp, oldinum, 0, &inodedep)) == 0) 9821 panic("handle_workitem_remove: lost inodedep"); 9822 if (dirrem->dm_state & ONDEPLIST) 9823 LIST_REMOVE(dirrem, dm_inonext); 9824 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd), 9825 ("handle_workitem_remove: Journal entries not written.")); 9826 9827 /* 9828 * Move all dependencies waiting on the remove to complete 9829 * from the dirrem to the inode inowait list to be completed 9830 * after the inode has been updated and written to disk. 9831 * 9832 * Any marked MKDIR_PARENT are saved to be completed when the 9833 * dotdot ref is removed unless DIRCHG is specified. For 9834 * directory change operations there will be no further 9835 * directory writes and the jsegdeps need to be moved along 9836 * with the rest to be completed when the inode is free or 9837 * stable in the inode free list. 9838 */ 9839 LIST_INIT(&dotdotwk); 9840 while ((wk = LIST_FIRST(&dirrem->dm_jwork)) != NULL) { 9841 WORKLIST_REMOVE(wk); 9842 if ((dirrem->dm_state & DIRCHG) == 0 && 9843 wk->wk_state & MKDIR_PARENT) { 9844 wk->wk_state &= ~MKDIR_PARENT; 9845 WORKLIST_INSERT(&dotdotwk, wk); 9846 continue; 9847 } 9848 WORKLIST_INSERT(&inodedep->id_inowait, wk); 9849 } 9850 LIST_SWAP(&dirrem->dm_jwork, &dotdotwk, worklist, wk_list); 9851 /* 9852 * Normal file deletion. 9853 */ 9854 if ((dirrem->dm_state & RMDIR) == 0) { 9855 ip->i_nlink--; 9856 DIP_SET(ip, i_nlink, ip->i_nlink); 9857 ip->i_flag |= IN_CHANGE; 9858 if (ip->i_nlink < ip->i_effnlink) 9859 panic("handle_workitem_remove: bad file delta"); 9860 if (ip->i_nlink == 0) 9861 unlinked_inodedep(mp, inodedep); 9862 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9863 KASSERT(LIST_EMPTY(&dirrem->dm_jwork), 9864 ("handle_workitem_remove: worklist not empty. %s", 9865 TYPENAME(LIST_FIRST(&dirrem->dm_jwork)->wk_type))); 9866 WORKITEM_FREE(dirrem, D_DIRREM); 9867 FREE_LOCK(ump); 9868 goto out; 9869 } 9870 /* 9871 * Directory deletion. Decrement reference count for both the 9872 * just deleted parent directory entry and the reference for ".". 9873 * Arrange to have the reference count on the parent decremented 9874 * to account for the loss of "..". 9875 */ 9876 ip->i_nlink -= 2; 9877 DIP_SET(ip, i_nlink, ip->i_nlink); 9878 ip->i_flag |= IN_CHANGE; 9879 if (ip->i_nlink < ip->i_effnlink) 9880 panic("handle_workitem_remove: bad dir delta"); 9881 if (ip->i_nlink == 0) 9882 unlinked_inodedep(mp, inodedep); 9883 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9884 /* 9885 * Rename a directory to a new parent. Since, we are both deleting 9886 * and creating a new directory entry, the link count on the new 9887 * directory should not change. Thus we skip the followup dirrem. 9888 */ 9889 if (dirrem->dm_state & DIRCHG) { 9890 KASSERT(LIST_EMPTY(&dirrem->dm_jwork), 9891 ("handle_workitem_remove: DIRCHG and worklist not empty.")); 9892 WORKITEM_FREE(dirrem, D_DIRREM); 9893 FREE_LOCK(ump); 9894 goto out; 9895 } 9896 dirrem->dm_state = ONDEPLIST; 9897 dirrem->dm_oldinum = dirrem->dm_dirinum; 9898 /* 9899 * Place the dirrem on the parent's diremhd list. 9900 */ 9901 if (inodedep_lookup(mp, dirrem->dm_oldinum, 0, &inodedep) == 0) 9902 panic("handle_workitem_remove: lost dir inodedep"); 9903 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 9904 /* 9905 * If the allocated inode has never been written to disk, then 9906 * the on-disk inode is zero'ed and we can remove the file 9907 * immediately. When journaling if the inode has been marked 9908 * unlinked and not DEPCOMPLETE we know it can never be written. 9909 */ 9910 inodedep_lookup(mp, oldinum, 0, &inodedep); 9911 if (inodedep == NULL || 9912 (inodedep->id_state & (DEPCOMPLETE | UNLINKED)) == UNLINKED || 9913 check_inode_unwritten(inodedep)) { 9914 FREE_LOCK(ump); 9915 vput(vp); 9916 return handle_workitem_remove(dirrem, flags); 9917 } 9918 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 9919 FREE_LOCK(ump); 9920 ip->i_flag |= IN_CHANGE; 9921out: 9922 ffs_update(vp, 0); 9923 vput(vp); 9924 return (0); 9925} 9926 9927/* 9928 * Inode de-allocation dependencies. 9929 * 9930 * When an inode's link count is reduced to zero, it can be de-allocated. We 9931 * found it convenient to postpone de-allocation until after the inode is 9932 * written to disk with its new link count (zero). At this point, all of the 9933 * on-disk inode's block pointers are nullified and, with careful dependency 9934 * list ordering, all dependencies related to the inode will be satisfied and 9935 * the corresponding dependency structures de-allocated. So, if/when the 9936 * inode is reused, there will be no mixing of old dependencies with new 9937 * ones. This artificial dependency is set up by the block de-allocation 9938 * procedure above (softdep_setup_freeblocks) and completed by the 9939 * following procedure. 9940 */ 9941static void 9942handle_workitem_freefile(freefile) 9943 struct freefile *freefile; 9944{ 9945 struct workhead wkhd; 9946 struct fs *fs; 9947 struct inodedep *idp; 9948 struct ufsmount *ump; 9949 int error; 9950 9951 ump = VFSTOUFS(freefile->fx_list.wk_mp); 9952 fs = ump->um_fs; 9953#ifdef DEBUG 9954 ACQUIRE_LOCK(ump); 9955 error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp); 9956 FREE_LOCK(ump); 9957 if (error) 9958 panic("handle_workitem_freefile: inodedep %p survived", idp); 9959#endif 9960 UFS_LOCK(ump); 9961 fs->fs_pendinginodes -= 1; 9962 UFS_UNLOCK(ump); 9963 LIST_INIT(&wkhd); 9964 LIST_SWAP(&freefile->fx_jwork, &wkhd, worklist, wk_list); 9965 if ((error = ffs_freefile(ump, fs, freefile->fx_devvp, 9966 freefile->fx_oldinum, freefile->fx_mode, &wkhd)) != 0) 9967 softdep_error("handle_workitem_freefile", error); 9968 ACQUIRE_LOCK(ump); 9969 WORKITEM_FREE(freefile, D_FREEFILE); 9970 FREE_LOCK(ump); 9971} 9972 9973 9974/* 9975 * Helper function which unlinks marker element from work list and returns 9976 * the next element on the list. 9977 */ 9978static __inline struct worklist * 9979markernext(struct worklist *marker) 9980{ 9981 struct worklist *next; 9982 9983 next = LIST_NEXT(marker, wk_list); 9984 LIST_REMOVE(marker, wk_list); 9985 return next; 9986} 9987 9988/* 9989 * Disk writes. 9990 * 9991 * The dependency structures constructed above are most actively used when file 9992 * system blocks are written to disk. No constraints are placed on when a 9993 * block can be written, but unsatisfied update dependencies are made safe by 9994 * modifying (or replacing) the source memory for the duration of the disk 9995 * write. When the disk write completes, the memory block is again brought 9996 * up-to-date. 9997 * 9998 * In-core inode structure reclamation. 9999 * 10000 * Because there are a finite number of "in-core" inode structures, they are 10001 * reused regularly. By transferring all inode-related dependencies to the 10002 * in-memory inode block and indexing them separately (via "inodedep"s), we 10003 * can allow "in-core" inode structures to be reused at any time and avoid 10004 * any increase in contention. 10005 * 10006 * Called just before entering the device driver to initiate a new disk I/O. 10007 * The buffer must be locked, thus, no I/O completion operations can occur 10008 * while we are manipulating its associated dependencies. 10009 */ 10010static void 10011softdep_disk_io_initiation(bp) 10012 struct buf *bp; /* structure describing disk write to occur */ 10013{ 10014 struct worklist *wk; 10015 struct worklist marker; 10016 struct inodedep *inodedep; 10017 struct freeblks *freeblks; 10018 struct jblkdep *jblkdep; 10019 struct newblk *newblk; 10020 struct ufsmount *ump; 10021 10022 /* 10023 * We only care about write operations. There should never 10024 * be dependencies for reads. 10025 */ 10026 if (bp->b_iocmd != BIO_WRITE) 10027 panic("softdep_disk_io_initiation: not write"); 10028 10029 if (bp->b_vflags & BV_BKGRDINPROG) 10030 panic("softdep_disk_io_initiation: Writing buffer with " 10031 "background write in progress: %p", bp); 10032 10033 ump = softdep_bp_to_mp(bp); 10034 if (ump == NULL) 10035 return; 10036 10037 marker.wk_type = D_LAST + 1; /* Not a normal workitem */ 10038 PHOLD(curproc); /* Don't swap out kernel stack */ 10039 ACQUIRE_LOCK(ump); 10040 /* 10041 * Do any necessary pre-I/O processing. 10042 */ 10043 for (wk = LIST_FIRST(&bp->b_dep); wk != NULL; 10044 wk = markernext(&marker)) { 10045 LIST_INSERT_AFTER(wk, &marker, wk_list); 10046 switch (wk->wk_type) { 10047 10048 case D_PAGEDEP: 10049 initiate_write_filepage(WK_PAGEDEP(wk), bp); 10050 continue; 10051 10052 case D_INODEDEP: 10053 inodedep = WK_INODEDEP(wk); 10054 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) 10055 initiate_write_inodeblock_ufs1(inodedep, bp); 10056 else 10057 initiate_write_inodeblock_ufs2(inodedep, bp); 10058 continue; 10059 10060 case D_INDIRDEP: 10061 initiate_write_indirdep(WK_INDIRDEP(wk), bp); 10062 continue; 10063 10064 case D_BMSAFEMAP: 10065 initiate_write_bmsafemap(WK_BMSAFEMAP(wk), bp); 10066 continue; 10067 10068 case D_JSEG: 10069 WK_JSEG(wk)->js_buf = NULL; 10070 continue; 10071 10072 case D_FREEBLKS: 10073 freeblks = WK_FREEBLKS(wk); 10074 jblkdep = LIST_FIRST(&freeblks->fb_jblkdephd); 10075 /* 10076 * We have to wait for the freeblks to be journaled 10077 * before we can write an inodeblock with updated 10078 * pointers. Be careful to arrange the marker so 10079 * we revisit the freeblks if it's not removed by 10080 * the first jwait(). 10081 */ 10082 if (jblkdep != NULL) { 10083 LIST_REMOVE(&marker, wk_list); 10084 LIST_INSERT_BEFORE(wk, &marker, wk_list); 10085 jwait(&jblkdep->jb_list, MNT_WAIT); 10086 } 10087 continue; 10088 case D_ALLOCDIRECT: 10089 case D_ALLOCINDIR: 10090 /* 10091 * We have to wait for the jnewblk to be journaled 10092 * before we can write to a block if the contents 10093 * may be confused with an earlier file's indirect 10094 * at recovery time. Handle the marker as described 10095 * above. 10096 */ 10097 newblk = WK_NEWBLK(wk); 10098 if (newblk->nb_jnewblk != NULL && 10099 indirblk_lookup(newblk->nb_list.wk_mp, 10100 newblk->nb_newblkno)) { 10101 LIST_REMOVE(&marker, wk_list); 10102 LIST_INSERT_BEFORE(wk, &marker, wk_list); 10103 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 10104 } 10105 continue; 10106 10107 case D_SBDEP: 10108 initiate_write_sbdep(WK_SBDEP(wk)); 10109 continue; 10110 10111 case D_MKDIR: 10112 case D_FREEWORK: 10113 case D_FREEDEP: 10114 case D_JSEGDEP: 10115 continue; 10116 10117 default: 10118 panic("handle_disk_io_initiation: Unexpected type %s", 10119 TYPENAME(wk->wk_type)); 10120 /* NOTREACHED */ 10121 } 10122 } 10123 FREE_LOCK(ump); 10124 PRELE(curproc); /* Allow swapout of kernel stack */ 10125} 10126 10127/* 10128 * Called from within the procedure above to deal with unsatisfied 10129 * allocation dependencies in a directory. The buffer must be locked, 10130 * thus, no I/O completion operations can occur while we are 10131 * manipulating its associated dependencies. 10132 */ 10133static void 10134initiate_write_filepage(pagedep, bp) 10135 struct pagedep *pagedep; 10136 struct buf *bp; 10137{ 10138 struct jremref *jremref; 10139 struct jmvref *jmvref; 10140 struct dirrem *dirrem; 10141 struct diradd *dap; 10142 struct direct *ep; 10143 int i; 10144 10145 if (pagedep->pd_state & IOSTARTED) { 10146 /* 10147 * This can only happen if there is a driver that does not 10148 * understand chaining. Here biodone will reissue the call 10149 * to strategy for the incomplete buffers. 10150 */ 10151 printf("initiate_write_filepage: already started\n"); 10152 return; 10153 } 10154 pagedep->pd_state |= IOSTARTED; 10155 /* 10156 * Wait for all journal remove dependencies to hit the disk. 10157 * We can not allow any potentially conflicting directory adds 10158 * to be visible before removes and rollback is too difficult. 10159 * The per-filesystem lock may be dropped and re-acquired, however 10160 * we hold the buf locked so the dependency can not go away. 10161 */ 10162 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) 10163 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) 10164 jwait(&jremref->jr_list, MNT_WAIT); 10165 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) 10166 jwait(&jmvref->jm_list, MNT_WAIT); 10167 for (i = 0; i < DAHASHSZ; i++) { 10168 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 10169 ep = (struct direct *) 10170 ((char *)bp->b_data + dap->da_offset); 10171 if (ep->d_ino != dap->da_newinum) 10172 panic("%s: dir inum %ju != new %ju", 10173 "initiate_write_filepage", 10174 (uintmax_t)ep->d_ino, 10175 (uintmax_t)dap->da_newinum); 10176 if (dap->da_state & DIRCHG) 10177 ep->d_ino = dap->da_previous->dm_oldinum; 10178 else 10179 ep->d_ino = 0; 10180 dap->da_state &= ~ATTACHED; 10181 dap->da_state |= UNDONE; 10182 } 10183 } 10184} 10185 10186/* 10187 * Version of initiate_write_inodeblock that handles UFS1 dinodes. 10188 * Note that any bug fixes made to this routine must be done in the 10189 * version found below. 10190 * 10191 * Called from within the procedure above to deal with unsatisfied 10192 * allocation dependencies in an inodeblock. The buffer must be 10193 * locked, thus, no I/O completion operations can occur while we 10194 * are manipulating its associated dependencies. 10195 */ 10196static void 10197initiate_write_inodeblock_ufs1(inodedep, bp) 10198 struct inodedep *inodedep; 10199 struct buf *bp; /* The inode block */ 10200{ 10201 struct allocdirect *adp, *lastadp; 10202 struct ufs1_dinode *dp; 10203 struct ufs1_dinode *sip; 10204 struct inoref *inoref; 10205 struct ufsmount *ump; 10206 struct fs *fs; 10207 ufs_lbn_t i; 10208#ifdef INVARIANTS 10209 ufs_lbn_t prevlbn = 0; 10210#endif 10211 int deplist; 10212 10213 if (inodedep->id_state & IOSTARTED) 10214 panic("initiate_write_inodeblock_ufs1: already started"); 10215 inodedep->id_state |= IOSTARTED; 10216 fs = inodedep->id_fs; 10217 ump = VFSTOUFS(inodedep->id_list.wk_mp); 10218 LOCK_OWNED(ump); 10219 dp = (struct ufs1_dinode *)bp->b_data + 10220 ino_to_fsbo(fs, inodedep->id_ino); 10221 10222 /* 10223 * If we're on the unlinked list but have not yet written our 10224 * next pointer initialize it here. 10225 */ 10226 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) { 10227 struct inodedep *inon; 10228 10229 inon = TAILQ_NEXT(inodedep, id_unlinked); 10230 dp->di_freelink = inon ? inon->id_ino : 0; 10231 } 10232 /* 10233 * If the bitmap is not yet written, then the allocated 10234 * inode cannot be written to disk. 10235 */ 10236 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 10237 if (inodedep->id_savedino1 != NULL) 10238 panic("initiate_write_inodeblock_ufs1: I/O underway"); 10239 FREE_LOCK(ump); 10240 sip = malloc(sizeof(struct ufs1_dinode), 10241 M_SAVEDINO, M_SOFTDEP_FLAGS); 10242 ACQUIRE_LOCK(ump); 10243 inodedep->id_savedino1 = sip; 10244 *inodedep->id_savedino1 = *dp; 10245 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 10246 dp->di_gen = inodedep->id_savedino1->di_gen; 10247 dp->di_freelink = inodedep->id_savedino1->di_freelink; 10248 return; 10249 } 10250 /* 10251 * If no dependencies, then there is nothing to roll back. 10252 */ 10253 inodedep->id_savedsize = dp->di_size; 10254 inodedep->id_savedextsize = 0; 10255 inodedep->id_savednlink = dp->di_nlink; 10256 if (TAILQ_EMPTY(&inodedep->id_inoupdt) && 10257 TAILQ_EMPTY(&inodedep->id_inoreflst)) 10258 return; 10259 /* 10260 * Revert the link count to that of the first unwritten journal entry. 10261 */ 10262 inoref = TAILQ_FIRST(&inodedep->id_inoreflst); 10263 if (inoref) 10264 dp->di_nlink = inoref->if_nlink; 10265 /* 10266 * Set the dependencies to busy. 10267 */ 10268 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10269 adp = TAILQ_NEXT(adp, ad_next)) { 10270#ifdef INVARIANTS 10271 if (deplist != 0 && prevlbn >= adp->ad_offset) 10272 panic("softdep_write_inodeblock: lbn order"); 10273 prevlbn = adp->ad_offset; 10274 if (adp->ad_offset < UFS_NDADDR && 10275 dp->di_db[adp->ad_offset] != adp->ad_newblkno) 10276 panic("initiate_write_inodeblock_ufs1: " 10277 "direct pointer #%jd mismatch %d != %jd", 10278 (intmax_t)adp->ad_offset, 10279 dp->di_db[adp->ad_offset], 10280 (intmax_t)adp->ad_newblkno); 10281 if (adp->ad_offset >= UFS_NDADDR && 10282 dp->di_ib[adp->ad_offset - UFS_NDADDR] != adp->ad_newblkno) 10283 panic("initiate_write_inodeblock_ufs1: " 10284 "indirect pointer #%jd mismatch %d != %jd", 10285 (intmax_t)adp->ad_offset - UFS_NDADDR, 10286 dp->di_ib[adp->ad_offset - UFS_NDADDR], 10287 (intmax_t)adp->ad_newblkno); 10288 deplist |= 1 << adp->ad_offset; 10289 if ((adp->ad_state & ATTACHED) == 0) 10290 panic("initiate_write_inodeblock_ufs1: " 10291 "Unknown state 0x%x", adp->ad_state); 10292#endif /* INVARIANTS */ 10293 adp->ad_state &= ~ATTACHED; 10294 adp->ad_state |= UNDONE; 10295 } 10296 /* 10297 * The on-disk inode cannot claim to be any larger than the last 10298 * fragment that has been written. Otherwise, the on-disk inode 10299 * might have fragments that were not the last block in the file 10300 * which would corrupt the filesystem. 10301 */ 10302 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10303 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10304 if (adp->ad_offset >= UFS_NDADDR) 10305 break; 10306 dp->di_db[adp->ad_offset] = adp->ad_oldblkno; 10307 /* keep going until hitting a rollback to a frag */ 10308 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10309 continue; 10310 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10311 for (i = adp->ad_offset + 1; i < UFS_NDADDR; i++) { 10312#ifdef INVARIANTS 10313 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 10314 panic("initiate_write_inodeblock_ufs1: " 10315 "lost dep1"); 10316#endif /* INVARIANTS */ 10317 dp->di_db[i] = 0; 10318 } 10319 for (i = 0; i < UFS_NIADDR; i++) { 10320#ifdef INVARIANTS 10321 if (dp->di_ib[i] != 0 && 10322 (deplist & ((1 << UFS_NDADDR) << i)) == 0) 10323 panic("initiate_write_inodeblock_ufs1: " 10324 "lost dep2"); 10325#endif /* INVARIANTS */ 10326 dp->di_ib[i] = 0; 10327 } 10328 return; 10329 } 10330 /* 10331 * If we have zero'ed out the last allocated block of the file, 10332 * roll back the size to the last currently allocated block. 10333 * We know that this last allocated block is a full-sized as 10334 * we already checked for fragments in the loop above. 10335 */ 10336 if (lastadp != NULL && 10337 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10338 for (i = lastadp->ad_offset; i >= 0; i--) 10339 if (dp->di_db[i] != 0) 10340 break; 10341 dp->di_size = (i + 1) * fs->fs_bsize; 10342 } 10343 /* 10344 * The only dependencies are for indirect blocks. 10345 * 10346 * The file size for indirect block additions is not guaranteed. 10347 * Such a guarantee would be non-trivial to achieve. The conventional 10348 * synchronous write implementation also does not make this guarantee. 10349 * Fsck should catch and fix discrepancies. Arguably, the file size 10350 * can be over-estimated without destroying integrity when the file 10351 * moves into the indirect blocks (i.e., is large). If we want to 10352 * postpone fsck, we are stuck with this argument. 10353 */ 10354 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 10355 dp->di_ib[adp->ad_offset - UFS_NDADDR] = 0; 10356} 10357 10358/* 10359 * Version of initiate_write_inodeblock that handles UFS2 dinodes. 10360 * Note that any bug fixes made to this routine must be done in the 10361 * version found above. 10362 * 10363 * Called from within the procedure above to deal with unsatisfied 10364 * allocation dependencies in an inodeblock. The buffer must be 10365 * locked, thus, no I/O completion operations can occur while we 10366 * are manipulating its associated dependencies. 10367 */ 10368static void 10369initiate_write_inodeblock_ufs2(inodedep, bp) 10370 struct inodedep *inodedep; 10371 struct buf *bp; /* The inode block */ 10372{ 10373 struct allocdirect *adp, *lastadp; 10374 struct ufs2_dinode *dp; 10375 struct ufs2_dinode *sip; 10376 struct inoref *inoref; 10377 struct ufsmount *ump; 10378 struct fs *fs; 10379 ufs_lbn_t i; 10380#ifdef INVARIANTS 10381 ufs_lbn_t prevlbn = 0; 10382#endif 10383 int deplist; 10384 10385 if (inodedep->id_state & IOSTARTED) 10386 panic("initiate_write_inodeblock_ufs2: already started"); 10387 inodedep->id_state |= IOSTARTED; 10388 fs = inodedep->id_fs; 10389 ump = VFSTOUFS(inodedep->id_list.wk_mp); 10390 LOCK_OWNED(ump); 10391 dp = (struct ufs2_dinode *)bp->b_data + 10392 ino_to_fsbo(fs, inodedep->id_ino); 10393 10394 /* 10395 * If we're on the unlinked list but have not yet written our 10396 * next pointer initialize it here. 10397 */ 10398 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) { 10399 struct inodedep *inon; 10400 10401 inon = TAILQ_NEXT(inodedep, id_unlinked); 10402 dp->di_freelink = inon ? inon->id_ino : 0; 10403 } 10404 /* 10405 * If the bitmap is not yet written, then the allocated 10406 * inode cannot be written to disk. 10407 */ 10408 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 10409 if (inodedep->id_savedino2 != NULL) 10410 panic("initiate_write_inodeblock_ufs2: I/O underway"); 10411 FREE_LOCK(ump); 10412 sip = malloc(sizeof(struct ufs2_dinode), 10413 M_SAVEDINO, M_SOFTDEP_FLAGS); 10414 ACQUIRE_LOCK(ump); 10415 inodedep->id_savedino2 = sip; 10416 *inodedep->id_savedino2 = *dp; 10417 bzero((caddr_t)dp, sizeof(struct ufs2_dinode)); 10418 dp->di_gen = inodedep->id_savedino2->di_gen; 10419 dp->di_freelink = inodedep->id_savedino2->di_freelink; 10420 return; 10421 } 10422 /* 10423 * If no dependencies, then there is nothing to roll back. 10424 */ 10425 inodedep->id_savedsize = dp->di_size; 10426 inodedep->id_savedextsize = dp->di_extsize; 10427 inodedep->id_savednlink = dp->di_nlink; 10428 if (TAILQ_EMPTY(&inodedep->id_inoupdt) && 10429 TAILQ_EMPTY(&inodedep->id_extupdt) && 10430 TAILQ_EMPTY(&inodedep->id_inoreflst)) 10431 return; 10432 /* 10433 * Revert the link count to that of the first unwritten journal entry. 10434 */ 10435 inoref = TAILQ_FIRST(&inodedep->id_inoreflst); 10436 if (inoref) 10437 dp->di_nlink = inoref->if_nlink; 10438 10439 /* 10440 * Set the ext data dependencies to busy. 10441 */ 10442 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 10443 adp = TAILQ_NEXT(adp, ad_next)) { 10444#ifdef INVARIANTS 10445 if (deplist != 0 && prevlbn >= adp->ad_offset) 10446 panic("initiate_write_inodeblock_ufs2: lbn order"); 10447 prevlbn = adp->ad_offset; 10448 if (dp->di_extb[adp->ad_offset] != adp->ad_newblkno) 10449 panic("initiate_write_inodeblock_ufs2: " 10450 "ext pointer #%jd mismatch %jd != %jd", 10451 (intmax_t)adp->ad_offset, 10452 (intmax_t)dp->di_extb[adp->ad_offset], 10453 (intmax_t)adp->ad_newblkno); 10454 deplist |= 1 << adp->ad_offset; 10455 if ((adp->ad_state & ATTACHED) == 0) 10456 panic("initiate_write_inodeblock_ufs2: Unknown " 10457 "state 0x%x", adp->ad_state); 10458#endif /* INVARIANTS */ 10459 adp->ad_state &= ~ATTACHED; 10460 adp->ad_state |= UNDONE; 10461 } 10462 /* 10463 * The on-disk inode cannot claim to be any larger than the last 10464 * fragment that has been written. Otherwise, the on-disk inode 10465 * might have fragments that were not the last block in the ext 10466 * data which would corrupt the filesystem. 10467 */ 10468 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 10469 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10470 dp->di_extb[adp->ad_offset] = adp->ad_oldblkno; 10471 /* keep going until hitting a rollback to a frag */ 10472 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10473 continue; 10474 dp->di_extsize = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10475 for (i = adp->ad_offset + 1; i < UFS_NXADDR; i++) { 10476#ifdef INVARIANTS 10477 if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0) 10478 panic("initiate_write_inodeblock_ufs2: " 10479 "lost dep1"); 10480#endif /* INVARIANTS */ 10481 dp->di_extb[i] = 0; 10482 } 10483 lastadp = NULL; 10484 break; 10485 } 10486 /* 10487 * If we have zero'ed out the last allocated block of the ext 10488 * data, roll back the size to the last currently allocated block. 10489 * We know that this last allocated block is a full-sized as 10490 * we already checked for fragments in the loop above. 10491 */ 10492 if (lastadp != NULL && 10493 dp->di_extsize <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10494 for (i = lastadp->ad_offset; i >= 0; i--) 10495 if (dp->di_extb[i] != 0) 10496 break; 10497 dp->di_extsize = (i + 1) * fs->fs_bsize; 10498 } 10499 /* 10500 * Set the file data dependencies to busy. 10501 */ 10502 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10503 adp = TAILQ_NEXT(adp, ad_next)) { 10504#ifdef INVARIANTS 10505 if (deplist != 0 && prevlbn >= adp->ad_offset) 10506 panic("softdep_write_inodeblock: lbn order"); 10507 if ((adp->ad_state & ATTACHED) == 0) 10508 panic("inodedep %p and adp %p not attached", inodedep, adp); 10509 prevlbn = adp->ad_offset; 10510 if (adp->ad_offset < UFS_NDADDR && 10511 dp->di_db[adp->ad_offset] != adp->ad_newblkno) 10512 panic("initiate_write_inodeblock_ufs2: " 10513 "direct pointer #%jd mismatch %jd != %jd", 10514 (intmax_t)adp->ad_offset, 10515 (intmax_t)dp->di_db[adp->ad_offset], 10516 (intmax_t)adp->ad_newblkno); 10517 if (adp->ad_offset >= UFS_NDADDR && 10518 dp->di_ib[adp->ad_offset - UFS_NDADDR] != adp->ad_newblkno) 10519 panic("initiate_write_inodeblock_ufs2: " 10520 "indirect pointer #%jd mismatch %jd != %jd", 10521 (intmax_t)adp->ad_offset - UFS_NDADDR, 10522 (intmax_t)dp->di_ib[adp->ad_offset - UFS_NDADDR], 10523 (intmax_t)adp->ad_newblkno); 10524 deplist |= 1 << adp->ad_offset; 10525 if ((adp->ad_state & ATTACHED) == 0) 10526 panic("initiate_write_inodeblock_ufs2: Unknown " 10527 "state 0x%x", adp->ad_state); 10528#endif /* INVARIANTS */ 10529 adp->ad_state &= ~ATTACHED; 10530 adp->ad_state |= UNDONE; 10531 } 10532 /* 10533 * The on-disk inode cannot claim to be any larger than the last 10534 * fragment that has been written. Otherwise, the on-disk inode 10535 * might have fragments that were not the last block in the file 10536 * which would corrupt the filesystem. 10537 */ 10538 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10539 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10540 if (adp->ad_offset >= UFS_NDADDR) 10541 break; 10542 dp->di_db[adp->ad_offset] = adp->ad_oldblkno; 10543 /* keep going until hitting a rollback to a frag */ 10544 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10545 continue; 10546 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10547 for (i = adp->ad_offset + 1; i < UFS_NDADDR; i++) { 10548#ifdef INVARIANTS 10549 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 10550 panic("initiate_write_inodeblock_ufs2: " 10551 "lost dep2"); 10552#endif /* INVARIANTS */ 10553 dp->di_db[i] = 0; 10554 } 10555 for (i = 0; i < UFS_NIADDR; i++) { 10556#ifdef INVARIANTS 10557 if (dp->di_ib[i] != 0 && 10558 (deplist & ((1 << UFS_NDADDR) << i)) == 0) 10559 panic("initiate_write_inodeblock_ufs2: " 10560 "lost dep3"); 10561#endif /* INVARIANTS */ 10562 dp->di_ib[i] = 0; 10563 } 10564 return; 10565 } 10566 /* 10567 * If we have zero'ed out the last allocated block of the file, 10568 * roll back the size to the last currently allocated block. 10569 * We know that this last allocated block is a full-sized as 10570 * we already checked for fragments in the loop above. 10571 */ 10572 if (lastadp != NULL && 10573 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10574 for (i = lastadp->ad_offset; i >= 0; i--) 10575 if (dp->di_db[i] != 0) 10576 break; 10577 dp->di_size = (i + 1) * fs->fs_bsize; 10578 } 10579 /* 10580 * The only dependencies are for indirect blocks. 10581 * 10582 * The file size for indirect block additions is not guaranteed. 10583 * Such a guarantee would be non-trivial to achieve. The conventional 10584 * synchronous write implementation also does not make this guarantee. 10585 * Fsck should catch and fix discrepancies. Arguably, the file size 10586 * can be over-estimated without destroying integrity when the file 10587 * moves into the indirect blocks (i.e., is large). If we want to 10588 * postpone fsck, we are stuck with this argument. 10589 */ 10590 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 10591 dp->di_ib[adp->ad_offset - UFS_NDADDR] = 0; 10592} 10593 10594/* 10595 * Cancel an indirdep as a result of truncation. Release all of the 10596 * children allocindirs and place their journal work on the appropriate 10597 * list. 10598 */ 10599static void 10600cancel_indirdep(indirdep, bp, freeblks) 10601 struct indirdep *indirdep; 10602 struct buf *bp; 10603 struct freeblks *freeblks; 10604{ 10605 struct allocindir *aip; 10606 10607 /* 10608 * None of the indirect pointers will ever be visible, 10609 * so they can simply be tossed. GOINGAWAY ensures 10610 * that allocated pointers will be saved in the buffer 10611 * cache until they are freed. Note that they will 10612 * only be able to be found by their physical address 10613 * since the inode mapping the logical address will 10614 * be gone. The save buffer used for the safe copy 10615 * was allocated in setup_allocindir_phase2 using 10616 * the physical address so it could be used for this 10617 * purpose. Hence we swap the safe copy with the real 10618 * copy, allowing the safe copy to be freed and holding 10619 * on to the real copy for later use in indir_trunc. 10620 */ 10621 if (indirdep->ir_state & GOINGAWAY) 10622 panic("cancel_indirdep: already gone"); 10623 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 10624 indirdep->ir_state |= DEPCOMPLETE; 10625 LIST_REMOVE(indirdep, ir_next); 10626 } 10627 indirdep->ir_state |= GOINGAWAY; 10628 /* 10629 * Pass in bp for blocks still have journal writes 10630 * pending so we can cancel them on their own. 10631 */ 10632 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != NULL) 10633 cancel_allocindir(aip, bp, freeblks, 0); 10634 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != NULL) 10635 cancel_allocindir(aip, NULL, freeblks, 0); 10636 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != NULL) 10637 cancel_allocindir(aip, NULL, freeblks, 0); 10638 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) 10639 cancel_allocindir(aip, NULL, freeblks, 0); 10640 /* 10641 * If there are pending partial truncations we need to keep the 10642 * old block copy around until they complete. This is because 10643 * the current b_data is not a perfect superset of the available 10644 * blocks. 10645 */ 10646 if (TAILQ_EMPTY(&indirdep->ir_trunc)) 10647 bcopy(bp->b_data, indirdep->ir_savebp->b_data, bp->b_bcount); 10648 else 10649 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 10650 WORKLIST_REMOVE(&indirdep->ir_list); 10651 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, &indirdep->ir_list); 10652 indirdep->ir_bp = NULL; 10653 indirdep->ir_freeblks = freeblks; 10654} 10655 10656/* 10657 * Free an indirdep once it no longer has new pointers to track. 10658 */ 10659static void 10660free_indirdep(indirdep) 10661 struct indirdep *indirdep; 10662{ 10663 10664 KASSERT(TAILQ_EMPTY(&indirdep->ir_trunc), 10665 ("free_indirdep: Indir trunc list not empty.")); 10666 KASSERT(LIST_EMPTY(&indirdep->ir_completehd), 10667 ("free_indirdep: Complete head not empty.")); 10668 KASSERT(LIST_EMPTY(&indirdep->ir_writehd), 10669 ("free_indirdep: write head not empty.")); 10670 KASSERT(LIST_EMPTY(&indirdep->ir_donehd), 10671 ("free_indirdep: done head not empty.")); 10672 KASSERT(LIST_EMPTY(&indirdep->ir_deplisthd), 10673 ("free_indirdep: deplist head not empty.")); 10674 KASSERT((indirdep->ir_state & DEPCOMPLETE), 10675 ("free_indirdep: %p still on newblk list.", indirdep)); 10676 KASSERT(indirdep->ir_saveddata == NULL, 10677 ("free_indirdep: %p still has saved data.", indirdep)); 10678 if (indirdep->ir_state & ONWORKLIST) 10679 WORKLIST_REMOVE(&indirdep->ir_list); 10680 WORKITEM_FREE(indirdep, D_INDIRDEP); 10681} 10682 10683/* 10684 * Called before a write to an indirdep. This routine is responsible for 10685 * rolling back pointers to a safe state which includes only those 10686 * allocindirs which have been completed. 10687 */ 10688static void 10689initiate_write_indirdep(indirdep, bp) 10690 struct indirdep *indirdep; 10691 struct buf *bp; 10692{ 10693 struct ufsmount *ump; 10694 10695 indirdep->ir_state |= IOSTARTED; 10696 if (indirdep->ir_state & GOINGAWAY) 10697 panic("disk_io_initiation: indirdep gone"); 10698 /* 10699 * If there are no remaining dependencies, this will be writing 10700 * the real pointers. 10701 */ 10702 if (LIST_EMPTY(&indirdep->ir_deplisthd) && 10703 TAILQ_EMPTY(&indirdep->ir_trunc)) 10704 return; 10705 /* 10706 * Replace up-to-date version with safe version. 10707 */ 10708 if (indirdep->ir_saveddata == NULL) { 10709 ump = VFSTOUFS(indirdep->ir_list.wk_mp); 10710 LOCK_OWNED(ump); 10711 FREE_LOCK(ump); 10712 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP, 10713 M_SOFTDEP_FLAGS); 10714 ACQUIRE_LOCK(ump); 10715 } 10716 indirdep->ir_state &= ~ATTACHED; 10717 indirdep->ir_state |= UNDONE; 10718 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 10719 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 10720 bp->b_bcount); 10721} 10722 10723/* 10724 * Called when an inode has been cleared in a cg bitmap. This finally 10725 * eliminates any canceled jaddrefs 10726 */ 10727void 10728softdep_setup_inofree(mp, bp, ino, wkhd) 10729 struct mount *mp; 10730 struct buf *bp; 10731 ino_t ino; 10732 struct workhead *wkhd; 10733{ 10734 struct worklist *wk, *wkn; 10735 struct inodedep *inodedep; 10736 struct ufsmount *ump; 10737 uint8_t *inosused; 10738 struct cg *cgp; 10739 struct fs *fs; 10740 10741 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 10742 ("softdep_setup_inofree called on non-softdep filesystem")); 10743 ump = VFSTOUFS(mp); 10744 ACQUIRE_LOCK(ump); 10745 fs = ump->um_fs; 10746 cgp = (struct cg *)bp->b_data; 10747 inosused = cg_inosused(cgp); 10748 if (isset(inosused, ino % fs->fs_ipg)) 10749 panic("softdep_setup_inofree: inode %ju not freed.", 10750 (uintmax_t)ino); 10751 if (inodedep_lookup(mp, ino, 0, &inodedep)) 10752 panic("softdep_setup_inofree: ino %ju has existing inodedep %p", 10753 (uintmax_t)ino, inodedep); 10754 if (wkhd) { 10755 LIST_FOREACH_SAFE(wk, wkhd, wk_list, wkn) { 10756 if (wk->wk_type != D_JADDREF) 10757 continue; 10758 WORKLIST_REMOVE(wk); 10759 /* 10760 * We can free immediately even if the jaddref 10761 * isn't attached in a background write as now 10762 * the bitmaps are reconciled. 10763 */ 10764 wk->wk_state |= COMPLETE | ATTACHED; 10765 free_jaddref(WK_JADDREF(wk)); 10766 } 10767 jwork_move(&bp->b_dep, wkhd); 10768 } 10769 FREE_LOCK(ump); 10770} 10771 10772 10773/* 10774 * Called via ffs_blkfree() after a set of frags has been cleared from a cg 10775 * map. Any dependencies waiting for the write to clear are added to the 10776 * buf's list and any jnewblks that are being canceled are discarded 10777 * immediately. 10778 */ 10779void 10780softdep_setup_blkfree(mp, bp, blkno, frags, wkhd) 10781 struct mount *mp; 10782 struct buf *bp; 10783 ufs2_daddr_t blkno; 10784 int frags; 10785 struct workhead *wkhd; 10786{ 10787 struct bmsafemap *bmsafemap; 10788 struct jnewblk *jnewblk; 10789 struct ufsmount *ump; 10790 struct worklist *wk; 10791 struct fs *fs; 10792#ifdef SUJ_DEBUG 10793 uint8_t *blksfree; 10794 struct cg *cgp; 10795 ufs2_daddr_t jstart; 10796 ufs2_daddr_t jend; 10797 ufs2_daddr_t end; 10798 long bno; 10799 int i; 10800#endif 10801 10802 CTR3(KTR_SUJ, 10803 "softdep_setup_blkfree: blkno %jd frags %d wk head %p", 10804 blkno, frags, wkhd); 10805 10806 ump = VFSTOUFS(mp); 10807 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 10808 ("softdep_setup_blkfree called on non-softdep filesystem")); 10809 ACQUIRE_LOCK(ump); 10810 /* Lookup the bmsafemap so we track when it is dirty. */ 10811 fs = ump->um_fs; 10812 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL); 10813 /* 10814 * Detach any jnewblks which have been canceled. They must linger 10815 * until the bitmap is cleared again by ffs_blkfree() to prevent 10816 * an unjournaled allocation from hitting the disk. 10817 */ 10818 if (wkhd) { 10819 while ((wk = LIST_FIRST(wkhd)) != NULL) { 10820 CTR2(KTR_SUJ, 10821 "softdep_setup_blkfree: blkno %jd wk type %d", 10822 blkno, wk->wk_type); 10823 WORKLIST_REMOVE(wk); 10824 if (wk->wk_type != D_JNEWBLK) { 10825 WORKLIST_INSERT(&bmsafemap->sm_freehd, wk); 10826 continue; 10827 } 10828 jnewblk = WK_JNEWBLK(wk); 10829 KASSERT(jnewblk->jn_state & GOINGAWAY, 10830 ("softdep_setup_blkfree: jnewblk not canceled.")); 10831#ifdef SUJ_DEBUG 10832 /* 10833 * Assert that this block is free in the bitmap 10834 * before we discard the jnewblk. 10835 */ 10836 cgp = (struct cg *)bp->b_data; 10837 blksfree = cg_blksfree(cgp); 10838 bno = dtogd(fs, jnewblk->jn_blkno); 10839 for (i = jnewblk->jn_oldfrags; 10840 i < jnewblk->jn_frags; i++) { 10841 if (isset(blksfree, bno + i)) 10842 continue; 10843 panic("softdep_setup_blkfree: not free"); 10844 } 10845#endif 10846 /* 10847 * Even if it's not attached we can free immediately 10848 * as the new bitmap is correct. 10849 */ 10850 wk->wk_state |= COMPLETE | ATTACHED; 10851 free_jnewblk(jnewblk); 10852 } 10853 } 10854 10855#ifdef SUJ_DEBUG 10856 /* 10857 * Assert that we are not freeing a block which has an outstanding 10858 * allocation dependency. 10859 */ 10860 fs = VFSTOUFS(mp)->um_fs; 10861 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL); 10862 end = blkno + frags; 10863 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) { 10864 /* 10865 * Don't match against blocks that will be freed when the 10866 * background write is done. 10867 */ 10868 if ((jnewblk->jn_state & (ATTACHED | COMPLETE | DEPCOMPLETE)) == 10869 (COMPLETE | DEPCOMPLETE)) 10870 continue; 10871 jstart = jnewblk->jn_blkno + jnewblk->jn_oldfrags; 10872 jend = jnewblk->jn_blkno + jnewblk->jn_frags; 10873 if ((blkno >= jstart && blkno < jend) || 10874 (end > jstart && end <= jend)) { 10875 printf("state 0x%X %jd - %d %d dep %p\n", 10876 jnewblk->jn_state, jnewblk->jn_blkno, 10877 jnewblk->jn_oldfrags, jnewblk->jn_frags, 10878 jnewblk->jn_dep); 10879 panic("softdep_setup_blkfree: " 10880 "%jd-%jd(%d) overlaps with %jd-%jd", 10881 blkno, end, frags, jstart, jend); 10882 } 10883 } 10884#endif 10885 FREE_LOCK(ump); 10886} 10887 10888/* 10889 * Revert a block allocation when the journal record that describes it 10890 * is not yet written. 10891 */ 10892static int 10893jnewblk_rollback(jnewblk, fs, cgp, blksfree) 10894 struct jnewblk *jnewblk; 10895 struct fs *fs; 10896 struct cg *cgp; 10897 uint8_t *blksfree; 10898{ 10899 ufs1_daddr_t fragno; 10900 long cgbno, bbase; 10901 int frags, blk; 10902 int i; 10903 10904 frags = 0; 10905 cgbno = dtogd(fs, jnewblk->jn_blkno); 10906 /* 10907 * We have to test which frags need to be rolled back. We may 10908 * be operating on a stale copy when doing background writes. 10909 */ 10910 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) 10911 if (isclr(blksfree, cgbno + i)) 10912 frags++; 10913 if (frags == 0) 10914 return (0); 10915 /* 10916 * This is mostly ffs_blkfree() sans some validation and 10917 * superblock updates. 10918 */ 10919 if (frags == fs->fs_frag) { 10920 fragno = fragstoblks(fs, cgbno); 10921 ffs_setblock(fs, blksfree, fragno); 10922 ffs_clusteracct(fs, cgp, fragno, 1); 10923 cgp->cg_cs.cs_nbfree++; 10924 } else { 10925 cgbno += jnewblk->jn_oldfrags; 10926 bbase = cgbno - fragnum(fs, cgbno); 10927 /* Decrement the old frags. */ 10928 blk = blkmap(fs, blksfree, bbase); 10929 ffs_fragacct(fs, blk, cgp->cg_frsum, -1); 10930 /* Deallocate the fragment */ 10931 for (i = 0; i < frags; i++) 10932 setbit(blksfree, cgbno + i); 10933 cgp->cg_cs.cs_nffree += frags; 10934 /* Add back in counts associated with the new frags */ 10935 blk = blkmap(fs, blksfree, bbase); 10936 ffs_fragacct(fs, blk, cgp->cg_frsum, 1); 10937 /* If a complete block has been reassembled, account for it. */ 10938 fragno = fragstoblks(fs, bbase); 10939 if (ffs_isblock(fs, blksfree, fragno)) { 10940 cgp->cg_cs.cs_nffree -= fs->fs_frag; 10941 ffs_clusteracct(fs, cgp, fragno, 1); 10942 cgp->cg_cs.cs_nbfree++; 10943 } 10944 } 10945 stat_jnewblk++; 10946 jnewblk->jn_state &= ~ATTACHED; 10947 jnewblk->jn_state |= UNDONE; 10948 10949 return (frags); 10950} 10951 10952static void 10953initiate_write_bmsafemap(bmsafemap, bp) 10954 struct bmsafemap *bmsafemap; 10955 struct buf *bp; /* The cg block. */ 10956{ 10957 struct jaddref *jaddref; 10958 struct jnewblk *jnewblk; 10959 uint8_t *inosused; 10960 uint8_t *blksfree; 10961 struct cg *cgp; 10962 struct fs *fs; 10963 ino_t ino; 10964 10965 /* 10966 * If this is a background write, we did this at the time that 10967 * the copy was made, so do not need to do it again. 10968 */ 10969 if (bmsafemap->sm_state & IOSTARTED) 10970 return; 10971 bmsafemap->sm_state |= IOSTARTED; 10972 /* 10973 * Clear any inode allocations which are pending journal writes. 10974 */ 10975 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd) != NULL) { 10976 cgp = (struct cg *)bp->b_data; 10977 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 10978 inosused = cg_inosused(cgp); 10979 LIST_FOREACH(jaddref, &bmsafemap->sm_jaddrefhd, ja_bmdeps) { 10980 ino = jaddref->ja_ino % fs->fs_ipg; 10981 if (isset(inosused, ino)) { 10982 if ((jaddref->ja_mode & IFMT) == IFDIR) 10983 cgp->cg_cs.cs_ndir--; 10984 cgp->cg_cs.cs_nifree++; 10985 clrbit(inosused, ino); 10986 jaddref->ja_state &= ~ATTACHED; 10987 jaddref->ja_state |= UNDONE; 10988 stat_jaddref++; 10989 } else 10990 panic("initiate_write_bmsafemap: inode %ju " 10991 "marked free", (uintmax_t)jaddref->ja_ino); 10992 } 10993 } 10994 /* 10995 * Clear any block allocations which are pending journal writes. 10996 */ 10997 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) { 10998 cgp = (struct cg *)bp->b_data; 10999 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 11000 blksfree = cg_blksfree(cgp); 11001 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) { 11002 if (jnewblk_rollback(jnewblk, fs, cgp, blksfree)) 11003 continue; 11004 panic("initiate_write_bmsafemap: block %jd " 11005 "marked free", jnewblk->jn_blkno); 11006 } 11007 } 11008 /* 11009 * Move allocation lists to the written lists so they can be 11010 * cleared once the block write is complete. 11011 */ 11012 LIST_SWAP(&bmsafemap->sm_inodedephd, &bmsafemap->sm_inodedepwr, 11013 inodedep, id_deps); 11014 LIST_SWAP(&bmsafemap->sm_newblkhd, &bmsafemap->sm_newblkwr, 11015 newblk, nb_deps); 11016 LIST_SWAP(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, worklist, 11017 wk_list); 11018} 11019 11020/* 11021 * This routine is called during the completion interrupt 11022 * service routine for a disk write (from the procedure called 11023 * by the device driver to inform the filesystem caches of 11024 * a request completion). It should be called early in this 11025 * procedure, before the block is made available to other 11026 * processes or other routines are called. 11027 * 11028 */ 11029static void 11030softdep_disk_write_complete(bp) 11031 struct buf *bp; /* describes the completed disk write */ 11032{ 11033 struct worklist *wk; 11034 struct worklist *owk; 11035 struct ufsmount *ump; 11036 struct workhead reattach; 11037 struct freeblks *freeblks; 11038 struct buf *sbp; 11039 11040 ump = softdep_bp_to_mp(bp); 11041 if (ump == NULL) 11042 return; 11043 11044 sbp = NULL; 11045 11046 /* 11047 * If an error occurred while doing the write, then the data 11048 * has not hit the disk and the dependencies cannot be processed. 11049 * But we do have to go through and roll forward any dependencies 11050 * that were rolled back before the disk write. 11051 */ 11052 ACQUIRE_LOCK(ump); 11053 if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0) { 11054 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 11055 switch (wk->wk_type) { 11056 11057 case D_PAGEDEP: 11058 handle_written_filepage(WK_PAGEDEP(wk), bp, 0); 11059 continue; 11060 11061 case D_INODEDEP: 11062 handle_written_inodeblock(WK_INODEDEP(wk), 11063 bp, 0); 11064 continue; 11065 11066 case D_BMSAFEMAP: 11067 handle_written_bmsafemap(WK_BMSAFEMAP(wk), 11068 bp, 0); 11069 continue; 11070 11071 case D_INDIRDEP: 11072 handle_written_indirdep(WK_INDIRDEP(wk), 11073 bp, &sbp, 0); 11074 continue; 11075 default: 11076 /* nothing to roll forward */ 11077 continue; 11078 } 11079 } 11080 FREE_LOCK(ump); 11081 return; 11082 } 11083 LIST_INIT(&reattach); 11084 11085 /* 11086 * Ump SU lock must not be released anywhere in this code segment. 11087 */ 11088 owk = NULL; 11089 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 11090 WORKLIST_REMOVE(wk); 11091 atomic_add_long(&dep_write[wk->wk_type], 1); 11092 if (wk == owk) 11093 panic("duplicate worklist: %p\n", wk); 11094 owk = wk; 11095 switch (wk->wk_type) { 11096 11097 case D_PAGEDEP: 11098 if (handle_written_filepage(WK_PAGEDEP(wk), bp, 11099 WRITESUCCEEDED)) 11100 WORKLIST_INSERT(&reattach, wk); 11101 continue; 11102 11103 case D_INODEDEP: 11104 if (handle_written_inodeblock(WK_INODEDEP(wk), bp, 11105 WRITESUCCEEDED)) 11106 WORKLIST_INSERT(&reattach, wk); 11107 continue; 11108 11109 case D_BMSAFEMAP: 11110 if (handle_written_bmsafemap(WK_BMSAFEMAP(wk), bp, 11111 WRITESUCCEEDED)) 11112 WORKLIST_INSERT(&reattach, wk); 11113 continue; 11114 11115 case D_MKDIR: 11116 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 11117 continue; 11118 11119 case D_ALLOCDIRECT: 11120 wk->wk_state |= COMPLETE; 11121 handle_allocdirect_partdone(WK_ALLOCDIRECT(wk), NULL); 11122 continue; 11123 11124 case D_ALLOCINDIR: 11125 wk->wk_state |= COMPLETE; 11126 handle_allocindir_partdone(WK_ALLOCINDIR(wk)); 11127 continue; 11128 11129 case D_INDIRDEP: 11130 if (handle_written_indirdep(WK_INDIRDEP(wk), bp, &sbp, 11131 WRITESUCCEEDED)) 11132 WORKLIST_INSERT(&reattach, wk); 11133 continue; 11134 11135 case D_FREEBLKS: 11136 wk->wk_state |= COMPLETE; 11137 freeblks = WK_FREEBLKS(wk); 11138 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE && 11139 LIST_EMPTY(&freeblks->fb_jblkdephd)) 11140 add_to_worklist(wk, WK_NODELAY); 11141 continue; 11142 11143 case D_FREEWORK: 11144 handle_written_freework(WK_FREEWORK(wk)); 11145 break; 11146 11147 case D_JSEGDEP: 11148 free_jsegdep(WK_JSEGDEP(wk)); 11149 continue; 11150 11151 case D_JSEG: 11152 handle_written_jseg(WK_JSEG(wk), bp); 11153 continue; 11154 11155 case D_SBDEP: 11156 if (handle_written_sbdep(WK_SBDEP(wk), bp)) 11157 WORKLIST_INSERT(&reattach, wk); 11158 continue; 11159 11160 case D_FREEDEP: 11161 free_freedep(WK_FREEDEP(wk)); 11162 continue; 11163 11164 default: 11165 panic("handle_disk_write_complete: Unknown type %s", 11166 TYPENAME(wk->wk_type)); 11167 /* NOTREACHED */ 11168 } 11169 } 11170 /* 11171 * Reattach any requests that must be redone. 11172 */ 11173 while ((wk = LIST_FIRST(&reattach)) != NULL) { 11174 WORKLIST_REMOVE(wk); 11175 WORKLIST_INSERT(&bp->b_dep, wk); 11176 } 11177 FREE_LOCK(ump); 11178 if (sbp) 11179 brelse(sbp); 11180} 11181 11182/* 11183 * Called from within softdep_disk_write_complete above. 11184 */ 11185static void 11186handle_allocdirect_partdone(adp, wkhd) 11187 struct allocdirect *adp; /* the completed allocdirect */ 11188 struct workhead *wkhd; /* Work to do when inode is writtne. */ 11189{ 11190 struct allocdirectlst *listhead; 11191 struct allocdirect *listadp; 11192 struct inodedep *inodedep; 11193 long bsize; 11194 11195 LOCK_OWNED(VFSTOUFS(adp->ad_block.nb_list.wk_mp)); 11196 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 11197 return; 11198 /* 11199 * The on-disk inode cannot claim to be any larger than the last 11200 * fragment that has been written. Otherwise, the on-disk inode 11201 * might have fragments that were not the last block in the file 11202 * which would corrupt the filesystem. Thus, we cannot free any 11203 * allocdirects after one whose ad_oldblkno claims a fragment as 11204 * these blocks must be rolled back to zero before writing the inode. 11205 * We check the currently active set of allocdirects in id_inoupdt 11206 * or id_extupdt as appropriate. 11207 */ 11208 inodedep = adp->ad_inodedep; 11209 bsize = inodedep->id_fs->fs_bsize; 11210 if (adp->ad_state & EXTDATA) 11211 listhead = &inodedep->id_extupdt; 11212 else 11213 listhead = &inodedep->id_inoupdt; 11214 TAILQ_FOREACH(listadp, listhead, ad_next) { 11215 /* found our block */ 11216 if (listadp == adp) 11217 break; 11218 /* continue if ad_oldlbn is not a fragment */ 11219 if (listadp->ad_oldsize == 0 || 11220 listadp->ad_oldsize == bsize) 11221 continue; 11222 /* hit a fragment */ 11223 return; 11224 } 11225 /* 11226 * If we have reached the end of the current list without 11227 * finding the just finished dependency, then it must be 11228 * on the future dependency list. Future dependencies cannot 11229 * be freed until they are moved to the current list. 11230 */ 11231 if (listadp == NULL) { 11232#ifdef DEBUG 11233 if (adp->ad_state & EXTDATA) 11234 listhead = &inodedep->id_newextupdt; 11235 else 11236 listhead = &inodedep->id_newinoupdt; 11237 TAILQ_FOREACH(listadp, listhead, ad_next) 11238 /* found our block */ 11239 if (listadp == adp) 11240 break; 11241 if (listadp == NULL) 11242 panic("handle_allocdirect_partdone: lost dep"); 11243#endif /* DEBUG */ 11244 return; 11245 } 11246 /* 11247 * If we have found the just finished dependency, then queue 11248 * it along with anything that follows it that is complete. 11249 * Since the pointer has not yet been written in the inode 11250 * as the dependency prevents it, place the allocdirect on the 11251 * bufwait list where it will be freed once the pointer is 11252 * valid. 11253 */ 11254 if (wkhd == NULL) 11255 wkhd = &inodedep->id_bufwait; 11256 for (; adp; adp = listadp) { 11257 listadp = TAILQ_NEXT(adp, ad_next); 11258 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 11259 return; 11260 TAILQ_REMOVE(listhead, adp, ad_next); 11261 WORKLIST_INSERT(wkhd, &adp->ad_block.nb_list); 11262 } 11263} 11264 11265/* 11266 * Called from within softdep_disk_write_complete above. This routine 11267 * completes successfully written allocindirs. 11268 */ 11269static void 11270handle_allocindir_partdone(aip) 11271 struct allocindir *aip; /* the completed allocindir */ 11272{ 11273 struct indirdep *indirdep; 11274 11275 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 11276 return; 11277 indirdep = aip->ai_indirdep; 11278 LIST_REMOVE(aip, ai_next); 11279 /* 11280 * Don't set a pointer while the buffer is undergoing IO or while 11281 * we have active truncations. 11282 */ 11283 if (indirdep->ir_state & UNDONE || !TAILQ_EMPTY(&indirdep->ir_trunc)) { 11284 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 11285 return; 11286 } 11287 if (indirdep->ir_state & UFS1FMT) 11288 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 11289 aip->ai_newblkno; 11290 else 11291 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 11292 aip->ai_newblkno; 11293 /* 11294 * Await the pointer write before freeing the allocindir. 11295 */ 11296 LIST_INSERT_HEAD(&indirdep->ir_writehd, aip, ai_next); 11297} 11298 11299/* 11300 * Release segments held on a jwork list. 11301 */ 11302static void 11303handle_jwork(wkhd) 11304 struct workhead *wkhd; 11305{ 11306 struct worklist *wk; 11307 11308 while ((wk = LIST_FIRST(wkhd)) != NULL) { 11309 WORKLIST_REMOVE(wk); 11310 switch (wk->wk_type) { 11311 case D_JSEGDEP: 11312 free_jsegdep(WK_JSEGDEP(wk)); 11313 continue; 11314 case D_FREEDEP: 11315 free_freedep(WK_FREEDEP(wk)); 11316 continue; 11317 case D_FREEFRAG: 11318 rele_jseg(WK_JSEG(WK_FREEFRAG(wk)->ff_jdep)); 11319 WORKITEM_FREE(wk, D_FREEFRAG); 11320 continue; 11321 case D_FREEWORK: 11322 handle_written_freework(WK_FREEWORK(wk)); 11323 continue; 11324 default: 11325 panic("handle_jwork: Unknown type %s\n", 11326 TYPENAME(wk->wk_type)); 11327 } 11328 } 11329} 11330 11331/* 11332 * Handle the bufwait list on an inode when it is safe to release items 11333 * held there. This normally happens after an inode block is written but 11334 * may be delayed and handled later if there are pending journal items that 11335 * are not yet safe to be released. 11336 */ 11337static struct freefile * 11338handle_bufwait(inodedep, refhd) 11339 struct inodedep *inodedep; 11340 struct workhead *refhd; 11341{ 11342 struct jaddref *jaddref; 11343 struct freefile *freefile; 11344 struct worklist *wk; 11345 11346 freefile = NULL; 11347 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 11348 WORKLIST_REMOVE(wk); 11349 switch (wk->wk_type) { 11350 case D_FREEFILE: 11351 /* 11352 * We defer adding freefile to the worklist 11353 * until all other additions have been made to 11354 * ensure that it will be done after all the 11355 * old blocks have been freed. 11356 */ 11357 if (freefile != NULL) 11358 panic("handle_bufwait: freefile"); 11359 freefile = WK_FREEFILE(wk); 11360 continue; 11361 11362 case D_MKDIR: 11363 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 11364 continue; 11365 11366 case D_DIRADD: 11367 diradd_inode_written(WK_DIRADD(wk), inodedep); 11368 continue; 11369 11370 case D_FREEFRAG: 11371 wk->wk_state |= COMPLETE; 11372 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE) 11373 add_to_worklist(wk, 0); 11374 continue; 11375 11376 case D_DIRREM: 11377 wk->wk_state |= COMPLETE; 11378 add_to_worklist(wk, 0); 11379 continue; 11380 11381 case D_ALLOCDIRECT: 11382 case D_ALLOCINDIR: 11383 free_newblk(WK_NEWBLK(wk)); 11384 continue; 11385 11386 case D_JNEWBLK: 11387 wk->wk_state |= COMPLETE; 11388 free_jnewblk(WK_JNEWBLK(wk)); 11389 continue; 11390 11391 /* 11392 * Save freed journal segments and add references on 11393 * the supplied list which will delay their release 11394 * until the cg bitmap is cleared on disk. 11395 */ 11396 case D_JSEGDEP: 11397 if (refhd == NULL) 11398 free_jsegdep(WK_JSEGDEP(wk)); 11399 else 11400 WORKLIST_INSERT(refhd, wk); 11401 continue; 11402 11403 case D_JADDREF: 11404 jaddref = WK_JADDREF(wk); 11405 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, 11406 if_deps); 11407 /* 11408 * Transfer any jaddrefs to the list to be freed with 11409 * the bitmap if we're handling a removed file. 11410 */ 11411 if (refhd == NULL) { 11412 wk->wk_state |= COMPLETE; 11413 free_jaddref(jaddref); 11414 } else 11415 WORKLIST_INSERT(refhd, wk); 11416 continue; 11417 11418 default: 11419 panic("handle_bufwait: Unknown type %p(%s)", 11420 wk, TYPENAME(wk->wk_type)); 11421 /* NOTREACHED */ 11422 } 11423 } 11424 return (freefile); 11425} 11426/* 11427 * Called from within softdep_disk_write_complete above to restore 11428 * in-memory inode block contents to their most up-to-date state. Note 11429 * that this routine is always called from interrupt level with further 11430 * interrupts from this device blocked. 11431 * 11432 * If the write did not succeed, we will do all the roll-forward 11433 * operations, but we will not take the actions that will allow its 11434 * dependencies to be processed. 11435 */ 11436static int 11437handle_written_inodeblock(inodedep, bp, flags) 11438 struct inodedep *inodedep; 11439 struct buf *bp; /* buffer containing the inode block */ 11440 int flags; 11441{ 11442 struct freefile *freefile; 11443 struct allocdirect *adp, *nextadp; 11444 struct ufs1_dinode *dp1 = NULL; 11445 struct ufs2_dinode *dp2 = NULL; 11446 struct workhead wkhd; 11447 int hadchanges, fstype; 11448 ino_t freelink; 11449 11450 LIST_INIT(&wkhd); 11451 hadchanges = 0; 11452 freefile = NULL; 11453 if ((inodedep->id_state & IOSTARTED) == 0) 11454 panic("handle_written_inodeblock: not started"); 11455 inodedep->id_state &= ~IOSTARTED; 11456 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) { 11457 fstype = UFS1; 11458 dp1 = (struct ufs1_dinode *)bp->b_data + 11459 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 11460 freelink = dp1->di_freelink; 11461 } else { 11462 fstype = UFS2; 11463 dp2 = (struct ufs2_dinode *)bp->b_data + 11464 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 11465 freelink = dp2->di_freelink; 11466 } 11467 /* 11468 * Leave this inodeblock dirty until it's in the list. 11469 */ 11470 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) == UNLINKED && 11471 (flags & WRITESUCCEEDED)) { 11472 struct inodedep *inon; 11473 11474 inon = TAILQ_NEXT(inodedep, id_unlinked); 11475 if ((inon == NULL && freelink == 0) || 11476 (inon && inon->id_ino == freelink)) { 11477 if (inon) 11478 inon->id_state |= UNLINKPREV; 11479 inodedep->id_state |= UNLINKNEXT; 11480 } 11481 hadchanges = 1; 11482 } 11483 /* 11484 * If we had to rollback the inode allocation because of 11485 * bitmaps being incomplete, then simply restore it. 11486 * Keep the block dirty so that it will not be reclaimed until 11487 * all associated dependencies have been cleared and the 11488 * corresponding updates written to disk. 11489 */ 11490 if (inodedep->id_savedino1 != NULL) { 11491 hadchanges = 1; 11492 if (fstype == UFS1) 11493 *dp1 = *inodedep->id_savedino1; 11494 else 11495 *dp2 = *inodedep->id_savedino2; 11496 free(inodedep->id_savedino1, M_SAVEDINO); 11497 inodedep->id_savedino1 = NULL; 11498 if ((bp->b_flags & B_DELWRI) == 0) 11499 stat_inode_bitmap++; 11500 bdirty(bp); 11501 /* 11502 * If the inode is clear here and GOINGAWAY it will never 11503 * be written. Process the bufwait and clear any pending 11504 * work which may include the freefile. 11505 */ 11506 if (inodedep->id_state & GOINGAWAY) 11507 goto bufwait; 11508 return (1); 11509 } 11510 if (flags & WRITESUCCEEDED) 11511 inodedep->id_state |= COMPLETE; 11512 /* 11513 * Roll forward anything that had to be rolled back before 11514 * the inode could be updated. 11515 */ 11516 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 11517 nextadp = TAILQ_NEXT(adp, ad_next); 11518 if (adp->ad_state & ATTACHED) 11519 panic("handle_written_inodeblock: new entry"); 11520 if (fstype == UFS1) { 11521 if (adp->ad_offset < UFS_NDADDR) { 11522 if (dp1->di_db[adp->ad_offset]!=adp->ad_oldblkno) 11523 panic("%s %s #%jd mismatch %d != %jd", 11524 "handle_written_inodeblock:", 11525 "direct pointer", 11526 (intmax_t)adp->ad_offset, 11527 dp1->di_db[adp->ad_offset], 11528 (intmax_t)adp->ad_oldblkno); 11529 dp1->di_db[adp->ad_offset] = adp->ad_newblkno; 11530 } else { 11531 if (dp1->di_ib[adp->ad_offset - UFS_NDADDR] != 11532 0) 11533 panic("%s: %s #%jd allocated as %d", 11534 "handle_written_inodeblock", 11535 "indirect pointer", 11536 (intmax_t)adp->ad_offset - 11537 UFS_NDADDR, 11538 dp1->di_ib[adp->ad_offset - 11539 UFS_NDADDR]); 11540 dp1->di_ib[adp->ad_offset - UFS_NDADDR] = 11541 adp->ad_newblkno; 11542 } 11543 } else { 11544 if (adp->ad_offset < UFS_NDADDR) { 11545 if (dp2->di_db[adp->ad_offset]!=adp->ad_oldblkno) 11546 panic("%s: %s #%jd %s %jd != %jd", 11547 "handle_written_inodeblock", 11548 "direct pointer", 11549 (intmax_t)adp->ad_offset, "mismatch", 11550 (intmax_t)dp2->di_db[adp->ad_offset], 11551 (intmax_t)adp->ad_oldblkno); 11552 dp2->di_db[adp->ad_offset] = adp->ad_newblkno; 11553 } else { 11554 if (dp2->di_ib[adp->ad_offset - UFS_NDADDR] != 11555 0) 11556 panic("%s: %s #%jd allocated as %jd", 11557 "handle_written_inodeblock", 11558 "indirect pointer", 11559 (intmax_t)adp->ad_offset - 11560 UFS_NDADDR, 11561 (intmax_t) 11562 dp2->di_ib[adp->ad_offset - 11563 UFS_NDADDR]); 11564 dp2->di_ib[adp->ad_offset - UFS_NDADDR] = 11565 adp->ad_newblkno; 11566 } 11567 } 11568 adp->ad_state &= ~UNDONE; 11569 adp->ad_state |= ATTACHED; 11570 hadchanges = 1; 11571 } 11572 for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) { 11573 nextadp = TAILQ_NEXT(adp, ad_next); 11574 if (adp->ad_state & ATTACHED) 11575 panic("handle_written_inodeblock: new entry"); 11576 if (dp2->di_extb[adp->ad_offset] != adp->ad_oldblkno) 11577 panic("%s: direct pointers #%jd %s %jd != %jd", 11578 "handle_written_inodeblock", 11579 (intmax_t)adp->ad_offset, "mismatch", 11580 (intmax_t)dp2->di_extb[adp->ad_offset], 11581 (intmax_t)adp->ad_oldblkno); 11582 dp2->di_extb[adp->ad_offset] = adp->ad_newblkno; 11583 adp->ad_state &= ~UNDONE; 11584 adp->ad_state |= ATTACHED; 11585 hadchanges = 1; 11586 } 11587 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 11588 stat_direct_blk_ptrs++; 11589 /* 11590 * Reset the file size to its most up-to-date value. 11591 */ 11592 if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1) 11593 panic("handle_written_inodeblock: bad size"); 11594 if (inodedep->id_savednlink > UFS_LINK_MAX) 11595 panic("handle_written_inodeblock: Invalid link count " 11596 "%jd for inodedep %p", (uintmax_t)inodedep->id_savednlink, 11597 inodedep); 11598 if (fstype == UFS1) { 11599 if (dp1->di_nlink != inodedep->id_savednlink) { 11600 dp1->di_nlink = inodedep->id_savednlink; 11601 hadchanges = 1; 11602 } 11603 if (dp1->di_size != inodedep->id_savedsize) { 11604 dp1->di_size = inodedep->id_savedsize; 11605 hadchanges = 1; 11606 } 11607 } else { 11608 if (dp2->di_nlink != inodedep->id_savednlink) { 11609 dp2->di_nlink = inodedep->id_savednlink; 11610 hadchanges = 1; 11611 } 11612 if (dp2->di_size != inodedep->id_savedsize) { 11613 dp2->di_size = inodedep->id_savedsize; 11614 hadchanges = 1; 11615 } 11616 if (dp2->di_extsize != inodedep->id_savedextsize) { 11617 dp2->di_extsize = inodedep->id_savedextsize; 11618 hadchanges = 1; 11619 } 11620 } 11621 inodedep->id_savedsize = -1; 11622 inodedep->id_savedextsize = -1; 11623 inodedep->id_savednlink = -1; 11624 /* 11625 * If there were any rollbacks in the inode block, then it must be 11626 * marked dirty so that its will eventually get written back in 11627 * its correct form. 11628 */ 11629 if (hadchanges) 11630 bdirty(bp); 11631bufwait: 11632 /* 11633 * If the write did not succeed, we have done all the roll-forward 11634 * operations, but we cannot take the actions that will allow its 11635 * dependencies to be processed. 11636 */ 11637 if ((flags & WRITESUCCEEDED) == 0) 11638 return (hadchanges); 11639 /* 11640 * Process any allocdirects that completed during the update. 11641 */ 11642 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 11643 handle_allocdirect_partdone(adp, &wkhd); 11644 if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 11645 handle_allocdirect_partdone(adp, &wkhd); 11646 /* 11647 * Process deallocations that were held pending until the 11648 * inode had been written to disk. Freeing of the inode 11649 * is delayed until after all blocks have been freed to 11650 * avoid creation of new <vfsid, inum, lbn> triples 11651 * before the old ones have been deleted. Completely 11652 * unlinked inodes are not processed until the unlinked 11653 * inode list is written or the last reference is removed. 11654 */ 11655 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) != UNLINKED) { 11656 freefile = handle_bufwait(inodedep, NULL); 11657 if (freefile && !LIST_EMPTY(&wkhd)) { 11658 WORKLIST_INSERT(&wkhd, &freefile->fx_list); 11659 freefile = NULL; 11660 } 11661 } 11662 /* 11663 * Move rolled forward dependency completions to the bufwait list 11664 * now that those that were already written have been processed. 11665 */ 11666 if (!LIST_EMPTY(&wkhd) && hadchanges == 0) 11667 panic("handle_written_inodeblock: bufwait but no changes"); 11668 jwork_move(&inodedep->id_bufwait, &wkhd); 11669 11670 if (freefile != NULL) { 11671 /* 11672 * If the inode is goingaway it was never written. Fake up 11673 * the state here so free_inodedep() can succeed. 11674 */ 11675 if (inodedep->id_state & GOINGAWAY) 11676 inodedep->id_state |= COMPLETE | DEPCOMPLETE; 11677 if (free_inodedep(inodedep) == 0) 11678 panic("handle_written_inodeblock: live inodedep %p", 11679 inodedep); 11680 add_to_worklist(&freefile->fx_list, 0); 11681 return (0); 11682 } 11683 11684 /* 11685 * If no outstanding dependencies, free it. 11686 */ 11687 if (free_inodedep(inodedep) || 11688 (TAILQ_FIRST(&inodedep->id_inoreflst) == 0 && 11689 TAILQ_FIRST(&inodedep->id_inoupdt) == 0 && 11690 TAILQ_FIRST(&inodedep->id_extupdt) == 0 && 11691 LIST_FIRST(&inodedep->id_bufwait) == 0)) 11692 return (0); 11693 return (hadchanges); 11694} 11695 11696/* 11697 * Perform needed roll-forwards and kick off any dependencies that 11698 * can now be processed. 11699 * 11700 * If the write did not succeed, we will do all the roll-forward 11701 * operations, but we will not take the actions that will allow its 11702 * dependencies to be processed. 11703 */ 11704static int 11705handle_written_indirdep(indirdep, bp, bpp, flags) 11706 struct indirdep *indirdep; 11707 struct buf *bp; 11708 struct buf **bpp; 11709 int flags; 11710{ 11711 struct allocindir *aip; 11712 struct buf *sbp; 11713 int chgs; 11714 11715 if (indirdep->ir_state & GOINGAWAY) 11716 panic("handle_written_indirdep: indirdep gone"); 11717 if ((indirdep->ir_state & IOSTARTED) == 0) 11718 panic("handle_written_indirdep: IO not started"); 11719 chgs = 0; 11720 /* 11721 * If there were rollbacks revert them here. 11722 */ 11723 if (indirdep->ir_saveddata) { 11724 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 11725 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 11726 free(indirdep->ir_saveddata, M_INDIRDEP); 11727 indirdep->ir_saveddata = NULL; 11728 } 11729 chgs = 1; 11730 } 11731 indirdep->ir_state &= ~(UNDONE | IOSTARTED); 11732 indirdep->ir_state |= ATTACHED; 11733 /* 11734 * If the write did not succeed, we have done all the roll-forward 11735 * operations, but we cannot take the actions that will allow its 11736 * dependencies to be processed. 11737 */ 11738 if ((flags & WRITESUCCEEDED) == 0) { 11739 stat_indir_blk_ptrs++; 11740 bdirty(bp); 11741 return (1); 11742 } 11743 /* 11744 * Move allocindirs with written pointers to the completehd if 11745 * the indirdep's pointer is not yet written. Otherwise 11746 * free them here. 11747 */ 11748 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != NULL) { 11749 LIST_REMOVE(aip, ai_next); 11750 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 11751 LIST_INSERT_HEAD(&indirdep->ir_completehd, aip, 11752 ai_next); 11753 newblk_freefrag(&aip->ai_block); 11754 continue; 11755 } 11756 free_newblk(&aip->ai_block); 11757 } 11758 /* 11759 * Move allocindirs that have finished dependency processing from 11760 * the done list to the write list after updating the pointers. 11761 */ 11762 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 11763 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != NULL) { 11764 handle_allocindir_partdone(aip); 11765 if (aip == LIST_FIRST(&indirdep->ir_donehd)) 11766 panic("disk_write_complete: not gone"); 11767 chgs = 1; 11768 } 11769 } 11770 /* 11771 * Preserve the indirdep if there were any changes or if it is not 11772 * yet valid on disk. 11773 */ 11774 if (chgs) { 11775 stat_indir_blk_ptrs++; 11776 bdirty(bp); 11777 return (1); 11778 } 11779 /* 11780 * If there were no changes we can discard the savedbp and detach 11781 * ourselves from the buf. We are only carrying completed pointers 11782 * in this case. 11783 */ 11784 sbp = indirdep->ir_savebp; 11785 sbp->b_flags |= B_INVAL | B_NOCACHE; 11786 indirdep->ir_savebp = NULL; 11787 indirdep->ir_bp = NULL; 11788 if (*bpp != NULL) 11789 panic("handle_written_indirdep: bp already exists."); 11790 *bpp = sbp; 11791 /* 11792 * The indirdep may not be freed until its parent points at it. 11793 */ 11794 if (indirdep->ir_state & DEPCOMPLETE) 11795 free_indirdep(indirdep); 11796 11797 return (0); 11798} 11799 11800/* 11801 * Process a diradd entry after its dependent inode has been written. 11802 */ 11803static void 11804diradd_inode_written(dap, inodedep) 11805 struct diradd *dap; 11806 struct inodedep *inodedep; 11807{ 11808 11809 LOCK_OWNED(VFSTOUFS(dap->da_list.wk_mp)); 11810 dap->da_state |= COMPLETE; 11811 complete_diradd(dap); 11812 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 11813} 11814 11815/* 11816 * Returns true if the bmsafemap will have rollbacks when written. Must only 11817 * be called with the per-filesystem lock and the buf lock on the cg held. 11818 */ 11819static int 11820bmsafemap_backgroundwrite(bmsafemap, bp) 11821 struct bmsafemap *bmsafemap; 11822 struct buf *bp; 11823{ 11824 int dirty; 11825 11826 LOCK_OWNED(VFSTOUFS(bmsafemap->sm_list.wk_mp)); 11827 dirty = !LIST_EMPTY(&bmsafemap->sm_jaddrefhd) | 11828 !LIST_EMPTY(&bmsafemap->sm_jnewblkhd); 11829 /* 11830 * If we're initiating a background write we need to process the 11831 * rollbacks as they exist now, not as they exist when IO starts. 11832 * No other consumers will look at the contents of the shadowed 11833 * buf so this is safe to do here. 11834 */ 11835 if (bp->b_xflags & BX_BKGRDMARKER) 11836 initiate_write_bmsafemap(bmsafemap, bp); 11837 11838 return (dirty); 11839} 11840 11841/* 11842 * Re-apply an allocation when a cg write is complete. 11843 */ 11844static int 11845jnewblk_rollforward(jnewblk, fs, cgp, blksfree) 11846 struct jnewblk *jnewblk; 11847 struct fs *fs; 11848 struct cg *cgp; 11849 uint8_t *blksfree; 11850{ 11851 ufs1_daddr_t fragno; 11852 ufs2_daddr_t blkno; 11853 long cgbno, bbase; 11854 int frags, blk; 11855 int i; 11856 11857 frags = 0; 11858 cgbno = dtogd(fs, jnewblk->jn_blkno); 11859 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) { 11860 if (isclr(blksfree, cgbno + i)) 11861 panic("jnewblk_rollforward: re-allocated fragment"); 11862 frags++; 11863 } 11864 if (frags == fs->fs_frag) { 11865 blkno = fragstoblks(fs, cgbno); 11866 ffs_clrblock(fs, blksfree, (long)blkno); 11867 ffs_clusteracct(fs, cgp, blkno, -1); 11868 cgp->cg_cs.cs_nbfree--; 11869 } else { 11870 bbase = cgbno - fragnum(fs, cgbno); 11871 cgbno += jnewblk->jn_oldfrags; 11872 /* If a complete block had been reassembled, account for it. */ 11873 fragno = fragstoblks(fs, bbase); 11874 if (ffs_isblock(fs, blksfree, fragno)) { 11875 cgp->cg_cs.cs_nffree += fs->fs_frag; 11876 ffs_clusteracct(fs, cgp, fragno, -1); 11877 cgp->cg_cs.cs_nbfree--; 11878 } 11879 /* Decrement the old frags. */ 11880 blk = blkmap(fs, blksfree, bbase); 11881 ffs_fragacct(fs, blk, cgp->cg_frsum, -1); 11882 /* Allocate the fragment */ 11883 for (i = 0; i < frags; i++) 11884 clrbit(blksfree, cgbno + i); 11885 cgp->cg_cs.cs_nffree -= frags; 11886 /* Add back in counts associated with the new frags */ 11887 blk = blkmap(fs, blksfree, bbase); 11888 ffs_fragacct(fs, blk, cgp->cg_frsum, 1); 11889 } 11890 return (frags); 11891} 11892 11893/* 11894 * Complete a write to a bmsafemap structure. Roll forward any bitmap 11895 * changes if it's not a background write. Set all written dependencies 11896 * to DEPCOMPLETE and free the structure if possible. 11897 * 11898 * If the write did not succeed, we will do all the roll-forward 11899 * operations, but we will not take the actions that will allow its 11900 * dependencies to be processed. 11901 */ 11902static int 11903handle_written_bmsafemap(bmsafemap, bp, flags) 11904 struct bmsafemap *bmsafemap; 11905 struct buf *bp; 11906 int flags; 11907{ 11908 struct newblk *newblk; 11909 struct inodedep *inodedep; 11910 struct jaddref *jaddref, *jatmp; 11911 struct jnewblk *jnewblk, *jntmp; 11912 struct ufsmount *ump; 11913 uint8_t *inosused; 11914 uint8_t *blksfree; 11915 struct cg *cgp; 11916 struct fs *fs; 11917 ino_t ino; 11918 int foreground; 11919 int chgs; 11920 11921 if ((bmsafemap->sm_state & IOSTARTED) == 0) 11922 panic("handle_written_bmsafemap: Not started\n"); 11923 ump = VFSTOUFS(bmsafemap->sm_list.wk_mp); 11924 chgs = 0; 11925 bmsafemap->sm_state &= ~IOSTARTED; 11926 foreground = (bp->b_xflags & BX_BKGRDMARKER) == 0; 11927 /* 11928 * If write was successful, release journal work that was waiting 11929 * on the write. Otherwise move the work back. 11930 */ 11931 if (flags & WRITESUCCEEDED) 11932 handle_jwork(&bmsafemap->sm_freewr); 11933 else 11934 LIST_CONCAT(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, 11935 worklist, wk_list); 11936 11937 /* 11938 * Restore unwritten inode allocation pending jaddref writes. 11939 */ 11940 if (!LIST_EMPTY(&bmsafemap->sm_jaddrefhd)) { 11941 cgp = (struct cg *)bp->b_data; 11942 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 11943 inosused = cg_inosused(cgp); 11944 LIST_FOREACH_SAFE(jaddref, &bmsafemap->sm_jaddrefhd, 11945 ja_bmdeps, jatmp) { 11946 if ((jaddref->ja_state & UNDONE) == 0) 11947 continue; 11948 ino = jaddref->ja_ino % fs->fs_ipg; 11949 if (isset(inosused, ino)) 11950 panic("handle_written_bmsafemap: " 11951 "re-allocated inode"); 11952 /* Do the roll-forward only if it's a real copy. */ 11953 if (foreground) { 11954 if ((jaddref->ja_mode & IFMT) == IFDIR) 11955 cgp->cg_cs.cs_ndir++; 11956 cgp->cg_cs.cs_nifree--; 11957 setbit(inosused, ino); 11958 chgs = 1; 11959 } 11960 jaddref->ja_state &= ~UNDONE; 11961 jaddref->ja_state |= ATTACHED; 11962 free_jaddref(jaddref); 11963 } 11964 } 11965 /* 11966 * Restore any block allocations which are pending journal writes. 11967 */ 11968 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) { 11969 cgp = (struct cg *)bp->b_data; 11970 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 11971 blksfree = cg_blksfree(cgp); 11972 LIST_FOREACH_SAFE(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps, 11973 jntmp) { 11974 if ((jnewblk->jn_state & UNDONE) == 0) 11975 continue; 11976 /* Do the roll-forward only if it's a real copy. */ 11977 if (foreground && 11978 jnewblk_rollforward(jnewblk, fs, cgp, blksfree)) 11979 chgs = 1; 11980 jnewblk->jn_state &= ~(UNDONE | NEWBLOCK); 11981 jnewblk->jn_state |= ATTACHED; 11982 free_jnewblk(jnewblk); 11983 } 11984 } 11985 /* 11986 * If the write did not succeed, we have done all the roll-forward 11987 * operations, but we cannot take the actions that will allow its 11988 * dependencies to be processed. 11989 */ 11990 if ((flags & WRITESUCCEEDED) == 0) { 11991 LIST_CONCAT(&bmsafemap->sm_newblkhd, &bmsafemap->sm_newblkwr, 11992 newblk, nb_deps); 11993 LIST_CONCAT(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, 11994 worklist, wk_list); 11995 if (foreground) 11996 bdirty(bp); 11997 return (1); 11998 } 11999 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkwr))) { 12000 newblk->nb_state |= DEPCOMPLETE; 12001 newblk->nb_state &= ~ONDEPLIST; 12002 newblk->nb_bmsafemap = NULL; 12003 LIST_REMOVE(newblk, nb_deps); 12004 if (newblk->nb_list.wk_type == D_ALLOCDIRECT) 12005 handle_allocdirect_partdone( 12006 WK_ALLOCDIRECT(&newblk->nb_list), NULL); 12007 else if (newblk->nb_list.wk_type == D_ALLOCINDIR) 12008 handle_allocindir_partdone( 12009 WK_ALLOCINDIR(&newblk->nb_list)); 12010 else if (newblk->nb_list.wk_type != D_NEWBLK) 12011 panic("handle_written_bmsafemap: Unexpected type: %s", 12012 TYPENAME(newblk->nb_list.wk_type)); 12013 } 12014 while ((inodedep = LIST_FIRST(&bmsafemap->sm_inodedepwr)) != NULL) { 12015 inodedep->id_state |= DEPCOMPLETE; 12016 inodedep->id_state &= ~ONDEPLIST; 12017 LIST_REMOVE(inodedep, id_deps); 12018 inodedep->id_bmsafemap = NULL; 12019 } 12020 LIST_REMOVE(bmsafemap, sm_next); 12021 if (chgs == 0 && LIST_EMPTY(&bmsafemap->sm_jaddrefhd) && 12022 LIST_EMPTY(&bmsafemap->sm_jnewblkhd) && 12023 LIST_EMPTY(&bmsafemap->sm_newblkhd) && 12024 LIST_EMPTY(&bmsafemap->sm_inodedephd) && 12025 LIST_EMPTY(&bmsafemap->sm_freehd)) { 12026 LIST_REMOVE(bmsafemap, sm_hash); 12027 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 12028 return (0); 12029 } 12030 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next); 12031 if (foreground) 12032 bdirty(bp); 12033 return (1); 12034} 12035 12036/* 12037 * Try to free a mkdir dependency. 12038 */ 12039static void 12040complete_mkdir(mkdir) 12041 struct mkdir *mkdir; 12042{ 12043 struct diradd *dap; 12044 12045 if ((mkdir->md_state & ALLCOMPLETE) != ALLCOMPLETE) 12046 return; 12047 LIST_REMOVE(mkdir, md_mkdirs); 12048 dap = mkdir->md_diradd; 12049 dap->da_state &= ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)); 12050 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) { 12051 dap->da_state |= DEPCOMPLETE; 12052 complete_diradd(dap); 12053 } 12054 WORKITEM_FREE(mkdir, D_MKDIR); 12055} 12056 12057/* 12058 * Handle the completion of a mkdir dependency. 12059 */ 12060static void 12061handle_written_mkdir(mkdir, type) 12062 struct mkdir *mkdir; 12063 int type; 12064{ 12065 12066 if ((mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)) != type) 12067 panic("handle_written_mkdir: bad type"); 12068 mkdir->md_state |= COMPLETE; 12069 complete_mkdir(mkdir); 12070} 12071 12072static int 12073free_pagedep(pagedep) 12074 struct pagedep *pagedep; 12075{ 12076 int i; 12077 12078 if (pagedep->pd_state & NEWBLOCK) 12079 return (0); 12080 if (!LIST_EMPTY(&pagedep->pd_dirremhd)) 12081 return (0); 12082 for (i = 0; i < DAHASHSZ; i++) 12083 if (!LIST_EMPTY(&pagedep->pd_diraddhd[i])) 12084 return (0); 12085 if (!LIST_EMPTY(&pagedep->pd_pendinghd)) 12086 return (0); 12087 if (!LIST_EMPTY(&pagedep->pd_jmvrefhd)) 12088 return (0); 12089 if (pagedep->pd_state & ONWORKLIST) 12090 WORKLIST_REMOVE(&pagedep->pd_list); 12091 LIST_REMOVE(pagedep, pd_hash); 12092 WORKITEM_FREE(pagedep, D_PAGEDEP); 12093 12094 return (1); 12095} 12096 12097/* 12098 * Called from within softdep_disk_write_complete above. 12099 * A write operation was just completed. Removed inodes can 12100 * now be freed and associated block pointers may be committed. 12101 * Note that this routine is always called from interrupt level 12102 * with further interrupts from this device blocked. 12103 * 12104 * If the write did not succeed, we will do all the roll-forward 12105 * operations, but we will not take the actions that will allow its 12106 * dependencies to be processed. 12107 */ 12108static int 12109handle_written_filepage(pagedep, bp, flags) 12110 struct pagedep *pagedep; 12111 struct buf *bp; /* buffer containing the written page */ 12112 int flags; 12113{ 12114 struct dirrem *dirrem; 12115 struct diradd *dap, *nextdap; 12116 struct direct *ep; 12117 int i, chgs; 12118 12119 if ((pagedep->pd_state & IOSTARTED) == 0) 12120 panic("handle_written_filepage: not started"); 12121 pagedep->pd_state &= ~IOSTARTED; 12122 if ((flags & WRITESUCCEEDED) == 0) 12123 goto rollforward; 12124 /* 12125 * Process any directory removals that have been committed. 12126 */ 12127 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 12128 LIST_REMOVE(dirrem, dm_next); 12129 dirrem->dm_state |= COMPLETE; 12130 dirrem->dm_dirinum = pagedep->pd_ino; 12131 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd), 12132 ("handle_written_filepage: Journal entries not written.")); 12133 add_to_worklist(&dirrem->dm_list, 0); 12134 } 12135 /* 12136 * Free any directory additions that have been committed. 12137 * If it is a newly allocated block, we have to wait until 12138 * the on-disk directory inode claims the new block. 12139 */ 12140 if ((pagedep->pd_state & NEWBLOCK) == 0) 12141 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 12142 free_diradd(dap, NULL); 12143rollforward: 12144 /* 12145 * Uncommitted directory entries must be restored. 12146 */ 12147 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 12148 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 12149 dap = nextdap) { 12150 nextdap = LIST_NEXT(dap, da_pdlist); 12151 if (dap->da_state & ATTACHED) 12152 panic("handle_written_filepage: attached"); 12153 ep = (struct direct *) 12154 ((char *)bp->b_data + dap->da_offset); 12155 ep->d_ino = dap->da_newinum; 12156 dap->da_state &= ~UNDONE; 12157 dap->da_state |= ATTACHED; 12158 chgs = 1; 12159 /* 12160 * If the inode referenced by the directory has 12161 * been written out, then the dependency can be 12162 * moved to the pending list. 12163 */ 12164 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 12165 LIST_REMOVE(dap, da_pdlist); 12166 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 12167 da_pdlist); 12168 } 12169 } 12170 } 12171 /* 12172 * If there were any rollbacks in the directory, then it must be 12173 * marked dirty so that its will eventually get written back in 12174 * its correct form. 12175 */ 12176 if (chgs || (flags & WRITESUCCEEDED) == 0) { 12177 if ((bp->b_flags & B_DELWRI) == 0) 12178 stat_dir_entry++; 12179 bdirty(bp); 12180 return (1); 12181 } 12182 /* 12183 * If we are not waiting for a new directory block to be 12184 * claimed by its inode, then the pagedep will be freed. 12185 * Otherwise it will remain to track any new entries on 12186 * the page in case they are fsync'ed. 12187 */ 12188 free_pagedep(pagedep); 12189 return (0); 12190} 12191 12192/* 12193 * Writing back in-core inode structures. 12194 * 12195 * The filesystem only accesses an inode's contents when it occupies an 12196 * "in-core" inode structure. These "in-core" structures are separate from 12197 * the page frames used to cache inode blocks. Only the latter are 12198 * transferred to/from the disk. So, when the updated contents of the 12199 * "in-core" inode structure are copied to the corresponding in-memory inode 12200 * block, the dependencies are also transferred. The following procedure is 12201 * called when copying a dirty "in-core" inode to a cached inode block. 12202 */ 12203 12204/* 12205 * Called when an inode is loaded from disk. If the effective link count 12206 * differed from the actual link count when it was last flushed, then we 12207 * need to ensure that the correct effective link count is put back. 12208 */ 12209void 12210softdep_load_inodeblock(ip) 12211 struct inode *ip; /* the "in_core" copy of the inode */ 12212{ 12213 struct inodedep *inodedep; 12214 struct ufsmount *ump; 12215 12216 ump = ITOUMP(ip); 12217 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 12218 ("softdep_load_inodeblock called on non-softdep filesystem")); 12219 /* 12220 * Check for alternate nlink count. 12221 */ 12222 ip->i_effnlink = ip->i_nlink; 12223 ACQUIRE_LOCK(ump); 12224 if (inodedep_lookup(UFSTOVFS(ump), ip->i_number, 0, &inodedep) == 0) { 12225 FREE_LOCK(ump); 12226 return; 12227 } 12228 ip->i_effnlink -= inodedep->id_nlinkdelta; 12229 FREE_LOCK(ump); 12230} 12231 12232/* 12233 * This routine is called just before the "in-core" inode 12234 * information is to be copied to the in-memory inode block. 12235 * Recall that an inode block contains several inodes. If 12236 * the force flag is set, then the dependencies will be 12237 * cleared so that the update can always be made. Note that 12238 * the buffer is locked when this routine is called, so we 12239 * will never be in the middle of writing the inode block 12240 * to disk. 12241 */ 12242void 12243softdep_update_inodeblock(ip, bp, waitfor) 12244 struct inode *ip; /* the "in_core" copy of the inode */ 12245 struct buf *bp; /* the buffer containing the inode block */ 12246 int waitfor; /* nonzero => update must be allowed */ 12247{ 12248 struct inodedep *inodedep; 12249 struct inoref *inoref; 12250 struct ufsmount *ump; 12251 struct worklist *wk; 12252 struct mount *mp; 12253 struct buf *ibp; 12254 struct fs *fs; 12255 int error; 12256 12257 ump = ITOUMP(ip); 12258 mp = UFSTOVFS(ump); 12259 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 12260 ("softdep_update_inodeblock called on non-softdep filesystem")); 12261 fs = ump->um_fs; 12262 /* 12263 * Preserve the freelink that is on disk. clear_unlinked_inodedep() 12264 * does not have access to the in-core ip so must write directly into 12265 * the inode block buffer when setting freelink. 12266 */ 12267 if (fs->fs_magic == FS_UFS1_MAGIC) 12268 DIP_SET(ip, i_freelink, ((struct ufs1_dinode *)bp->b_data + 12269 ino_to_fsbo(fs, ip->i_number))->di_freelink); 12270 else 12271 DIP_SET(ip, i_freelink, ((struct ufs2_dinode *)bp->b_data + 12272 ino_to_fsbo(fs, ip->i_number))->di_freelink); 12273 /* 12274 * If the effective link count is not equal to the actual link 12275 * count, then we must track the difference in an inodedep while 12276 * the inode is (potentially) tossed out of the cache. Otherwise, 12277 * if there is no existing inodedep, then there are no dependencies 12278 * to track. 12279 */ 12280 ACQUIRE_LOCK(ump); 12281again: 12282 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 12283 FREE_LOCK(ump); 12284 if (ip->i_effnlink != ip->i_nlink) 12285 panic("softdep_update_inodeblock: bad link count"); 12286 return; 12287 } 12288 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) 12289 panic("softdep_update_inodeblock: bad delta"); 12290 /* 12291 * If we're flushing all dependencies we must also move any waiting 12292 * for journal writes onto the bufwait list prior to I/O. 12293 */ 12294 if (waitfor) { 12295 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12296 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12297 == DEPCOMPLETE) { 12298 jwait(&inoref->if_list, MNT_WAIT); 12299 goto again; 12300 } 12301 } 12302 } 12303 /* 12304 * Changes have been initiated. Anything depending on these 12305 * changes cannot occur until this inode has been written. 12306 */ 12307 inodedep->id_state &= ~COMPLETE; 12308 if ((inodedep->id_state & ONWORKLIST) == 0) 12309 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 12310 /* 12311 * Any new dependencies associated with the incore inode must 12312 * now be moved to the list associated with the buffer holding 12313 * the in-memory copy of the inode. Once merged process any 12314 * allocdirects that are completed by the merger. 12315 */ 12316 merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt); 12317 if (!TAILQ_EMPTY(&inodedep->id_inoupdt)) 12318 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt), 12319 NULL); 12320 merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt); 12321 if (!TAILQ_EMPTY(&inodedep->id_extupdt)) 12322 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt), 12323 NULL); 12324 /* 12325 * Now that the inode has been pushed into the buffer, the 12326 * operations dependent on the inode being written to disk 12327 * can be moved to the id_bufwait so that they will be 12328 * processed when the buffer I/O completes. 12329 */ 12330 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 12331 WORKLIST_REMOVE(wk); 12332 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 12333 } 12334 /* 12335 * Newly allocated inodes cannot be written until the bitmap 12336 * that allocates them have been written (indicated by 12337 * DEPCOMPLETE being set in id_state). If we are doing a 12338 * forced sync (e.g., an fsync on a file), we force the bitmap 12339 * to be written so that the update can be done. 12340 */ 12341 if (waitfor == 0) { 12342 FREE_LOCK(ump); 12343 return; 12344 } 12345retry: 12346 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) != 0) { 12347 FREE_LOCK(ump); 12348 return; 12349 } 12350 ibp = inodedep->id_bmsafemap->sm_buf; 12351 ibp = getdirtybuf(ibp, LOCK_PTR(ump), MNT_WAIT); 12352 if (ibp == NULL) { 12353 /* 12354 * If ibp came back as NULL, the dependency could have been 12355 * freed while we slept. Look it up again, and check to see 12356 * that it has completed. 12357 */ 12358 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 12359 goto retry; 12360 FREE_LOCK(ump); 12361 return; 12362 } 12363 FREE_LOCK(ump); 12364 if ((error = bwrite(ibp)) != 0) 12365 softdep_error("softdep_update_inodeblock: bwrite", error); 12366} 12367 12368/* 12369 * Merge the a new inode dependency list (such as id_newinoupdt) into an 12370 * old inode dependency list (such as id_inoupdt). 12371 */ 12372static void 12373merge_inode_lists(newlisthead, oldlisthead) 12374 struct allocdirectlst *newlisthead; 12375 struct allocdirectlst *oldlisthead; 12376{ 12377 struct allocdirect *listadp, *newadp; 12378 12379 newadp = TAILQ_FIRST(newlisthead); 12380 if (newadp != NULL) 12381 LOCK_OWNED(VFSTOUFS(newadp->ad_block.nb_list.wk_mp)); 12382 for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) { 12383 if (listadp->ad_offset < newadp->ad_offset) { 12384 listadp = TAILQ_NEXT(listadp, ad_next); 12385 continue; 12386 } 12387 TAILQ_REMOVE(newlisthead, newadp, ad_next); 12388 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 12389 if (listadp->ad_offset == newadp->ad_offset) { 12390 allocdirect_merge(oldlisthead, newadp, 12391 listadp); 12392 listadp = newadp; 12393 } 12394 newadp = TAILQ_FIRST(newlisthead); 12395 } 12396 while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) { 12397 TAILQ_REMOVE(newlisthead, newadp, ad_next); 12398 TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next); 12399 } 12400} 12401 12402/* 12403 * If we are doing an fsync, then we must ensure that any directory 12404 * entries for the inode have been written after the inode gets to disk. 12405 */ 12406int 12407softdep_fsync(vp) 12408 struct vnode *vp; /* the "in_core" copy of the inode */ 12409{ 12410 struct inodedep *inodedep; 12411 struct pagedep *pagedep; 12412 struct inoref *inoref; 12413 struct ufsmount *ump; 12414 struct worklist *wk; 12415 struct diradd *dap; 12416 struct mount *mp; 12417 struct vnode *pvp; 12418 struct inode *ip; 12419 struct buf *bp; 12420 struct fs *fs; 12421 struct thread *td = curthread; 12422 int error, flushparent, pagedep_new_block; 12423 ino_t parentino; 12424 ufs_lbn_t lbn; 12425 12426 ip = VTOI(vp); 12427 mp = vp->v_mount; 12428 ump = VFSTOUFS(mp); 12429 fs = ump->um_fs; 12430 if (MOUNTEDSOFTDEP(mp) == 0) 12431 return (0); 12432 ACQUIRE_LOCK(ump); 12433restart: 12434 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 12435 FREE_LOCK(ump); 12436 return (0); 12437 } 12438 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12439 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12440 == DEPCOMPLETE) { 12441 jwait(&inoref->if_list, MNT_WAIT); 12442 goto restart; 12443 } 12444 } 12445 if (!LIST_EMPTY(&inodedep->id_inowait) || 12446 !TAILQ_EMPTY(&inodedep->id_extupdt) || 12447 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 12448 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 12449 !TAILQ_EMPTY(&inodedep->id_newinoupdt)) 12450 panic("softdep_fsync: pending ops %p", inodedep); 12451 for (error = 0, flushparent = 0; ; ) { 12452 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 12453 break; 12454 if (wk->wk_type != D_DIRADD) 12455 panic("softdep_fsync: Unexpected type %s", 12456 TYPENAME(wk->wk_type)); 12457 dap = WK_DIRADD(wk); 12458 /* 12459 * Flush our parent if this directory entry has a MKDIR_PARENT 12460 * dependency or is contained in a newly allocated block. 12461 */ 12462 if (dap->da_state & DIRCHG) 12463 pagedep = dap->da_previous->dm_pagedep; 12464 else 12465 pagedep = dap->da_pagedep; 12466 parentino = pagedep->pd_ino; 12467 lbn = pagedep->pd_lbn; 12468 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) 12469 panic("softdep_fsync: dirty"); 12470 if ((dap->da_state & MKDIR_PARENT) || 12471 (pagedep->pd_state & NEWBLOCK)) 12472 flushparent = 1; 12473 else 12474 flushparent = 0; 12475 /* 12476 * If we are being fsync'ed as part of vgone'ing this vnode, 12477 * then we will not be able to release and recover the 12478 * vnode below, so we just have to give up on writing its 12479 * directory entry out. It will eventually be written, just 12480 * not now, but then the user was not asking to have it 12481 * written, so we are not breaking any promises. 12482 */ 12483 if (vp->v_iflag & VI_DOOMED) 12484 break; 12485 /* 12486 * We prevent deadlock by always fetching inodes from the 12487 * root, moving down the directory tree. Thus, when fetching 12488 * our parent directory, we first try to get the lock. If 12489 * that fails, we must unlock ourselves before requesting 12490 * the lock on our parent. See the comment in ufs_lookup 12491 * for details on possible races. 12492 */ 12493 FREE_LOCK(ump); 12494 if (ffs_vgetf(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp, 12495 FFSV_FORCEINSMQ)) { 12496 /* 12497 * Unmount cannot proceed after unlock because 12498 * caller must have called vn_start_write(). 12499 */ 12500 VOP_UNLOCK(vp, 0); 12501 error = ffs_vgetf(mp, parentino, LK_EXCLUSIVE, 12502 &pvp, FFSV_FORCEINSMQ); 12503 MPASS(VTOI(pvp)->i_mode != 0); 12504 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 12505 if (vp->v_iflag & VI_DOOMED) { 12506 if (error == 0) 12507 vput(pvp); 12508 error = ENOENT; 12509 } 12510 if (error != 0) 12511 return (error); 12512 } 12513 /* 12514 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps 12515 * that are contained in direct blocks will be resolved by 12516 * doing a ffs_update. Pagedeps contained in indirect blocks 12517 * may require a complete sync'ing of the directory. So, we 12518 * try the cheap and fast ffs_update first, and if that fails, 12519 * then we do the slower ffs_syncvnode of the directory. 12520 */ 12521 if (flushparent) { 12522 int locked; 12523 12524 if ((error = ffs_update(pvp, 1)) != 0) { 12525 vput(pvp); 12526 return (error); 12527 } 12528 ACQUIRE_LOCK(ump); 12529 locked = 1; 12530 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) { 12531 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) != NULL) { 12532 if (wk->wk_type != D_DIRADD) 12533 panic("softdep_fsync: Unexpected type %s", 12534 TYPENAME(wk->wk_type)); 12535 dap = WK_DIRADD(wk); 12536 if (dap->da_state & DIRCHG) 12537 pagedep = dap->da_previous->dm_pagedep; 12538 else 12539 pagedep = dap->da_pagedep; 12540 pagedep_new_block = pagedep->pd_state & NEWBLOCK; 12541 FREE_LOCK(ump); 12542 locked = 0; 12543 if (pagedep_new_block && (error = 12544 ffs_syncvnode(pvp, MNT_WAIT, 0))) { 12545 vput(pvp); 12546 return (error); 12547 } 12548 } 12549 } 12550 if (locked) 12551 FREE_LOCK(ump); 12552 } 12553 /* 12554 * Flush directory page containing the inode's name. 12555 */ 12556 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred, 12557 &bp); 12558 if (error == 0) 12559 error = bwrite(bp); 12560 else 12561 brelse(bp); 12562 vput(pvp); 12563 if (error != 0) 12564 return (error); 12565 ACQUIRE_LOCK(ump); 12566 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 12567 break; 12568 } 12569 FREE_LOCK(ump); 12570 return (0); 12571} 12572 12573/* 12574 * Flush all the dirty bitmaps associated with the block device 12575 * before flushing the rest of the dirty blocks so as to reduce 12576 * the number of dependencies that will have to be rolled back. 12577 * 12578 * XXX Unused? 12579 */ 12580void 12581softdep_fsync_mountdev(vp) 12582 struct vnode *vp; 12583{ 12584 struct buf *bp, *nbp; 12585 struct worklist *wk; 12586 struct bufobj *bo; 12587 12588 if (!vn_isdisk(vp, NULL)) 12589 panic("softdep_fsync_mountdev: vnode not a disk"); 12590 bo = &vp->v_bufobj; 12591restart: 12592 BO_LOCK(bo); 12593 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 12594 /* 12595 * If it is already scheduled, skip to the next buffer. 12596 */ 12597 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 12598 continue; 12599 12600 if ((bp->b_flags & B_DELWRI) == 0) 12601 panic("softdep_fsync_mountdev: not dirty"); 12602 /* 12603 * We are only interested in bitmaps with outstanding 12604 * dependencies. 12605 */ 12606 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 12607 wk->wk_type != D_BMSAFEMAP || 12608 (bp->b_vflags & BV_BKGRDINPROG)) { 12609 BUF_UNLOCK(bp); 12610 continue; 12611 } 12612 BO_UNLOCK(bo); 12613 bremfree(bp); 12614 (void) bawrite(bp); 12615 goto restart; 12616 } 12617 drain_output(vp); 12618 BO_UNLOCK(bo); 12619} 12620 12621/* 12622 * Sync all cylinder groups that were dirty at the time this function is 12623 * called. Newly dirtied cgs will be inserted before the sentinel. This 12624 * is used to flush freedep activity that may be holding up writes to a 12625 * indirect block. 12626 */ 12627static int 12628sync_cgs(mp, waitfor) 12629 struct mount *mp; 12630 int waitfor; 12631{ 12632 struct bmsafemap *bmsafemap; 12633 struct bmsafemap *sentinel; 12634 struct ufsmount *ump; 12635 struct buf *bp; 12636 int error; 12637 12638 sentinel = malloc(sizeof(*sentinel), M_BMSAFEMAP, M_ZERO | M_WAITOK); 12639 sentinel->sm_cg = -1; 12640 ump = VFSTOUFS(mp); 12641 error = 0; 12642 ACQUIRE_LOCK(ump); 12643 LIST_INSERT_HEAD(&ump->softdep_dirtycg, sentinel, sm_next); 12644 for (bmsafemap = LIST_NEXT(sentinel, sm_next); bmsafemap != NULL; 12645 bmsafemap = LIST_NEXT(sentinel, sm_next)) { 12646 /* Skip sentinels and cgs with no work to release. */ 12647 if (bmsafemap->sm_cg == -1 || 12648 (LIST_EMPTY(&bmsafemap->sm_freehd) && 12649 LIST_EMPTY(&bmsafemap->sm_freewr))) { 12650 LIST_REMOVE(sentinel, sm_next); 12651 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next); 12652 continue; 12653 } 12654 /* 12655 * If we don't get the lock and we're waiting try again, if 12656 * not move on to the next buf and try to sync it. 12657 */ 12658 bp = getdirtybuf(bmsafemap->sm_buf, LOCK_PTR(ump), waitfor); 12659 if (bp == NULL && waitfor == MNT_WAIT) 12660 continue; 12661 LIST_REMOVE(sentinel, sm_next); 12662 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next); 12663 if (bp == NULL) 12664 continue; 12665 FREE_LOCK(ump); 12666 if (waitfor == MNT_NOWAIT) 12667 bawrite(bp); 12668 else 12669 error = bwrite(bp); 12670 ACQUIRE_LOCK(ump); 12671 if (error) 12672 break; 12673 } 12674 LIST_REMOVE(sentinel, sm_next); 12675 FREE_LOCK(ump); 12676 free(sentinel, M_BMSAFEMAP); 12677 return (error); 12678} 12679 12680/* 12681 * This routine is called when we are trying to synchronously flush a 12682 * file. This routine must eliminate any filesystem metadata dependencies 12683 * so that the syncing routine can succeed. 12684 */ 12685int 12686softdep_sync_metadata(struct vnode *vp) 12687{ 12688 struct inode *ip; 12689 int error; 12690 12691 ip = VTOI(vp); 12692 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 12693 ("softdep_sync_metadata called on non-softdep filesystem")); 12694 /* 12695 * Ensure that any direct block dependencies have been cleared, 12696 * truncations are started, and inode references are journaled. 12697 */ 12698 ACQUIRE_LOCK(VFSTOUFS(vp->v_mount)); 12699 /* 12700 * Write all journal records to prevent rollbacks on devvp. 12701 */ 12702 if (vp->v_type == VCHR) 12703 softdep_flushjournal(vp->v_mount); 12704 error = flush_inodedep_deps(vp, vp->v_mount, ip->i_number); 12705 /* 12706 * Ensure that all truncates are written so we won't find deps on 12707 * indirect blocks. 12708 */ 12709 process_truncates(vp); 12710 FREE_LOCK(VFSTOUFS(vp->v_mount)); 12711 12712 return (error); 12713} 12714 12715/* 12716 * This routine is called when we are attempting to sync a buf with 12717 * dependencies. If waitfor is MNT_NOWAIT it attempts to schedule any 12718 * other IO it can but returns EBUSY if the buffer is not yet able to 12719 * be written. Dependencies which will not cause rollbacks will always 12720 * return 0. 12721 */ 12722int 12723softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor) 12724{ 12725 struct indirdep *indirdep; 12726 struct pagedep *pagedep; 12727 struct allocindir *aip; 12728 struct newblk *newblk; 12729 struct ufsmount *ump; 12730 struct buf *nbp; 12731 struct worklist *wk; 12732 int i, error; 12733 12734 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 12735 ("softdep_sync_buf called on non-softdep filesystem")); 12736 /* 12737 * For VCHR we just don't want to force flush any dependencies that 12738 * will cause rollbacks. 12739 */ 12740 if (vp->v_type == VCHR) { 12741 if (waitfor == MNT_NOWAIT && softdep_count_dependencies(bp, 0)) 12742 return (EBUSY); 12743 return (0); 12744 } 12745 ump = VFSTOUFS(vp->v_mount); 12746 ACQUIRE_LOCK(ump); 12747 /* 12748 * As we hold the buffer locked, none of its dependencies 12749 * will disappear. 12750 */ 12751 error = 0; 12752top: 12753 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 12754 switch (wk->wk_type) { 12755 12756 case D_ALLOCDIRECT: 12757 case D_ALLOCINDIR: 12758 newblk = WK_NEWBLK(wk); 12759 if (newblk->nb_jnewblk != NULL) { 12760 if (waitfor == MNT_NOWAIT) { 12761 error = EBUSY; 12762 goto out_unlock; 12763 } 12764 jwait(&newblk->nb_jnewblk->jn_list, waitfor); 12765 goto top; 12766 } 12767 if (newblk->nb_state & DEPCOMPLETE || 12768 waitfor == MNT_NOWAIT) 12769 continue; 12770 nbp = newblk->nb_bmsafemap->sm_buf; 12771 nbp = getdirtybuf(nbp, LOCK_PTR(ump), waitfor); 12772 if (nbp == NULL) 12773 goto top; 12774 FREE_LOCK(ump); 12775 if ((error = bwrite(nbp)) != 0) 12776 goto out; 12777 ACQUIRE_LOCK(ump); 12778 continue; 12779 12780 case D_INDIRDEP: 12781 indirdep = WK_INDIRDEP(wk); 12782 if (waitfor == MNT_NOWAIT) { 12783 if (!TAILQ_EMPTY(&indirdep->ir_trunc) || 12784 !LIST_EMPTY(&indirdep->ir_deplisthd)) { 12785 error = EBUSY; 12786 goto out_unlock; 12787 } 12788 } 12789 if (!TAILQ_EMPTY(&indirdep->ir_trunc)) 12790 panic("softdep_sync_buf: truncation pending."); 12791 restart: 12792 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 12793 newblk = (struct newblk *)aip; 12794 if (newblk->nb_jnewblk != NULL) { 12795 jwait(&newblk->nb_jnewblk->jn_list, 12796 waitfor); 12797 goto restart; 12798 } 12799 if (newblk->nb_state & DEPCOMPLETE) 12800 continue; 12801 nbp = newblk->nb_bmsafemap->sm_buf; 12802 nbp = getdirtybuf(nbp, LOCK_PTR(ump), waitfor); 12803 if (nbp == NULL) 12804 goto restart; 12805 FREE_LOCK(ump); 12806 if ((error = bwrite(nbp)) != 0) 12807 goto out; 12808 ACQUIRE_LOCK(ump); 12809 goto restart; 12810 } 12811 continue; 12812 12813 case D_PAGEDEP: 12814 /* 12815 * Only flush directory entries in synchronous passes. 12816 */ 12817 if (waitfor != MNT_WAIT) { 12818 error = EBUSY; 12819 goto out_unlock; 12820 } 12821 /* 12822 * While syncing snapshots, we must allow recursive 12823 * lookups. 12824 */ 12825 BUF_AREC(bp); 12826 /* 12827 * We are trying to sync a directory that may 12828 * have dependencies on both its own metadata 12829 * and/or dependencies on the inodes of any 12830 * recently allocated files. We walk its diradd 12831 * lists pushing out the associated inode. 12832 */ 12833 pagedep = WK_PAGEDEP(wk); 12834 for (i = 0; i < DAHASHSZ; i++) { 12835 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 12836 continue; 12837 if ((error = flush_pagedep_deps(vp, wk->wk_mp, 12838 &pagedep->pd_diraddhd[i]))) { 12839 BUF_NOREC(bp); 12840 goto out_unlock; 12841 } 12842 } 12843 BUF_NOREC(bp); 12844 continue; 12845 12846 case D_FREEWORK: 12847 case D_FREEDEP: 12848 case D_JSEGDEP: 12849 case D_JNEWBLK: 12850 continue; 12851 12852 default: 12853 panic("softdep_sync_buf: Unknown type %s", 12854 TYPENAME(wk->wk_type)); 12855 /* NOTREACHED */ 12856 } 12857 } 12858out_unlock: 12859 FREE_LOCK(ump); 12860out: 12861 return (error); 12862} 12863 12864/* 12865 * Flush the dependencies associated with an inodedep. 12866 */ 12867static int 12868flush_inodedep_deps(vp, mp, ino) 12869 struct vnode *vp; 12870 struct mount *mp; 12871 ino_t ino; 12872{ 12873 struct inodedep *inodedep; 12874 struct inoref *inoref; 12875 struct ufsmount *ump; 12876 int error, waitfor; 12877 12878 /* 12879 * This work is done in two passes. The first pass grabs most 12880 * of the buffers and begins asynchronously writing them. The 12881 * only way to wait for these asynchronous writes is to sleep 12882 * on the filesystem vnode which may stay busy for a long time 12883 * if the filesystem is active. So, instead, we make a second 12884 * pass over the dependencies blocking on each write. In the 12885 * usual case we will be blocking against a write that we 12886 * initiated, so when it is done the dependency will have been 12887 * resolved. Thus the second pass is expected to end quickly. 12888 * We give a brief window at the top of the loop to allow 12889 * any pending I/O to complete. 12890 */ 12891 ump = VFSTOUFS(mp); 12892 LOCK_OWNED(ump); 12893 for (error = 0, waitfor = MNT_NOWAIT; ; ) { 12894 if (error) 12895 return (error); 12896 FREE_LOCK(ump); 12897 ACQUIRE_LOCK(ump); 12898restart: 12899 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 12900 return (0); 12901 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12902 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12903 == DEPCOMPLETE) { 12904 jwait(&inoref->if_list, MNT_WAIT); 12905 goto restart; 12906 } 12907 } 12908 if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) || 12909 flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) || 12910 flush_deplist(&inodedep->id_extupdt, waitfor, &error) || 12911 flush_deplist(&inodedep->id_newextupdt, waitfor, &error)) 12912 continue; 12913 /* 12914 * If pass2, we are done, otherwise do pass 2. 12915 */ 12916 if (waitfor == MNT_WAIT) 12917 break; 12918 waitfor = MNT_WAIT; 12919 } 12920 /* 12921 * Try freeing inodedep in case all dependencies have been removed. 12922 */ 12923 if (inodedep_lookup(mp, ino, 0, &inodedep) != 0) 12924 (void) free_inodedep(inodedep); 12925 return (0); 12926} 12927 12928/* 12929 * Flush an inode dependency list. 12930 */ 12931static int 12932flush_deplist(listhead, waitfor, errorp) 12933 struct allocdirectlst *listhead; 12934 int waitfor; 12935 int *errorp; 12936{ 12937 struct allocdirect *adp; 12938 struct newblk *newblk; 12939 struct ufsmount *ump; 12940 struct buf *bp; 12941 12942 if ((adp = TAILQ_FIRST(listhead)) == NULL) 12943 return (0); 12944 ump = VFSTOUFS(adp->ad_list.wk_mp); 12945 LOCK_OWNED(ump); 12946 TAILQ_FOREACH(adp, listhead, ad_next) { 12947 newblk = (struct newblk *)adp; 12948 if (newblk->nb_jnewblk != NULL) { 12949 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 12950 return (1); 12951 } 12952 if (newblk->nb_state & DEPCOMPLETE) 12953 continue; 12954 bp = newblk->nb_bmsafemap->sm_buf; 12955 bp = getdirtybuf(bp, LOCK_PTR(ump), waitfor); 12956 if (bp == NULL) { 12957 if (waitfor == MNT_NOWAIT) 12958 continue; 12959 return (1); 12960 } 12961 FREE_LOCK(ump); 12962 if (waitfor == MNT_NOWAIT) 12963 bawrite(bp); 12964 else 12965 *errorp = bwrite(bp); 12966 ACQUIRE_LOCK(ump); 12967 return (1); 12968 } 12969 return (0); 12970} 12971 12972/* 12973 * Flush dependencies associated with an allocdirect block. 12974 */ 12975static int 12976flush_newblk_dep(vp, mp, lbn) 12977 struct vnode *vp; 12978 struct mount *mp; 12979 ufs_lbn_t lbn; 12980{ 12981 struct newblk *newblk; 12982 struct ufsmount *ump; 12983 struct bufobj *bo; 12984 struct inode *ip; 12985 struct buf *bp; 12986 ufs2_daddr_t blkno; 12987 int error; 12988 12989 error = 0; 12990 bo = &vp->v_bufobj; 12991 ip = VTOI(vp); 12992 blkno = DIP(ip, i_db[lbn]); 12993 if (blkno == 0) 12994 panic("flush_newblk_dep: Missing block"); 12995 ump = VFSTOUFS(mp); 12996 ACQUIRE_LOCK(ump); 12997 /* 12998 * Loop until all dependencies related to this block are satisfied. 12999 * We must be careful to restart after each sleep in case a write 13000 * completes some part of this process for us. 13001 */ 13002 for (;;) { 13003 if (newblk_lookup(mp, blkno, 0, &newblk) == 0) { 13004 FREE_LOCK(ump); 13005 break; 13006 } 13007 if (newblk->nb_list.wk_type != D_ALLOCDIRECT) 13008 panic("flush_newblk_dep: Bad newblk %p", newblk); 13009 /* 13010 * Flush the journal. 13011 */ 13012 if (newblk->nb_jnewblk != NULL) { 13013 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 13014 continue; 13015 } 13016 /* 13017 * Write the bitmap dependency. 13018 */ 13019 if ((newblk->nb_state & DEPCOMPLETE) == 0) { 13020 bp = newblk->nb_bmsafemap->sm_buf; 13021 bp = getdirtybuf(bp, LOCK_PTR(ump), MNT_WAIT); 13022 if (bp == NULL) 13023 continue; 13024 FREE_LOCK(ump); 13025 error = bwrite(bp); 13026 if (error) 13027 break; 13028 ACQUIRE_LOCK(ump); 13029 continue; 13030 } 13031 /* 13032 * Write the buffer. 13033 */ 13034 FREE_LOCK(ump); 13035 BO_LOCK(bo); 13036 bp = gbincore(bo, lbn); 13037 if (bp != NULL) { 13038 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | 13039 LK_INTERLOCK, BO_LOCKPTR(bo)); 13040 if (error == ENOLCK) { 13041 ACQUIRE_LOCK(ump); 13042 error = 0; 13043 continue; /* Slept, retry */ 13044 } 13045 if (error != 0) 13046 break; /* Failed */ 13047 if (bp->b_flags & B_DELWRI) { 13048 bremfree(bp); 13049 error = bwrite(bp); 13050 if (error) 13051 break; 13052 } else 13053 BUF_UNLOCK(bp); 13054 } else 13055 BO_UNLOCK(bo); 13056 /* 13057 * We have to wait for the direct pointers to 13058 * point at the newdirblk before the dependency 13059 * will go away. 13060 */ 13061 error = ffs_update(vp, 1); 13062 if (error) 13063 break; 13064 ACQUIRE_LOCK(ump); 13065 } 13066 return (error); 13067} 13068 13069/* 13070 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 13071 */ 13072static int 13073flush_pagedep_deps(pvp, mp, diraddhdp) 13074 struct vnode *pvp; 13075 struct mount *mp; 13076 struct diraddhd *diraddhdp; 13077{ 13078 struct inodedep *inodedep; 13079 struct inoref *inoref; 13080 struct ufsmount *ump; 13081 struct diradd *dap; 13082 struct vnode *vp; 13083 int error = 0; 13084 struct buf *bp; 13085 ino_t inum; 13086 struct diraddhd unfinished; 13087 13088 LIST_INIT(&unfinished); 13089 ump = VFSTOUFS(mp); 13090 LOCK_OWNED(ump); 13091restart: 13092 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 13093 /* 13094 * Flush ourselves if this directory entry 13095 * has a MKDIR_PARENT dependency. 13096 */ 13097 if (dap->da_state & MKDIR_PARENT) { 13098 FREE_LOCK(ump); 13099 if ((error = ffs_update(pvp, 1)) != 0) 13100 break; 13101 ACQUIRE_LOCK(ump); 13102 /* 13103 * If that cleared dependencies, go on to next. 13104 */ 13105 if (dap != LIST_FIRST(diraddhdp)) 13106 continue; 13107 /* 13108 * All MKDIR_PARENT dependencies and all the 13109 * NEWBLOCK pagedeps that are contained in direct 13110 * blocks were resolved by doing above ffs_update. 13111 * Pagedeps contained in indirect blocks may 13112 * require a complete sync'ing of the directory. 13113 * We are in the midst of doing a complete sync, 13114 * so if they are not resolved in this pass we 13115 * defer them for now as they will be sync'ed by 13116 * our caller shortly. 13117 */ 13118 LIST_REMOVE(dap, da_pdlist); 13119 LIST_INSERT_HEAD(&unfinished, dap, da_pdlist); 13120 continue; 13121 } 13122 /* 13123 * A newly allocated directory must have its "." and 13124 * ".." entries written out before its name can be 13125 * committed in its parent. 13126 */ 13127 inum = dap->da_newinum; 13128 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 13129 panic("flush_pagedep_deps: lost inode1"); 13130 /* 13131 * Wait for any pending journal adds to complete so we don't 13132 * cause rollbacks while syncing. 13133 */ 13134 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 13135 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 13136 == DEPCOMPLETE) { 13137 jwait(&inoref->if_list, MNT_WAIT); 13138 goto restart; 13139 } 13140 } 13141 if (dap->da_state & MKDIR_BODY) { 13142 FREE_LOCK(ump); 13143 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp, 13144 FFSV_FORCEINSMQ))) 13145 break; 13146 MPASS(VTOI(vp)->i_mode != 0); 13147 error = flush_newblk_dep(vp, mp, 0); 13148 /* 13149 * If we still have the dependency we might need to 13150 * update the vnode to sync the new link count to 13151 * disk. 13152 */ 13153 if (error == 0 && dap == LIST_FIRST(diraddhdp)) 13154 error = ffs_update(vp, 1); 13155 vput(vp); 13156 if (error != 0) 13157 break; 13158 ACQUIRE_LOCK(ump); 13159 /* 13160 * If that cleared dependencies, go on to next. 13161 */ 13162 if (dap != LIST_FIRST(diraddhdp)) 13163 continue; 13164 if (dap->da_state & MKDIR_BODY) { 13165 inodedep_lookup(UFSTOVFS(ump), inum, 0, 13166 &inodedep); 13167 panic("flush_pagedep_deps: MKDIR_BODY " 13168 "inodedep %p dap %p vp %p", 13169 inodedep, dap, vp); 13170 } 13171 } 13172 /* 13173 * Flush the inode on which the directory entry depends. 13174 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 13175 * the only remaining dependency is that the updated inode 13176 * count must get pushed to disk. The inode has already 13177 * been pushed into its inode buffer (via VOP_UPDATE) at 13178 * the time of the reference count change. So we need only 13179 * locate that buffer, ensure that there will be no rollback 13180 * caused by a bitmap dependency, then write the inode buffer. 13181 */ 13182retry: 13183 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 13184 panic("flush_pagedep_deps: lost inode"); 13185 /* 13186 * If the inode still has bitmap dependencies, 13187 * push them to disk. 13188 */ 13189 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) == 0) { 13190 bp = inodedep->id_bmsafemap->sm_buf; 13191 bp = getdirtybuf(bp, LOCK_PTR(ump), MNT_WAIT); 13192 if (bp == NULL) 13193 goto retry; 13194 FREE_LOCK(ump); 13195 if ((error = bwrite(bp)) != 0) 13196 break; 13197 ACQUIRE_LOCK(ump); 13198 if (dap != LIST_FIRST(diraddhdp)) 13199 continue; 13200 } 13201 /* 13202 * If the inode is still sitting in a buffer waiting 13203 * to be written or waiting for the link count to be 13204 * adjusted update it here to flush it to disk. 13205 */ 13206 if (dap == LIST_FIRST(diraddhdp)) { 13207 FREE_LOCK(ump); 13208 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp, 13209 FFSV_FORCEINSMQ))) 13210 break; 13211 MPASS(VTOI(vp)->i_mode != 0); 13212 error = ffs_update(vp, 1); 13213 vput(vp); 13214 if (error) 13215 break; 13216 ACQUIRE_LOCK(ump); 13217 } 13218 /* 13219 * If we have failed to get rid of all the dependencies 13220 * then something is seriously wrong. 13221 */ 13222 if (dap == LIST_FIRST(diraddhdp)) { 13223 inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep); 13224 panic("flush_pagedep_deps: failed to flush " 13225 "inodedep %p ino %ju dap %p", 13226 inodedep, (uintmax_t)inum, dap); 13227 } 13228 } 13229 if (error) 13230 ACQUIRE_LOCK(ump); 13231 while ((dap = LIST_FIRST(&unfinished)) != NULL) { 13232 LIST_REMOVE(dap, da_pdlist); 13233 LIST_INSERT_HEAD(diraddhdp, dap, da_pdlist); 13234 } 13235 return (error); 13236} 13237 13238/* 13239 * A large burst of file addition or deletion activity can drive the 13240 * memory load excessively high. First attempt to slow things down 13241 * using the techniques below. If that fails, this routine requests 13242 * the offending operations to fall back to running synchronously 13243 * until the memory load returns to a reasonable level. 13244 */ 13245int 13246softdep_slowdown(vp) 13247 struct vnode *vp; 13248{ 13249 struct ufsmount *ump; 13250 int jlow; 13251 int max_softdeps_hard; 13252 13253 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 13254 ("softdep_slowdown called on non-softdep filesystem")); 13255 ump = VFSTOUFS(vp->v_mount); 13256 ACQUIRE_LOCK(ump); 13257 jlow = 0; 13258 /* 13259 * Check for journal space if needed. 13260 */ 13261 if (DOINGSUJ(vp)) { 13262 if (journal_space(ump, 0) == 0) 13263 jlow = 1; 13264 } 13265 /* 13266 * If the system is under its limits and our filesystem is 13267 * not responsible for more than our share of the usage and 13268 * we are not low on journal space, then no need to slow down. 13269 */ 13270 max_softdeps_hard = max_softdeps * 11 / 10; 13271 if (dep_current[D_DIRREM] < max_softdeps_hard / 2 && 13272 dep_current[D_INODEDEP] < max_softdeps_hard && 13273 dep_current[D_INDIRDEP] < max_softdeps_hard / 1000 && 13274 dep_current[D_FREEBLKS] < max_softdeps_hard && jlow == 0 && 13275 ump->softdep_curdeps[D_DIRREM] < 13276 (max_softdeps_hard / 2) / stat_flush_threads && 13277 ump->softdep_curdeps[D_INODEDEP] < 13278 max_softdeps_hard / stat_flush_threads && 13279 ump->softdep_curdeps[D_INDIRDEP] < 13280 (max_softdeps_hard / 1000) / stat_flush_threads && 13281 ump->softdep_curdeps[D_FREEBLKS] < 13282 max_softdeps_hard / stat_flush_threads) { 13283 FREE_LOCK(ump); 13284 return (0); 13285 } 13286 /* 13287 * If the journal is low or our filesystem is over its limit 13288 * then speedup the cleanup. 13289 */ 13290 if (ump->softdep_curdeps[D_INDIRDEP] < 13291 (max_softdeps_hard / 1000) / stat_flush_threads || jlow) 13292 softdep_speedup(ump); 13293 stat_sync_limit_hit += 1; 13294 FREE_LOCK(ump); 13295 /* 13296 * We only slow down the rate at which new dependencies are 13297 * generated if we are not using journaling. With journaling, 13298 * the cleanup should always be sufficient to keep things 13299 * under control. 13300 */ 13301 if (DOINGSUJ(vp)) 13302 return (0); 13303 return (1); 13304} 13305 13306/* 13307 * Called by the allocation routines when they are about to fail 13308 * in the hope that we can free up the requested resource (inodes 13309 * or disk space). 13310 * 13311 * First check to see if the work list has anything on it. If it has, 13312 * clean up entries until we successfully free the requested resource. 13313 * Because this process holds inodes locked, we cannot handle any remove 13314 * requests that might block on a locked inode as that could lead to 13315 * deadlock. If the worklist yields none of the requested resource, 13316 * start syncing out vnodes to free up the needed space. 13317 */ 13318int 13319softdep_request_cleanup(fs, vp, cred, resource) 13320 struct fs *fs; 13321 struct vnode *vp; 13322 struct ucred *cred; 13323 int resource; 13324{ 13325 struct ufsmount *ump; 13326 struct mount *mp; 13327 long starttime; 13328 ufs2_daddr_t needed; 13329 int error, failed_vnode; 13330 13331 /* 13332 * If we are being called because of a process doing a 13333 * copy-on-write, then it is not safe to process any 13334 * worklist items as we will recurse into the copyonwrite 13335 * routine. This will result in an incoherent snapshot. 13336 * If the vnode that we hold is a snapshot, we must avoid 13337 * handling other resources that could cause deadlock. 13338 */ 13339 if ((curthread->td_pflags & TDP_COWINPROGRESS) || IS_SNAPSHOT(VTOI(vp))) 13340 return (0); 13341 13342 if (resource == FLUSH_BLOCKS_WAIT) 13343 stat_cleanup_blkrequests += 1; 13344 else 13345 stat_cleanup_inorequests += 1; 13346 13347 mp = vp->v_mount; 13348 ump = VFSTOUFS(mp); 13349 mtx_assert(UFS_MTX(ump), MA_OWNED); 13350 UFS_UNLOCK(ump); 13351 error = ffs_update(vp, 1); 13352 if (error != 0 || MOUNTEDSOFTDEP(mp) == 0) { 13353 UFS_LOCK(ump); 13354 return (0); 13355 } 13356 /* 13357 * If we are in need of resources, start by cleaning up 13358 * any block removals associated with our inode. 13359 */ 13360 ACQUIRE_LOCK(ump); 13361 process_removes(vp); 13362 process_truncates(vp); 13363 FREE_LOCK(ump); 13364 /* 13365 * Now clean up at least as many resources as we will need. 13366 * 13367 * When requested to clean up inodes, the number that are needed 13368 * is set by the number of simultaneous writers (mnt_writeopcount) 13369 * plus a bit of slop (2) in case some more writers show up while 13370 * we are cleaning. 13371 * 13372 * When requested to free up space, the amount of space that 13373 * we need is enough blocks to allocate a full-sized segment 13374 * (fs_contigsumsize). The number of such segments that will 13375 * be needed is set by the number of simultaneous writers 13376 * (mnt_writeopcount) plus a bit of slop (2) in case some more 13377 * writers show up while we are cleaning. 13378 * 13379 * Additionally, if we are unpriviledged and allocating space, 13380 * we need to ensure that we clean up enough blocks to get the 13381 * needed number of blocks over the threshold of the minimum 13382 * number of blocks required to be kept free by the filesystem 13383 * (fs_minfree). 13384 */ 13385 if (resource == FLUSH_INODES_WAIT) { 13386 needed = vp->v_mount->mnt_writeopcount + 2; 13387 } else if (resource == FLUSH_BLOCKS_WAIT) { 13388 needed = (vp->v_mount->mnt_writeopcount + 2) * 13389 fs->fs_contigsumsize; 13390 if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE, 0)) 13391 needed += fragstoblks(fs, 13392 roundup((fs->fs_dsize * fs->fs_minfree / 100) - 13393 fs->fs_cstotal.cs_nffree, fs->fs_frag)); 13394 } else { 13395 UFS_LOCK(ump); 13396 printf("softdep_request_cleanup: Unknown resource type %d\n", 13397 resource); 13398 return (0); 13399 } 13400 starttime = time_second; 13401retry: 13402 if ((resource == FLUSH_BLOCKS_WAIT && ump->softdep_on_worklist > 0 && 13403 fs->fs_cstotal.cs_nbfree <= needed) || 13404 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 && 13405 fs->fs_cstotal.cs_nifree <= needed)) { 13406 ACQUIRE_LOCK(ump); 13407 if (ump->softdep_on_worklist > 0 && 13408 process_worklist_item(UFSTOVFS(ump), 13409 ump->softdep_on_worklist, LK_NOWAIT) != 0) 13410 stat_worklist_push += 1; 13411 FREE_LOCK(ump); 13412 } 13413 /* 13414 * If we still need resources and there are no more worklist 13415 * entries to process to obtain them, we have to start flushing 13416 * the dirty vnodes to force the release of additional requests 13417 * to the worklist that we can then process to reap addition 13418 * resources. We walk the vnodes associated with the mount point 13419 * until we get the needed worklist requests that we can reap. 13420 * 13421 * If there are several threads all needing to clean the same 13422 * mount point, only one is allowed to walk the mount list. 13423 * When several threads all try to walk the same mount list, 13424 * they end up competing with each other and often end up in 13425 * livelock. This approach ensures that forward progress is 13426 * made at the cost of occational ENOSPC errors being returned 13427 * that might otherwise have been avoided. 13428 */ 13429 error = 1; 13430 if ((resource == FLUSH_BLOCKS_WAIT && 13431 fs->fs_cstotal.cs_nbfree <= needed) || 13432 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 && 13433 fs->fs_cstotal.cs_nifree <= needed)) { 13434 ACQUIRE_LOCK(ump); 13435 if ((ump->um_softdep->sd_flags & FLUSH_RC_ACTIVE) == 0) { 13436 ump->um_softdep->sd_flags |= FLUSH_RC_ACTIVE; 13437 FREE_LOCK(ump); 13438 failed_vnode = softdep_request_cleanup_flush(mp, ump); 13439 ACQUIRE_LOCK(ump); 13440 ump->um_softdep->sd_flags &= ~FLUSH_RC_ACTIVE; 13441 FREE_LOCK(ump); 13442 if (ump->softdep_on_worklist > 0) { 13443 stat_cleanup_retries += 1; 13444 if (!failed_vnode) 13445 goto retry; 13446 } 13447 } else { 13448 FREE_LOCK(ump); 13449 error = 0; 13450 } 13451 stat_cleanup_failures += 1; 13452 } 13453 if (time_second - starttime > stat_cleanup_high_delay) 13454 stat_cleanup_high_delay = time_second - starttime; 13455 UFS_LOCK(ump); 13456 return (error); 13457} 13458 13459/* 13460 * Scan the vnodes for the specified mount point flushing out any 13461 * vnodes that can be locked without waiting. Finally, try to flush 13462 * the device associated with the mount point if it can be locked 13463 * without waiting. 13464 * 13465 * We return 0 if we were able to lock every vnode in our scan. 13466 * If we had to skip one or more vnodes, we return 1. 13467 */ 13468static int 13469softdep_request_cleanup_flush(mp, ump) 13470 struct mount *mp; 13471 struct ufsmount *ump; 13472{ 13473 struct thread *td; 13474 struct vnode *lvp, *mvp; 13475 int failed_vnode; 13476 13477 failed_vnode = 0; 13478 td = curthread; 13479 MNT_VNODE_FOREACH_ALL(lvp, mp, mvp) { 13480 if (TAILQ_FIRST(&lvp->v_bufobj.bo_dirty.bv_hd) == 0) { 13481 VI_UNLOCK(lvp); 13482 continue; 13483 } 13484 if (vget(lvp, LK_EXCLUSIVE | LK_INTERLOCK | LK_NOWAIT, 13485 td) != 0) { 13486 failed_vnode = 1; 13487 continue; 13488 } 13489 if (lvp->v_vflag & VV_NOSYNC) { /* unlinked */ 13490 vput(lvp); 13491 continue; 13492 } 13493 (void) ffs_syncvnode(lvp, MNT_NOWAIT, 0); 13494 vput(lvp); 13495 } 13496 lvp = ump->um_devvp; 13497 if (vn_lock(lvp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 13498 VOP_FSYNC(lvp, MNT_NOWAIT, td); 13499 VOP_UNLOCK(lvp, 0); 13500 } 13501 return (failed_vnode); 13502} 13503 13504static bool 13505softdep_excess_items(struct ufsmount *ump, int item) 13506{ 13507 13508 KASSERT(item >= 0 && item < D_LAST, ("item %d", item)); 13509 return (dep_current[item] > max_softdeps && 13510 ump->softdep_curdeps[item] > max_softdeps / 13511 stat_flush_threads); 13512} 13513 13514static void 13515schedule_cleanup(struct mount *mp) 13516{ 13517 struct ufsmount *ump; 13518 struct thread *td; 13519 13520 ump = VFSTOUFS(mp); 13521 LOCK_OWNED(ump); 13522 FREE_LOCK(ump); 13523 td = curthread; 13524 if ((td->td_pflags & TDP_KTHREAD) != 0 && 13525 (td->td_proc->p_flag2 & P2_AST_SU) == 0) { 13526 /* 13527 * No ast is delivered to kernel threads, so nobody 13528 * would deref the mp. Some kernel threads 13529 * explicitely check for AST, e.g. NFS daemon does 13530 * this in the serving loop. 13531 */ 13532 return; 13533 } 13534 if (td->td_su != NULL) 13535 vfs_rel(td->td_su); 13536 vfs_ref(mp); 13537 td->td_su = mp; 13538 thread_lock(td); 13539 td->td_flags |= TDF_ASTPENDING; 13540 thread_unlock(td); 13541} 13542 13543static void 13544softdep_ast_cleanup_proc(struct thread *td) 13545{ 13546 struct mount *mp; 13547 struct ufsmount *ump; 13548 int error; 13549 bool req; 13550 13551 while ((mp = td->td_su) != NULL) { 13552 td->td_su = NULL; 13553 error = vfs_busy(mp, MBF_NOWAIT); 13554 vfs_rel(mp); 13555 if (error != 0) 13556 return; 13557 if (ffs_own_mount(mp) && MOUNTEDSOFTDEP(mp)) { 13558 ump = VFSTOUFS(mp); 13559 for (;;) { 13560 req = false; 13561 ACQUIRE_LOCK(ump); 13562 if (softdep_excess_items(ump, D_INODEDEP)) { 13563 req = true; 13564 request_cleanup(mp, FLUSH_INODES); 13565 } 13566 if (softdep_excess_items(ump, D_DIRREM)) { 13567 req = true; 13568 request_cleanup(mp, FLUSH_BLOCKS); 13569 } 13570 FREE_LOCK(ump); 13571 if (softdep_excess_items(ump, D_NEWBLK) || 13572 softdep_excess_items(ump, D_ALLOCDIRECT) || 13573 softdep_excess_items(ump, D_ALLOCINDIR)) { 13574 error = vn_start_write(NULL, &mp, 13575 V_WAIT); 13576 if (error == 0) { 13577 req = true; 13578 VFS_SYNC(mp, MNT_WAIT); 13579 vn_finished_write(mp); 13580 } 13581 } 13582 if ((td->td_pflags & TDP_KTHREAD) != 0 || !req) 13583 break; 13584 } 13585 } 13586 vfs_unbusy(mp); 13587 } 13588 if ((mp = td->td_su) != NULL) { 13589 td->td_su = NULL; 13590 vfs_rel(mp); 13591 } 13592} 13593 13594/* 13595 * If memory utilization has gotten too high, deliberately slow things 13596 * down and speed up the I/O processing. 13597 */ 13598static int 13599request_cleanup(mp, resource) 13600 struct mount *mp; 13601 int resource; 13602{ 13603 struct thread *td = curthread; 13604 struct ufsmount *ump; 13605 13606 ump = VFSTOUFS(mp); 13607 LOCK_OWNED(ump); 13608 /* 13609 * We never hold up the filesystem syncer or buf daemon. 13610 */ 13611 if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF)) 13612 return (0); 13613 /* 13614 * First check to see if the work list has gotten backlogged. 13615 * If it has, co-opt this process to help clean up two entries. 13616 * Because this process may hold inodes locked, we cannot 13617 * handle any remove requests that might block on a locked 13618 * inode as that could lead to deadlock. We set TDP_SOFTDEP 13619 * to avoid recursively processing the worklist. 13620 */ 13621 if (ump->softdep_on_worklist > max_softdeps / 10) { 13622 td->td_pflags |= TDP_SOFTDEP; 13623 process_worklist_item(mp, 2, LK_NOWAIT); 13624 td->td_pflags &= ~TDP_SOFTDEP; 13625 stat_worklist_push += 2; 13626 return(1); 13627 } 13628 /* 13629 * Next, we attempt to speed up the syncer process. If that 13630 * is successful, then we allow the process to continue. 13631 */ 13632 if (softdep_speedup(ump) && 13633 resource != FLUSH_BLOCKS_WAIT && 13634 resource != FLUSH_INODES_WAIT) 13635 return(0); 13636 /* 13637 * If we are resource constrained on inode dependencies, try 13638 * flushing some dirty inodes. Otherwise, we are constrained 13639 * by file deletions, so try accelerating flushes of directories 13640 * with removal dependencies. We would like to do the cleanup 13641 * here, but we probably hold an inode locked at this point and 13642 * that might deadlock against one that we try to clean. So, 13643 * the best that we can do is request the syncer daemon to do 13644 * the cleanup for us. 13645 */ 13646 switch (resource) { 13647 13648 case FLUSH_INODES: 13649 case FLUSH_INODES_WAIT: 13650 ACQUIRE_GBLLOCK(&lk); 13651 stat_ino_limit_push += 1; 13652 req_clear_inodedeps += 1; 13653 FREE_GBLLOCK(&lk); 13654 stat_countp = &stat_ino_limit_hit; 13655 break; 13656 13657 case FLUSH_BLOCKS: 13658 case FLUSH_BLOCKS_WAIT: 13659 ACQUIRE_GBLLOCK(&lk); 13660 stat_blk_limit_push += 1; 13661 req_clear_remove += 1; 13662 FREE_GBLLOCK(&lk); 13663 stat_countp = &stat_blk_limit_hit; 13664 break; 13665 13666 default: 13667 panic("request_cleanup: unknown type"); 13668 } 13669 /* 13670 * Hopefully the syncer daemon will catch up and awaken us. 13671 * We wait at most tickdelay before proceeding in any case. 13672 */ 13673 ACQUIRE_GBLLOCK(&lk); 13674 FREE_LOCK(ump); 13675 proc_waiting += 1; 13676 if (callout_pending(&softdep_callout) == FALSE) 13677 callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2, 13678 pause_timer, 0); 13679 13680 if ((td->td_pflags & TDP_KTHREAD) == 0) 13681 msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0); 13682 proc_waiting -= 1; 13683 FREE_GBLLOCK(&lk); 13684 ACQUIRE_LOCK(ump); 13685 return (1); 13686} 13687 13688/* 13689 * Awaken processes pausing in request_cleanup and clear proc_waiting 13690 * to indicate that there is no longer a timer running. Pause_timer 13691 * will be called with the global softdep mutex (&lk) locked. 13692 */ 13693static void 13694pause_timer(arg) 13695 void *arg; 13696{ 13697 13698 GBLLOCK_OWNED(&lk); 13699 /* 13700 * The callout_ API has acquired mtx and will hold it around this 13701 * function call. 13702 */ 13703 *stat_countp += proc_waiting; 13704 wakeup(&proc_waiting); 13705} 13706 13707/* 13708 * If requested, try removing inode or removal dependencies. 13709 */ 13710static void 13711check_clear_deps(mp) 13712 struct mount *mp; 13713{ 13714 13715 /* 13716 * If we are suspended, it may be because of our using 13717 * too many inodedeps, so help clear them out. 13718 */ 13719 if (MOUNTEDSUJ(mp) && VFSTOUFS(mp)->softdep_jblocks->jb_suspended) 13720 clear_inodedeps(mp); 13721 /* 13722 * General requests for cleanup of backed up dependencies 13723 */ 13724 ACQUIRE_GBLLOCK(&lk); 13725 if (req_clear_inodedeps) { 13726 req_clear_inodedeps -= 1; 13727 FREE_GBLLOCK(&lk); 13728 clear_inodedeps(mp); 13729 ACQUIRE_GBLLOCK(&lk); 13730 wakeup(&proc_waiting); 13731 } 13732 if (req_clear_remove) { 13733 req_clear_remove -= 1; 13734 FREE_GBLLOCK(&lk); 13735 clear_remove(mp); 13736 ACQUIRE_GBLLOCK(&lk); 13737 wakeup(&proc_waiting); 13738 } 13739 FREE_GBLLOCK(&lk); 13740} 13741 13742/* 13743 * Flush out a directory with at least one removal dependency in an effort to 13744 * reduce the number of dirrem, freefile, and freeblks dependency structures. 13745 */ 13746static void 13747clear_remove(mp) 13748 struct mount *mp; 13749{ 13750 struct pagedep_hashhead *pagedephd; 13751 struct pagedep *pagedep; 13752 struct ufsmount *ump; 13753 struct vnode *vp; 13754 struct bufobj *bo; 13755 int error, cnt; 13756 ino_t ino; 13757 13758 ump = VFSTOUFS(mp); 13759 LOCK_OWNED(ump); 13760 13761 for (cnt = 0; cnt <= ump->pagedep_hash_size; cnt++) { 13762 pagedephd = &ump->pagedep_hashtbl[ump->pagedep_nextclean++]; 13763 if (ump->pagedep_nextclean > ump->pagedep_hash_size) 13764 ump->pagedep_nextclean = 0; 13765 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 13766 if (LIST_EMPTY(&pagedep->pd_dirremhd)) 13767 continue; 13768 ino = pagedep->pd_ino; 13769 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 13770 continue; 13771 FREE_LOCK(ump); 13772 13773 /* 13774 * Let unmount clear deps 13775 */ 13776 error = vfs_busy(mp, MBF_NOWAIT); 13777 if (error != 0) 13778 goto finish_write; 13779 error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp, 13780 FFSV_FORCEINSMQ); 13781 vfs_unbusy(mp); 13782 if (error != 0) { 13783 softdep_error("clear_remove: vget", error); 13784 goto finish_write; 13785 } 13786 MPASS(VTOI(vp)->i_mode != 0); 13787 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0))) 13788 softdep_error("clear_remove: fsync", error); 13789 bo = &vp->v_bufobj; 13790 BO_LOCK(bo); 13791 drain_output(vp); 13792 BO_UNLOCK(bo); 13793 vput(vp); 13794 finish_write: 13795 vn_finished_write(mp); 13796 ACQUIRE_LOCK(ump); 13797 return; 13798 } 13799 } 13800} 13801 13802/* 13803 * Clear out a block of dirty inodes in an effort to reduce 13804 * the number of inodedep dependency structures. 13805 */ 13806static void 13807clear_inodedeps(mp) 13808 struct mount *mp; 13809{ 13810 struct inodedep_hashhead *inodedephd; 13811 struct inodedep *inodedep; 13812 struct ufsmount *ump; 13813 struct vnode *vp; 13814 struct fs *fs; 13815 int error, cnt; 13816 ino_t firstino, lastino, ino; 13817 13818 ump = VFSTOUFS(mp); 13819 fs = ump->um_fs; 13820 LOCK_OWNED(ump); 13821 /* 13822 * Pick a random inode dependency to be cleared. 13823 * We will then gather up all the inodes in its block 13824 * that have dependencies and flush them out. 13825 */ 13826 for (cnt = 0; cnt <= ump->inodedep_hash_size; cnt++) { 13827 inodedephd = &ump->inodedep_hashtbl[ump->inodedep_nextclean++]; 13828 if (ump->inodedep_nextclean > ump->inodedep_hash_size) 13829 ump->inodedep_nextclean = 0; 13830 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 13831 break; 13832 } 13833 if (inodedep == NULL) 13834 return; 13835 /* 13836 * Find the last inode in the block with dependencies. 13837 */ 13838 firstino = rounddown2(inodedep->id_ino, INOPB(fs)); 13839 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 13840 if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0) 13841 break; 13842 /* 13843 * Asynchronously push all but the last inode with dependencies. 13844 * Synchronously push the last inode with dependencies to ensure 13845 * that the inode block gets written to free up the inodedeps. 13846 */ 13847 for (ino = firstino; ino <= lastino; ino++) { 13848 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 13849 continue; 13850 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 13851 continue; 13852 FREE_LOCK(ump); 13853 error = vfs_busy(mp, MBF_NOWAIT); /* Let unmount clear deps */ 13854 if (error != 0) { 13855 vn_finished_write(mp); 13856 ACQUIRE_LOCK(ump); 13857 return; 13858 } 13859 if ((error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp, 13860 FFSV_FORCEINSMQ)) != 0) { 13861 softdep_error("clear_inodedeps: vget", error); 13862 vfs_unbusy(mp); 13863 vn_finished_write(mp); 13864 ACQUIRE_LOCK(ump); 13865 return; 13866 } 13867 vfs_unbusy(mp); 13868 if (VTOI(vp)->i_mode == 0) { 13869 vgone(vp); 13870 } else if (ino == lastino) { 13871 if ((error = ffs_syncvnode(vp, MNT_WAIT, 0))) 13872 softdep_error("clear_inodedeps: fsync1", error); 13873 } else { 13874 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0))) 13875 softdep_error("clear_inodedeps: fsync2", error); 13876 BO_LOCK(&vp->v_bufobj); 13877 drain_output(vp); 13878 BO_UNLOCK(&vp->v_bufobj); 13879 } 13880 vput(vp); 13881 vn_finished_write(mp); 13882 ACQUIRE_LOCK(ump); 13883 } 13884} 13885 13886void 13887softdep_buf_append(bp, wkhd) 13888 struct buf *bp; 13889 struct workhead *wkhd; 13890{ 13891 struct worklist *wk; 13892 struct ufsmount *ump; 13893 13894 if ((wk = LIST_FIRST(wkhd)) == NULL) 13895 return; 13896 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0, 13897 ("softdep_buf_append called on non-softdep filesystem")); 13898 ump = VFSTOUFS(wk->wk_mp); 13899 ACQUIRE_LOCK(ump); 13900 while ((wk = LIST_FIRST(wkhd)) != NULL) { 13901 WORKLIST_REMOVE(wk); 13902 WORKLIST_INSERT(&bp->b_dep, wk); 13903 } 13904 FREE_LOCK(ump); 13905 13906} 13907 13908void 13909softdep_inode_append(ip, cred, wkhd) 13910 struct inode *ip; 13911 struct ucred *cred; 13912 struct workhead *wkhd; 13913{ 13914 struct buf *bp; 13915 struct fs *fs; 13916 struct ufsmount *ump; 13917 int error; 13918 13919 ump = ITOUMP(ip); 13920 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 13921 ("softdep_inode_append called on non-softdep filesystem")); 13922 fs = ump->um_fs; 13923 error = bread(ump->um_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 13924 (int)fs->fs_bsize, cred, &bp); 13925 if (error) { 13926 bqrelse(bp); 13927 softdep_freework(wkhd); 13928 return; 13929 } 13930 softdep_buf_append(bp, wkhd); 13931 bqrelse(bp); 13932} 13933 13934void 13935softdep_freework(wkhd) 13936 struct workhead *wkhd; 13937{ 13938 struct worklist *wk; 13939 struct ufsmount *ump; 13940 13941 if ((wk = LIST_FIRST(wkhd)) == NULL) 13942 return; 13943 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0, 13944 ("softdep_freework called on non-softdep filesystem")); 13945 ump = VFSTOUFS(wk->wk_mp); 13946 ACQUIRE_LOCK(ump); 13947 handle_jwork(wkhd); 13948 FREE_LOCK(ump); 13949} 13950 13951static struct ufsmount * 13952softdep_bp_to_mp(bp) 13953 struct buf *bp; 13954{ 13955 struct mount *mp; 13956 struct vnode *vp; 13957 13958 if (LIST_EMPTY(&bp->b_dep)) 13959 return (NULL); 13960 vp = bp->b_vp; 13961 KASSERT(vp != NULL, 13962 ("%s, buffer with dependencies lacks vnode", __func__)); 13963 13964 /* 13965 * The ump mount point is stable after we get a correct 13966 * pointer, since bp is locked and this prevents unmount from 13967 * proceeding. But to get to it, we cannot dereference bp->b_dep 13968 * head wk_mp, because we do not yet own SU ump lock and 13969 * workitem might be freed while dereferenced. 13970 */ 13971retry: 13972 switch (vp->v_type) { 13973 case VCHR: 13974 VI_LOCK(vp); 13975 mp = vp->v_type == VCHR ? vp->v_rdev->si_mountpt : NULL; 13976 VI_UNLOCK(vp); 13977 if (mp == NULL) 13978 goto retry; 13979 break; 13980 case VREG: 13981 case VDIR: 13982 case VLNK: 13983 case VFIFO: 13984 case VSOCK: 13985 mp = vp->v_mount; 13986 break; 13987 case VBLK: 13988 vn_printf(vp, "softdep_bp_to_mp: unexpected block device\n"); 13989 /* FALLTHROUGH */ 13990 case VNON: 13991 case VBAD: 13992 case VMARKER: 13993 mp = NULL; 13994 break; 13995 default: 13996 vn_printf(vp, "unknown vnode type"); 13997 mp = NULL; 13998 break; 13999 } 14000 return (VFSTOUFS(mp)); 14001} 14002 14003/* 14004 * Function to determine if the buffer has outstanding dependencies 14005 * that will cause a roll-back if the buffer is written. If wantcount 14006 * is set, return number of dependencies, otherwise just yes or no. 14007 */ 14008static int 14009softdep_count_dependencies(bp, wantcount) 14010 struct buf *bp; 14011 int wantcount; 14012{ 14013 struct worklist *wk; 14014 struct ufsmount *ump; 14015 struct bmsafemap *bmsafemap; 14016 struct freework *freework; 14017 struct inodedep *inodedep; 14018 struct indirdep *indirdep; 14019 struct freeblks *freeblks; 14020 struct allocindir *aip; 14021 struct pagedep *pagedep; 14022 struct dirrem *dirrem; 14023 struct newblk *newblk; 14024 struct mkdir *mkdir; 14025 struct diradd *dap; 14026 int i, retval; 14027 14028 ump = softdep_bp_to_mp(bp); 14029 if (ump == NULL) 14030 return (0); 14031 retval = 0; 14032 ACQUIRE_LOCK(ump); 14033 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 14034 switch (wk->wk_type) { 14035 14036 case D_INODEDEP: 14037 inodedep = WK_INODEDEP(wk); 14038 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 14039 /* bitmap allocation dependency */ 14040 retval += 1; 14041 if (!wantcount) 14042 goto out; 14043 } 14044 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 14045 /* direct block pointer dependency */ 14046 retval += 1; 14047 if (!wantcount) 14048 goto out; 14049 } 14050 if (TAILQ_FIRST(&inodedep->id_extupdt)) { 14051 /* direct block pointer dependency */ 14052 retval += 1; 14053 if (!wantcount) 14054 goto out; 14055 } 14056 if (TAILQ_FIRST(&inodedep->id_inoreflst)) { 14057 /* Add reference dependency. */ 14058 retval += 1; 14059 if (!wantcount) 14060 goto out; 14061 } 14062 continue; 14063 14064 case D_INDIRDEP: 14065 indirdep = WK_INDIRDEP(wk); 14066 14067 TAILQ_FOREACH(freework, &indirdep->ir_trunc, fw_next) { 14068 /* indirect truncation dependency */ 14069 retval += 1; 14070 if (!wantcount) 14071 goto out; 14072 } 14073 14074 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 14075 /* indirect block pointer dependency */ 14076 retval += 1; 14077 if (!wantcount) 14078 goto out; 14079 } 14080 continue; 14081 14082 case D_PAGEDEP: 14083 pagedep = WK_PAGEDEP(wk); 14084 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 14085 if (LIST_FIRST(&dirrem->dm_jremrefhd)) { 14086 /* Journal remove ref dependency. */ 14087 retval += 1; 14088 if (!wantcount) 14089 goto out; 14090 } 14091 } 14092 for (i = 0; i < DAHASHSZ; i++) { 14093 14094 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 14095 /* directory entry dependency */ 14096 retval += 1; 14097 if (!wantcount) 14098 goto out; 14099 } 14100 } 14101 continue; 14102 14103 case D_BMSAFEMAP: 14104 bmsafemap = WK_BMSAFEMAP(wk); 14105 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd)) { 14106 /* Add reference dependency. */ 14107 retval += 1; 14108 if (!wantcount) 14109 goto out; 14110 } 14111 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd)) { 14112 /* Allocate block dependency. */ 14113 retval += 1; 14114 if (!wantcount) 14115 goto out; 14116 } 14117 continue; 14118 14119 case D_FREEBLKS: 14120 freeblks = WK_FREEBLKS(wk); 14121 if (LIST_FIRST(&freeblks->fb_jblkdephd)) { 14122 /* Freeblk journal dependency. */ 14123 retval += 1; 14124 if (!wantcount) 14125 goto out; 14126 } 14127 continue; 14128 14129 case D_ALLOCDIRECT: 14130 case D_ALLOCINDIR: 14131 newblk = WK_NEWBLK(wk); 14132 if (newblk->nb_jnewblk) { 14133 /* Journal allocate dependency. */ 14134 retval += 1; 14135 if (!wantcount) 14136 goto out; 14137 } 14138 continue; 14139 14140 case D_MKDIR: 14141 mkdir = WK_MKDIR(wk); 14142 if (mkdir->md_jaddref) { 14143 /* Journal reference dependency. */ 14144 retval += 1; 14145 if (!wantcount) 14146 goto out; 14147 } 14148 continue; 14149 14150 case D_FREEWORK: 14151 case D_FREEDEP: 14152 case D_JSEGDEP: 14153 case D_JSEG: 14154 case D_SBDEP: 14155 /* never a dependency on these blocks */ 14156 continue; 14157 14158 default: 14159 panic("softdep_count_dependencies: Unexpected type %s", 14160 TYPENAME(wk->wk_type)); 14161 /* NOTREACHED */ 14162 } 14163 } 14164out: 14165 FREE_LOCK(ump); 14166 return (retval); 14167} 14168 14169/* 14170 * Acquire exclusive access to a buffer. 14171 * Must be called with a locked mtx parameter. 14172 * Return acquired buffer or NULL on failure. 14173 */ 14174static struct buf * 14175getdirtybuf(bp, lock, waitfor) 14176 struct buf *bp; 14177 struct rwlock *lock; 14178 int waitfor; 14179{ 14180 int error; 14181 14182 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) { 14183 if (waitfor != MNT_WAIT) 14184 return (NULL); 14185 error = BUF_LOCK(bp, 14186 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, lock); 14187 /* 14188 * Even if we successfully acquire bp here, we have dropped 14189 * lock, which may violates our guarantee. 14190 */ 14191 if (error == 0) 14192 BUF_UNLOCK(bp); 14193 else if (error != ENOLCK) 14194 panic("getdirtybuf: inconsistent lock: %d", error); 14195 rw_wlock(lock); 14196 return (NULL); 14197 } 14198 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 14199 if (lock != BO_LOCKPTR(bp->b_bufobj) && waitfor == MNT_WAIT) { 14200 rw_wunlock(lock); 14201 BO_LOCK(bp->b_bufobj); 14202 BUF_UNLOCK(bp); 14203 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 14204 bp->b_vflags |= BV_BKGRDWAIT; 14205 msleep(&bp->b_xflags, BO_LOCKPTR(bp->b_bufobj), 14206 PRIBIO | PDROP, "getbuf", 0); 14207 } else 14208 BO_UNLOCK(bp->b_bufobj); 14209 rw_wlock(lock); 14210 return (NULL); 14211 } 14212 BUF_UNLOCK(bp); 14213 if (waitfor != MNT_WAIT) 14214 return (NULL); 14215#ifdef DEBUG_VFS_LOCKS 14216 if (bp->b_vp->v_type != VCHR) 14217 ASSERT_BO_WLOCKED(bp->b_bufobj); 14218#endif 14219 bp->b_vflags |= BV_BKGRDWAIT; 14220 rw_sleep(&bp->b_xflags, lock, PRIBIO, "getbuf", 0); 14221 return (NULL); 14222 } 14223 if ((bp->b_flags & B_DELWRI) == 0) { 14224 BUF_UNLOCK(bp); 14225 return (NULL); 14226 } 14227 bremfree(bp); 14228 return (bp); 14229} 14230 14231 14232/* 14233 * Check if it is safe to suspend the file system now. On entry, 14234 * the vnode interlock for devvp should be held. Return 0 with 14235 * the mount interlock held if the file system can be suspended now, 14236 * otherwise return EAGAIN with the mount interlock held. 14237 */ 14238int 14239softdep_check_suspend(struct mount *mp, 14240 struct vnode *devvp, 14241 int softdep_depcnt, 14242 int softdep_accdepcnt, 14243 int secondary_writes, 14244 int secondary_accwrites) 14245{ 14246 struct bufobj *bo; 14247 struct ufsmount *ump; 14248 struct inodedep *inodedep; 14249 int error, unlinked; 14250 14251 bo = &devvp->v_bufobj; 14252 ASSERT_BO_WLOCKED(bo); 14253 14254 /* 14255 * If we are not running with soft updates, then we need only 14256 * deal with secondary writes as we try to suspend. 14257 */ 14258 if (MOUNTEDSOFTDEP(mp) == 0) { 14259 MNT_ILOCK(mp); 14260 while (mp->mnt_secondary_writes != 0) { 14261 BO_UNLOCK(bo); 14262 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp), 14263 (PUSER - 1) | PDROP, "secwr", 0); 14264 BO_LOCK(bo); 14265 MNT_ILOCK(mp); 14266 } 14267 14268 /* 14269 * Reasons for needing more work before suspend: 14270 * - Dirty buffers on devvp. 14271 * - Secondary writes occurred after start of vnode sync loop 14272 */ 14273 error = 0; 14274 if (bo->bo_numoutput > 0 || 14275 bo->bo_dirty.bv_cnt > 0 || 14276 secondary_writes != 0 || 14277 mp->mnt_secondary_writes != 0 || 14278 secondary_accwrites != mp->mnt_secondary_accwrites) 14279 error = EAGAIN; 14280 BO_UNLOCK(bo); 14281 return (error); 14282 } 14283 14284 /* 14285 * If we are running with soft updates, then we need to coordinate 14286 * with them as we try to suspend. 14287 */ 14288 ump = VFSTOUFS(mp); 14289 for (;;) { 14290 if (!TRY_ACQUIRE_LOCK(ump)) { 14291 BO_UNLOCK(bo); 14292 ACQUIRE_LOCK(ump); 14293 FREE_LOCK(ump); 14294 BO_LOCK(bo); 14295 continue; 14296 } 14297 MNT_ILOCK(mp); 14298 if (mp->mnt_secondary_writes != 0) { 14299 FREE_LOCK(ump); 14300 BO_UNLOCK(bo); 14301 msleep(&mp->mnt_secondary_writes, 14302 MNT_MTX(mp), 14303 (PUSER - 1) | PDROP, "secwr", 0); 14304 BO_LOCK(bo); 14305 continue; 14306 } 14307 break; 14308 } 14309 14310 unlinked = 0; 14311 if (MOUNTEDSUJ(mp)) { 14312 for (inodedep = TAILQ_FIRST(&ump->softdep_unlinked); 14313 inodedep != NULL; 14314 inodedep = TAILQ_NEXT(inodedep, id_unlinked)) { 14315 if ((inodedep->id_state & (UNLINKED | UNLINKLINKS | 14316 UNLINKONLIST)) != (UNLINKED | UNLINKLINKS | 14317 UNLINKONLIST) || 14318 !check_inodedep_free(inodedep)) 14319 continue; 14320 unlinked++; 14321 } 14322 } 14323 14324 /* 14325 * Reasons for needing more work before suspend: 14326 * - Dirty buffers on devvp. 14327 * - Softdep activity occurred after start of vnode sync loop 14328 * - Secondary writes occurred after start of vnode sync loop 14329 */ 14330 error = 0; 14331 if (bo->bo_numoutput > 0 || 14332 bo->bo_dirty.bv_cnt > 0 || 14333 softdep_depcnt != unlinked || 14334 ump->softdep_deps != unlinked || 14335 softdep_accdepcnt != ump->softdep_accdeps || 14336 secondary_writes != 0 || 14337 mp->mnt_secondary_writes != 0 || 14338 secondary_accwrites != mp->mnt_secondary_accwrites) 14339 error = EAGAIN; 14340 FREE_LOCK(ump); 14341 BO_UNLOCK(bo); 14342 return (error); 14343} 14344 14345 14346/* 14347 * Get the number of dependency structures for the file system, both 14348 * the current number and the total number allocated. These will 14349 * later be used to detect that softdep processing has occurred. 14350 */ 14351void 14352softdep_get_depcounts(struct mount *mp, 14353 int *softdep_depsp, 14354 int *softdep_accdepsp) 14355{ 14356 struct ufsmount *ump; 14357 14358 if (MOUNTEDSOFTDEP(mp) == 0) { 14359 *softdep_depsp = 0; 14360 *softdep_accdepsp = 0; 14361 return; 14362 } 14363 ump = VFSTOUFS(mp); 14364 ACQUIRE_LOCK(ump); 14365 *softdep_depsp = ump->softdep_deps; 14366 *softdep_accdepsp = ump->softdep_accdeps; 14367 FREE_LOCK(ump); 14368} 14369 14370/* 14371 * Wait for pending output on a vnode to complete. 14372 */ 14373static void 14374drain_output(vp) 14375 struct vnode *vp; 14376{ 14377 14378 ASSERT_VOP_LOCKED(vp, "drain_output"); 14379 (void)bufobj_wwait(&vp->v_bufobj, 0, 0); 14380} 14381 14382/* 14383 * Called whenever a buffer that is being invalidated or reallocated 14384 * contains dependencies. This should only happen if an I/O error has 14385 * occurred. The routine is called with the buffer locked. 14386 */ 14387static void 14388softdep_deallocate_dependencies(bp) 14389 struct buf *bp; 14390{ 14391 14392 if ((bp->b_ioflags & BIO_ERROR) == 0) 14393 panic("softdep_deallocate_dependencies: dangling deps"); 14394 if (bp->b_vp != NULL && bp->b_vp->v_mount != NULL) 14395 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); 14396 else 14397 printf("softdep_deallocate_dependencies: " 14398 "got error %d while accessing filesystem\n", bp->b_error); 14399 if (bp->b_error != ENXIO) 14400 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 14401} 14402 14403/* 14404 * Function to handle asynchronous write errors in the filesystem. 14405 */ 14406static void 14407softdep_error(func, error) 14408 char *func; 14409 int error; 14410{ 14411 14412 /* XXX should do something better! */ 14413 printf("%s: got error %d while accessing filesystem\n", func, error); 14414} 14415 14416#ifdef DDB 14417 14418static void 14419inodedep_print(struct inodedep *inodedep, int verbose) 14420{ 14421 db_printf("%p fs %p st %x ino %jd inoblk %jd delta %jd nlink %jd" 14422 " saveino %p\n", 14423 inodedep, inodedep->id_fs, inodedep->id_state, 14424 (intmax_t)inodedep->id_ino, 14425 (intmax_t)fsbtodb(inodedep->id_fs, 14426 ino_to_fsba(inodedep->id_fs, inodedep->id_ino)), 14427 (intmax_t)inodedep->id_nlinkdelta, 14428 (intmax_t)inodedep->id_savednlink, 14429 inodedep->id_savedino1); 14430 14431 if (verbose == 0) 14432 return; 14433 14434 db_printf("\tpendinghd %p, bufwait %p, inowait %p, inoreflst %p, " 14435 "mkdiradd %p\n", 14436 LIST_FIRST(&inodedep->id_pendinghd), 14437 LIST_FIRST(&inodedep->id_bufwait), 14438 LIST_FIRST(&inodedep->id_inowait), 14439 TAILQ_FIRST(&inodedep->id_inoreflst), 14440 inodedep->id_mkdiradd); 14441 db_printf("\tinoupdt %p, newinoupdt %p, extupdt %p, newextupdt %p\n", 14442 TAILQ_FIRST(&inodedep->id_inoupdt), 14443 TAILQ_FIRST(&inodedep->id_newinoupdt), 14444 TAILQ_FIRST(&inodedep->id_extupdt), 14445 TAILQ_FIRST(&inodedep->id_newextupdt)); 14446} 14447 14448DB_SHOW_COMMAND(inodedep, db_show_inodedep) 14449{ 14450 14451 if (have_addr == 0) { 14452 db_printf("Address required\n"); 14453 return; 14454 } 14455 inodedep_print((struct inodedep*)addr, 1); 14456} 14457 14458DB_SHOW_COMMAND(inodedeps, db_show_inodedeps) 14459{ 14460 struct inodedep_hashhead *inodedephd; 14461 struct inodedep *inodedep; 14462 struct ufsmount *ump; 14463 int cnt; 14464 14465 if (have_addr == 0) { 14466 db_printf("Address required\n"); 14467 return; 14468 } 14469 ump = (struct ufsmount *)addr; 14470 for (cnt = 0; cnt < ump->inodedep_hash_size; cnt++) { 14471 inodedephd = &ump->inodedep_hashtbl[cnt]; 14472 LIST_FOREACH(inodedep, inodedephd, id_hash) { 14473 inodedep_print(inodedep, 0); 14474 } 14475 } 14476} 14477 14478DB_SHOW_COMMAND(worklist, db_show_worklist) 14479{ 14480 struct worklist *wk; 14481 14482 if (have_addr == 0) { 14483 db_printf("Address required\n"); 14484 return; 14485 } 14486 wk = (struct worklist *)addr; 14487 printf("worklist: %p type %s state 0x%X\n", 14488 wk, TYPENAME(wk->wk_type), wk->wk_state); 14489} 14490 14491DB_SHOW_COMMAND(workhead, db_show_workhead) 14492{ 14493 struct workhead *wkhd; 14494 struct worklist *wk; 14495 int i; 14496 14497 if (have_addr == 0) { 14498 db_printf("Address required\n"); 14499 return; 14500 } 14501 wkhd = (struct workhead *)addr; 14502 wk = LIST_FIRST(wkhd); 14503 for (i = 0; i < 100 && wk != NULL; i++, wk = LIST_NEXT(wk, wk_list)) 14504 db_printf("worklist: %p type %s state 0x%X", 14505 wk, TYPENAME(wk->wk_type), wk->wk_state); 14506 if (i == 100) 14507 db_printf("workhead overflow"); 14508 printf("\n"); 14509} 14510 14511 14512DB_SHOW_COMMAND(mkdirs, db_show_mkdirs) 14513{ 14514 struct mkdirlist *mkdirlisthd; 14515 struct jaddref *jaddref; 14516 struct diradd *diradd; 14517 struct mkdir *mkdir; 14518 14519 if (have_addr == 0) { 14520 db_printf("Address required\n"); 14521 return; 14522 } 14523 mkdirlisthd = (struct mkdirlist *)addr; 14524 LIST_FOREACH(mkdir, mkdirlisthd, md_mkdirs) { 14525 diradd = mkdir->md_diradd; 14526 db_printf("mkdir: %p state 0x%X dap %p state 0x%X", 14527 mkdir, mkdir->md_state, diradd, diradd->da_state); 14528 if ((jaddref = mkdir->md_jaddref) != NULL) 14529 db_printf(" jaddref %p jaddref state 0x%X", 14530 jaddref, jaddref->ja_state); 14531 db_printf("\n"); 14532 } 14533} 14534 14535/* exported to ffs_vfsops.c */ 14536extern void db_print_ffs(struct ufsmount *ump); 14537void 14538db_print_ffs(struct ufsmount *ump) 14539{ 14540 db_printf("mp %p %s devvp %p fs %p su_wl %d su_deps %d su_req %d\n", 14541 ump->um_mountp, ump->um_mountp->mnt_stat.f_mntonname, 14542 ump->um_devvp, ump->um_fs, ump->softdep_on_worklist, 14543 ump->softdep_deps, ump->softdep_req); 14544} 14545 14546#endif /* DDB */ 14547 14548#endif /* SOFTUPDATES */ 14549