ffs_softdep.c revision 278257
1/*- 2 * Copyright 1998, 2000 Marshall Kirk McKusick. 3 * Copyright 2009, 2010 Jeffrey W. Roberson <jeff@FreeBSD.org> 4 * All rights reserved. 5 * 6 * The soft updates code is derived from the appendix of a University 7 * of Michigan technical report (Gregory R. Ganger and Yale N. Patt, 8 * "Soft Updates: A Solution to the Metadata Update Problem in File 9 * Systems", CSE-TR-254-95, August 1995). 10 * 11 * Further information about soft updates can be obtained from: 12 * 13 * Marshall Kirk McKusick http://www.mckusick.com/softdep/ 14 * 1614 Oxford Street mckusick@mckusick.com 15 * Berkeley, CA 94709-1608 +1-510-843-9542 16 * USA 17 * 18 * Redistribution and use in source and binary forms, with or without 19 * modification, are permitted provided that the following conditions 20 * are met: 21 * 22 * 1. Redistributions of source code must retain the above copyright 23 * notice, this list of conditions and the following disclaimer. 24 * 2. Redistributions in binary form must reproduce the above copyright 25 * notice, this list of conditions and the following disclaimer in the 26 * documentation and/or other materials provided with the distribution. 27 * 28 * THIS SOFTWARE IS PROVIDED BY THE AUTHORS ``AS IS'' AND ANY EXPRESS OR 29 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 30 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 31 * IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY DIRECT, INDIRECT, 32 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, 33 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS 34 * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND 35 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR 36 * TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE 37 * USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 38 * 39 * from: @(#)ffs_softdep.c 9.59 (McKusick) 6/21/00 40 */ 41 42#include <sys/cdefs.h> 43__FBSDID("$FreeBSD: head/sys/ufs/ffs/ffs_softdep.c 278257 2015-02-05 13:00:27Z kib $"); 44 45#include "opt_ffs.h" 46#include "opt_quota.h" 47#include "opt_ddb.h" 48 49/* 50 * For now we want the safety net that the DEBUG flag provides. 51 */ 52#ifndef DEBUG 53#define DEBUG 54#endif 55 56#include <sys/param.h> 57#include <sys/kernel.h> 58#include <sys/systm.h> 59#include <sys/bio.h> 60#include <sys/buf.h> 61#include <sys/kdb.h> 62#include <sys/kthread.h> 63#include <sys/ktr.h> 64#include <sys/limits.h> 65#include <sys/lock.h> 66#include <sys/malloc.h> 67#include <sys/mount.h> 68#include <sys/mutex.h> 69#include <sys/namei.h> 70#include <sys/priv.h> 71#include <sys/proc.h> 72#include <sys/rwlock.h> 73#include <sys/stat.h> 74#include <sys/sysctl.h> 75#include <sys/syslog.h> 76#include <sys/vnode.h> 77#include <sys/conf.h> 78 79#include <ufs/ufs/dir.h> 80#include <ufs/ufs/extattr.h> 81#include <ufs/ufs/quota.h> 82#include <ufs/ufs/inode.h> 83#include <ufs/ufs/ufsmount.h> 84#include <ufs/ffs/fs.h> 85#include <ufs/ffs/softdep.h> 86#include <ufs/ffs/ffs_extern.h> 87#include <ufs/ufs/ufs_extern.h> 88 89#include <vm/vm.h> 90#include <vm/vm_extern.h> 91#include <vm/vm_object.h> 92 93#include <geom/geom.h> 94 95#include <ddb/ddb.h> 96 97#define KTR_SUJ 0 /* Define to KTR_SPARE. */ 98 99#ifndef SOFTUPDATES 100 101int 102softdep_flushfiles(oldmnt, flags, td) 103 struct mount *oldmnt; 104 int flags; 105 struct thread *td; 106{ 107 108 panic("softdep_flushfiles called"); 109} 110 111int 112softdep_mount(devvp, mp, fs, cred) 113 struct vnode *devvp; 114 struct mount *mp; 115 struct fs *fs; 116 struct ucred *cred; 117{ 118 119 return (0); 120} 121 122void 123softdep_initialize() 124{ 125 126 return; 127} 128 129void 130softdep_uninitialize() 131{ 132 133 return; 134} 135 136void 137softdep_unmount(mp) 138 struct mount *mp; 139{ 140 141 panic("softdep_unmount called"); 142} 143 144void 145softdep_setup_sbupdate(ump, fs, bp) 146 struct ufsmount *ump; 147 struct fs *fs; 148 struct buf *bp; 149{ 150 151 panic("softdep_setup_sbupdate called"); 152} 153 154void 155softdep_setup_inomapdep(bp, ip, newinum, mode) 156 struct buf *bp; 157 struct inode *ip; 158 ino_t newinum; 159 int mode; 160{ 161 162 panic("softdep_setup_inomapdep called"); 163} 164 165void 166softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags) 167 struct buf *bp; 168 struct mount *mp; 169 ufs2_daddr_t newblkno; 170 int frags; 171 int oldfrags; 172{ 173 174 panic("softdep_setup_blkmapdep called"); 175} 176 177void 178softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 179 struct inode *ip; 180 ufs_lbn_t lbn; 181 ufs2_daddr_t newblkno; 182 ufs2_daddr_t oldblkno; 183 long newsize; 184 long oldsize; 185 struct buf *bp; 186{ 187 188 panic("softdep_setup_allocdirect called"); 189} 190 191void 192softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 193 struct inode *ip; 194 ufs_lbn_t lbn; 195 ufs2_daddr_t newblkno; 196 ufs2_daddr_t oldblkno; 197 long newsize; 198 long oldsize; 199 struct buf *bp; 200{ 201 202 panic("softdep_setup_allocext called"); 203} 204 205void 206softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 207 struct inode *ip; 208 ufs_lbn_t lbn; 209 struct buf *bp; 210 int ptrno; 211 ufs2_daddr_t newblkno; 212 ufs2_daddr_t oldblkno; 213 struct buf *nbp; 214{ 215 216 panic("softdep_setup_allocindir_page called"); 217} 218 219void 220softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 221 struct buf *nbp; 222 struct inode *ip; 223 struct buf *bp; 224 int ptrno; 225 ufs2_daddr_t newblkno; 226{ 227 228 panic("softdep_setup_allocindir_meta called"); 229} 230 231void 232softdep_journal_freeblocks(ip, cred, length, flags) 233 struct inode *ip; 234 struct ucred *cred; 235 off_t length; 236 int flags; 237{ 238 239 panic("softdep_journal_freeblocks called"); 240} 241 242void 243softdep_journal_fsync(ip) 244 struct inode *ip; 245{ 246 247 panic("softdep_journal_fsync called"); 248} 249 250void 251softdep_setup_freeblocks(ip, length, flags) 252 struct inode *ip; 253 off_t length; 254 int flags; 255{ 256 257 panic("softdep_setup_freeblocks called"); 258} 259 260void 261softdep_freefile(pvp, ino, mode) 262 struct vnode *pvp; 263 ino_t ino; 264 int mode; 265{ 266 267 panic("softdep_freefile called"); 268} 269 270int 271softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 272 struct buf *bp; 273 struct inode *dp; 274 off_t diroffset; 275 ino_t newinum; 276 struct buf *newdirbp; 277 int isnewblk; 278{ 279 280 panic("softdep_setup_directory_add called"); 281} 282 283void 284softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize) 285 struct buf *bp; 286 struct inode *dp; 287 caddr_t base; 288 caddr_t oldloc; 289 caddr_t newloc; 290 int entrysize; 291{ 292 293 panic("softdep_change_directoryentry_offset called"); 294} 295 296void 297softdep_setup_remove(bp, dp, ip, isrmdir) 298 struct buf *bp; 299 struct inode *dp; 300 struct inode *ip; 301 int isrmdir; 302{ 303 304 panic("softdep_setup_remove called"); 305} 306 307void 308softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 309 struct buf *bp; 310 struct inode *dp; 311 struct inode *ip; 312 ino_t newinum; 313 int isrmdir; 314{ 315 316 panic("softdep_setup_directory_change called"); 317} 318 319void 320softdep_setup_blkfree(mp, bp, blkno, frags, wkhd) 321 struct mount *mp; 322 struct buf *bp; 323 ufs2_daddr_t blkno; 324 int frags; 325 struct workhead *wkhd; 326{ 327 328 panic("%s called", __FUNCTION__); 329} 330 331void 332softdep_setup_inofree(mp, bp, ino, wkhd) 333 struct mount *mp; 334 struct buf *bp; 335 ino_t ino; 336 struct workhead *wkhd; 337{ 338 339 panic("%s called", __FUNCTION__); 340} 341 342void 343softdep_setup_unlink(dp, ip) 344 struct inode *dp; 345 struct inode *ip; 346{ 347 348 panic("%s called", __FUNCTION__); 349} 350 351void 352softdep_setup_link(dp, ip) 353 struct inode *dp; 354 struct inode *ip; 355{ 356 357 panic("%s called", __FUNCTION__); 358} 359 360void 361softdep_revert_link(dp, ip) 362 struct inode *dp; 363 struct inode *ip; 364{ 365 366 panic("%s called", __FUNCTION__); 367} 368 369void 370softdep_setup_rmdir(dp, ip) 371 struct inode *dp; 372 struct inode *ip; 373{ 374 375 panic("%s called", __FUNCTION__); 376} 377 378void 379softdep_revert_rmdir(dp, ip) 380 struct inode *dp; 381 struct inode *ip; 382{ 383 384 panic("%s called", __FUNCTION__); 385} 386 387void 388softdep_setup_create(dp, ip) 389 struct inode *dp; 390 struct inode *ip; 391{ 392 393 panic("%s called", __FUNCTION__); 394} 395 396void 397softdep_revert_create(dp, ip) 398 struct inode *dp; 399 struct inode *ip; 400{ 401 402 panic("%s called", __FUNCTION__); 403} 404 405void 406softdep_setup_mkdir(dp, ip) 407 struct inode *dp; 408 struct inode *ip; 409{ 410 411 panic("%s called", __FUNCTION__); 412} 413 414void 415softdep_revert_mkdir(dp, ip) 416 struct inode *dp; 417 struct inode *ip; 418{ 419 420 panic("%s called", __FUNCTION__); 421} 422 423void 424softdep_setup_dotdot_link(dp, ip) 425 struct inode *dp; 426 struct inode *ip; 427{ 428 429 panic("%s called", __FUNCTION__); 430} 431 432int 433softdep_prealloc(vp, waitok) 434 struct vnode *vp; 435 int waitok; 436{ 437 438 panic("%s called", __FUNCTION__); 439} 440 441int 442softdep_journal_lookup(mp, vpp) 443 struct mount *mp; 444 struct vnode **vpp; 445{ 446 447 return (ENOENT); 448} 449 450void 451softdep_change_linkcnt(ip) 452 struct inode *ip; 453{ 454 455 panic("softdep_change_linkcnt called"); 456} 457 458void 459softdep_load_inodeblock(ip) 460 struct inode *ip; 461{ 462 463 panic("softdep_load_inodeblock called"); 464} 465 466void 467softdep_update_inodeblock(ip, bp, waitfor) 468 struct inode *ip; 469 struct buf *bp; 470 int waitfor; 471{ 472 473 panic("softdep_update_inodeblock called"); 474} 475 476int 477softdep_fsync(vp) 478 struct vnode *vp; /* the "in_core" copy of the inode */ 479{ 480 481 return (0); 482} 483 484void 485softdep_fsync_mountdev(vp) 486 struct vnode *vp; 487{ 488 489 return; 490} 491 492int 493softdep_flushworklist(oldmnt, countp, td) 494 struct mount *oldmnt; 495 int *countp; 496 struct thread *td; 497{ 498 499 *countp = 0; 500 return (0); 501} 502 503int 504softdep_sync_metadata(struct vnode *vp) 505{ 506 507 panic("softdep_sync_metadata called"); 508} 509 510int 511softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor) 512{ 513 514 panic("softdep_sync_buf called"); 515} 516 517int 518softdep_slowdown(vp) 519 struct vnode *vp; 520{ 521 522 panic("softdep_slowdown called"); 523} 524 525int 526softdep_request_cleanup(fs, vp, cred, resource) 527 struct fs *fs; 528 struct vnode *vp; 529 struct ucred *cred; 530 int resource; 531{ 532 533 return (0); 534} 535 536int 537softdep_check_suspend(struct mount *mp, 538 struct vnode *devvp, 539 int softdep_depcnt, 540 int softdep_accdepcnt, 541 int secondary_writes, 542 int secondary_accwrites) 543{ 544 struct bufobj *bo; 545 int error; 546 547 (void) softdep_depcnt, 548 (void) softdep_accdepcnt; 549 550 bo = &devvp->v_bufobj; 551 ASSERT_BO_WLOCKED(bo); 552 553 MNT_ILOCK(mp); 554 while (mp->mnt_secondary_writes != 0) { 555 BO_UNLOCK(bo); 556 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp), 557 (PUSER - 1) | PDROP, "secwr", 0); 558 BO_LOCK(bo); 559 MNT_ILOCK(mp); 560 } 561 562 /* 563 * Reasons for needing more work before suspend: 564 * - Dirty buffers on devvp. 565 * - Secondary writes occurred after start of vnode sync loop 566 */ 567 error = 0; 568 if (bo->bo_numoutput > 0 || 569 bo->bo_dirty.bv_cnt > 0 || 570 secondary_writes != 0 || 571 mp->mnt_secondary_writes != 0 || 572 secondary_accwrites != mp->mnt_secondary_accwrites) 573 error = EAGAIN; 574 BO_UNLOCK(bo); 575 return (error); 576} 577 578void 579softdep_get_depcounts(struct mount *mp, 580 int *softdepactivep, 581 int *softdepactiveaccp) 582{ 583 (void) mp; 584 *softdepactivep = 0; 585 *softdepactiveaccp = 0; 586} 587 588void 589softdep_buf_append(bp, wkhd) 590 struct buf *bp; 591 struct workhead *wkhd; 592{ 593 594 panic("softdep_buf_appendwork called"); 595} 596 597void 598softdep_inode_append(ip, cred, wkhd) 599 struct inode *ip; 600 struct ucred *cred; 601 struct workhead *wkhd; 602{ 603 604 panic("softdep_inode_appendwork called"); 605} 606 607void 608softdep_freework(wkhd) 609 struct workhead *wkhd; 610{ 611 612 panic("softdep_freework called"); 613} 614 615#else 616 617FEATURE(softupdates, "FFS soft-updates support"); 618 619static SYSCTL_NODE(_debug, OID_AUTO, softdep, CTLFLAG_RW, 0, 620 "soft updates stats"); 621static SYSCTL_NODE(_debug_softdep, OID_AUTO, total, CTLFLAG_RW, 0, 622 "total dependencies allocated"); 623static SYSCTL_NODE(_debug_softdep, OID_AUTO, highuse, CTLFLAG_RW, 0, 624 "high use dependencies allocated"); 625static SYSCTL_NODE(_debug_softdep, OID_AUTO, current, CTLFLAG_RW, 0, 626 "current dependencies allocated"); 627static SYSCTL_NODE(_debug_softdep, OID_AUTO, write, CTLFLAG_RW, 0, 628 "current dependencies written"); 629 630unsigned long dep_current[D_LAST + 1]; 631unsigned long dep_highuse[D_LAST + 1]; 632unsigned long dep_total[D_LAST + 1]; 633unsigned long dep_write[D_LAST + 1]; 634 635#define SOFTDEP_TYPE(type, str, long) \ 636 static MALLOC_DEFINE(M_ ## type, #str, long); \ 637 SYSCTL_ULONG(_debug_softdep_total, OID_AUTO, str, CTLFLAG_RD, \ 638 &dep_total[D_ ## type], 0, ""); \ 639 SYSCTL_ULONG(_debug_softdep_current, OID_AUTO, str, CTLFLAG_RD, \ 640 &dep_current[D_ ## type], 0, ""); \ 641 SYSCTL_ULONG(_debug_softdep_highuse, OID_AUTO, str, CTLFLAG_RD, \ 642 &dep_highuse[D_ ## type], 0, ""); \ 643 SYSCTL_ULONG(_debug_softdep_write, OID_AUTO, str, CTLFLAG_RD, \ 644 &dep_write[D_ ## type], 0, ""); 645 646SOFTDEP_TYPE(PAGEDEP, pagedep, "File page dependencies"); 647SOFTDEP_TYPE(INODEDEP, inodedep, "Inode dependencies"); 648SOFTDEP_TYPE(BMSAFEMAP, bmsafemap, 649 "Block or frag allocated from cyl group map"); 650SOFTDEP_TYPE(NEWBLK, newblk, "New block or frag allocation dependency"); 651SOFTDEP_TYPE(ALLOCDIRECT, allocdirect, "Block or frag dependency for an inode"); 652SOFTDEP_TYPE(INDIRDEP, indirdep, "Indirect block dependencies"); 653SOFTDEP_TYPE(ALLOCINDIR, allocindir, "Block dependency for an indirect block"); 654SOFTDEP_TYPE(FREEFRAG, freefrag, "Previously used frag for an inode"); 655SOFTDEP_TYPE(FREEBLKS, freeblks, "Blocks freed from an inode"); 656SOFTDEP_TYPE(FREEFILE, freefile, "Inode deallocated"); 657SOFTDEP_TYPE(DIRADD, diradd, "New directory entry"); 658SOFTDEP_TYPE(MKDIR, mkdir, "New directory"); 659SOFTDEP_TYPE(DIRREM, dirrem, "Directory entry deleted"); 660SOFTDEP_TYPE(NEWDIRBLK, newdirblk, "Unclaimed new directory block"); 661SOFTDEP_TYPE(FREEWORK, freework, "free an inode block"); 662SOFTDEP_TYPE(FREEDEP, freedep, "track a block free"); 663SOFTDEP_TYPE(JADDREF, jaddref, "Journal inode ref add"); 664SOFTDEP_TYPE(JREMREF, jremref, "Journal inode ref remove"); 665SOFTDEP_TYPE(JMVREF, jmvref, "Journal inode ref move"); 666SOFTDEP_TYPE(JNEWBLK, jnewblk, "Journal new block"); 667SOFTDEP_TYPE(JFREEBLK, jfreeblk, "Journal free block"); 668SOFTDEP_TYPE(JFREEFRAG, jfreefrag, "Journal free frag"); 669SOFTDEP_TYPE(JSEG, jseg, "Journal segment"); 670SOFTDEP_TYPE(JSEGDEP, jsegdep, "Journal segment complete"); 671SOFTDEP_TYPE(SBDEP, sbdep, "Superblock write dependency"); 672SOFTDEP_TYPE(JTRUNC, jtrunc, "Journal inode truncation"); 673SOFTDEP_TYPE(JFSYNC, jfsync, "Journal fsync complete"); 674 675static MALLOC_DEFINE(M_SENTINEL, "sentinel", "Worklist sentinel"); 676 677static MALLOC_DEFINE(M_SAVEDINO, "savedino", "Saved inodes"); 678static MALLOC_DEFINE(M_JBLOCKS, "jblocks", "Journal block locations"); 679static MALLOC_DEFINE(M_MOUNTDATA, "softdep", "Softdep per-mount data"); 680 681#define M_SOFTDEP_FLAGS (M_WAITOK) 682 683/* 684 * translate from workitem type to memory type 685 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 686 */ 687static struct malloc_type *memtype[] = { 688 M_PAGEDEP, 689 M_INODEDEP, 690 M_BMSAFEMAP, 691 M_NEWBLK, 692 M_ALLOCDIRECT, 693 M_INDIRDEP, 694 M_ALLOCINDIR, 695 M_FREEFRAG, 696 M_FREEBLKS, 697 M_FREEFILE, 698 M_DIRADD, 699 M_MKDIR, 700 M_DIRREM, 701 M_NEWDIRBLK, 702 M_FREEWORK, 703 M_FREEDEP, 704 M_JADDREF, 705 M_JREMREF, 706 M_JMVREF, 707 M_JNEWBLK, 708 M_JFREEBLK, 709 M_JFREEFRAG, 710 M_JSEG, 711 M_JSEGDEP, 712 M_SBDEP, 713 M_JTRUNC, 714 M_JFSYNC, 715 M_SENTINEL 716}; 717 718#define DtoM(type) (memtype[type]) 719 720/* 721 * Names of malloc types. 722 */ 723#define TYPENAME(type) \ 724 ((unsigned)(type) <= D_LAST ? memtype[type]->ks_shortdesc : "???") 725/* 726 * End system adaptation definitions. 727 */ 728 729#define DOTDOT_OFFSET offsetof(struct dirtemplate, dotdot_ino) 730#define DOT_OFFSET offsetof(struct dirtemplate, dot_ino) 731 732/* 733 * Internal function prototypes. 734 */ 735static void check_clear_deps(struct mount *); 736static void softdep_error(char *, int); 737static int softdep_process_worklist(struct mount *, int); 738static int softdep_waitidle(struct mount *, int); 739static void drain_output(struct vnode *); 740static struct buf *getdirtybuf(struct buf *, struct rwlock *, int); 741static void clear_remove(struct mount *); 742static void clear_inodedeps(struct mount *); 743static void unlinked_inodedep(struct mount *, struct inodedep *); 744static void clear_unlinked_inodedep(struct inodedep *); 745static struct inodedep *first_unlinked_inodedep(struct ufsmount *); 746static int flush_pagedep_deps(struct vnode *, struct mount *, 747 struct diraddhd *); 748static int free_pagedep(struct pagedep *); 749static int flush_newblk_dep(struct vnode *, struct mount *, ufs_lbn_t); 750static int flush_inodedep_deps(struct vnode *, struct mount *, ino_t); 751static int flush_deplist(struct allocdirectlst *, int, int *); 752static int sync_cgs(struct mount *, int); 753static int handle_written_filepage(struct pagedep *, struct buf *); 754static int handle_written_sbdep(struct sbdep *, struct buf *); 755static void initiate_write_sbdep(struct sbdep *); 756static void diradd_inode_written(struct diradd *, struct inodedep *); 757static int handle_written_indirdep(struct indirdep *, struct buf *, 758 struct buf**); 759static int handle_written_inodeblock(struct inodedep *, struct buf *); 760static int jnewblk_rollforward(struct jnewblk *, struct fs *, struct cg *, 761 uint8_t *); 762static int handle_written_bmsafemap(struct bmsafemap *, struct buf *); 763static void handle_written_jaddref(struct jaddref *); 764static void handle_written_jremref(struct jremref *); 765static void handle_written_jseg(struct jseg *, struct buf *); 766static void handle_written_jnewblk(struct jnewblk *); 767static void handle_written_jblkdep(struct jblkdep *); 768static void handle_written_jfreefrag(struct jfreefrag *); 769static void complete_jseg(struct jseg *); 770static void complete_jsegs(struct jseg *); 771static void jseg_write(struct ufsmount *ump, struct jseg *, uint8_t *); 772static void jaddref_write(struct jaddref *, struct jseg *, uint8_t *); 773static void jremref_write(struct jremref *, struct jseg *, uint8_t *); 774static void jmvref_write(struct jmvref *, struct jseg *, uint8_t *); 775static void jtrunc_write(struct jtrunc *, struct jseg *, uint8_t *); 776static void jfsync_write(struct jfsync *, struct jseg *, uint8_t *data); 777static void jnewblk_write(struct jnewblk *, struct jseg *, uint8_t *); 778static void jfreeblk_write(struct jfreeblk *, struct jseg *, uint8_t *); 779static void jfreefrag_write(struct jfreefrag *, struct jseg *, uint8_t *); 780static inline void inoref_write(struct inoref *, struct jseg *, 781 struct jrefrec *); 782static void handle_allocdirect_partdone(struct allocdirect *, 783 struct workhead *); 784static struct jnewblk *cancel_newblk(struct newblk *, struct worklist *, 785 struct workhead *); 786static void indirdep_complete(struct indirdep *); 787static int indirblk_lookup(struct mount *, ufs2_daddr_t); 788static void indirblk_insert(struct freework *); 789static void indirblk_remove(struct freework *); 790static void handle_allocindir_partdone(struct allocindir *); 791static void initiate_write_filepage(struct pagedep *, struct buf *); 792static void initiate_write_indirdep(struct indirdep*, struct buf *); 793static void handle_written_mkdir(struct mkdir *, int); 794static int jnewblk_rollback(struct jnewblk *, struct fs *, struct cg *, 795 uint8_t *); 796static void initiate_write_bmsafemap(struct bmsafemap *, struct buf *); 797static void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *); 798static void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *); 799static void handle_workitem_freefile(struct freefile *); 800static int handle_workitem_remove(struct dirrem *, int); 801static struct dirrem *newdirrem(struct buf *, struct inode *, 802 struct inode *, int, struct dirrem **); 803static struct indirdep *indirdep_lookup(struct mount *, struct inode *, 804 struct buf *); 805static void cancel_indirdep(struct indirdep *, struct buf *, 806 struct freeblks *); 807static void free_indirdep(struct indirdep *); 808static void free_diradd(struct diradd *, struct workhead *); 809static void merge_diradd(struct inodedep *, struct diradd *); 810static void complete_diradd(struct diradd *); 811static struct diradd *diradd_lookup(struct pagedep *, int); 812static struct jremref *cancel_diradd_dotdot(struct inode *, struct dirrem *, 813 struct jremref *); 814static struct jremref *cancel_mkdir_dotdot(struct inode *, struct dirrem *, 815 struct jremref *); 816static void cancel_diradd(struct diradd *, struct dirrem *, struct jremref *, 817 struct jremref *, struct jremref *); 818static void dirrem_journal(struct dirrem *, struct jremref *, struct jremref *, 819 struct jremref *); 820static void cancel_allocindir(struct allocindir *, struct buf *bp, 821 struct freeblks *, int); 822static int setup_trunc_indir(struct freeblks *, struct inode *, 823 ufs_lbn_t, ufs_lbn_t, ufs2_daddr_t); 824static void complete_trunc_indir(struct freework *); 825static void trunc_indirdep(struct indirdep *, struct freeblks *, struct buf *, 826 int); 827static void complete_mkdir(struct mkdir *); 828static void free_newdirblk(struct newdirblk *); 829static void free_jremref(struct jremref *); 830static void free_jaddref(struct jaddref *); 831static void free_jsegdep(struct jsegdep *); 832static void free_jsegs(struct jblocks *); 833static void rele_jseg(struct jseg *); 834static void free_jseg(struct jseg *, struct jblocks *); 835static void free_jnewblk(struct jnewblk *); 836static void free_jblkdep(struct jblkdep *); 837static void free_jfreefrag(struct jfreefrag *); 838static void free_freedep(struct freedep *); 839static void journal_jremref(struct dirrem *, struct jremref *, 840 struct inodedep *); 841static void cancel_jnewblk(struct jnewblk *, struct workhead *); 842static int cancel_jaddref(struct jaddref *, struct inodedep *, 843 struct workhead *); 844static void cancel_jfreefrag(struct jfreefrag *); 845static inline void setup_freedirect(struct freeblks *, struct inode *, 846 int, int); 847static inline void setup_freeext(struct freeblks *, struct inode *, int, int); 848static inline void setup_freeindir(struct freeblks *, struct inode *, int, 849 ufs_lbn_t, int); 850static inline struct freeblks *newfreeblks(struct mount *, struct inode *); 851static void freeblks_free(struct ufsmount *, struct freeblks *, int); 852static void indir_trunc(struct freework *, ufs2_daddr_t, ufs_lbn_t); 853static ufs2_daddr_t blkcount(struct fs *, ufs2_daddr_t, off_t); 854static int trunc_check_buf(struct buf *, int *, ufs_lbn_t, int, int); 855static void trunc_dependencies(struct inode *, struct freeblks *, ufs_lbn_t, 856 int, int); 857static void trunc_pages(struct inode *, off_t, ufs2_daddr_t, int); 858static int cancel_pagedep(struct pagedep *, struct freeblks *, int); 859static int deallocate_dependencies(struct buf *, struct freeblks *, int); 860static void newblk_freefrag(struct newblk*); 861static void free_newblk(struct newblk *); 862static void cancel_allocdirect(struct allocdirectlst *, 863 struct allocdirect *, struct freeblks *); 864static int check_inode_unwritten(struct inodedep *); 865static int free_inodedep(struct inodedep *); 866static void freework_freeblock(struct freework *); 867static void freework_enqueue(struct freework *); 868static int handle_workitem_freeblocks(struct freeblks *, int); 869static int handle_complete_freeblocks(struct freeblks *, int); 870static void handle_workitem_indirblk(struct freework *); 871static void handle_written_freework(struct freework *); 872static void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *); 873static struct worklist *jnewblk_merge(struct worklist *, struct worklist *, 874 struct workhead *); 875static struct freefrag *setup_allocindir_phase2(struct buf *, struct inode *, 876 struct inodedep *, struct allocindir *, ufs_lbn_t); 877static struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t, 878 ufs2_daddr_t, ufs_lbn_t); 879static void handle_workitem_freefrag(struct freefrag *); 880static struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long, 881 ufs_lbn_t); 882static void allocdirect_merge(struct allocdirectlst *, 883 struct allocdirect *, struct allocdirect *); 884static struct freefrag *allocindir_merge(struct allocindir *, 885 struct allocindir *); 886static int bmsafemap_find(struct bmsafemap_hashhead *, int, 887 struct bmsafemap **); 888static struct bmsafemap *bmsafemap_lookup(struct mount *, struct buf *, 889 int cg, struct bmsafemap *); 890static int newblk_find(struct newblk_hashhead *, ufs2_daddr_t, int, 891 struct newblk **); 892static int newblk_lookup(struct mount *, ufs2_daddr_t, int, struct newblk **); 893static int inodedep_find(struct inodedep_hashhead *, ino_t, 894 struct inodedep **); 895static int inodedep_lookup(struct mount *, ino_t, int, struct inodedep **); 896static int pagedep_lookup(struct mount *, struct buf *bp, ino_t, ufs_lbn_t, 897 int, struct pagedep **); 898static int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t, 899 struct pagedep **); 900static void pause_timer(void *); 901static int request_cleanup(struct mount *, int); 902static int process_worklist_item(struct mount *, int, int); 903static void process_removes(struct vnode *); 904static void process_truncates(struct vnode *); 905static void jwork_move(struct workhead *, struct workhead *); 906static void jwork_insert(struct workhead *, struct jsegdep *); 907static void add_to_worklist(struct worklist *, int); 908static void wake_worklist(struct worklist *); 909static void wait_worklist(struct worklist *, char *); 910static void remove_from_worklist(struct worklist *); 911static void softdep_flush(void *); 912static void softdep_flushjournal(struct mount *); 913static int softdep_speedup(struct ufsmount *); 914static void worklist_speedup(struct mount *); 915static int journal_mount(struct mount *, struct fs *, struct ucred *); 916static void journal_unmount(struct ufsmount *); 917static int journal_space(struct ufsmount *, int); 918static void journal_suspend(struct ufsmount *); 919static int journal_unsuspend(struct ufsmount *ump); 920static void softdep_prelink(struct vnode *, struct vnode *); 921static void add_to_journal(struct worklist *); 922static void remove_from_journal(struct worklist *); 923static void softdep_process_journal(struct mount *, struct worklist *, int); 924static struct jremref *newjremref(struct dirrem *, struct inode *, 925 struct inode *ip, off_t, nlink_t); 926static struct jaddref *newjaddref(struct inode *, ino_t, off_t, int16_t, 927 uint16_t); 928static inline void newinoref(struct inoref *, ino_t, ino_t, off_t, nlink_t, 929 uint16_t); 930static inline struct jsegdep *inoref_jseg(struct inoref *); 931static struct jmvref *newjmvref(struct inode *, ino_t, off_t, off_t); 932static struct jfreeblk *newjfreeblk(struct freeblks *, ufs_lbn_t, 933 ufs2_daddr_t, int); 934static void adjust_newfreework(struct freeblks *, int); 935static struct jtrunc *newjtrunc(struct freeblks *, off_t, int); 936static void move_newblock_dep(struct jaddref *, struct inodedep *); 937static void cancel_jfreeblk(struct freeblks *, ufs2_daddr_t); 938static struct jfreefrag *newjfreefrag(struct freefrag *, struct inode *, 939 ufs2_daddr_t, long, ufs_lbn_t); 940static struct freework *newfreework(struct ufsmount *, struct freeblks *, 941 struct freework *, ufs_lbn_t, ufs2_daddr_t, int, int, int); 942static int jwait(struct worklist *, int); 943static struct inodedep *inodedep_lookup_ip(struct inode *); 944static int bmsafemap_backgroundwrite(struct bmsafemap *, struct buf *); 945static struct freefile *handle_bufwait(struct inodedep *, struct workhead *); 946static void handle_jwork(struct workhead *); 947static struct mkdir *setup_newdir(struct diradd *, ino_t, ino_t, struct buf *, 948 struct mkdir **); 949static struct jblocks *jblocks_create(void); 950static ufs2_daddr_t jblocks_alloc(struct jblocks *, int, int *); 951static void jblocks_free(struct jblocks *, struct mount *, int); 952static void jblocks_destroy(struct jblocks *); 953static void jblocks_add(struct jblocks *, ufs2_daddr_t, int); 954 955/* 956 * Exported softdep operations. 957 */ 958static void softdep_disk_io_initiation(struct buf *); 959static void softdep_disk_write_complete(struct buf *); 960static void softdep_deallocate_dependencies(struct buf *); 961static int softdep_count_dependencies(struct buf *bp, int); 962 963/* 964 * Global lock over all of soft updates. 965 */ 966static struct mtx lk; 967MTX_SYSINIT(softdep_lock, &lk, "Global Softdep Lock", MTX_DEF); 968 969#define ACQUIRE_GBLLOCK(lk) mtx_lock(lk) 970#define FREE_GBLLOCK(lk) mtx_unlock(lk) 971#define GBLLOCK_OWNED(lk) mtx_assert((lk), MA_OWNED) 972 973/* 974 * Per-filesystem soft-updates locking. 975 */ 976#define LOCK_PTR(ump) (&(ump)->um_softdep->sd_fslock) 977#define TRY_ACQUIRE_LOCK(ump) rw_try_wlock(&(ump)->um_softdep->sd_fslock) 978#define ACQUIRE_LOCK(ump) rw_wlock(&(ump)->um_softdep->sd_fslock) 979#define FREE_LOCK(ump) rw_wunlock(&(ump)->um_softdep->sd_fslock) 980#define LOCK_OWNED(ump) rw_assert(&(ump)->um_softdep->sd_fslock, \ 981 RA_WLOCKED) 982 983#define BUF_AREC(bp) lockallowrecurse(&(bp)->b_lock) 984#define BUF_NOREC(bp) lockdisablerecurse(&(bp)->b_lock) 985 986/* 987 * Worklist queue management. 988 * These routines require that the lock be held. 989 */ 990#ifndef /* NOT */ DEBUG 991#define WORKLIST_INSERT(head, item) do { \ 992 (item)->wk_state |= ONWORKLIST; \ 993 LIST_INSERT_HEAD(head, item, wk_list); \ 994} while (0) 995#define WORKLIST_REMOVE(item) do { \ 996 (item)->wk_state &= ~ONWORKLIST; \ 997 LIST_REMOVE(item, wk_list); \ 998} while (0) 999#define WORKLIST_INSERT_UNLOCKED WORKLIST_INSERT 1000#define WORKLIST_REMOVE_UNLOCKED WORKLIST_REMOVE 1001 1002#else /* DEBUG */ 1003static void worklist_insert(struct workhead *, struct worklist *, int); 1004static void worklist_remove(struct worklist *, int); 1005 1006#define WORKLIST_INSERT(head, item) worklist_insert(head, item, 1) 1007#define WORKLIST_INSERT_UNLOCKED(head, item) worklist_insert(head, item, 0) 1008#define WORKLIST_REMOVE(item) worklist_remove(item, 1) 1009#define WORKLIST_REMOVE_UNLOCKED(item) worklist_remove(item, 0) 1010 1011static void 1012worklist_insert(head, item, locked) 1013 struct workhead *head; 1014 struct worklist *item; 1015 int locked; 1016{ 1017 1018 if (locked) 1019 LOCK_OWNED(VFSTOUFS(item->wk_mp)); 1020 if (item->wk_state & ONWORKLIST) 1021 panic("worklist_insert: %p %s(0x%X) already on list", 1022 item, TYPENAME(item->wk_type), item->wk_state); 1023 item->wk_state |= ONWORKLIST; 1024 LIST_INSERT_HEAD(head, item, wk_list); 1025} 1026 1027static void 1028worklist_remove(item, locked) 1029 struct worklist *item; 1030 int locked; 1031{ 1032 1033 if (locked) 1034 LOCK_OWNED(VFSTOUFS(item->wk_mp)); 1035 if ((item->wk_state & ONWORKLIST) == 0) 1036 panic("worklist_remove: %p %s(0x%X) not on list", 1037 item, TYPENAME(item->wk_type), item->wk_state); 1038 item->wk_state &= ~ONWORKLIST; 1039 LIST_REMOVE(item, wk_list); 1040} 1041#endif /* DEBUG */ 1042 1043/* 1044 * Merge two jsegdeps keeping only the oldest one as newer references 1045 * can't be discarded until after older references. 1046 */ 1047static inline struct jsegdep * 1048jsegdep_merge(struct jsegdep *one, struct jsegdep *two) 1049{ 1050 struct jsegdep *swp; 1051 1052 if (two == NULL) 1053 return (one); 1054 1055 if (one->jd_seg->js_seq > two->jd_seg->js_seq) { 1056 swp = one; 1057 one = two; 1058 two = swp; 1059 } 1060 WORKLIST_REMOVE(&two->jd_list); 1061 free_jsegdep(two); 1062 1063 return (one); 1064} 1065 1066/* 1067 * If two freedeps are compatible free one to reduce list size. 1068 */ 1069static inline struct freedep * 1070freedep_merge(struct freedep *one, struct freedep *two) 1071{ 1072 if (two == NULL) 1073 return (one); 1074 1075 if (one->fd_freework == two->fd_freework) { 1076 WORKLIST_REMOVE(&two->fd_list); 1077 free_freedep(two); 1078 } 1079 return (one); 1080} 1081 1082/* 1083 * Move journal work from one list to another. Duplicate freedeps and 1084 * jsegdeps are coalesced to keep the lists as small as possible. 1085 */ 1086static void 1087jwork_move(dst, src) 1088 struct workhead *dst; 1089 struct workhead *src; 1090{ 1091 struct freedep *freedep; 1092 struct jsegdep *jsegdep; 1093 struct worklist *wkn; 1094 struct worklist *wk; 1095 1096 KASSERT(dst != src, 1097 ("jwork_move: dst == src")); 1098 freedep = NULL; 1099 jsegdep = NULL; 1100 LIST_FOREACH_SAFE(wk, dst, wk_list, wkn) { 1101 if (wk->wk_type == D_JSEGDEP) 1102 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep); 1103 else if (wk->wk_type == D_FREEDEP) 1104 freedep = freedep_merge(WK_FREEDEP(wk), freedep); 1105 } 1106 1107 while ((wk = LIST_FIRST(src)) != NULL) { 1108 WORKLIST_REMOVE(wk); 1109 WORKLIST_INSERT(dst, wk); 1110 if (wk->wk_type == D_JSEGDEP) { 1111 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep); 1112 continue; 1113 } 1114 if (wk->wk_type == D_FREEDEP) 1115 freedep = freedep_merge(WK_FREEDEP(wk), freedep); 1116 } 1117} 1118 1119static void 1120jwork_insert(dst, jsegdep) 1121 struct workhead *dst; 1122 struct jsegdep *jsegdep; 1123{ 1124 struct jsegdep *jsegdepn; 1125 struct worklist *wk; 1126 1127 LIST_FOREACH(wk, dst, wk_list) 1128 if (wk->wk_type == D_JSEGDEP) 1129 break; 1130 if (wk == NULL) { 1131 WORKLIST_INSERT(dst, &jsegdep->jd_list); 1132 return; 1133 } 1134 jsegdepn = WK_JSEGDEP(wk); 1135 if (jsegdep->jd_seg->js_seq < jsegdepn->jd_seg->js_seq) { 1136 WORKLIST_REMOVE(wk); 1137 free_jsegdep(jsegdepn); 1138 WORKLIST_INSERT(dst, &jsegdep->jd_list); 1139 } else 1140 free_jsegdep(jsegdep); 1141} 1142 1143/* 1144 * Routines for tracking and managing workitems. 1145 */ 1146static void workitem_free(struct worklist *, int); 1147static void workitem_alloc(struct worklist *, int, struct mount *); 1148static void workitem_reassign(struct worklist *, int); 1149 1150#define WORKITEM_FREE(item, type) \ 1151 workitem_free((struct worklist *)(item), (type)) 1152#define WORKITEM_REASSIGN(item, type) \ 1153 workitem_reassign((struct worklist *)(item), (type)) 1154 1155static void 1156workitem_free(item, type) 1157 struct worklist *item; 1158 int type; 1159{ 1160 struct ufsmount *ump; 1161 1162#ifdef DEBUG 1163 if (item->wk_state & ONWORKLIST) 1164 panic("workitem_free: %s(0x%X) still on list", 1165 TYPENAME(item->wk_type), item->wk_state); 1166 if (item->wk_type != type && type != D_NEWBLK) 1167 panic("workitem_free: type mismatch %s != %s", 1168 TYPENAME(item->wk_type), TYPENAME(type)); 1169#endif 1170 if (item->wk_state & IOWAITING) 1171 wakeup(item); 1172 ump = VFSTOUFS(item->wk_mp); 1173 LOCK_OWNED(ump); 1174 KASSERT(ump->softdep_deps > 0, 1175 ("workitem_free: %s: softdep_deps going negative", 1176 ump->um_fs->fs_fsmnt)); 1177 if (--ump->softdep_deps == 0 && ump->softdep_req) 1178 wakeup(&ump->softdep_deps); 1179 KASSERT(dep_current[item->wk_type] > 0, 1180 ("workitem_free: %s: dep_current[%s] going negative", 1181 ump->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1182 KASSERT(ump->softdep_curdeps[item->wk_type] > 0, 1183 ("workitem_free: %s: softdep_curdeps[%s] going negative", 1184 ump->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1185 atomic_subtract_long(&dep_current[item->wk_type], 1); 1186 ump->softdep_curdeps[item->wk_type] -= 1; 1187 free(item, DtoM(type)); 1188} 1189 1190static void 1191workitem_alloc(item, type, mp) 1192 struct worklist *item; 1193 int type; 1194 struct mount *mp; 1195{ 1196 struct ufsmount *ump; 1197 1198 item->wk_type = type; 1199 item->wk_mp = mp; 1200 item->wk_state = 0; 1201 1202 ump = VFSTOUFS(mp); 1203 ACQUIRE_GBLLOCK(&lk); 1204 dep_current[type]++; 1205 if (dep_current[type] > dep_highuse[type]) 1206 dep_highuse[type] = dep_current[type]; 1207 dep_total[type]++; 1208 FREE_GBLLOCK(&lk); 1209 ACQUIRE_LOCK(ump); 1210 ump->softdep_curdeps[type] += 1; 1211 ump->softdep_deps++; 1212 ump->softdep_accdeps++; 1213 FREE_LOCK(ump); 1214} 1215 1216static void 1217workitem_reassign(item, newtype) 1218 struct worklist *item; 1219 int newtype; 1220{ 1221 struct ufsmount *ump; 1222 1223 ump = VFSTOUFS(item->wk_mp); 1224 LOCK_OWNED(ump); 1225 KASSERT(ump->softdep_curdeps[item->wk_type] > 0, 1226 ("workitem_reassign: %s: softdep_curdeps[%s] going negative", 1227 VFSTOUFS(item->wk_mp)->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1228 ump->softdep_curdeps[item->wk_type] -= 1; 1229 ump->softdep_curdeps[newtype] += 1; 1230 KASSERT(dep_current[item->wk_type] > 0, 1231 ("workitem_reassign: %s: dep_current[%s] going negative", 1232 VFSTOUFS(item->wk_mp)->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1233 ACQUIRE_GBLLOCK(&lk); 1234 dep_current[newtype]++; 1235 dep_current[item->wk_type]--; 1236 if (dep_current[newtype] > dep_highuse[newtype]) 1237 dep_highuse[newtype] = dep_current[newtype]; 1238 dep_total[newtype]++; 1239 FREE_GBLLOCK(&lk); 1240 item->wk_type = newtype; 1241} 1242 1243/* 1244 * Workitem queue management 1245 */ 1246static int max_softdeps; /* maximum number of structs before slowdown */ 1247static int tickdelay = 2; /* number of ticks to pause during slowdown */ 1248static int proc_waiting; /* tracks whether we have a timeout posted */ 1249static int *stat_countp; /* statistic to count in proc_waiting timeout */ 1250static struct callout softdep_callout; 1251static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 1252static int req_clear_remove; /* syncer process flush some freeblks */ 1253static int softdep_flushcache = 0; /* Should we do BIO_FLUSH? */ 1254 1255/* 1256 * runtime statistics 1257 */ 1258static int stat_flush_threads; /* number of softdep flushing threads */ 1259static int stat_worklist_push; /* number of worklist cleanups */ 1260static int stat_blk_limit_push; /* number of times block limit neared */ 1261static int stat_ino_limit_push; /* number of times inode limit neared */ 1262static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 1263static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 1264static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 1265static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 1266static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 1267static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 1268static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 1269static int stat_jaddref; /* bufs redirtied as ino bitmap can not write */ 1270static int stat_jnewblk; /* bufs redirtied as blk bitmap can not write */ 1271static int stat_journal_min; /* Times hit journal min threshold */ 1272static int stat_journal_low; /* Times hit journal low threshold */ 1273static int stat_journal_wait; /* Times blocked in jwait(). */ 1274static int stat_jwait_filepage; /* Times blocked in jwait() for filepage. */ 1275static int stat_jwait_freeblks; /* Times blocked in jwait() for freeblks. */ 1276static int stat_jwait_inode; /* Times blocked in jwait() for inodes. */ 1277static int stat_jwait_newblk; /* Times blocked in jwait() for newblks. */ 1278static int stat_cleanup_high_delay; /* Maximum cleanup delay (in ticks) */ 1279static int stat_cleanup_blkrequests; /* Number of block cleanup requests */ 1280static int stat_cleanup_inorequests; /* Number of inode cleanup requests */ 1281static int stat_cleanup_retries; /* Number of cleanups that needed to flush */ 1282static int stat_cleanup_failures; /* Number of cleanup requests that failed */ 1283static int stat_emptyjblocks; /* Number of potentially empty journal blocks */ 1284 1285SYSCTL_INT(_debug_softdep, OID_AUTO, max_softdeps, CTLFLAG_RW, 1286 &max_softdeps, 0, ""); 1287SYSCTL_INT(_debug_softdep, OID_AUTO, tickdelay, CTLFLAG_RW, 1288 &tickdelay, 0, ""); 1289SYSCTL_INT(_debug_softdep, OID_AUTO, flush_threads, CTLFLAG_RD, 1290 &stat_flush_threads, 0, ""); 1291SYSCTL_INT(_debug_softdep, OID_AUTO, worklist_push, CTLFLAG_RW, 1292 &stat_worklist_push, 0,""); 1293SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_push, CTLFLAG_RW, 1294 &stat_blk_limit_push, 0,""); 1295SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_push, CTLFLAG_RW, 1296 &stat_ino_limit_push, 0,""); 1297SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_hit, CTLFLAG_RW, 1298 &stat_blk_limit_hit, 0, ""); 1299SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_hit, CTLFLAG_RW, 1300 &stat_ino_limit_hit, 0, ""); 1301SYSCTL_INT(_debug_softdep, OID_AUTO, sync_limit_hit, CTLFLAG_RW, 1302 &stat_sync_limit_hit, 0, ""); 1303SYSCTL_INT(_debug_softdep, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, 1304 &stat_indir_blk_ptrs, 0, ""); 1305SYSCTL_INT(_debug_softdep, OID_AUTO, inode_bitmap, CTLFLAG_RW, 1306 &stat_inode_bitmap, 0, ""); 1307SYSCTL_INT(_debug_softdep, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, 1308 &stat_direct_blk_ptrs, 0, ""); 1309SYSCTL_INT(_debug_softdep, OID_AUTO, dir_entry, CTLFLAG_RW, 1310 &stat_dir_entry, 0, ""); 1311SYSCTL_INT(_debug_softdep, OID_AUTO, jaddref_rollback, CTLFLAG_RW, 1312 &stat_jaddref, 0, ""); 1313SYSCTL_INT(_debug_softdep, OID_AUTO, jnewblk_rollback, CTLFLAG_RW, 1314 &stat_jnewblk, 0, ""); 1315SYSCTL_INT(_debug_softdep, OID_AUTO, journal_low, CTLFLAG_RW, 1316 &stat_journal_low, 0, ""); 1317SYSCTL_INT(_debug_softdep, OID_AUTO, journal_min, CTLFLAG_RW, 1318 &stat_journal_min, 0, ""); 1319SYSCTL_INT(_debug_softdep, OID_AUTO, journal_wait, CTLFLAG_RW, 1320 &stat_journal_wait, 0, ""); 1321SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_filepage, CTLFLAG_RW, 1322 &stat_jwait_filepage, 0, ""); 1323SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_freeblks, CTLFLAG_RW, 1324 &stat_jwait_freeblks, 0, ""); 1325SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_inode, CTLFLAG_RW, 1326 &stat_jwait_inode, 0, ""); 1327SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_newblk, CTLFLAG_RW, 1328 &stat_jwait_newblk, 0, ""); 1329SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_blkrequests, CTLFLAG_RW, 1330 &stat_cleanup_blkrequests, 0, ""); 1331SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_inorequests, CTLFLAG_RW, 1332 &stat_cleanup_inorequests, 0, ""); 1333SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_high_delay, CTLFLAG_RW, 1334 &stat_cleanup_high_delay, 0, ""); 1335SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_retries, CTLFLAG_RW, 1336 &stat_cleanup_retries, 0, ""); 1337SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_failures, CTLFLAG_RW, 1338 &stat_cleanup_failures, 0, ""); 1339SYSCTL_INT(_debug_softdep, OID_AUTO, flushcache, CTLFLAG_RW, 1340 &softdep_flushcache, 0, ""); 1341SYSCTL_INT(_debug_softdep, OID_AUTO, emptyjblocks, CTLFLAG_RD, 1342 &stat_emptyjblocks, 0, ""); 1343 1344SYSCTL_DECL(_vfs_ffs); 1345 1346/* Whether to recompute the summary at mount time */ 1347static int compute_summary_at_mount = 0; 1348SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW, 1349 &compute_summary_at_mount, 0, "Recompute summary at mount"); 1350static int print_threads = 0; 1351SYSCTL_INT(_debug_softdep, OID_AUTO, print_threads, CTLFLAG_RW, 1352 &print_threads, 0, "Notify flusher thread start/stop"); 1353 1354/* List of all filesystems mounted with soft updates */ 1355static TAILQ_HEAD(, mount_softdeps) softdepmounts; 1356 1357/* 1358 * This function cleans the worklist for a filesystem. 1359 * Each filesystem running with soft dependencies gets its own 1360 * thread to run in this function. The thread is started up in 1361 * softdep_mount and shutdown in softdep_unmount. They show up 1362 * as part of the kernel "bufdaemon" process whose process 1363 * entry is available in bufdaemonproc. 1364 */ 1365static int searchfailed; 1366extern struct proc *bufdaemonproc; 1367static void 1368softdep_flush(addr) 1369 void *addr; 1370{ 1371 struct mount *mp; 1372 struct thread *td; 1373 struct ufsmount *ump; 1374 1375 td = curthread; 1376 td->td_pflags |= TDP_NORUNNINGBUF; 1377 mp = (struct mount *)addr; 1378 ump = VFSTOUFS(mp); 1379 atomic_add_int(&stat_flush_threads, 1); 1380 ACQUIRE_LOCK(ump); 1381 ump->softdep_flags &= ~FLUSH_STARTING; 1382 wakeup(&ump->softdep_flushtd); 1383 FREE_LOCK(ump); 1384 if (print_threads) { 1385 if (stat_flush_threads == 1) 1386 printf("Running %s at pid %d\n", bufdaemonproc->p_comm, 1387 bufdaemonproc->p_pid); 1388 printf("Start thread %s\n", td->td_name); 1389 } 1390 for (;;) { 1391 while (softdep_process_worklist(mp, 0) > 0 || 1392 (MOUNTEDSUJ(mp) && 1393 VFSTOUFS(mp)->softdep_jblocks->jb_suspended)) 1394 kthread_suspend_check(); 1395 ACQUIRE_LOCK(ump); 1396 if ((ump->softdep_flags & (FLUSH_CLEANUP | FLUSH_EXIT)) == 0) 1397 msleep(&ump->softdep_flushtd, LOCK_PTR(ump), PVM, 1398 "sdflush", hz / 2); 1399 ump->softdep_flags &= ~FLUSH_CLEANUP; 1400 /* 1401 * Check to see if we are done and need to exit. 1402 */ 1403 if ((ump->softdep_flags & FLUSH_EXIT) == 0) { 1404 FREE_LOCK(ump); 1405 continue; 1406 } 1407 ump->softdep_flags &= ~FLUSH_EXIT; 1408 FREE_LOCK(ump); 1409 wakeup(&ump->softdep_flags); 1410 if (print_threads) 1411 printf("Stop thread %s: searchfailed %d, did cleanups %d\n", td->td_name, searchfailed, ump->um_softdep->sd_cleanups); 1412 atomic_subtract_int(&stat_flush_threads, 1); 1413 kthread_exit(); 1414 panic("kthread_exit failed\n"); 1415 } 1416} 1417 1418static void 1419worklist_speedup(mp) 1420 struct mount *mp; 1421{ 1422 struct ufsmount *ump; 1423 1424 ump = VFSTOUFS(mp); 1425 LOCK_OWNED(ump); 1426 if ((ump->softdep_flags & (FLUSH_CLEANUP | FLUSH_EXIT)) == 0) 1427 ump->softdep_flags |= FLUSH_CLEANUP; 1428 wakeup(&ump->softdep_flushtd); 1429} 1430 1431static int 1432softdep_speedup(ump) 1433 struct ufsmount *ump; 1434{ 1435 struct ufsmount *altump; 1436 struct mount_softdeps *sdp; 1437 1438 LOCK_OWNED(ump); 1439 worklist_speedup(ump->um_mountp); 1440 bd_speedup(); 1441 /* 1442 * If we have global shortages, then we need other 1443 * filesystems to help with the cleanup. Here we wakeup a 1444 * flusher thread for a filesystem that is over its fair 1445 * share of resources. 1446 */ 1447 if (req_clear_inodedeps || req_clear_remove) { 1448 ACQUIRE_GBLLOCK(&lk); 1449 TAILQ_FOREACH(sdp, &softdepmounts, sd_next) { 1450 if ((altump = sdp->sd_ump) == ump) 1451 continue; 1452 if (((req_clear_inodedeps && 1453 altump->softdep_curdeps[D_INODEDEP] > 1454 max_softdeps / stat_flush_threads) || 1455 (req_clear_remove && 1456 altump->softdep_curdeps[D_DIRREM] > 1457 (max_softdeps / 2) / stat_flush_threads)) && 1458 TRY_ACQUIRE_LOCK(altump)) 1459 break; 1460 } 1461 if (sdp == NULL) { 1462 searchfailed++; 1463 FREE_GBLLOCK(&lk); 1464 } else { 1465 /* 1466 * Move to the end of the list so we pick a 1467 * different one on out next try. 1468 */ 1469 TAILQ_REMOVE(&softdepmounts, sdp, sd_next); 1470 TAILQ_INSERT_TAIL(&softdepmounts, sdp, sd_next); 1471 FREE_GBLLOCK(&lk); 1472 if ((altump->softdep_flags & 1473 (FLUSH_CLEANUP | FLUSH_EXIT)) == 0) 1474 altump->softdep_flags |= FLUSH_CLEANUP; 1475 altump->um_softdep->sd_cleanups++; 1476 wakeup(&altump->softdep_flushtd); 1477 FREE_LOCK(altump); 1478 } 1479 } 1480 return (speedup_syncer()); 1481} 1482 1483/* 1484 * Add an item to the end of the work queue. 1485 * This routine requires that the lock be held. 1486 * This is the only routine that adds items to the list. 1487 * The following routine is the only one that removes items 1488 * and does so in order from first to last. 1489 */ 1490 1491#define WK_HEAD 0x0001 /* Add to HEAD. */ 1492#define WK_NODELAY 0x0002 /* Process immediately. */ 1493 1494static void 1495add_to_worklist(wk, flags) 1496 struct worklist *wk; 1497 int flags; 1498{ 1499 struct ufsmount *ump; 1500 1501 ump = VFSTOUFS(wk->wk_mp); 1502 LOCK_OWNED(ump); 1503 if (wk->wk_state & ONWORKLIST) 1504 panic("add_to_worklist: %s(0x%X) already on list", 1505 TYPENAME(wk->wk_type), wk->wk_state); 1506 wk->wk_state |= ONWORKLIST; 1507 if (ump->softdep_on_worklist == 0) { 1508 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); 1509 ump->softdep_worklist_tail = wk; 1510 } else if (flags & WK_HEAD) { 1511 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); 1512 } else { 1513 LIST_INSERT_AFTER(ump->softdep_worklist_tail, wk, wk_list); 1514 ump->softdep_worklist_tail = wk; 1515 } 1516 ump->softdep_on_worklist += 1; 1517 if (flags & WK_NODELAY) 1518 worklist_speedup(wk->wk_mp); 1519} 1520 1521/* 1522 * Remove the item to be processed. If we are removing the last 1523 * item on the list, we need to recalculate the tail pointer. 1524 */ 1525static void 1526remove_from_worklist(wk) 1527 struct worklist *wk; 1528{ 1529 struct ufsmount *ump; 1530 1531 ump = VFSTOUFS(wk->wk_mp); 1532 WORKLIST_REMOVE(wk); 1533 if (ump->softdep_worklist_tail == wk) 1534 ump->softdep_worklist_tail = 1535 (struct worklist *)wk->wk_list.le_prev; 1536 ump->softdep_on_worklist -= 1; 1537} 1538 1539static void 1540wake_worklist(wk) 1541 struct worklist *wk; 1542{ 1543 if (wk->wk_state & IOWAITING) { 1544 wk->wk_state &= ~IOWAITING; 1545 wakeup(wk); 1546 } 1547} 1548 1549static void 1550wait_worklist(wk, wmesg) 1551 struct worklist *wk; 1552 char *wmesg; 1553{ 1554 struct ufsmount *ump; 1555 1556 ump = VFSTOUFS(wk->wk_mp); 1557 wk->wk_state |= IOWAITING; 1558 msleep(wk, LOCK_PTR(ump), PVM, wmesg, 0); 1559} 1560 1561/* 1562 * Process that runs once per second to handle items in the background queue. 1563 * 1564 * Note that we ensure that everything is done in the order in which they 1565 * appear in the queue. The code below depends on this property to ensure 1566 * that blocks of a file are freed before the inode itself is freed. This 1567 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 1568 * until all the old ones have been purged from the dependency lists. 1569 */ 1570static int 1571softdep_process_worklist(mp, full) 1572 struct mount *mp; 1573 int full; 1574{ 1575 int cnt, matchcnt; 1576 struct ufsmount *ump; 1577 long starttime; 1578 1579 KASSERT(mp != NULL, ("softdep_process_worklist: NULL mp")); 1580 if (MOUNTEDSOFTDEP(mp) == 0) 1581 return (0); 1582 matchcnt = 0; 1583 ump = VFSTOUFS(mp); 1584 ACQUIRE_LOCK(ump); 1585 starttime = time_second; 1586 softdep_process_journal(mp, NULL, full ? MNT_WAIT : 0); 1587 check_clear_deps(mp); 1588 while (ump->softdep_on_worklist > 0) { 1589 if ((cnt = process_worklist_item(mp, 10, LK_NOWAIT)) == 0) 1590 break; 1591 else 1592 matchcnt += cnt; 1593 check_clear_deps(mp); 1594 /* 1595 * We do not generally want to stop for buffer space, but if 1596 * we are really being a buffer hog, we will stop and wait. 1597 */ 1598 if (should_yield()) { 1599 FREE_LOCK(ump); 1600 kern_yield(PRI_USER); 1601 bwillwrite(); 1602 ACQUIRE_LOCK(ump); 1603 } 1604 /* 1605 * Never allow processing to run for more than one 1606 * second. This gives the syncer thread the opportunity 1607 * to pause if appropriate. 1608 */ 1609 if (!full && starttime != time_second) 1610 break; 1611 } 1612 if (full == 0) 1613 journal_unsuspend(ump); 1614 FREE_LOCK(ump); 1615 return (matchcnt); 1616} 1617 1618/* 1619 * Process all removes associated with a vnode if we are running out of 1620 * journal space. Any other process which attempts to flush these will 1621 * be unable as we have the vnodes locked. 1622 */ 1623static void 1624process_removes(vp) 1625 struct vnode *vp; 1626{ 1627 struct inodedep *inodedep; 1628 struct dirrem *dirrem; 1629 struct ufsmount *ump; 1630 struct mount *mp; 1631 ino_t inum; 1632 1633 mp = vp->v_mount; 1634 ump = VFSTOUFS(mp); 1635 LOCK_OWNED(ump); 1636 inum = VTOI(vp)->i_number; 1637 for (;;) { 1638top: 1639 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0) 1640 return; 1641 LIST_FOREACH(dirrem, &inodedep->id_dirremhd, dm_inonext) { 1642 /* 1643 * If another thread is trying to lock this vnode 1644 * it will fail but we must wait for it to do so 1645 * before we can proceed. 1646 */ 1647 if (dirrem->dm_state & INPROGRESS) { 1648 wait_worklist(&dirrem->dm_list, "pwrwait"); 1649 goto top; 1650 } 1651 if ((dirrem->dm_state & (COMPLETE | ONWORKLIST)) == 1652 (COMPLETE | ONWORKLIST)) 1653 break; 1654 } 1655 if (dirrem == NULL) 1656 return; 1657 remove_from_worklist(&dirrem->dm_list); 1658 FREE_LOCK(ump); 1659 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT)) 1660 panic("process_removes: suspended filesystem"); 1661 handle_workitem_remove(dirrem, 0); 1662 vn_finished_secondary_write(mp); 1663 ACQUIRE_LOCK(ump); 1664 } 1665} 1666 1667/* 1668 * Process all truncations associated with a vnode if we are running out 1669 * of journal space. This is called when the vnode lock is already held 1670 * and no other process can clear the truncation. This function returns 1671 * a value greater than zero if it did any work. 1672 */ 1673static void 1674process_truncates(vp) 1675 struct vnode *vp; 1676{ 1677 struct inodedep *inodedep; 1678 struct freeblks *freeblks; 1679 struct ufsmount *ump; 1680 struct mount *mp; 1681 ino_t inum; 1682 int cgwait; 1683 1684 mp = vp->v_mount; 1685 ump = VFSTOUFS(mp); 1686 LOCK_OWNED(ump); 1687 inum = VTOI(vp)->i_number; 1688 for (;;) { 1689 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0) 1690 return; 1691 cgwait = 0; 1692 TAILQ_FOREACH(freeblks, &inodedep->id_freeblklst, fb_next) { 1693 /* Journal entries not yet written. */ 1694 if (!LIST_EMPTY(&freeblks->fb_jblkdephd)) { 1695 jwait(&LIST_FIRST( 1696 &freeblks->fb_jblkdephd)->jb_list, 1697 MNT_WAIT); 1698 break; 1699 } 1700 /* Another thread is executing this item. */ 1701 if (freeblks->fb_state & INPROGRESS) { 1702 wait_worklist(&freeblks->fb_list, "ptrwait"); 1703 break; 1704 } 1705 /* Freeblks is waiting on a inode write. */ 1706 if ((freeblks->fb_state & COMPLETE) == 0) { 1707 FREE_LOCK(ump); 1708 ffs_update(vp, 1); 1709 ACQUIRE_LOCK(ump); 1710 break; 1711 } 1712 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST)) == 1713 (ALLCOMPLETE | ONWORKLIST)) { 1714 remove_from_worklist(&freeblks->fb_list); 1715 freeblks->fb_state |= INPROGRESS; 1716 FREE_LOCK(ump); 1717 if (vn_start_secondary_write(NULL, &mp, 1718 V_NOWAIT)) 1719 panic("process_truncates: " 1720 "suspended filesystem"); 1721 handle_workitem_freeblocks(freeblks, 0); 1722 vn_finished_secondary_write(mp); 1723 ACQUIRE_LOCK(ump); 1724 break; 1725 } 1726 if (freeblks->fb_cgwait) 1727 cgwait++; 1728 } 1729 if (cgwait) { 1730 FREE_LOCK(ump); 1731 sync_cgs(mp, MNT_WAIT); 1732 ffs_sync_snap(mp, MNT_WAIT); 1733 ACQUIRE_LOCK(ump); 1734 continue; 1735 } 1736 if (freeblks == NULL) 1737 break; 1738 } 1739 return; 1740} 1741 1742/* 1743 * Process one item on the worklist. 1744 */ 1745static int 1746process_worklist_item(mp, target, flags) 1747 struct mount *mp; 1748 int target; 1749 int flags; 1750{ 1751 struct worklist sentinel; 1752 struct worklist *wk; 1753 struct ufsmount *ump; 1754 int matchcnt; 1755 int error; 1756 1757 KASSERT(mp != NULL, ("process_worklist_item: NULL mp")); 1758 /* 1759 * If we are being called because of a process doing a 1760 * copy-on-write, then it is not safe to write as we may 1761 * recurse into the copy-on-write routine. 1762 */ 1763 if (curthread->td_pflags & TDP_COWINPROGRESS) 1764 return (-1); 1765 PHOLD(curproc); /* Don't let the stack go away. */ 1766 ump = VFSTOUFS(mp); 1767 LOCK_OWNED(ump); 1768 matchcnt = 0; 1769 sentinel.wk_mp = NULL; 1770 sentinel.wk_type = D_SENTINEL; 1771 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, &sentinel, wk_list); 1772 for (wk = LIST_NEXT(&sentinel, wk_list); wk != NULL; 1773 wk = LIST_NEXT(&sentinel, wk_list)) { 1774 if (wk->wk_type == D_SENTINEL) { 1775 LIST_REMOVE(&sentinel, wk_list); 1776 LIST_INSERT_AFTER(wk, &sentinel, wk_list); 1777 continue; 1778 } 1779 if (wk->wk_state & INPROGRESS) 1780 panic("process_worklist_item: %p already in progress.", 1781 wk); 1782 wk->wk_state |= INPROGRESS; 1783 remove_from_worklist(wk); 1784 FREE_LOCK(ump); 1785 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT)) 1786 panic("process_worklist_item: suspended filesystem"); 1787 switch (wk->wk_type) { 1788 case D_DIRREM: 1789 /* removal of a directory entry */ 1790 error = handle_workitem_remove(WK_DIRREM(wk), flags); 1791 break; 1792 1793 case D_FREEBLKS: 1794 /* releasing blocks and/or fragments from a file */ 1795 error = handle_workitem_freeblocks(WK_FREEBLKS(wk), 1796 flags); 1797 break; 1798 1799 case D_FREEFRAG: 1800 /* releasing a fragment when replaced as a file grows */ 1801 handle_workitem_freefrag(WK_FREEFRAG(wk)); 1802 error = 0; 1803 break; 1804 1805 case D_FREEFILE: 1806 /* releasing an inode when its link count drops to 0 */ 1807 handle_workitem_freefile(WK_FREEFILE(wk)); 1808 error = 0; 1809 break; 1810 1811 default: 1812 panic("%s_process_worklist: Unknown type %s", 1813 "softdep", TYPENAME(wk->wk_type)); 1814 /* NOTREACHED */ 1815 } 1816 vn_finished_secondary_write(mp); 1817 ACQUIRE_LOCK(ump); 1818 if (error == 0) { 1819 if (++matchcnt == target) 1820 break; 1821 continue; 1822 } 1823 /* 1824 * We have to retry the worklist item later. Wake up any 1825 * waiters who may be able to complete it immediately and 1826 * add the item back to the head so we don't try to execute 1827 * it again. 1828 */ 1829 wk->wk_state &= ~INPROGRESS; 1830 wake_worklist(wk); 1831 add_to_worklist(wk, WK_HEAD); 1832 } 1833 LIST_REMOVE(&sentinel, wk_list); 1834 /* Sentinal could've become the tail from remove_from_worklist. */ 1835 if (ump->softdep_worklist_tail == &sentinel) 1836 ump->softdep_worklist_tail = 1837 (struct worklist *)sentinel.wk_list.le_prev; 1838 PRELE(curproc); 1839 return (matchcnt); 1840} 1841 1842/* 1843 * Move dependencies from one buffer to another. 1844 */ 1845int 1846softdep_move_dependencies(oldbp, newbp) 1847 struct buf *oldbp; 1848 struct buf *newbp; 1849{ 1850 struct worklist *wk, *wktail; 1851 struct ufsmount *ump; 1852 int dirty; 1853 1854 if ((wk = LIST_FIRST(&oldbp->b_dep)) == NULL) 1855 return (0); 1856 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0, 1857 ("softdep_move_dependencies called on non-softdep filesystem")); 1858 dirty = 0; 1859 wktail = NULL; 1860 ump = VFSTOUFS(wk->wk_mp); 1861 ACQUIRE_LOCK(ump); 1862 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 1863 LIST_REMOVE(wk, wk_list); 1864 if (wk->wk_type == D_BMSAFEMAP && 1865 bmsafemap_backgroundwrite(WK_BMSAFEMAP(wk), newbp)) 1866 dirty = 1; 1867 if (wktail == 0) 1868 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 1869 else 1870 LIST_INSERT_AFTER(wktail, wk, wk_list); 1871 wktail = wk; 1872 } 1873 FREE_LOCK(ump); 1874 1875 return (dirty); 1876} 1877 1878/* 1879 * Purge the work list of all items associated with a particular mount point. 1880 */ 1881int 1882softdep_flushworklist(oldmnt, countp, td) 1883 struct mount *oldmnt; 1884 int *countp; 1885 struct thread *td; 1886{ 1887 struct vnode *devvp; 1888 int count, error = 0; 1889 struct ufsmount *ump; 1890 1891 /* 1892 * Alternately flush the block device associated with the mount 1893 * point and process any dependencies that the flushing 1894 * creates. We continue until no more worklist dependencies 1895 * are found. 1896 */ 1897 *countp = 0; 1898 ump = VFSTOUFS(oldmnt); 1899 devvp = ump->um_devvp; 1900 while ((count = softdep_process_worklist(oldmnt, 1)) > 0) { 1901 *countp += count; 1902 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1903 error = VOP_FSYNC(devvp, MNT_WAIT, td); 1904 VOP_UNLOCK(devvp, 0); 1905 if (error) 1906 break; 1907 } 1908 return (error); 1909} 1910 1911static int 1912softdep_waitidle(struct mount *mp, int flags __unused) 1913{ 1914 struct ufsmount *ump; 1915 int error; 1916 int i; 1917 1918 ump = VFSTOUFS(mp); 1919 ACQUIRE_LOCK(ump); 1920 for (i = 0; i < 10 && ump->softdep_deps; i++) { 1921 ump->softdep_req = 1; 1922 KASSERT((flags & FORCECLOSE) == 0 || 1923 ump->softdep_on_worklist == 0, 1924 ("softdep_waitidle: work added after flush")); 1925 msleep(&ump->softdep_deps, LOCK_PTR(ump), PVM, "softdeps", 1); 1926 } 1927 ump->softdep_req = 0; 1928 FREE_LOCK(ump); 1929 error = 0; 1930 if (i == 10) { 1931 error = EBUSY; 1932 printf("softdep_waitidle: Failed to flush worklist for %p\n", 1933 mp); 1934 } 1935 1936 return (error); 1937} 1938 1939/* 1940 * Flush all vnodes and worklist items associated with a specified mount point. 1941 */ 1942int 1943softdep_flushfiles(oldmnt, flags, td) 1944 struct mount *oldmnt; 1945 int flags; 1946 struct thread *td; 1947{ 1948#ifdef QUOTA 1949 struct ufsmount *ump; 1950 int i; 1951#endif 1952 int error, early, depcount, loopcnt, retry_flush_count, retry; 1953 int morework; 1954 1955 KASSERT(MOUNTEDSOFTDEP(oldmnt) != 0, 1956 ("softdep_flushfiles called on non-softdep filesystem")); 1957 loopcnt = 10; 1958 retry_flush_count = 3; 1959retry_flush: 1960 error = 0; 1961 1962 /* 1963 * Alternately flush the vnodes associated with the mount 1964 * point and process any dependencies that the flushing 1965 * creates. In theory, this loop can happen at most twice, 1966 * but we give it a few extra just to be sure. 1967 */ 1968 for (; loopcnt > 0; loopcnt--) { 1969 /* 1970 * Do another flush in case any vnodes were brought in 1971 * as part of the cleanup operations. 1972 */ 1973 early = retry_flush_count == 1 || (oldmnt->mnt_kern_flag & 1974 MNTK_UNMOUNT) == 0 ? 0 : EARLYFLUSH; 1975 if ((error = ffs_flushfiles(oldmnt, flags | early, td)) != 0) 1976 break; 1977 if ((error = softdep_flushworklist(oldmnt, &depcount, td)) != 0 || 1978 depcount == 0) 1979 break; 1980 } 1981 /* 1982 * If we are unmounting then it is an error to fail. If we 1983 * are simply trying to downgrade to read-only, then filesystem 1984 * activity can keep us busy forever, so we just fail with EBUSY. 1985 */ 1986 if (loopcnt == 0) { 1987 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 1988 panic("softdep_flushfiles: looping"); 1989 error = EBUSY; 1990 } 1991 if (!error) 1992 error = softdep_waitidle(oldmnt, flags); 1993 if (!error) { 1994 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) { 1995 retry = 0; 1996 MNT_ILOCK(oldmnt); 1997 KASSERT((oldmnt->mnt_kern_flag & MNTK_NOINSMNTQ) != 0, 1998 ("softdep_flushfiles: !MNTK_NOINSMNTQ")); 1999 morework = oldmnt->mnt_nvnodelistsize > 0; 2000#ifdef QUOTA 2001 ump = VFSTOUFS(oldmnt); 2002 UFS_LOCK(ump); 2003 for (i = 0; i < MAXQUOTAS; i++) { 2004 if (ump->um_quotas[i] != NULLVP) 2005 morework = 1; 2006 } 2007 UFS_UNLOCK(ump); 2008#endif 2009 if (morework) { 2010 if (--retry_flush_count > 0) { 2011 retry = 1; 2012 loopcnt = 3; 2013 } else 2014 error = EBUSY; 2015 } 2016 MNT_IUNLOCK(oldmnt); 2017 if (retry) 2018 goto retry_flush; 2019 } 2020 } 2021 return (error); 2022} 2023 2024/* 2025 * Structure hashing. 2026 * 2027 * There are four types of structures that can be looked up: 2028 * 1) pagedep structures identified by mount point, inode number, 2029 * and logical block. 2030 * 2) inodedep structures identified by mount point and inode number. 2031 * 3) newblk structures identified by mount point and 2032 * physical block number. 2033 * 4) bmsafemap structures identified by mount point and 2034 * cylinder group number. 2035 * 2036 * The "pagedep" and "inodedep" dependency structures are hashed 2037 * separately from the file blocks and inodes to which they correspond. 2038 * This separation helps when the in-memory copy of an inode or 2039 * file block must be replaced. It also obviates the need to access 2040 * an inode or file page when simply updating (or de-allocating) 2041 * dependency structures. Lookup of newblk structures is needed to 2042 * find newly allocated blocks when trying to associate them with 2043 * their allocdirect or allocindir structure. 2044 * 2045 * The lookup routines optionally create and hash a new instance when 2046 * an existing entry is not found. The bmsafemap lookup routine always 2047 * allocates a new structure if an existing one is not found. 2048 */ 2049#define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 2050#define NODELAY 0x0002 /* cannot do background work */ 2051 2052/* 2053 * Structures and routines associated with pagedep caching. 2054 */ 2055#define PAGEDEP_HASH(ump, inum, lbn) \ 2056 (&(ump)->pagedep_hashtbl[((inum) + (lbn)) & (ump)->pagedep_hash_size]) 2057 2058static int 2059pagedep_find(pagedephd, ino, lbn, pagedeppp) 2060 struct pagedep_hashhead *pagedephd; 2061 ino_t ino; 2062 ufs_lbn_t lbn; 2063 struct pagedep **pagedeppp; 2064{ 2065 struct pagedep *pagedep; 2066 2067 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 2068 if (ino == pagedep->pd_ino && lbn == pagedep->pd_lbn) { 2069 *pagedeppp = pagedep; 2070 return (1); 2071 } 2072 } 2073 *pagedeppp = NULL; 2074 return (0); 2075} 2076/* 2077 * Look up a pagedep. Return 1 if found, 0 otherwise. 2078 * If not found, allocate if DEPALLOC flag is passed. 2079 * Found or allocated entry is returned in pagedeppp. 2080 * This routine must be called with splbio interrupts blocked. 2081 */ 2082static int 2083pagedep_lookup(mp, bp, ino, lbn, flags, pagedeppp) 2084 struct mount *mp; 2085 struct buf *bp; 2086 ino_t ino; 2087 ufs_lbn_t lbn; 2088 int flags; 2089 struct pagedep **pagedeppp; 2090{ 2091 struct pagedep *pagedep; 2092 struct pagedep_hashhead *pagedephd; 2093 struct worklist *wk; 2094 struct ufsmount *ump; 2095 int ret; 2096 int i; 2097 2098 ump = VFSTOUFS(mp); 2099 LOCK_OWNED(ump); 2100 if (bp) { 2101 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 2102 if (wk->wk_type == D_PAGEDEP) { 2103 *pagedeppp = WK_PAGEDEP(wk); 2104 return (1); 2105 } 2106 } 2107 } 2108 pagedephd = PAGEDEP_HASH(ump, ino, lbn); 2109 ret = pagedep_find(pagedephd, ino, lbn, pagedeppp); 2110 if (ret) { 2111 if (((*pagedeppp)->pd_state & ONWORKLIST) == 0 && bp) 2112 WORKLIST_INSERT(&bp->b_dep, &(*pagedeppp)->pd_list); 2113 return (1); 2114 } 2115 if ((flags & DEPALLOC) == 0) 2116 return (0); 2117 FREE_LOCK(ump); 2118 pagedep = malloc(sizeof(struct pagedep), 2119 M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO); 2120 workitem_alloc(&pagedep->pd_list, D_PAGEDEP, mp); 2121 ACQUIRE_LOCK(ump); 2122 ret = pagedep_find(pagedephd, ino, lbn, pagedeppp); 2123 if (*pagedeppp) { 2124 /* 2125 * This should never happen since we only create pagedeps 2126 * with the vnode lock held. Could be an assert. 2127 */ 2128 WORKITEM_FREE(pagedep, D_PAGEDEP); 2129 return (ret); 2130 } 2131 pagedep->pd_ino = ino; 2132 pagedep->pd_lbn = lbn; 2133 LIST_INIT(&pagedep->pd_dirremhd); 2134 LIST_INIT(&pagedep->pd_pendinghd); 2135 for (i = 0; i < DAHASHSZ; i++) 2136 LIST_INIT(&pagedep->pd_diraddhd[i]); 2137 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 2138 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2139 *pagedeppp = pagedep; 2140 return (0); 2141} 2142 2143/* 2144 * Structures and routines associated with inodedep caching. 2145 */ 2146#define INODEDEP_HASH(ump, inum) \ 2147 (&(ump)->inodedep_hashtbl[(inum) & (ump)->inodedep_hash_size]) 2148 2149static int 2150inodedep_find(inodedephd, inum, inodedeppp) 2151 struct inodedep_hashhead *inodedephd; 2152 ino_t inum; 2153 struct inodedep **inodedeppp; 2154{ 2155 struct inodedep *inodedep; 2156 2157 LIST_FOREACH(inodedep, inodedephd, id_hash) 2158 if (inum == inodedep->id_ino) 2159 break; 2160 if (inodedep) { 2161 *inodedeppp = inodedep; 2162 return (1); 2163 } 2164 *inodedeppp = NULL; 2165 2166 return (0); 2167} 2168/* 2169 * Look up an inodedep. Return 1 if found, 0 if not found. 2170 * If not found, allocate if DEPALLOC flag is passed. 2171 * Found or allocated entry is returned in inodedeppp. 2172 * This routine must be called with splbio interrupts blocked. 2173 */ 2174static int 2175inodedep_lookup(mp, inum, flags, inodedeppp) 2176 struct mount *mp; 2177 ino_t inum; 2178 int flags; 2179 struct inodedep **inodedeppp; 2180{ 2181 struct inodedep *inodedep; 2182 struct inodedep_hashhead *inodedephd; 2183 struct ufsmount *ump; 2184 struct fs *fs; 2185 2186 ump = VFSTOUFS(mp); 2187 LOCK_OWNED(ump); 2188 fs = ump->um_fs; 2189 inodedephd = INODEDEP_HASH(ump, inum); 2190 2191 if (inodedep_find(inodedephd, inum, inodedeppp)) 2192 return (1); 2193 if ((flags & DEPALLOC) == 0) 2194 return (0); 2195 /* 2196 * If the system is over its limit and our filesystem is 2197 * responsible for more than our share of that usage and 2198 * we are not in a rush, request some inodedep cleanup. 2199 */ 2200 while (dep_current[D_INODEDEP] > max_softdeps && 2201 (flags & NODELAY) == 0 && 2202 ump->softdep_curdeps[D_INODEDEP] > 2203 max_softdeps / stat_flush_threads) 2204 request_cleanup(mp, FLUSH_INODES); 2205 FREE_LOCK(ump); 2206 inodedep = malloc(sizeof(struct inodedep), 2207 M_INODEDEP, M_SOFTDEP_FLAGS); 2208 workitem_alloc(&inodedep->id_list, D_INODEDEP, mp); 2209 ACQUIRE_LOCK(ump); 2210 if (inodedep_find(inodedephd, inum, inodedeppp)) { 2211 WORKITEM_FREE(inodedep, D_INODEDEP); 2212 return (1); 2213 } 2214 inodedep->id_fs = fs; 2215 inodedep->id_ino = inum; 2216 inodedep->id_state = ALLCOMPLETE; 2217 inodedep->id_nlinkdelta = 0; 2218 inodedep->id_savedino1 = NULL; 2219 inodedep->id_savedsize = -1; 2220 inodedep->id_savedextsize = -1; 2221 inodedep->id_savednlink = -1; 2222 inodedep->id_bmsafemap = NULL; 2223 inodedep->id_mkdiradd = NULL; 2224 LIST_INIT(&inodedep->id_dirremhd); 2225 LIST_INIT(&inodedep->id_pendinghd); 2226 LIST_INIT(&inodedep->id_inowait); 2227 LIST_INIT(&inodedep->id_bufwait); 2228 TAILQ_INIT(&inodedep->id_inoreflst); 2229 TAILQ_INIT(&inodedep->id_inoupdt); 2230 TAILQ_INIT(&inodedep->id_newinoupdt); 2231 TAILQ_INIT(&inodedep->id_extupdt); 2232 TAILQ_INIT(&inodedep->id_newextupdt); 2233 TAILQ_INIT(&inodedep->id_freeblklst); 2234 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 2235 *inodedeppp = inodedep; 2236 return (0); 2237} 2238 2239/* 2240 * Structures and routines associated with newblk caching. 2241 */ 2242#define NEWBLK_HASH(ump, inum) \ 2243 (&(ump)->newblk_hashtbl[(inum) & (ump)->newblk_hash_size]) 2244 2245static int 2246newblk_find(newblkhd, newblkno, flags, newblkpp) 2247 struct newblk_hashhead *newblkhd; 2248 ufs2_daddr_t newblkno; 2249 int flags; 2250 struct newblk **newblkpp; 2251{ 2252 struct newblk *newblk; 2253 2254 LIST_FOREACH(newblk, newblkhd, nb_hash) { 2255 if (newblkno != newblk->nb_newblkno) 2256 continue; 2257 /* 2258 * If we're creating a new dependency don't match those that 2259 * have already been converted to allocdirects. This is for 2260 * a frag extend. 2261 */ 2262 if ((flags & DEPALLOC) && newblk->nb_list.wk_type != D_NEWBLK) 2263 continue; 2264 break; 2265 } 2266 if (newblk) { 2267 *newblkpp = newblk; 2268 return (1); 2269 } 2270 *newblkpp = NULL; 2271 return (0); 2272} 2273 2274/* 2275 * Look up a newblk. Return 1 if found, 0 if not found. 2276 * If not found, allocate if DEPALLOC flag is passed. 2277 * Found or allocated entry is returned in newblkpp. 2278 */ 2279static int 2280newblk_lookup(mp, newblkno, flags, newblkpp) 2281 struct mount *mp; 2282 ufs2_daddr_t newblkno; 2283 int flags; 2284 struct newblk **newblkpp; 2285{ 2286 struct newblk *newblk; 2287 struct newblk_hashhead *newblkhd; 2288 struct ufsmount *ump; 2289 2290 ump = VFSTOUFS(mp); 2291 LOCK_OWNED(ump); 2292 newblkhd = NEWBLK_HASH(ump, newblkno); 2293 if (newblk_find(newblkhd, newblkno, flags, newblkpp)) 2294 return (1); 2295 if ((flags & DEPALLOC) == 0) 2296 return (0); 2297 FREE_LOCK(ump); 2298 newblk = malloc(sizeof(union allblk), M_NEWBLK, 2299 M_SOFTDEP_FLAGS | M_ZERO); 2300 workitem_alloc(&newblk->nb_list, D_NEWBLK, mp); 2301 ACQUIRE_LOCK(ump); 2302 if (newblk_find(newblkhd, newblkno, flags, newblkpp)) { 2303 WORKITEM_FREE(newblk, D_NEWBLK); 2304 return (1); 2305 } 2306 newblk->nb_freefrag = NULL; 2307 LIST_INIT(&newblk->nb_indirdeps); 2308 LIST_INIT(&newblk->nb_newdirblk); 2309 LIST_INIT(&newblk->nb_jwork); 2310 newblk->nb_state = ATTACHED; 2311 newblk->nb_newblkno = newblkno; 2312 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 2313 *newblkpp = newblk; 2314 return (0); 2315} 2316 2317/* 2318 * Structures and routines associated with freed indirect block caching. 2319 */ 2320#define INDIR_HASH(ump, blkno) \ 2321 (&(ump)->indir_hashtbl[(blkno) & (ump)->indir_hash_size]) 2322 2323/* 2324 * Lookup an indirect block in the indir hash table. The freework is 2325 * removed and potentially freed. The caller must do a blocking journal 2326 * write before writing to the blkno. 2327 */ 2328static int 2329indirblk_lookup(mp, blkno) 2330 struct mount *mp; 2331 ufs2_daddr_t blkno; 2332{ 2333 struct freework *freework; 2334 struct indir_hashhead *wkhd; 2335 struct ufsmount *ump; 2336 2337 ump = VFSTOUFS(mp); 2338 wkhd = INDIR_HASH(ump, blkno); 2339 TAILQ_FOREACH(freework, wkhd, fw_next) { 2340 if (freework->fw_blkno != blkno) 2341 continue; 2342 indirblk_remove(freework); 2343 return (1); 2344 } 2345 return (0); 2346} 2347 2348/* 2349 * Insert an indirect block represented by freework into the indirblk 2350 * hash table so that it may prevent the block from being re-used prior 2351 * to the journal being written. 2352 */ 2353static void 2354indirblk_insert(freework) 2355 struct freework *freework; 2356{ 2357 struct jblocks *jblocks; 2358 struct jseg *jseg; 2359 struct ufsmount *ump; 2360 2361 ump = VFSTOUFS(freework->fw_list.wk_mp); 2362 jblocks = ump->softdep_jblocks; 2363 jseg = TAILQ_LAST(&jblocks->jb_segs, jseglst); 2364 if (jseg == NULL) 2365 return; 2366 2367 LIST_INSERT_HEAD(&jseg->js_indirs, freework, fw_segs); 2368 TAILQ_INSERT_HEAD(INDIR_HASH(ump, freework->fw_blkno), freework, 2369 fw_next); 2370 freework->fw_state &= ~DEPCOMPLETE; 2371} 2372 2373static void 2374indirblk_remove(freework) 2375 struct freework *freework; 2376{ 2377 struct ufsmount *ump; 2378 2379 ump = VFSTOUFS(freework->fw_list.wk_mp); 2380 LIST_REMOVE(freework, fw_segs); 2381 TAILQ_REMOVE(INDIR_HASH(ump, freework->fw_blkno), freework, fw_next); 2382 freework->fw_state |= DEPCOMPLETE; 2383 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE) 2384 WORKITEM_FREE(freework, D_FREEWORK); 2385} 2386 2387/* 2388 * Executed during filesystem system initialization before 2389 * mounting any filesystems. 2390 */ 2391void 2392softdep_initialize() 2393{ 2394 2395 TAILQ_INIT(&softdepmounts); 2396 max_softdeps = desiredvnodes * 4; 2397 2398 /* initialise bioops hack */ 2399 bioops.io_start = softdep_disk_io_initiation; 2400 bioops.io_complete = softdep_disk_write_complete; 2401 bioops.io_deallocate = softdep_deallocate_dependencies; 2402 bioops.io_countdeps = softdep_count_dependencies; 2403 2404 /* Initialize the callout with an mtx. */ 2405 callout_init_mtx(&softdep_callout, &lk, 0); 2406} 2407 2408/* 2409 * Executed after all filesystems have been unmounted during 2410 * filesystem module unload. 2411 */ 2412void 2413softdep_uninitialize() 2414{ 2415 2416 /* clear bioops hack */ 2417 bioops.io_start = NULL; 2418 bioops.io_complete = NULL; 2419 bioops.io_deallocate = NULL; 2420 bioops.io_countdeps = NULL; 2421 2422 callout_drain(&softdep_callout); 2423} 2424 2425/* 2426 * Called at mount time to notify the dependency code that a 2427 * filesystem wishes to use it. 2428 */ 2429int 2430softdep_mount(devvp, mp, fs, cred) 2431 struct vnode *devvp; 2432 struct mount *mp; 2433 struct fs *fs; 2434 struct ucred *cred; 2435{ 2436 struct csum_total cstotal; 2437 struct mount_softdeps *sdp; 2438 struct ufsmount *ump; 2439 struct cg *cgp; 2440 struct buf *bp; 2441 int i, error, cyl; 2442 2443 sdp = malloc(sizeof(struct mount_softdeps), M_MOUNTDATA, 2444 M_WAITOK | M_ZERO); 2445 MNT_ILOCK(mp); 2446 mp->mnt_flag = (mp->mnt_flag & ~MNT_ASYNC) | MNT_SOFTDEP; 2447 if ((mp->mnt_kern_flag & MNTK_SOFTDEP) == 0) { 2448 mp->mnt_kern_flag = (mp->mnt_kern_flag & ~MNTK_ASYNC) | 2449 MNTK_SOFTDEP | MNTK_NOASYNC; 2450 } 2451 ump = VFSTOUFS(mp); 2452 ump->um_softdep = sdp; 2453 MNT_IUNLOCK(mp); 2454 rw_init(LOCK_PTR(ump), "Per-Filesystem Softdep Lock"); 2455 sdp->sd_ump = ump; 2456 LIST_INIT(&ump->softdep_workitem_pending); 2457 LIST_INIT(&ump->softdep_journal_pending); 2458 TAILQ_INIT(&ump->softdep_unlinked); 2459 LIST_INIT(&ump->softdep_dirtycg); 2460 ump->softdep_worklist_tail = NULL; 2461 ump->softdep_on_worklist = 0; 2462 ump->softdep_deps = 0; 2463 LIST_INIT(&ump->softdep_mkdirlisthd); 2464 ump->pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 2465 &ump->pagedep_hash_size); 2466 ump->pagedep_nextclean = 0; 2467 ump->inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, 2468 &ump->inodedep_hash_size); 2469 ump->inodedep_nextclean = 0; 2470 ump->newblk_hashtbl = hashinit(max_softdeps / 2, M_NEWBLK, 2471 &ump->newblk_hash_size); 2472 ump->bmsafemap_hashtbl = hashinit(1024, M_BMSAFEMAP, 2473 &ump->bmsafemap_hash_size); 2474 i = 1 << (ffs(desiredvnodes / 10) - 1); 2475 ump->indir_hashtbl = malloc(i * sizeof(struct indir_hashhead), 2476 M_FREEWORK, M_WAITOK); 2477 ump->indir_hash_size = i - 1; 2478 for (i = 0; i <= ump->indir_hash_size; i++) 2479 TAILQ_INIT(&ump->indir_hashtbl[i]); 2480 ACQUIRE_GBLLOCK(&lk); 2481 TAILQ_INSERT_TAIL(&softdepmounts, sdp, sd_next); 2482 FREE_GBLLOCK(&lk); 2483 if ((fs->fs_flags & FS_SUJ) && 2484 (error = journal_mount(mp, fs, cred)) != 0) { 2485 printf("Failed to start journal: %d\n", error); 2486 softdep_unmount(mp); 2487 return (error); 2488 } 2489 /* 2490 * Start our flushing thread in the bufdaemon process. 2491 */ 2492 ACQUIRE_LOCK(ump); 2493 ump->softdep_flags |= FLUSH_STARTING; 2494 FREE_LOCK(ump); 2495 kproc_kthread_add(&softdep_flush, mp, &bufdaemonproc, 2496 &ump->softdep_flushtd, 0, 0, "softdepflush", "%s worker", 2497 mp->mnt_stat.f_mntonname); 2498 ACQUIRE_LOCK(ump); 2499 while ((ump->softdep_flags & FLUSH_STARTING) != 0) { 2500 msleep(&ump->softdep_flushtd, LOCK_PTR(ump), PVM, "sdstart", 2501 hz / 2); 2502 } 2503 FREE_LOCK(ump); 2504 /* 2505 * When doing soft updates, the counters in the 2506 * superblock may have gotten out of sync. Recomputation 2507 * can take a long time and can be deferred for background 2508 * fsck. However, the old behavior of scanning the cylinder 2509 * groups and recalculating them at mount time is available 2510 * by setting vfs.ffs.compute_summary_at_mount to one. 2511 */ 2512 if (compute_summary_at_mount == 0 || fs->fs_clean != 0) 2513 return (0); 2514 bzero(&cstotal, sizeof cstotal); 2515 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 2516 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 2517 fs->fs_cgsize, cred, &bp)) != 0) { 2518 brelse(bp); 2519 softdep_unmount(mp); 2520 return (error); 2521 } 2522 cgp = (struct cg *)bp->b_data; 2523 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 2524 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 2525 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 2526 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 2527 fs->fs_cs(fs, cyl) = cgp->cg_cs; 2528 brelse(bp); 2529 } 2530#ifdef DEBUG 2531 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 2532 printf("%s: superblock summary recomputed\n", fs->fs_fsmnt); 2533#endif 2534 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 2535 return (0); 2536} 2537 2538void 2539softdep_unmount(mp) 2540 struct mount *mp; 2541{ 2542 struct ufsmount *ump; 2543#ifdef INVARIANTS 2544 int i; 2545#endif 2546 2547 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 2548 ("softdep_unmount called on non-softdep filesystem")); 2549 ump = VFSTOUFS(mp); 2550 MNT_ILOCK(mp); 2551 mp->mnt_flag &= ~MNT_SOFTDEP; 2552 if (MOUNTEDSUJ(mp) == 0) { 2553 MNT_IUNLOCK(mp); 2554 } else { 2555 mp->mnt_flag &= ~MNT_SUJ; 2556 MNT_IUNLOCK(mp); 2557 journal_unmount(ump); 2558 } 2559 /* 2560 * Shut down our flushing thread. Check for NULL is if 2561 * softdep_mount errors out before the thread has been created. 2562 */ 2563 if (ump->softdep_flushtd != NULL) { 2564 ACQUIRE_LOCK(ump); 2565 ump->softdep_flags |= FLUSH_EXIT; 2566 wakeup(&ump->softdep_flushtd); 2567 msleep(&ump->softdep_flags, LOCK_PTR(ump), PVM | PDROP, 2568 "sdwait", 0); 2569 KASSERT((ump->softdep_flags & FLUSH_EXIT) == 0, 2570 ("Thread shutdown failed")); 2571 } 2572 /* 2573 * Free up our resources. 2574 */ 2575 ACQUIRE_GBLLOCK(&lk); 2576 TAILQ_REMOVE(&softdepmounts, ump->um_softdep, sd_next); 2577 FREE_GBLLOCK(&lk); 2578 rw_destroy(LOCK_PTR(ump)); 2579 hashdestroy(ump->pagedep_hashtbl, M_PAGEDEP, ump->pagedep_hash_size); 2580 hashdestroy(ump->inodedep_hashtbl, M_INODEDEP, ump->inodedep_hash_size); 2581 hashdestroy(ump->newblk_hashtbl, M_NEWBLK, ump->newblk_hash_size); 2582 hashdestroy(ump->bmsafemap_hashtbl, M_BMSAFEMAP, 2583 ump->bmsafemap_hash_size); 2584 free(ump->indir_hashtbl, M_FREEWORK); 2585#ifdef INVARIANTS 2586 for (i = 0; i <= D_LAST; i++) 2587 KASSERT(ump->softdep_curdeps[i] == 0, 2588 ("Unmount %s: Dep type %s != 0 (%ld)", ump->um_fs->fs_fsmnt, 2589 TYPENAME(i), ump->softdep_curdeps[i])); 2590#endif 2591 free(ump->um_softdep, M_MOUNTDATA); 2592} 2593 2594static struct jblocks * 2595jblocks_create(void) 2596{ 2597 struct jblocks *jblocks; 2598 2599 jblocks = malloc(sizeof(*jblocks), M_JBLOCKS, M_WAITOK | M_ZERO); 2600 TAILQ_INIT(&jblocks->jb_segs); 2601 jblocks->jb_avail = 10; 2602 jblocks->jb_extent = malloc(sizeof(struct jextent) * jblocks->jb_avail, 2603 M_JBLOCKS, M_WAITOK | M_ZERO); 2604 2605 return (jblocks); 2606} 2607 2608static ufs2_daddr_t 2609jblocks_alloc(jblocks, bytes, actual) 2610 struct jblocks *jblocks; 2611 int bytes; 2612 int *actual; 2613{ 2614 ufs2_daddr_t daddr; 2615 struct jextent *jext; 2616 int freecnt; 2617 int blocks; 2618 2619 blocks = bytes / DEV_BSIZE; 2620 jext = &jblocks->jb_extent[jblocks->jb_head]; 2621 freecnt = jext->je_blocks - jblocks->jb_off; 2622 if (freecnt == 0) { 2623 jblocks->jb_off = 0; 2624 if (++jblocks->jb_head > jblocks->jb_used) 2625 jblocks->jb_head = 0; 2626 jext = &jblocks->jb_extent[jblocks->jb_head]; 2627 freecnt = jext->je_blocks; 2628 } 2629 if (freecnt > blocks) 2630 freecnt = blocks; 2631 *actual = freecnt * DEV_BSIZE; 2632 daddr = jext->je_daddr + jblocks->jb_off; 2633 jblocks->jb_off += freecnt; 2634 jblocks->jb_free -= freecnt; 2635 2636 return (daddr); 2637} 2638 2639static void 2640jblocks_free(jblocks, mp, bytes) 2641 struct jblocks *jblocks; 2642 struct mount *mp; 2643 int bytes; 2644{ 2645 2646 LOCK_OWNED(VFSTOUFS(mp)); 2647 jblocks->jb_free += bytes / DEV_BSIZE; 2648 if (jblocks->jb_suspended) 2649 worklist_speedup(mp); 2650 wakeup(jblocks); 2651} 2652 2653static void 2654jblocks_destroy(jblocks) 2655 struct jblocks *jblocks; 2656{ 2657 2658 if (jblocks->jb_extent) 2659 free(jblocks->jb_extent, M_JBLOCKS); 2660 free(jblocks, M_JBLOCKS); 2661} 2662 2663static void 2664jblocks_add(jblocks, daddr, blocks) 2665 struct jblocks *jblocks; 2666 ufs2_daddr_t daddr; 2667 int blocks; 2668{ 2669 struct jextent *jext; 2670 2671 jblocks->jb_blocks += blocks; 2672 jblocks->jb_free += blocks; 2673 jext = &jblocks->jb_extent[jblocks->jb_used]; 2674 /* Adding the first block. */ 2675 if (jext->je_daddr == 0) { 2676 jext->je_daddr = daddr; 2677 jext->je_blocks = blocks; 2678 return; 2679 } 2680 /* Extending the last extent. */ 2681 if (jext->je_daddr + jext->je_blocks == daddr) { 2682 jext->je_blocks += blocks; 2683 return; 2684 } 2685 /* Adding a new extent. */ 2686 if (++jblocks->jb_used == jblocks->jb_avail) { 2687 jblocks->jb_avail *= 2; 2688 jext = malloc(sizeof(struct jextent) * jblocks->jb_avail, 2689 M_JBLOCKS, M_WAITOK | M_ZERO); 2690 memcpy(jext, jblocks->jb_extent, 2691 sizeof(struct jextent) * jblocks->jb_used); 2692 free(jblocks->jb_extent, M_JBLOCKS); 2693 jblocks->jb_extent = jext; 2694 } 2695 jext = &jblocks->jb_extent[jblocks->jb_used]; 2696 jext->je_daddr = daddr; 2697 jext->je_blocks = blocks; 2698 return; 2699} 2700 2701int 2702softdep_journal_lookup(mp, vpp) 2703 struct mount *mp; 2704 struct vnode **vpp; 2705{ 2706 struct componentname cnp; 2707 struct vnode *dvp; 2708 ino_t sujournal; 2709 int error; 2710 2711 error = VFS_VGET(mp, ROOTINO, LK_EXCLUSIVE, &dvp); 2712 if (error) 2713 return (error); 2714 bzero(&cnp, sizeof(cnp)); 2715 cnp.cn_nameiop = LOOKUP; 2716 cnp.cn_flags = ISLASTCN; 2717 cnp.cn_thread = curthread; 2718 cnp.cn_cred = curthread->td_ucred; 2719 cnp.cn_pnbuf = SUJ_FILE; 2720 cnp.cn_nameptr = SUJ_FILE; 2721 cnp.cn_namelen = strlen(SUJ_FILE); 2722 error = ufs_lookup_ino(dvp, NULL, &cnp, &sujournal); 2723 vput(dvp); 2724 if (error != 0) 2725 return (error); 2726 error = VFS_VGET(mp, sujournal, LK_EXCLUSIVE, vpp); 2727 return (error); 2728} 2729 2730/* 2731 * Open and verify the journal file. 2732 */ 2733static int 2734journal_mount(mp, fs, cred) 2735 struct mount *mp; 2736 struct fs *fs; 2737 struct ucred *cred; 2738{ 2739 struct jblocks *jblocks; 2740 struct ufsmount *ump; 2741 struct vnode *vp; 2742 struct inode *ip; 2743 ufs2_daddr_t blkno; 2744 int bcount; 2745 int error; 2746 int i; 2747 2748 ump = VFSTOUFS(mp); 2749 ump->softdep_journal_tail = NULL; 2750 ump->softdep_on_journal = 0; 2751 ump->softdep_accdeps = 0; 2752 ump->softdep_req = 0; 2753 ump->softdep_jblocks = NULL; 2754 error = softdep_journal_lookup(mp, &vp); 2755 if (error != 0) { 2756 printf("Failed to find journal. Use tunefs to create one\n"); 2757 return (error); 2758 } 2759 ip = VTOI(vp); 2760 if (ip->i_size < SUJ_MIN) { 2761 error = ENOSPC; 2762 goto out; 2763 } 2764 bcount = lblkno(fs, ip->i_size); /* Only use whole blocks. */ 2765 jblocks = jblocks_create(); 2766 for (i = 0; i < bcount; i++) { 2767 error = ufs_bmaparray(vp, i, &blkno, NULL, NULL, NULL); 2768 if (error) 2769 break; 2770 jblocks_add(jblocks, blkno, fsbtodb(fs, fs->fs_frag)); 2771 } 2772 if (error) { 2773 jblocks_destroy(jblocks); 2774 goto out; 2775 } 2776 jblocks->jb_low = jblocks->jb_free / 3; /* Reserve 33%. */ 2777 jblocks->jb_min = jblocks->jb_free / 10; /* Suspend at 10%. */ 2778 ump->softdep_jblocks = jblocks; 2779out: 2780 if (error == 0) { 2781 MNT_ILOCK(mp); 2782 mp->mnt_flag |= MNT_SUJ; 2783 mp->mnt_flag &= ~MNT_SOFTDEP; 2784 MNT_IUNLOCK(mp); 2785 /* 2786 * Only validate the journal contents if the 2787 * filesystem is clean, otherwise we write the logs 2788 * but they'll never be used. If the filesystem was 2789 * still dirty when we mounted it the journal is 2790 * invalid and a new journal can only be valid if it 2791 * starts from a clean mount. 2792 */ 2793 if (fs->fs_clean) { 2794 DIP_SET(ip, i_modrev, fs->fs_mtime); 2795 ip->i_flags |= IN_MODIFIED; 2796 ffs_update(vp, 1); 2797 } 2798 } 2799 vput(vp); 2800 return (error); 2801} 2802 2803static void 2804journal_unmount(ump) 2805 struct ufsmount *ump; 2806{ 2807 2808 if (ump->softdep_jblocks) 2809 jblocks_destroy(ump->softdep_jblocks); 2810 ump->softdep_jblocks = NULL; 2811} 2812 2813/* 2814 * Called when a journal record is ready to be written. Space is allocated 2815 * and the journal entry is created when the journal is flushed to stable 2816 * store. 2817 */ 2818static void 2819add_to_journal(wk) 2820 struct worklist *wk; 2821{ 2822 struct ufsmount *ump; 2823 2824 ump = VFSTOUFS(wk->wk_mp); 2825 LOCK_OWNED(ump); 2826 if (wk->wk_state & ONWORKLIST) 2827 panic("add_to_journal: %s(0x%X) already on list", 2828 TYPENAME(wk->wk_type), wk->wk_state); 2829 wk->wk_state |= ONWORKLIST | DEPCOMPLETE; 2830 if (LIST_EMPTY(&ump->softdep_journal_pending)) { 2831 ump->softdep_jblocks->jb_age = ticks; 2832 LIST_INSERT_HEAD(&ump->softdep_journal_pending, wk, wk_list); 2833 } else 2834 LIST_INSERT_AFTER(ump->softdep_journal_tail, wk, wk_list); 2835 ump->softdep_journal_tail = wk; 2836 ump->softdep_on_journal += 1; 2837} 2838 2839/* 2840 * Remove an arbitrary item for the journal worklist maintain the tail 2841 * pointer. This happens when a new operation obviates the need to 2842 * journal an old operation. 2843 */ 2844static void 2845remove_from_journal(wk) 2846 struct worklist *wk; 2847{ 2848 struct ufsmount *ump; 2849 2850 ump = VFSTOUFS(wk->wk_mp); 2851 LOCK_OWNED(ump); 2852#ifdef SUJ_DEBUG 2853 { 2854 struct worklist *wkn; 2855 2856 LIST_FOREACH(wkn, &ump->softdep_journal_pending, wk_list) 2857 if (wkn == wk) 2858 break; 2859 if (wkn == NULL) 2860 panic("remove_from_journal: %p is not in journal", wk); 2861 } 2862#endif 2863 /* 2864 * We emulate a TAILQ to save space in most structures which do not 2865 * require TAILQ semantics. Here we must update the tail position 2866 * when removing the tail which is not the final entry. This works 2867 * only if the worklist linkage are at the beginning of the structure. 2868 */ 2869 if (ump->softdep_journal_tail == wk) 2870 ump->softdep_journal_tail = 2871 (struct worklist *)wk->wk_list.le_prev; 2872 2873 WORKLIST_REMOVE(wk); 2874 ump->softdep_on_journal -= 1; 2875} 2876 2877/* 2878 * Check for journal space as well as dependency limits so the prelink 2879 * code can throttle both journaled and non-journaled filesystems. 2880 * Threshold is 0 for low and 1 for min. 2881 */ 2882static int 2883journal_space(ump, thresh) 2884 struct ufsmount *ump; 2885 int thresh; 2886{ 2887 struct jblocks *jblocks; 2888 int limit, avail; 2889 2890 jblocks = ump->softdep_jblocks; 2891 if (jblocks == NULL) 2892 return (1); 2893 /* 2894 * We use a tighter restriction here to prevent request_cleanup() 2895 * running in threads from running into locks we currently hold. 2896 * We have to be over the limit and our filesystem has to be 2897 * responsible for more than our share of that usage. 2898 */ 2899 limit = (max_softdeps / 10) * 9; 2900 if (dep_current[D_INODEDEP] > limit && 2901 ump->softdep_curdeps[D_INODEDEP] > limit / stat_flush_threads) 2902 return (0); 2903 if (thresh) 2904 thresh = jblocks->jb_min; 2905 else 2906 thresh = jblocks->jb_low; 2907 avail = (ump->softdep_on_journal * JREC_SIZE) / DEV_BSIZE; 2908 avail = jblocks->jb_free - avail; 2909 2910 return (avail > thresh); 2911} 2912 2913static void 2914journal_suspend(ump) 2915 struct ufsmount *ump; 2916{ 2917 struct jblocks *jblocks; 2918 struct mount *mp; 2919 2920 mp = UFSTOVFS(ump); 2921 jblocks = ump->softdep_jblocks; 2922 MNT_ILOCK(mp); 2923 if ((mp->mnt_kern_flag & MNTK_SUSPEND) == 0) { 2924 stat_journal_min++; 2925 mp->mnt_kern_flag |= MNTK_SUSPEND; 2926 mp->mnt_susp_owner = ump->softdep_flushtd; 2927 } 2928 jblocks->jb_suspended = 1; 2929 MNT_IUNLOCK(mp); 2930} 2931 2932static int 2933journal_unsuspend(struct ufsmount *ump) 2934{ 2935 struct jblocks *jblocks; 2936 struct mount *mp; 2937 2938 mp = UFSTOVFS(ump); 2939 jblocks = ump->softdep_jblocks; 2940 2941 if (jblocks != NULL && jblocks->jb_suspended && 2942 journal_space(ump, jblocks->jb_min)) { 2943 jblocks->jb_suspended = 0; 2944 FREE_LOCK(ump); 2945 mp->mnt_susp_owner = curthread; 2946 vfs_write_resume(mp, 0); 2947 ACQUIRE_LOCK(ump); 2948 return (1); 2949 } 2950 return (0); 2951} 2952 2953/* 2954 * Called before any allocation function to be certain that there is 2955 * sufficient space in the journal prior to creating any new records. 2956 * Since in the case of block allocation we may have multiple locked 2957 * buffers at the time of the actual allocation we can not block 2958 * when the journal records are created. Doing so would create a deadlock 2959 * if any of these buffers needed to be flushed to reclaim space. Instead 2960 * we require a sufficiently large amount of available space such that 2961 * each thread in the system could have passed this allocation check and 2962 * still have sufficient free space. With 20% of a minimum journal size 2963 * of 1MB we have 6553 records available. 2964 */ 2965int 2966softdep_prealloc(vp, waitok) 2967 struct vnode *vp; 2968 int waitok; 2969{ 2970 struct ufsmount *ump; 2971 2972 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 2973 ("softdep_prealloc called on non-softdep filesystem")); 2974 /* 2975 * Nothing to do if we are not running journaled soft updates. 2976 * If we currently hold the snapshot lock, we must avoid handling 2977 * other resources that could cause deadlock. 2978 */ 2979 if (DOINGSUJ(vp) == 0 || IS_SNAPSHOT(VTOI(vp))) 2980 return (0); 2981 ump = VFSTOUFS(vp->v_mount); 2982 ACQUIRE_LOCK(ump); 2983 if (journal_space(ump, 0)) { 2984 FREE_LOCK(ump); 2985 return (0); 2986 } 2987 stat_journal_low++; 2988 FREE_LOCK(ump); 2989 if (waitok == MNT_NOWAIT) 2990 return (ENOSPC); 2991 /* 2992 * Attempt to sync this vnode once to flush any journal 2993 * work attached to it. 2994 */ 2995 if ((curthread->td_pflags & TDP_COWINPROGRESS) == 0) 2996 ffs_syncvnode(vp, waitok, 0); 2997 ACQUIRE_LOCK(ump); 2998 process_removes(vp); 2999 process_truncates(vp); 3000 if (journal_space(ump, 0) == 0) { 3001 softdep_speedup(ump); 3002 if (journal_space(ump, 1) == 0) 3003 journal_suspend(ump); 3004 } 3005 FREE_LOCK(ump); 3006 3007 return (0); 3008} 3009 3010/* 3011 * Before adjusting a link count on a vnode verify that we have sufficient 3012 * journal space. If not, process operations that depend on the currently 3013 * locked pair of vnodes to try to flush space as the syncer, buf daemon, 3014 * and softdep flush threads can not acquire these locks to reclaim space. 3015 */ 3016static void 3017softdep_prelink(dvp, vp) 3018 struct vnode *dvp; 3019 struct vnode *vp; 3020{ 3021 struct ufsmount *ump; 3022 3023 ump = VFSTOUFS(dvp->v_mount); 3024 LOCK_OWNED(ump); 3025 /* 3026 * Nothing to do if we have sufficient journal space. 3027 * If we currently hold the snapshot lock, we must avoid 3028 * handling other resources that could cause deadlock. 3029 */ 3030 if (journal_space(ump, 0) || (vp && IS_SNAPSHOT(VTOI(vp)))) 3031 return; 3032 stat_journal_low++; 3033 FREE_LOCK(ump); 3034 if (vp) 3035 ffs_syncvnode(vp, MNT_NOWAIT, 0); 3036 ffs_syncvnode(dvp, MNT_WAIT, 0); 3037 ACQUIRE_LOCK(ump); 3038 /* Process vp before dvp as it may create .. removes. */ 3039 if (vp) { 3040 process_removes(vp); 3041 process_truncates(vp); 3042 } 3043 process_removes(dvp); 3044 process_truncates(dvp); 3045 softdep_speedup(ump); 3046 process_worklist_item(UFSTOVFS(ump), 2, LK_NOWAIT); 3047 if (journal_space(ump, 0) == 0) { 3048 softdep_speedup(ump); 3049 if (journal_space(ump, 1) == 0) 3050 journal_suspend(ump); 3051 } 3052} 3053 3054static void 3055jseg_write(ump, jseg, data) 3056 struct ufsmount *ump; 3057 struct jseg *jseg; 3058 uint8_t *data; 3059{ 3060 struct jsegrec *rec; 3061 3062 rec = (struct jsegrec *)data; 3063 rec->jsr_seq = jseg->js_seq; 3064 rec->jsr_oldest = jseg->js_oldseq; 3065 rec->jsr_cnt = jseg->js_cnt; 3066 rec->jsr_blocks = jseg->js_size / ump->um_devvp->v_bufobj.bo_bsize; 3067 rec->jsr_crc = 0; 3068 rec->jsr_time = ump->um_fs->fs_mtime; 3069} 3070 3071static inline void 3072inoref_write(inoref, jseg, rec) 3073 struct inoref *inoref; 3074 struct jseg *jseg; 3075 struct jrefrec *rec; 3076{ 3077 3078 inoref->if_jsegdep->jd_seg = jseg; 3079 rec->jr_ino = inoref->if_ino; 3080 rec->jr_parent = inoref->if_parent; 3081 rec->jr_nlink = inoref->if_nlink; 3082 rec->jr_mode = inoref->if_mode; 3083 rec->jr_diroff = inoref->if_diroff; 3084} 3085 3086static void 3087jaddref_write(jaddref, jseg, data) 3088 struct jaddref *jaddref; 3089 struct jseg *jseg; 3090 uint8_t *data; 3091{ 3092 struct jrefrec *rec; 3093 3094 rec = (struct jrefrec *)data; 3095 rec->jr_op = JOP_ADDREF; 3096 inoref_write(&jaddref->ja_ref, jseg, rec); 3097} 3098 3099static void 3100jremref_write(jremref, jseg, data) 3101 struct jremref *jremref; 3102 struct jseg *jseg; 3103 uint8_t *data; 3104{ 3105 struct jrefrec *rec; 3106 3107 rec = (struct jrefrec *)data; 3108 rec->jr_op = JOP_REMREF; 3109 inoref_write(&jremref->jr_ref, jseg, rec); 3110} 3111 3112static void 3113jmvref_write(jmvref, jseg, data) 3114 struct jmvref *jmvref; 3115 struct jseg *jseg; 3116 uint8_t *data; 3117{ 3118 struct jmvrec *rec; 3119 3120 rec = (struct jmvrec *)data; 3121 rec->jm_op = JOP_MVREF; 3122 rec->jm_ino = jmvref->jm_ino; 3123 rec->jm_parent = jmvref->jm_parent; 3124 rec->jm_oldoff = jmvref->jm_oldoff; 3125 rec->jm_newoff = jmvref->jm_newoff; 3126} 3127 3128static void 3129jnewblk_write(jnewblk, jseg, data) 3130 struct jnewblk *jnewblk; 3131 struct jseg *jseg; 3132 uint8_t *data; 3133{ 3134 struct jblkrec *rec; 3135 3136 jnewblk->jn_jsegdep->jd_seg = jseg; 3137 rec = (struct jblkrec *)data; 3138 rec->jb_op = JOP_NEWBLK; 3139 rec->jb_ino = jnewblk->jn_ino; 3140 rec->jb_blkno = jnewblk->jn_blkno; 3141 rec->jb_lbn = jnewblk->jn_lbn; 3142 rec->jb_frags = jnewblk->jn_frags; 3143 rec->jb_oldfrags = jnewblk->jn_oldfrags; 3144} 3145 3146static void 3147jfreeblk_write(jfreeblk, jseg, data) 3148 struct jfreeblk *jfreeblk; 3149 struct jseg *jseg; 3150 uint8_t *data; 3151{ 3152 struct jblkrec *rec; 3153 3154 jfreeblk->jf_dep.jb_jsegdep->jd_seg = jseg; 3155 rec = (struct jblkrec *)data; 3156 rec->jb_op = JOP_FREEBLK; 3157 rec->jb_ino = jfreeblk->jf_ino; 3158 rec->jb_blkno = jfreeblk->jf_blkno; 3159 rec->jb_lbn = jfreeblk->jf_lbn; 3160 rec->jb_frags = jfreeblk->jf_frags; 3161 rec->jb_oldfrags = 0; 3162} 3163 3164static void 3165jfreefrag_write(jfreefrag, jseg, data) 3166 struct jfreefrag *jfreefrag; 3167 struct jseg *jseg; 3168 uint8_t *data; 3169{ 3170 struct jblkrec *rec; 3171 3172 jfreefrag->fr_jsegdep->jd_seg = jseg; 3173 rec = (struct jblkrec *)data; 3174 rec->jb_op = JOP_FREEBLK; 3175 rec->jb_ino = jfreefrag->fr_ino; 3176 rec->jb_blkno = jfreefrag->fr_blkno; 3177 rec->jb_lbn = jfreefrag->fr_lbn; 3178 rec->jb_frags = jfreefrag->fr_frags; 3179 rec->jb_oldfrags = 0; 3180} 3181 3182static void 3183jtrunc_write(jtrunc, jseg, data) 3184 struct jtrunc *jtrunc; 3185 struct jseg *jseg; 3186 uint8_t *data; 3187{ 3188 struct jtrncrec *rec; 3189 3190 jtrunc->jt_dep.jb_jsegdep->jd_seg = jseg; 3191 rec = (struct jtrncrec *)data; 3192 rec->jt_op = JOP_TRUNC; 3193 rec->jt_ino = jtrunc->jt_ino; 3194 rec->jt_size = jtrunc->jt_size; 3195 rec->jt_extsize = jtrunc->jt_extsize; 3196} 3197 3198static void 3199jfsync_write(jfsync, jseg, data) 3200 struct jfsync *jfsync; 3201 struct jseg *jseg; 3202 uint8_t *data; 3203{ 3204 struct jtrncrec *rec; 3205 3206 rec = (struct jtrncrec *)data; 3207 rec->jt_op = JOP_SYNC; 3208 rec->jt_ino = jfsync->jfs_ino; 3209 rec->jt_size = jfsync->jfs_size; 3210 rec->jt_extsize = jfsync->jfs_extsize; 3211} 3212 3213static void 3214softdep_flushjournal(mp) 3215 struct mount *mp; 3216{ 3217 struct jblocks *jblocks; 3218 struct ufsmount *ump; 3219 3220 if (MOUNTEDSUJ(mp) == 0) 3221 return; 3222 ump = VFSTOUFS(mp); 3223 jblocks = ump->softdep_jblocks; 3224 ACQUIRE_LOCK(ump); 3225 while (ump->softdep_on_journal) { 3226 jblocks->jb_needseg = 1; 3227 softdep_process_journal(mp, NULL, MNT_WAIT); 3228 } 3229 FREE_LOCK(ump); 3230} 3231 3232static void softdep_synchronize_completed(struct bio *); 3233static void softdep_synchronize(struct bio *, struct ufsmount *, void *); 3234 3235static void 3236softdep_synchronize_completed(bp) 3237 struct bio *bp; 3238{ 3239 struct jseg *oldest; 3240 struct jseg *jseg; 3241 struct ufsmount *ump; 3242 3243 /* 3244 * caller1 marks the last segment written before we issued the 3245 * synchronize cache. 3246 */ 3247 jseg = bp->bio_caller1; 3248 if (jseg == NULL) { 3249 g_destroy_bio(bp); 3250 return; 3251 } 3252 ump = VFSTOUFS(jseg->js_list.wk_mp); 3253 ACQUIRE_LOCK(ump); 3254 oldest = NULL; 3255 /* 3256 * Mark all the journal entries waiting on the synchronize cache 3257 * as completed so they may continue on. 3258 */ 3259 while (jseg != NULL && (jseg->js_state & COMPLETE) == 0) { 3260 jseg->js_state |= COMPLETE; 3261 oldest = jseg; 3262 jseg = TAILQ_PREV(jseg, jseglst, js_next); 3263 } 3264 /* 3265 * Restart deferred journal entry processing from the oldest 3266 * completed jseg. 3267 */ 3268 if (oldest) 3269 complete_jsegs(oldest); 3270 3271 FREE_LOCK(ump); 3272 g_destroy_bio(bp); 3273} 3274 3275/* 3276 * Send BIO_FLUSH/SYNCHRONIZE CACHE to the device to enforce write ordering 3277 * barriers. The journal must be written prior to any blocks that depend 3278 * on it and the journal can not be released until the blocks have be 3279 * written. This code handles both barriers simultaneously. 3280 */ 3281static void 3282softdep_synchronize(bp, ump, caller1) 3283 struct bio *bp; 3284 struct ufsmount *ump; 3285 void *caller1; 3286{ 3287 3288 bp->bio_cmd = BIO_FLUSH; 3289 bp->bio_flags |= BIO_ORDERED; 3290 bp->bio_data = NULL; 3291 bp->bio_offset = ump->um_cp->provider->mediasize; 3292 bp->bio_length = 0; 3293 bp->bio_done = softdep_synchronize_completed; 3294 bp->bio_caller1 = caller1; 3295 g_io_request(bp, 3296 (struct g_consumer *)ump->um_devvp->v_bufobj.bo_private); 3297} 3298 3299/* 3300 * Flush some journal records to disk. 3301 */ 3302static void 3303softdep_process_journal(mp, needwk, flags) 3304 struct mount *mp; 3305 struct worklist *needwk; 3306 int flags; 3307{ 3308 struct jblocks *jblocks; 3309 struct ufsmount *ump; 3310 struct worklist *wk; 3311 struct jseg *jseg; 3312 struct buf *bp; 3313 struct bio *bio; 3314 uint8_t *data; 3315 struct fs *fs; 3316 int shouldflush; 3317 int segwritten; 3318 int jrecmin; /* Minimum records per block. */ 3319 int jrecmax; /* Maximum records per block. */ 3320 int size; 3321 int cnt; 3322 int off; 3323 int devbsize; 3324 3325 if (MOUNTEDSUJ(mp) == 0) 3326 return; 3327 shouldflush = softdep_flushcache; 3328 bio = NULL; 3329 jseg = NULL; 3330 ump = VFSTOUFS(mp); 3331 LOCK_OWNED(ump); 3332 fs = ump->um_fs; 3333 jblocks = ump->softdep_jblocks; 3334 devbsize = ump->um_devvp->v_bufobj.bo_bsize; 3335 /* 3336 * We write anywhere between a disk block and fs block. The upper 3337 * bound is picked to prevent buffer cache fragmentation and limit 3338 * processing time per I/O. 3339 */ 3340 jrecmin = (devbsize / JREC_SIZE) - 1; /* -1 for seg header */ 3341 jrecmax = (fs->fs_bsize / devbsize) * jrecmin; 3342 segwritten = 0; 3343 for (;;) { 3344 cnt = ump->softdep_on_journal; 3345 /* 3346 * Criteria for writing a segment: 3347 * 1) We have a full block. 3348 * 2) We're called from jwait() and haven't found the 3349 * journal item yet. 3350 * 3) Always write if needseg is set. 3351 * 4) If we are called from process_worklist and have 3352 * not yet written anything we write a partial block 3353 * to enforce a 1 second maximum latency on journal 3354 * entries. 3355 */ 3356 if (cnt < (jrecmax - 1) && needwk == NULL && 3357 jblocks->jb_needseg == 0 && (segwritten || cnt == 0)) 3358 break; 3359 cnt++; 3360 /* 3361 * Verify some free journal space. softdep_prealloc() should 3362 * guarantee that we don't run out so this is indicative of 3363 * a problem with the flow control. Try to recover 3364 * gracefully in any event. 3365 */ 3366 while (jblocks->jb_free == 0) { 3367 if (flags != MNT_WAIT) 3368 break; 3369 printf("softdep: Out of journal space!\n"); 3370 softdep_speedup(ump); 3371 msleep(jblocks, LOCK_PTR(ump), PRIBIO, "jblocks", hz); 3372 } 3373 FREE_LOCK(ump); 3374 jseg = malloc(sizeof(*jseg), M_JSEG, M_SOFTDEP_FLAGS); 3375 workitem_alloc(&jseg->js_list, D_JSEG, mp); 3376 LIST_INIT(&jseg->js_entries); 3377 LIST_INIT(&jseg->js_indirs); 3378 jseg->js_state = ATTACHED; 3379 if (shouldflush == 0) 3380 jseg->js_state |= COMPLETE; 3381 else if (bio == NULL) 3382 bio = g_alloc_bio(); 3383 jseg->js_jblocks = jblocks; 3384 bp = geteblk(fs->fs_bsize, 0); 3385 ACQUIRE_LOCK(ump); 3386 /* 3387 * If there was a race while we were allocating the block 3388 * and jseg the entry we care about was likely written. 3389 * We bail out in both the WAIT and NOWAIT case and assume 3390 * the caller will loop if the entry it cares about is 3391 * not written. 3392 */ 3393 cnt = ump->softdep_on_journal; 3394 if (cnt + jblocks->jb_needseg == 0 || jblocks->jb_free == 0) { 3395 bp->b_flags |= B_INVAL | B_NOCACHE; 3396 WORKITEM_FREE(jseg, D_JSEG); 3397 FREE_LOCK(ump); 3398 brelse(bp); 3399 ACQUIRE_LOCK(ump); 3400 break; 3401 } 3402 /* 3403 * Calculate the disk block size required for the available 3404 * records rounded to the min size. 3405 */ 3406 if (cnt == 0) 3407 size = devbsize; 3408 else if (cnt < jrecmax) 3409 size = howmany(cnt, jrecmin) * devbsize; 3410 else 3411 size = fs->fs_bsize; 3412 /* 3413 * Allocate a disk block for this journal data and account 3414 * for truncation of the requested size if enough contiguous 3415 * space was not available. 3416 */ 3417 bp->b_blkno = jblocks_alloc(jblocks, size, &size); 3418 bp->b_lblkno = bp->b_blkno; 3419 bp->b_offset = bp->b_blkno * DEV_BSIZE; 3420 bp->b_bcount = size; 3421 bp->b_flags &= ~B_INVAL; 3422 bp->b_flags |= B_VALIDSUSPWRT | B_NOCOPY; 3423 /* 3424 * Initialize our jseg with cnt records. Assign the next 3425 * sequence number to it and link it in-order. 3426 */ 3427 cnt = MIN(cnt, (size / devbsize) * jrecmin); 3428 jseg->js_buf = bp; 3429 jseg->js_cnt = cnt; 3430 jseg->js_refs = cnt + 1; /* Self ref. */ 3431 jseg->js_size = size; 3432 jseg->js_seq = jblocks->jb_nextseq++; 3433 if (jblocks->jb_oldestseg == NULL) 3434 jblocks->jb_oldestseg = jseg; 3435 jseg->js_oldseq = jblocks->jb_oldestseg->js_seq; 3436 TAILQ_INSERT_TAIL(&jblocks->jb_segs, jseg, js_next); 3437 if (jblocks->jb_writeseg == NULL) 3438 jblocks->jb_writeseg = jseg; 3439 /* 3440 * Start filling in records from the pending list. 3441 */ 3442 data = bp->b_data; 3443 off = 0; 3444 3445 /* 3446 * Always put a header on the first block. 3447 * XXX As with below, there might not be a chance to get 3448 * into the loop. Ensure that something valid is written. 3449 */ 3450 jseg_write(ump, jseg, data); 3451 off += JREC_SIZE; 3452 data = bp->b_data + off; 3453 3454 /* 3455 * XXX Something is wrong here. There's no work to do, 3456 * but we need to perform and I/O and allow it to complete 3457 * anyways. 3458 */ 3459 if (LIST_EMPTY(&ump->softdep_journal_pending)) 3460 stat_emptyjblocks++; 3461 3462 while ((wk = LIST_FIRST(&ump->softdep_journal_pending)) 3463 != NULL) { 3464 if (cnt == 0) 3465 break; 3466 /* Place a segment header on every device block. */ 3467 if ((off % devbsize) == 0) { 3468 jseg_write(ump, jseg, data); 3469 off += JREC_SIZE; 3470 data = bp->b_data + off; 3471 } 3472 if (wk == needwk) 3473 needwk = NULL; 3474 remove_from_journal(wk); 3475 wk->wk_state |= INPROGRESS; 3476 WORKLIST_INSERT(&jseg->js_entries, wk); 3477 switch (wk->wk_type) { 3478 case D_JADDREF: 3479 jaddref_write(WK_JADDREF(wk), jseg, data); 3480 break; 3481 case D_JREMREF: 3482 jremref_write(WK_JREMREF(wk), jseg, data); 3483 break; 3484 case D_JMVREF: 3485 jmvref_write(WK_JMVREF(wk), jseg, data); 3486 break; 3487 case D_JNEWBLK: 3488 jnewblk_write(WK_JNEWBLK(wk), jseg, data); 3489 break; 3490 case D_JFREEBLK: 3491 jfreeblk_write(WK_JFREEBLK(wk), jseg, data); 3492 break; 3493 case D_JFREEFRAG: 3494 jfreefrag_write(WK_JFREEFRAG(wk), jseg, data); 3495 break; 3496 case D_JTRUNC: 3497 jtrunc_write(WK_JTRUNC(wk), jseg, data); 3498 break; 3499 case D_JFSYNC: 3500 jfsync_write(WK_JFSYNC(wk), jseg, data); 3501 break; 3502 default: 3503 panic("process_journal: Unknown type %s", 3504 TYPENAME(wk->wk_type)); 3505 /* NOTREACHED */ 3506 } 3507 off += JREC_SIZE; 3508 data = bp->b_data + off; 3509 cnt--; 3510 } 3511 3512 /* Clear any remaining space so we don't leak kernel data */ 3513 if (size > off) 3514 bzero(data, size - off); 3515 3516 /* 3517 * Write this one buffer and continue. 3518 */ 3519 segwritten = 1; 3520 jblocks->jb_needseg = 0; 3521 WORKLIST_INSERT(&bp->b_dep, &jseg->js_list); 3522 FREE_LOCK(ump); 3523 pbgetvp(ump->um_devvp, bp); 3524 /* 3525 * We only do the blocking wait once we find the journal 3526 * entry we're looking for. 3527 */ 3528 if (needwk == NULL && flags == MNT_WAIT) 3529 bwrite(bp); 3530 else 3531 bawrite(bp); 3532 ACQUIRE_LOCK(ump); 3533 } 3534 /* 3535 * If we wrote a segment issue a synchronize cache so the journal 3536 * is reflected on disk before the data is written. Since reclaiming 3537 * journal space also requires writing a journal record this 3538 * process also enforces a barrier before reclamation. 3539 */ 3540 if (segwritten && shouldflush) { 3541 softdep_synchronize(bio, ump, 3542 TAILQ_LAST(&jblocks->jb_segs, jseglst)); 3543 } else if (bio) 3544 g_destroy_bio(bio); 3545 /* 3546 * If we've suspended the filesystem because we ran out of journal 3547 * space either try to sync it here to make some progress or 3548 * unsuspend it if we already have. 3549 */ 3550 if (flags == 0 && jblocks->jb_suspended) { 3551 if (journal_unsuspend(ump)) 3552 return; 3553 FREE_LOCK(ump); 3554 VFS_SYNC(mp, MNT_NOWAIT); 3555 ffs_sbupdate(ump, MNT_WAIT, 0); 3556 ACQUIRE_LOCK(ump); 3557 } 3558} 3559 3560/* 3561 * Complete a jseg, allowing all dependencies awaiting journal writes 3562 * to proceed. Each journal dependency also attaches a jsegdep to dependent 3563 * structures so that the journal segment can be freed to reclaim space. 3564 */ 3565static void 3566complete_jseg(jseg) 3567 struct jseg *jseg; 3568{ 3569 struct worklist *wk; 3570 struct jmvref *jmvref; 3571 int waiting; 3572#ifdef INVARIANTS 3573 int i = 0; 3574#endif 3575 3576 while ((wk = LIST_FIRST(&jseg->js_entries)) != NULL) { 3577 WORKLIST_REMOVE(wk); 3578 waiting = wk->wk_state & IOWAITING; 3579 wk->wk_state &= ~(INPROGRESS | IOWAITING); 3580 wk->wk_state |= COMPLETE; 3581 KASSERT(i++ < jseg->js_cnt, 3582 ("handle_written_jseg: overflow %d >= %d", 3583 i - 1, jseg->js_cnt)); 3584 switch (wk->wk_type) { 3585 case D_JADDREF: 3586 handle_written_jaddref(WK_JADDREF(wk)); 3587 break; 3588 case D_JREMREF: 3589 handle_written_jremref(WK_JREMREF(wk)); 3590 break; 3591 case D_JMVREF: 3592 rele_jseg(jseg); /* No jsegdep. */ 3593 jmvref = WK_JMVREF(wk); 3594 LIST_REMOVE(jmvref, jm_deps); 3595 if ((jmvref->jm_pagedep->pd_state & ONWORKLIST) == 0) 3596 free_pagedep(jmvref->jm_pagedep); 3597 WORKITEM_FREE(jmvref, D_JMVREF); 3598 break; 3599 case D_JNEWBLK: 3600 handle_written_jnewblk(WK_JNEWBLK(wk)); 3601 break; 3602 case D_JFREEBLK: 3603 handle_written_jblkdep(&WK_JFREEBLK(wk)->jf_dep); 3604 break; 3605 case D_JTRUNC: 3606 handle_written_jblkdep(&WK_JTRUNC(wk)->jt_dep); 3607 break; 3608 case D_JFSYNC: 3609 rele_jseg(jseg); /* No jsegdep. */ 3610 WORKITEM_FREE(wk, D_JFSYNC); 3611 break; 3612 case D_JFREEFRAG: 3613 handle_written_jfreefrag(WK_JFREEFRAG(wk)); 3614 break; 3615 default: 3616 panic("handle_written_jseg: Unknown type %s", 3617 TYPENAME(wk->wk_type)); 3618 /* NOTREACHED */ 3619 } 3620 if (waiting) 3621 wakeup(wk); 3622 } 3623 /* Release the self reference so the structure may be freed. */ 3624 rele_jseg(jseg); 3625} 3626 3627/* 3628 * Determine which jsegs are ready for completion processing. Waits for 3629 * synchronize cache to complete as well as forcing in-order completion 3630 * of journal entries. 3631 */ 3632static void 3633complete_jsegs(jseg) 3634 struct jseg *jseg; 3635{ 3636 struct jblocks *jblocks; 3637 struct jseg *jsegn; 3638 3639 jblocks = jseg->js_jblocks; 3640 /* 3641 * Don't allow out of order completions. If this isn't the first 3642 * block wait for it to write before we're done. 3643 */ 3644 if (jseg != jblocks->jb_writeseg) 3645 return; 3646 /* Iterate through available jsegs processing their entries. */ 3647 while (jseg && (jseg->js_state & ALLCOMPLETE) == ALLCOMPLETE) { 3648 jblocks->jb_oldestwrseq = jseg->js_oldseq; 3649 jsegn = TAILQ_NEXT(jseg, js_next); 3650 complete_jseg(jseg); 3651 jseg = jsegn; 3652 } 3653 jblocks->jb_writeseg = jseg; 3654 /* 3655 * Attempt to free jsegs now that oldestwrseq may have advanced. 3656 */ 3657 free_jsegs(jblocks); 3658} 3659 3660/* 3661 * Mark a jseg as DEPCOMPLETE and throw away the buffer. Attempt to handle 3662 * the final completions. 3663 */ 3664static void 3665handle_written_jseg(jseg, bp) 3666 struct jseg *jseg; 3667 struct buf *bp; 3668{ 3669 3670 if (jseg->js_refs == 0) 3671 panic("handle_written_jseg: No self-reference on %p", jseg); 3672 jseg->js_state |= DEPCOMPLETE; 3673 /* 3674 * We'll never need this buffer again, set flags so it will be 3675 * discarded. 3676 */ 3677 bp->b_flags |= B_INVAL | B_NOCACHE; 3678 pbrelvp(bp); 3679 complete_jsegs(jseg); 3680} 3681 3682static inline struct jsegdep * 3683inoref_jseg(inoref) 3684 struct inoref *inoref; 3685{ 3686 struct jsegdep *jsegdep; 3687 3688 jsegdep = inoref->if_jsegdep; 3689 inoref->if_jsegdep = NULL; 3690 3691 return (jsegdep); 3692} 3693 3694/* 3695 * Called once a jremref has made it to stable store. The jremref is marked 3696 * complete and we attempt to free it. Any pagedeps writes sleeping waiting 3697 * for the jremref to complete will be awoken by free_jremref. 3698 */ 3699static void 3700handle_written_jremref(jremref) 3701 struct jremref *jremref; 3702{ 3703 struct inodedep *inodedep; 3704 struct jsegdep *jsegdep; 3705 struct dirrem *dirrem; 3706 3707 /* Grab the jsegdep. */ 3708 jsegdep = inoref_jseg(&jremref->jr_ref); 3709 /* 3710 * Remove us from the inoref list. 3711 */ 3712 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 3713 0, &inodedep) == 0) 3714 panic("handle_written_jremref: Lost inodedep"); 3715 TAILQ_REMOVE(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps); 3716 /* 3717 * Complete the dirrem. 3718 */ 3719 dirrem = jremref->jr_dirrem; 3720 jremref->jr_dirrem = NULL; 3721 LIST_REMOVE(jremref, jr_deps); 3722 jsegdep->jd_state |= jremref->jr_state & MKDIR_PARENT; 3723 jwork_insert(&dirrem->dm_jwork, jsegdep); 3724 if (LIST_EMPTY(&dirrem->dm_jremrefhd) && 3725 (dirrem->dm_state & COMPLETE) != 0) 3726 add_to_worklist(&dirrem->dm_list, 0); 3727 free_jremref(jremref); 3728} 3729 3730/* 3731 * Called once a jaddref has made it to stable store. The dependency is 3732 * marked complete and any dependent structures are added to the inode 3733 * bufwait list to be completed as soon as it is written. If a bitmap write 3734 * depends on this entry we move the inode into the inodedephd of the 3735 * bmsafemap dependency and attempt to remove the jaddref from the bmsafemap. 3736 */ 3737static void 3738handle_written_jaddref(jaddref) 3739 struct jaddref *jaddref; 3740{ 3741 struct jsegdep *jsegdep; 3742 struct inodedep *inodedep; 3743 struct diradd *diradd; 3744 struct mkdir *mkdir; 3745 3746 /* Grab the jsegdep. */ 3747 jsegdep = inoref_jseg(&jaddref->ja_ref); 3748 mkdir = NULL; 3749 diradd = NULL; 3750 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino, 3751 0, &inodedep) == 0) 3752 panic("handle_written_jaddref: Lost inodedep."); 3753 if (jaddref->ja_diradd == NULL) 3754 panic("handle_written_jaddref: No dependency"); 3755 if (jaddref->ja_diradd->da_list.wk_type == D_DIRADD) { 3756 diradd = jaddref->ja_diradd; 3757 WORKLIST_INSERT(&inodedep->id_bufwait, &diradd->da_list); 3758 } else if (jaddref->ja_state & MKDIR_PARENT) { 3759 mkdir = jaddref->ja_mkdir; 3760 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir->md_list); 3761 } else if (jaddref->ja_state & MKDIR_BODY) 3762 mkdir = jaddref->ja_mkdir; 3763 else 3764 panic("handle_written_jaddref: Unknown dependency %p", 3765 jaddref->ja_diradd); 3766 jaddref->ja_diradd = NULL; /* also clears ja_mkdir */ 3767 /* 3768 * Remove us from the inode list. 3769 */ 3770 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, if_deps); 3771 /* 3772 * The mkdir may be waiting on the jaddref to clear before freeing. 3773 */ 3774 if (mkdir) { 3775 KASSERT(mkdir->md_list.wk_type == D_MKDIR, 3776 ("handle_written_jaddref: Incorrect type for mkdir %s", 3777 TYPENAME(mkdir->md_list.wk_type))); 3778 mkdir->md_jaddref = NULL; 3779 diradd = mkdir->md_diradd; 3780 mkdir->md_state |= DEPCOMPLETE; 3781 complete_mkdir(mkdir); 3782 } 3783 jwork_insert(&diradd->da_jwork, jsegdep); 3784 if (jaddref->ja_state & NEWBLOCK) { 3785 inodedep->id_state |= ONDEPLIST; 3786 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_inodedephd, 3787 inodedep, id_deps); 3788 } 3789 free_jaddref(jaddref); 3790} 3791 3792/* 3793 * Called once a jnewblk journal is written. The allocdirect or allocindir 3794 * is placed in the bmsafemap to await notification of a written bitmap. If 3795 * the operation was canceled we add the segdep to the appropriate 3796 * dependency to free the journal space once the canceling operation 3797 * completes. 3798 */ 3799static void 3800handle_written_jnewblk(jnewblk) 3801 struct jnewblk *jnewblk; 3802{ 3803 struct bmsafemap *bmsafemap; 3804 struct freefrag *freefrag; 3805 struct freework *freework; 3806 struct jsegdep *jsegdep; 3807 struct newblk *newblk; 3808 3809 /* Grab the jsegdep. */ 3810 jsegdep = jnewblk->jn_jsegdep; 3811 jnewblk->jn_jsegdep = NULL; 3812 if (jnewblk->jn_dep == NULL) 3813 panic("handle_written_jnewblk: No dependency for the segdep."); 3814 switch (jnewblk->jn_dep->wk_type) { 3815 case D_NEWBLK: 3816 case D_ALLOCDIRECT: 3817 case D_ALLOCINDIR: 3818 /* 3819 * Add the written block to the bmsafemap so it can 3820 * be notified when the bitmap is on disk. 3821 */ 3822 newblk = WK_NEWBLK(jnewblk->jn_dep); 3823 newblk->nb_jnewblk = NULL; 3824 if ((newblk->nb_state & GOINGAWAY) == 0) { 3825 bmsafemap = newblk->nb_bmsafemap; 3826 newblk->nb_state |= ONDEPLIST; 3827 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, 3828 nb_deps); 3829 } 3830 jwork_insert(&newblk->nb_jwork, jsegdep); 3831 break; 3832 case D_FREEFRAG: 3833 /* 3834 * A newblock being removed by a freefrag when replaced by 3835 * frag extension. 3836 */ 3837 freefrag = WK_FREEFRAG(jnewblk->jn_dep); 3838 freefrag->ff_jdep = NULL; 3839 jwork_insert(&freefrag->ff_jwork, jsegdep); 3840 break; 3841 case D_FREEWORK: 3842 /* 3843 * A direct block was removed by truncate. 3844 */ 3845 freework = WK_FREEWORK(jnewblk->jn_dep); 3846 freework->fw_jnewblk = NULL; 3847 jwork_insert(&freework->fw_freeblks->fb_jwork, jsegdep); 3848 break; 3849 default: 3850 panic("handle_written_jnewblk: Unknown type %d.", 3851 jnewblk->jn_dep->wk_type); 3852 } 3853 jnewblk->jn_dep = NULL; 3854 free_jnewblk(jnewblk); 3855} 3856 3857/* 3858 * Cancel a jfreefrag that won't be needed, probably due to colliding with 3859 * an in-flight allocation that has not yet been committed. Divorce us 3860 * from the freefrag and mark it DEPCOMPLETE so that it may be added 3861 * to the worklist. 3862 */ 3863static void 3864cancel_jfreefrag(jfreefrag) 3865 struct jfreefrag *jfreefrag; 3866{ 3867 struct freefrag *freefrag; 3868 3869 if (jfreefrag->fr_jsegdep) { 3870 free_jsegdep(jfreefrag->fr_jsegdep); 3871 jfreefrag->fr_jsegdep = NULL; 3872 } 3873 freefrag = jfreefrag->fr_freefrag; 3874 jfreefrag->fr_freefrag = NULL; 3875 free_jfreefrag(jfreefrag); 3876 freefrag->ff_state |= DEPCOMPLETE; 3877 CTR1(KTR_SUJ, "cancel_jfreefrag: blkno %jd", freefrag->ff_blkno); 3878} 3879 3880/* 3881 * Free a jfreefrag when the parent freefrag is rendered obsolete. 3882 */ 3883static void 3884free_jfreefrag(jfreefrag) 3885 struct jfreefrag *jfreefrag; 3886{ 3887 3888 if (jfreefrag->fr_state & INPROGRESS) 3889 WORKLIST_REMOVE(&jfreefrag->fr_list); 3890 else if (jfreefrag->fr_state & ONWORKLIST) 3891 remove_from_journal(&jfreefrag->fr_list); 3892 if (jfreefrag->fr_freefrag != NULL) 3893 panic("free_jfreefrag: Still attached to a freefrag."); 3894 WORKITEM_FREE(jfreefrag, D_JFREEFRAG); 3895} 3896 3897/* 3898 * Called when the journal write for a jfreefrag completes. The parent 3899 * freefrag is added to the worklist if this completes its dependencies. 3900 */ 3901static void 3902handle_written_jfreefrag(jfreefrag) 3903 struct jfreefrag *jfreefrag; 3904{ 3905 struct jsegdep *jsegdep; 3906 struct freefrag *freefrag; 3907 3908 /* Grab the jsegdep. */ 3909 jsegdep = jfreefrag->fr_jsegdep; 3910 jfreefrag->fr_jsegdep = NULL; 3911 freefrag = jfreefrag->fr_freefrag; 3912 if (freefrag == NULL) 3913 panic("handle_written_jfreefrag: No freefrag."); 3914 freefrag->ff_state |= DEPCOMPLETE; 3915 freefrag->ff_jdep = NULL; 3916 jwork_insert(&freefrag->ff_jwork, jsegdep); 3917 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE) 3918 add_to_worklist(&freefrag->ff_list, 0); 3919 jfreefrag->fr_freefrag = NULL; 3920 free_jfreefrag(jfreefrag); 3921} 3922 3923/* 3924 * Called when the journal write for a jfreeblk completes. The jfreeblk 3925 * is removed from the freeblks list of pending journal writes and the 3926 * jsegdep is moved to the freeblks jwork to be completed when all blocks 3927 * have been reclaimed. 3928 */ 3929static void 3930handle_written_jblkdep(jblkdep) 3931 struct jblkdep *jblkdep; 3932{ 3933 struct freeblks *freeblks; 3934 struct jsegdep *jsegdep; 3935 3936 /* Grab the jsegdep. */ 3937 jsegdep = jblkdep->jb_jsegdep; 3938 jblkdep->jb_jsegdep = NULL; 3939 freeblks = jblkdep->jb_freeblks; 3940 LIST_REMOVE(jblkdep, jb_deps); 3941 jwork_insert(&freeblks->fb_jwork, jsegdep); 3942 /* 3943 * If the freeblks is all journaled, we can add it to the worklist. 3944 */ 3945 if (LIST_EMPTY(&freeblks->fb_jblkdephd) && 3946 (freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 3947 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 3948 3949 free_jblkdep(jblkdep); 3950} 3951 3952static struct jsegdep * 3953newjsegdep(struct worklist *wk) 3954{ 3955 struct jsegdep *jsegdep; 3956 3957 jsegdep = malloc(sizeof(*jsegdep), M_JSEGDEP, M_SOFTDEP_FLAGS); 3958 workitem_alloc(&jsegdep->jd_list, D_JSEGDEP, wk->wk_mp); 3959 jsegdep->jd_seg = NULL; 3960 3961 return (jsegdep); 3962} 3963 3964static struct jmvref * 3965newjmvref(dp, ino, oldoff, newoff) 3966 struct inode *dp; 3967 ino_t ino; 3968 off_t oldoff; 3969 off_t newoff; 3970{ 3971 struct jmvref *jmvref; 3972 3973 jmvref = malloc(sizeof(*jmvref), M_JMVREF, M_SOFTDEP_FLAGS); 3974 workitem_alloc(&jmvref->jm_list, D_JMVREF, UFSTOVFS(dp->i_ump)); 3975 jmvref->jm_list.wk_state = ATTACHED | DEPCOMPLETE; 3976 jmvref->jm_parent = dp->i_number; 3977 jmvref->jm_ino = ino; 3978 jmvref->jm_oldoff = oldoff; 3979 jmvref->jm_newoff = newoff; 3980 3981 return (jmvref); 3982} 3983 3984/* 3985 * Allocate a new jremref that tracks the removal of ip from dp with the 3986 * directory entry offset of diroff. Mark the entry as ATTACHED and 3987 * DEPCOMPLETE as we have all the information required for the journal write 3988 * and the directory has already been removed from the buffer. The caller 3989 * is responsible for linking the jremref into the pagedep and adding it 3990 * to the journal to write. The MKDIR_PARENT flag is set if we're doing 3991 * a DOTDOT addition so handle_workitem_remove() can properly assign 3992 * the jsegdep when we're done. 3993 */ 3994static struct jremref * 3995newjremref(struct dirrem *dirrem, struct inode *dp, struct inode *ip, 3996 off_t diroff, nlink_t nlink) 3997{ 3998 struct jremref *jremref; 3999 4000 jremref = malloc(sizeof(*jremref), M_JREMREF, M_SOFTDEP_FLAGS); 4001 workitem_alloc(&jremref->jr_list, D_JREMREF, UFSTOVFS(dp->i_ump)); 4002 jremref->jr_state = ATTACHED; 4003 newinoref(&jremref->jr_ref, ip->i_number, dp->i_number, diroff, 4004 nlink, ip->i_mode); 4005 jremref->jr_dirrem = dirrem; 4006 4007 return (jremref); 4008} 4009 4010static inline void 4011newinoref(struct inoref *inoref, ino_t ino, ino_t parent, off_t diroff, 4012 nlink_t nlink, uint16_t mode) 4013{ 4014 4015 inoref->if_jsegdep = newjsegdep(&inoref->if_list); 4016 inoref->if_diroff = diroff; 4017 inoref->if_ino = ino; 4018 inoref->if_parent = parent; 4019 inoref->if_nlink = nlink; 4020 inoref->if_mode = mode; 4021} 4022 4023/* 4024 * Allocate a new jaddref to track the addition of ino to dp at diroff. The 4025 * directory offset may not be known until later. The caller is responsible 4026 * adding the entry to the journal when this information is available. nlink 4027 * should be the link count prior to the addition and mode is only required 4028 * to have the correct FMT. 4029 */ 4030static struct jaddref * 4031newjaddref(struct inode *dp, ino_t ino, off_t diroff, int16_t nlink, 4032 uint16_t mode) 4033{ 4034 struct jaddref *jaddref; 4035 4036 jaddref = malloc(sizeof(*jaddref), M_JADDREF, M_SOFTDEP_FLAGS); 4037 workitem_alloc(&jaddref->ja_list, D_JADDREF, UFSTOVFS(dp->i_ump)); 4038 jaddref->ja_state = ATTACHED; 4039 jaddref->ja_mkdir = NULL; 4040 newinoref(&jaddref->ja_ref, ino, dp->i_number, diroff, nlink, mode); 4041 4042 return (jaddref); 4043} 4044 4045/* 4046 * Create a new free dependency for a freework. The caller is responsible 4047 * for adjusting the reference count when it has the lock held. The freedep 4048 * will track an outstanding bitmap write that will ultimately clear the 4049 * freework to continue. 4050 */ 4051static struct freedep * 4052newfreedep(struct freework *freework) 4053{ 4054 struct freedep *freedep; 4055 4056 freedep = malloc(sizeof(*freedep), M_FREEDEP, M_SOFTDEP_FLAGS); 4057 workitem_alloc(&freedep->fd_list, D_FREEDEP, freework->fw_list.wk_mp); 4058 freedep->fd_freework = freework; 4059 4060 return (freedep); 4061} 4062 4063/* 4064 * Free a freedep structure once the buffer it is linked to is written. If 4065 * this is the last reference to the freework schedule it for completion. 4066 */ 4067static void 4068free_freedep(freedep) 4069 struct freedep *freedep; 4070{ 4071 struct freework *freework; 4072 4073 freework = freedep->fd_freework; 4074 freework->fw_freeblks->fb_cgwait--; 4075 if (--freework->fw_ref == 0) 4076 freework_enqueue(freework); 4077 WORKITEM_FREE(freedep, D_FREEDEP); 4078} 4079 4080/* 4081 * Allocate a new freework structure that may be a level in an indirect 4082 * when parent is not NULL or a top level block when it is. The top level 4083 * freework structures are allocated without the per-filesystem lock held 4084 * and before the freeblks is visible outside of softdep_setup_freeblocks(). 4085 */ 4086static struct freework * 4087newfreework(ump, freeblks, parent, lbn, nb, frags, off, journal) 4088 struct ufsmount *ump; 4089 struct freeblks *freeblks; 4090 struct freework *parent; 4091 ufs_lbn_t lbn; 4092 ufs2_daddr_t nb; 4093 int frags; 4094 int off; 4095 int journal; 4096{ 4097 struct freework *freework; 4098 4099 freework = malloc(sizeof(*freework), M_FREEWORK, M_SOFTDEP_FLAGS); 4100 workitem_alloc(&freework->fw_list, D_FREEWORK, freeblks->fb_list.wk_mp); 4101 freework->fw_state = ATTACHED; 4102 freework->fw_jnewblk = NULL; 4103 freework->fw_freeblks = freeblks; 4104 freework->fw_parent = parent; 4105 freework->fw_lbn = lbn; 4106 freework->fw_blkno = nb; 4107 freework->fw_frags = frags; 4108 freework->fw_indir = NULL; 4109 freework->fw_ref = (MOUNTEDSUJ(UFSTOVFS(ump)) == 0 || lbn >= -NXADDR) 4110 ? 0 : NINDIR(ump->um_fs) + 1; 4111 freework->fw_start = freework->fw_off = off; 4112 if (journal) 4113 newjfreeblk(freeblks, lbn, nb, frags); 4114 if (parent == NULL) { 4115 ACQUIRE_LOCK(ump); 4116 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list); 4117 freeblks->fb_ref++; 4118 FREE_LOCK(ump); 4119 } 4120 4121 return (freework); 4122} 4123 4124/* 4125 * Eliminate a jfreeblk for a block that does not need journaling. 4126 */ 4127static void 4128cancel_jfreeblk(freeblks, blkno) 4129 struct freeblks *freeblks; 4130 ufs2_daddr_t blkno; 4131{ 4132 struct jfreeblk *jfreeblk; 4133 struct jblkdep *jblkdep; 4134 4135 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) { 4136 if (jblkdep->jb_list.wk_type != D_JFREEBLK) 4137 continue; 4138 jfreeblk = WK_JFREEBLK(&jblkdep->jb_list); 4139 if (jfreeblk->jf_blkno == blkno) 4140 break; 4141 } 4142 if (jblkdep == NULL) 4143 return; 4144 CTR1(KTR_SUJ, "cancel_jfreeblk: blkno %jd", blkno); 4145 free_jsegdep(jblkdep->jb_jsegdep); 4146 LIST_REMOVE(jblkdep, jb_deps); 4147 WORKITEM_FREE(jfreeblk, D_JFREEBLK); 4148} 4149 4150/* 4151 * Allocate a new jfreeblk to journal top level block pointer when truncating 4152 * a file. The caller must add this to the worklist when the per-filesystem 4153 * lock is held. 4154 */ 4155static struct jfreeblk * 4156newjfreeblk(freeblks, lbn, blkno, frags) 4157 struct freeblks *freeblks; 4158 ufs_lbn_t lbn; 4159 ufs2_daddr_t blkno; 4160 int frags; 4161{ 4162 struct jfreeblk *jfreeblk; 4163 4164 jfreeblk = malloc(sizeof(*jfreeblk), M_JFREEBLK, M_SOFTDEP_FLAGS); 4165 workitem_alloc(&jfreeblk->jf_dep.jb_list, D_JFREEBLK, 4166 freeblks->fb_list.wk_mp); 4167 jfreeblk->jf_dep.jb_jsegdep = newjsegdep(&jfreeblk->jf_dep.jb_list); 4168 jfreeblk->jf_dep.jb_freeblks = freeblks; 4169 jfreeblk->jf_ino = freeblks->fb_inum; 4170 jfreeblk->jf_lbn = lbn; 4171 jfreeblk->jf_blkno = blkno; 4172 jfreeblk->jf_frags = frags; 4173 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jfreeblk->jf_dep, jb_deps); 4174 4175 return (jfreeblk); 4176} 4177 4178/* 4179 * The journal is only prepared to handle full-size block numbers, so we 4180 * have to adjust the record to reflect the change to a full-size block. 4181 * For example, suppose we have a block made up of fragments 8-15 and 4182 * want to free its last two fragments. We are given a request that says: 4183 * FREEBLK ino=5, blkno=14, lbn=0, frags=2, oldfrags=0 4184 * where frags are the number of fragments to free and oldfrags are the 4185 * number of fragments to keep. To block align it, we have to change it to 4186 * have a valid full-size blkno, so it becomes: 4187 * FREEBLK ino=5, blkno=8, lbn=0, frags=2, oldfrags=6 4188 */ 4189static void 4190adjust_newfreework(freeblks, frag_offset) 4191 struct freeblks *freeblks; 4192 int frag_offset; 4193{ 4194 struct jfreeblk *jfreeblk; 4195 4196 KASSERT((LIST_FIRST(&freeblks->fb_jblkdephd) != NULL && 4197 LIST_FIRST(&freeblks->fb_jblkdephd)->jb_list.wk_type == D_JFREEBLK), 4198 ("adjust_newfreework: Missing freeblks dependency")); 4199 4200 jfreeblk = WK_JFREEBLK(LIST_FIRST(&freeblks->fb_jblkdephd)); 4201 jfreeblk->jf_blkno -= frag_offset; 4202 jfreeblk->jf_frags += frag_offset; 4203} 4204 4205/* 4206 * Allocate a new jtrunc to track a partial truncation. 4207 */ 4208static struct jtrunc * 4209newjtrunc(freeblks, size, extsize) 4210 struct freeblks *freeblks; 4211 off_t size; 4212 int extsize; 4213{ 4214 struct jtrunc *jtrunc; 4215 4216 jtrunc = malloc(sizeof(*jtrunc), M_JTRUNC, M_SOFTDEP_FLAGS); 4217 workitem_alloc(&jtrunc->jt_dep.jb_list, D_JTRUNC, 4218 freeblks->fb_list.wk_mp); 4219 jtrunc->jt_dep.jb_jsegdep = newjsegdep(&jtrunc->jt_dep.jb_list); 4220 jtrunc->jt_dep.jb_freeblks = freeblks; 4221 jtrunc->jt_ino = freeblks->fb_inum; 4222 jtrunc->jt_size = size; 4223 jtrunc->jt_extsize = extsize; 4224 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jtrunc->jt_dep, jb_deps); 4225 4226 return (jtrunc); 4227} 4228 4229/* 4230 * If we're canceling a new bitmap we have to search for another ref 4231 * to move into the bmsafemap dep. This might be better expressed 4232 * with another structure. 4233 */ 4234static void 4235move_newblock_dep(jaddref, inodedep) 4236 struct jaddref *jaddref; 4237 struct inodedep *inodedep; 4238{ 4239 struct inoref *inoref; 4240 struct jaddref *jaddrefn; 4241 4242 jaddrefn = NULL; 4243 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref; 4244 inoref = TAILQ_NEXT(inoref, if_deps)) { 4245 if ((jaddref->ja_state & NEWBLOCK) && 4246 inoref->if_list.wk_type == D_JADDREF) { 4247 jaddrefn = (struct jaddref *)inoref; 4248 break; 4249 } 4250 } 4251 if (jaddrefn == NULL) 4252 return; 4253 jaddrefn->ja_state &= ~(ATTACHED | UNDONE); 4254 jaddrefn->ja_state |= jaddref->ja_state & 4255 (ATTACHED | UNDONE | NEWBLOCK); 4256 jaddref->ja_state &= ~(ATTACHED | UNDONE | NEWBLOCK); 4257 jaddref->ja_state |= ATTACHED; 4258 LIST_REMOVE(jaddref, ja_bmdeps); 4259 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_jaddrefhd, jaddrefn, 4260 ja_bmdeps); 4261} 4262 4263/* 4264 * Cancel a jaddref either before it has been written or while it is being 4265 * written. This happens when a link is removed before the add reaches 4266 * the disk. The jaddref dependency is kept linked into the bmsafemap 4267 * and inode to prevent the link count or bitmap from reaching the disk 4268 * until handle_workitem_remove() re-adjusts the counts and bitmaps as 4269 * required. 4270 * 4271 * Returns 1 if the canceled addref requires journaling of the remove and 4272 * 0 otherwise. 4273 */ 4274static int 4275cancel_jaddref(jaddref, inodedep, wkhd) 4276 struct jaddref *jaddref; 4277 struct inodedep *inodedep; 4278 struct workhead *wkhd; 4279{ 4280 struct inoref *inoref; 4281 struct jsegdep *jsegdep; 4282 int needsj; 4283 4284 KASSERT((jaddref->ja_state & COMPLETE) == 0, 4285 ("cancel_jaddref: Canceling complete jaddref")); 4286 if (jaddref->ja_state & (INPROGRESS | COMPLETE)) 4287 needsj = 1; 4288 else 4289 needsj = 0; 4290 if (inodedep == NULL) 4291 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino, 4292 0, &inodedep) == 0) 4293 panic("cancel_jaddref: Lost inodedep"); 4294 /* 4295 * We must adjust the nlink of any reference operation that follows 4296 * us so that it is consistent with the in-memory reference. This 4297 * ensures that inode nlink rollbacks always have the correct link. 4298 */ 4299 if (needsj == 0) { 4300 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref; 4301 inoref = TAILQ_NEXT(inoref, if_deps)) { 4302 if (inoref->if_state & GOINGAWAY) 4303 break; 4304 inoref->if_nlink--; 4305 } 4306 } 4307 jsegdep = inoref_jseg(&jaddref->ja_ref); 4308 if (jaddref->ja_state & NEWBLOCK) 4309 move_newblock_dep(jaddref, inodedep); 4310 wake_worklist(&jaddref->ja_list); 4311 jaddref->ja_mkdir = NULL; 4312 if (jaddref->ja_state & INPROGRESS) { 4313 jaddref->ja_state &= ~INPROGRESS; 4314 WORKLIST_REMOVE(&jaddref->ja_list); 4315 jwork_insert(wkhd, jsegdep); 4316 } else { 4317 free_jsegdep(jsegdep); 4318 if (jaddref->ja_state & DEPCOMPLETE) 4319 remove_from_journal(&jaddref->ja_list); 4320 } 4321 jaddref->ja_state |= (GOINGAWAY | DEPCOMPLETE); 4322 /* 4323 * Leave NEWBLOCK jaddrefs on the inodedep so handle_workitem_remove 4324 * can arrange for them to be freed with the bitmap. Otherwise we 4325 * no longer need this addref attached to the inoreflst and it 4326 * will incorrectly adjust nlink if we leave it. 4327 */ 4328 if ((jaddref->ja_state & NEWBLOCK) == 0) { 4329 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, 4330 if_deps); 4331 jaddref->ja_state |= COMPLETE; 4332 free_jaddref(jaddref); 4333 return (needsj); 4334 } 4335 /* 4336 * Leave the head of the list for jsegdeps for fast merging. 4337 */ 4338 if (LIST_FIRST(wkhd) != NULL) { 4339 jaddref->ja_state |= ONWORKLIST; 4340 LIST_INSERT_AFTER(LIST_FIRST(wkhd), &jaddref->ja_list, wk_list); 4341 } else 4342 WORKLIST_INSERT(wkhd, &jaddref->ja_list); 4343 4344 return (needsj); 4345} 4346 4347/* 4348 * Attempt to free a jaddref structure when some work completes. This 4349 * should only succeed once the entry is written and all dependencies have 4350 * been notified. 4351 */ 4352static void 4353free_jaddref(jaddref) 4354 struct jaddref *jaddref; 4355{ 4356 4357 if ((jaddref->ja_state & ALLCOMPLETE) != ALLCOMPLETE) 4358 return; 4359 if (jaddref->ja_ref.if_jsegdep) 4360 panic("free_jaddref: segdep attached to jaddref %p(0x%X)\n", 4361 jaddref, jaddref->ja_state); 4362 if (jaddref->ja_state & NEWBLOCK) 4363 LIST_REMOVE(jaddref, ja_bmdeps); 4364 if (jaddref->ja_state & (INPROGRESS | ONWORKLIST)) 4365 panic("free_jaddref: Bad state %p(0x%X)", 4366 jaddref, jaddref->ja_state); 4367 if (jaddref->ja_mkdir != NULL) 4368 panic("free_jaddref: Work pending, 0x%X\n", jaddref->ja_state); 4369 WORKITEM_FREE(jaddref, D_JADDREF); 4370} 4371 4372/* 4373 * Free a jremref structure once it has been written or discarded. 4374 */ 4375static void 4376free_jremref(jremref) 4377 struct jremref *jremref; 4378{ 4379 4380 if (jremref->jr_ref.if_jsegdep) 4381 free_jsegdep(jremref->jr_ref.if_jsegdep); 4382 if (jremref->jr_state & INPROGRESS) 4383 panic("free_jremref: IO still pending"); 4384 WORKITEM_FREE(jremref, D_JREMREF); 4385} 4386 4387/* 4388 * Free a jnewblk structure. 4389 */ 4390static void 4391free_jnewblk(jnewblk) 4392 struct jnewblk *jnewblk; 4393{ 4394 4395 if ((jnewblk->jn_state & ALLCOMPLETE) != ALLCOMPLETE) 4396 return; 4397 LIST_REMOVE(jnewblk, jn_deps); 4398 if (jnewblk->jn_dep != NULL) 4399 panic("free_jnewblk: Dependency still attached."); 4400 WORKITEM_FREE(jnewblk, D_JNEWBLK); 4401} 4402 4403/* 4404 * Cancel a jnewblk which has been been made redundant by frag extension. 4405 */ 4406static void 4407cancel_jnewblk(jnewblk, wkhd) 4408 struct jnewblk *jnewblk; 4409 struct workhead *wkhd; 4410{ 4411 struct jsegdep *jsegdep; 4412 4413 CTR1(KTR_SUJ, "cancel_jnewblk: blkno %jd", jnewblk->jn_blkno); 4414 jsegdep = jnewblk->jn_jsegdep; 4415 if (jnewblk->jn_jsegdep == NULL || jnewblk->jn_dep == NULL) 4416 panic("cancel_jnewblk: Invalid state"); 4417 jnewblk->jn_jsegdep = NULL; 4418 jnewblk->jn_dep = NULL; 4419 jnewblk->jn_state |= GOINGAWAY; 4420 if (jnewblk->jn_state & INPROGRESS) { 4421 jnewblk->jn_state &= ~INPROGRESS; 4422 WORKLIST_REMOVE(&jnewblk->jn_list); 4423 jwork_insert(wkhd, jsegdep); 4424 } else { 4425 free_jsegdep(jsegdep); 4426 remove_from_journal(&jnewblk->jn_list); 4427 } 4428 wake_worklist(&jnewblk->jn_list); 4429 WORKLIST_INSERT(wkhd, &jnewblk->jn_list); 4430} 4431 4432static void 4433free_jblkdep(jblkdep) 4434 struct jblkdep *jblkdep; 4435{ 4436 4437 if (jblkdep->jb_list.wk_type == D_JFREEBLK) 4438 WORKITEM_FREE(jblkdep, D_JFREEBLK); 4439 else if (jblkdep->jb_list.wk_type == D_JTRUNC) 4440 WORKITEM_FREE(jblkdep, D_JTRUNC); 4441 else 4442 panic("free_jblkdep: Unexpected type %s", 4443 TYPENAME(jblkdep->jb_list.wk_type)); 4444} 4445 4446/* 4447 * Free a single jseg once it is no longer referenced in memory or on 4448 * disk. Reclaim journal blocks and dependencies waiting for the segment 4449 * to disappear. 4450 */ 4451static void 4452free_jseg(jseg, jblocks) 4453 struct jseg *jseg; 4454 struct jblocks *jblocks; 4455{ 4456 struct freework *freework; 4457 4458 /* 4459 * Free freework structures that were lingering to indicate freed 4460 * indirect blocks that forced journal write ordering on reallocate. 4461 */ 4462 while ((freework = LIST_FIRST(&jseg->js_indirs)) != NULL) 4463 indirblk_remove(freework); 4464 if (jblocks->jb_oldestseg == jseg) 4465 jblocks->jb_oldestseg = TAILQ_NEXT(jseg, js_next); 4466 TAILQ_REMOVE(&jblocks->jb_segs, jseg, js_next); 4467 jblocks_free(jblocks, jseg->js_list.wk_mp, jseg->js_size); 4468 KASSERT(LIST_EMPTY(&jseg->js_entries), 4469 ("free_jseg: Freed jseg has valid entries.")); 4470 WORKITEM_FREE(jseg, D_JSEG); 4471} 4472 4473/* 4474 * Free all jsegs that meet the criteria for being reclaimed and update 4475 * oldestseg. 4476 */ 4477static void 4478free_jsegs(jblocks) 4479 struct jblocks *jblocks; 4480{ 4481 struct jseg *jseg; 4482 4483 /* 4484 * Free only those jsegs which have none allocated before them to 4485 * preserve the journal space ordering. 4486 */ 4487 while ((jseg = TAILQ_FIRST(&jblocks->jb_segs)) != NULL) { 4488 /* 4489 * Only reclaim space when nothing depends on this journal 4490 * set and another set has written that it is no longer 4491 * valid. 4492 */ 4493 if (jseg->js_refs != 0) { 4494 jblocks->jb_oldestseg = jseg; 4495 return; 4496 } 4497 if ((jseg->js_state & ALLCOMPLETE) != ALLCOMPLETE) 4498 break; 4499 if (jseg->js_seq > jblocks->jb_oldestwrseq) 4500 break; 4501 /* 4502 * We can free jsegs that didn't write entries when 4503 * oldestwrseq == js_seq. 4504 */ 4505 if (jseg->js_seq == jblocks->jb_oldestwrseq && 4506 jseg->js_cnt != 0) 4507 break; 4508 free_jseg(jseg, jblocks); 4509 } 4510 /* 4511 * If we exited the loop above we still must discover the 4512 * oldest valid segment. 4513 */ 4514 if (jseg) 4515 for (jseg = jblocks->jb_oldestseg; jseg != NULL; 4516 jseg = TAILQ_NEXT(jseg, js_next)) 4517 if (jseg->js_refs != 0) 4518 break; 4519 jblocks->jb_oldestseg = jseg; 4520 /* 4521 * The journal has no valid records but some jsegs may still be 4522 * waiting on oldestwrseq to advance. We force a small record 4523 * out to permit these lingering records to be reclaimed. 4524 */ 4525 if (jblocks->jb_oldestseg == NULL && !TAILQ_EMPTY(&jblocks->jb_segs)) 4526 jblocks->jb_needseg = 1; 4527} 4528 4529/* 4530 * Release one reference to a jseg and free it if the count reaches 0. This 4531 * should eventually reclaim journal space as well. 4532 */ 4533static void 4534rele_jseg(jseg) 4535 struct jseg *jseg; 4536{ 4537 4538 KASSERT(jseg->js_refs > 0, 4539 ("free_jseg: Invalid refcnt %d", jseg->js_refs)); 4540 if (--jseg->js_refs != 0) 4541 return; 4542 free_jsegs(jseg->js_jblocks); 4543} 4544 4545/* 4546 * Release a jsegdep and decrement the jseg count. 4547 */ 4548static void 4549free_jsegdep(jsegdep) 4550 struct jsegdep *jsegdep; 4551{ 4552 4553 if (jsegdep->jd_seg) 4554 rele_jseg(jsegdep->jd_seg); 4555 WORKITEM_FREE(jsegdep, D_JSEGDEP); 4556} 4557 4558/* 4559 * Wait for a journal item to make it to disk. Initiate journal processing 4560 * if required. 4561 */ 4562static int 4563jwait(wk, waitfor) 4564 struct worklist *wk; 4565 int waitfor; 4566{ 4567 4568 LOCK_OWNED(VFSTOUFS(wk->wk_mp)); 4569 /* 4570 * Blocking journal waits cause slow synchronous behavior. Record 4571 * stats on the frequency of these blocking operations. 4572 */ 4573 if (waitfor == MNT_WAIT) { 4574 stat_journal_wait++; 4575 switch (wk->wk_type) { 4576 case D_JREMREF: 4577 case D_JMVREF: 4578 stat_jwait_filepage++; 4579 break; 4580 case D_JTRUNC: 4581 case D_JFREEBLK: 4582 stat_jwait_freeblks++; 4583 break; 4584 case D_JNEWBLK: 4585 stat_jwait_newblk++; 4586 break; 4587 case D_JADDREF: 4588 stat_jwait_inode++; 4589 break; 4590 default: 4591 break; 4592 } 4593 } 4594 /* 4595 * If IO has not started we process the journal. We can't mark the 4596 * worklist item as IOWAITING because we drop the lock while 4597 * processing the journal and the worklist entry may be freed after 4598 * this point. The caller may call back in and re-issue the request. 4599 */ 4600 if ((wk->wk_state & INPROGRESS) == 0) { 4601 softdep_process_journal(wk->wk_mp, wk, waitfor); 4602 if (waitfor != MNT_WAIT) 4603 return (EBUSY); 4604 return (0); 4605 } 4606 if (waitfor != MNT_WAIT) 4607 return (EBUSY); 4608 wait_worklist(wk, "jwait"); 4609 return (0); 4610} 4611 4612/* 4613 * Lookup an inodedep based on an inode pointer and set the nlinkdelta as 4614 * appropriate. This is a convenience function to reduce duplicate code 4615 * for the setup and revert functions below. 4616 */ 4617static struct inodedep * 4618inodedep_lookup_ip(ip) 4619 struct inode *ip; 4620{ 4621 struct inodedep *inodedep; 4622 int dflags; 4623 4624 KASSERT(ip->i_nlink >= ip->i_effnlink, 4625 ("inodedep_lookup_ip: bad delta")); 4626 dflags = DEPALLOC; 4627 if (IS_SNAPSHOT(ip)) 4628 dflags |= NODELAY; 4629 (void) inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, dflags, 4630 &inodedep); 4631 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 4632 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked")); 4633 4634 return (inodedep); 4635} 4636 4637/* 4638 * Called prior to creating a new inode and linking it to a directory. The 4639 * jaddref structure must already be allocated by softdep_setup_inomapdep 4640 * and it is discovered here so we can initialize the mode and update 4641 * nlinkdelta. 4642 */ 4643void 4644softdep_setup_create(dp, ip) 4645 struct inode *dp; 4646 struct inode *ip; 4647{ 4648 struct inodedep *inodedep; 4649 struct jaddref *jaddref; 4650 struct vnode *dvp; 4651 4652 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0, 4653 ("softdep_setup_create called on non-softdep filesystem")); 4654 KASSERT(ip->i_nlink == 1, 4655 ("softdep_setup_create: Invalid link count.")); 4656 dvp = ITOV(dp); 4657 ACQUIRE_LOCK(dp->i_ump); 4658 inodedep = inodedep_lookup_ip(ip); 4659 if (DOINGSUJ(dvp)) { 4660 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4661 inoreflst); 4662 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 4663 ("softdep_setup_create: No addref structure present.")); 4664 } 4665 softdep_prelink(dvp, NULL); 4666 FREE_LOCK(dp->i_ump); 4667} 4668 4669/* 4670 * Create a jaddref structure to track the addition of a DOTDOT link when 4671 * we are reparenting an inode as part of a rename. This jaddref will be 4672 * found by softdep_setup_directory_change. Adjusts nlinkdelta for 4673 * non-journaling softdep. 4674 */ 4675void 4676softdep_setup_dotdot_link(dp, ip) 4677 struct inode *dp; 4678 struct inode *ip; 4679{ 4680 struct inodedep *inodedep; 4681 struct jaddref *jaddref; 4682 struct vnode *dvp; 4683 struct vnode *vp; 4684 4685 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0, 4686 ("softdep_setup_dotdot_link called on non-softdep filesystem")); 4687 dvp = ITOV(dp); 4688 vp = ITOV(ip); 4689 jaddref = NULL; 4690 /* 4691 * We don't set MKDIR_PARENT as this is not tied to a mkdir and 4692 * is used as a normal link would be. 4693 */ 4694 if (DOINGSUJ(dvp)) 4695 jaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET, 4696 dp->i_effnlink - 1, dp->i_mode); 4697 ACQUIRE_LOCK(dp->i_ump); 4698 inodedep = inodedep_lookup_ip(dp); 4699 if (jaddref) 4700 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 4701 if_deps); 4702 softdep_prelink(dvp, ITOV(ip)); 4703 FREE_LOCK(dp->i_ump); 4704} 4705 4706/* 4707 * Create a jaddref structure to track a new link to an inode. The directory 4708 * offset is not known until softdep_setup_directory_add or 4709 * softdep_setup_directory_change. Adjusts nlinkdelta for non-journaling 4710 * softdep. 4711 */ 4712void 4713softdep_setup_link(dp, ip) 4714 struct inode *dp; 4715 struct inode *ip; 4716{ 4717 struct inodedep *inodedep; 4718 struct jaddref *jaddref; 4719 struct vnode *dvp; 4720 4721 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0, 4722 ("softdep_setup_link called on non-softdep filesystem")); 4723 dvp = ITOV(dp); 4724 jaddref = NULL; 4725 if (DOINGSUJ(dvp)) 4726 jaddref = newjaddref(dp, ip->i_number, 0, ip->i_effnlink - 1, 4727 ip->i_mode); 4728 ACQUIRE_LOCK(dp->i_ump); 4729 inodedep = inodedep_lookup_ip(ip); 4730 if (jaddref) 4731 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 4732 if_deps); 4733 softdep_prelink(dvp, ITOV(ip)); 4734 FREE_LOCK(dp->i_ump); 4735} 4736 4737/* 4738 * Called to create the jaddref structures to track . and .. references as 4739 * well as lookup and further initialize the incomplete jaddref created 4740 * by softdep_setup_inomapdep when the inode was allocated. Adjusts 4741 * nlinkdelta for non-journaling softdep. 4742 */ 4743void 4744softdep_setup_mkdir(dp, ip) 4745 struct inode *dp; 4746 struct inode *ip; 4747{ 4748 struct inodedep *inodedep; 4749 struct jaddref *dotdotaddref; 4750 struct jaddref *dotaddref; 4751 struct jaddref *jaddref; 4752 struct vnode *dvp; 4753 4754 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0, 4755 ("softdep_setup_mkdir called on non-softdep filesystem")); 4756 dvp = ITOV(dp); 4757 dotaddref = dotdotaddref = NULL; 4758 if (DOINGSUJ(dvp)) { 4759 dotaddref = newjaddref(ip, ip->i_number, DOT_OFFSET, 1, 4760 ip->i_mode); 4761 dotaddref->ja_state |= MKDIR_BODY; 4762 dotdotaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET, 4763 dp->i_effnlink - 1, dp->i_mode); 4764 dotdotaddref->ja_state |= MKDIR_PARENT; 4765 } 4766 ACQUIRE_LOCK(dp->i_ump); 4767 inodedep = inodedep_lookup_ip(ip); 4768 if (DOINGSUJ(dvp)) { 4769 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4770 inoreflst); 4771 KASSERT(jaddref != NULL, 4772 ("softdep_setup_mkdir: No addref structure present.")); 4773 KASSERT(jaddref->ja_parent == dp->i_number, 4774 ("softdep_setup_mkdir: bad parent %ju", 4775 (uintmax_t)jaddref->ja_parent)); 4776 TAILQ_INSERT_BEFORE(&jaddref->ja_ref, &dotaddref->ja_ref, 4777 if_deps); 4778 } 4779 inodedep = inodedep_lookup_ip(dp); 4780 if (DOINGSUJ(dvp)) 4781 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, 4782 &dotdotaddref->ja_ref, if_deps); 4783 softdep_prelink(ITOV(dp), NULL); 4784 FREE_LOCK(dp->i_ump); 4785} 4786 4787/* 4788 * Called to track nlinkdelta of the inode and parent directories prior to 4789 * unlinking a directory. 4790 */ 4791void 4792softdep_setup_rmdir(dp, ip) 4793 struct inode *dp; 4794 struct inode *ip; 4795{ 4796 struct vnode *dvp; 4797 4798 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0, 4799 ("softdep_setup_rmdir called on non-softdep filesystem")); 4800 dvp = ITOV(dp); 4801 ACQUIRE_LOCK(dp->i_ump); 4802 (void) inodedep_lookup_ip(ip); 4803 (void) inodedep_lookup_ip(dp); 4804 softdep_prelink(dvp, ITOV(ip)); 4805 FREE_LOCK(dp->i_ump); 4806} 4807 4808/* 4809 * Called to track nlinkdelta of the inode and parent directories prior to 4810 * unlink. 4811 */ 4812void 4813softdep_setup_unlink(dp, ip) 4814 struct inode *dp; 4815 struct inode *ip; 4816{ 4817 struct vnode *dvp; 4818 4819 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0, 4820 ("softdep_setup_unlink called on non-softdep filesystem")); 4821 dvp = ITOV(dp); 4822 ACQUIRE_LOCK(dp->i_ump); 4823 (void) inodedep_lookup_ip(ip); 4824 (void) inodedep_lookup_ip(dp); 4825 softdep_prelink(dvp, ITOV(ip)); 4826 FREE_LOCK(dp->i_ump); 4827} 4828 4829/* 4830 * Called to release the journal structures created by a failed non-directory 4831 * creation. Adjusts nlinkdelta for non-journaling softdep. 4832 */ 4833void 4834softdep_revert_create(dp, ip) 4835 struct inode *dp; 4836 struct inode *ip; 4837{ 4838 struct inodedep *inodedep; 4839 struct jaddref *jaddref; 4840 struct vnode *dvp; 4841 4842 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0, 4843 ("softdep_revert_create called on non-softdep filesystem")); 4844 dvp = ITOV(dp); 4845 ACQUIRE_LOCK(dp->i_ump); 4846 inodedep = inodedep_lookup_ip(ip); 4847 if (DOINGSUJ(dvp)) { 4848 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4849 inoreflst); 4850 KASSERT(jaddref->ja_parent == dp->i_number, 4851 ("softdep_revert_create: addref parent mismatch")); 4852 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4853 } 4854 FREE_LOCK(dp->i_ump); 4855} 4856 4857/* 4858 * Called to release the journal structures created by a failed link 4859 * addition. Adjusts nlinkdelta for non-journaling softdep. 4860 */ 4861void 4862softdep_revert_link(dp, ip) 4863 struct inode *dp; 4864 struct inode *ip; 4865{ 4866 struct inodedep *inodedep; 4867 struct jaddref *jaddref; 4868 struct vnode *dvp; 4869 4870 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0, 4871 ("softdep_revert_link called on non-softdep filesystem")); 4872 dvp = ITOV(dp); 4873 ACQUIRE_LOCK(dp->i_ump); 4874 inodedep = inodedep_lookup_ip(ip); 4875 if (DOINGSUJ(dvp)) { 4876 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4877 inoreflst); 4878 KASSERT(jaddref->ja_parent == dp->i_number, 4879 ("softdep_revert_link: addref parent mismatch")); 4880 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4881 } 4882 FREE_LOCK(dp->i_ump); 4883} 4884 4885/* 4886 * Called to release the journal structures created by a failed mkdir 4887 * attempt. Adjusts nlinkdelta for non-journaling softdep. 4888 */ 4889void 4890softdep_revert_mkdir(dp, ip) 4891 struct inode *dp; 4892 struct inode *ip; 4893{ 4894 struct inodedep *inodedep; 4895 struct jaddref *jaddref; 4896 struct jaddref *dotaddref; 4897 struct vnode *dvp; 4898 4899 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0, 4900 ("softdep_revert_mkdir called on non-softdep filesystem")); 4901 dvp = ITOV(dp); 4902 4903 ACQUIRE_LOCK(dp->i_ump); 4904 inodedep = inodedep_lookup_ip(dp); 4905 if (DOINGSUJ(dvp)) { 4906 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4907 inoreflst); 4908 KASSERT(jaddref->ja_parent == ip->i_number, 4909 ("softdep_revert_mkdir: dotdot addref parent mismatch")); 4910 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4911 } 4912 inodedep = inodedep_lookup_ip(ip); 4913 if (DOINGSUJ(dvp)) { 4914 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4915 inoreflst); 4916 KASSERT(jaddref->ja_parent == dp->i_number, 4917 ("softdep_revert_mkdir: addref parent mismatch")); 4918 dotaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref, 4919 inoreflst, if_deps); 4920 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4921 KASSERT(dotaddref->ja_parent == ip->i_number, 4922 ("softdep_revert_mkdir: dot addref parent mismatch")); 4923 cancel_jaddref(dotaddref, inodedep, &inodedep->id_inowait); 4924 } 4925 FREE_LOCK(dp->i_ump); 4926} 4927 4928/* 4929 * Called to correct nlinkdelta after a failed rmdir. 4930 */ 4931void 4932softdep_revert_rmdir(dp, ip) 4933 struct inode *dp; 4934 struct inode *ip; 4935{ 4936 4937 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0, 4938 ("softdep_revert_rmdir called on non-softdep filesystem")); 4939 ACQUIRE_LOCK(dp->i_ump); 4940 (void) inodedep_lookup_ip(ip); 4941 (void) inodedep_lookup_ip(dp); 4942 FREE_LOCK(dp->i_ump); 4943} 4944 4945/* 4946 * Protecting the freemaps (or bitmaps). 4947 * 4948 * To eliminate the need to execute fsck before mounting a filesystem 4949 * after a power failure, one must (conservatively) guarantee that the 4950 * on-disk copy of the bitmaps never indicate that a live inode or block is 4951 * free. So, when a block or inode is allocated, the bitmap should be 4952 * updated (on disk) before any new pointers. When a block or inode is 4953 * freed, the bitmap should not be updated until all pointers have been 4954 * reset. The latter dependency is handled by the delayed de-allocation 4955 * approach described below for block and inode de-allocation. The former 4956 * dependency is handled by calling the following procedure when a block or 4957 * inode is allocated. When an inode is allocated an "inodedep" is created 4958 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 4959 * Each "inodedep" is also inserted into the hash indexing structure so 4960 * that any additional link additions can be made dependent on the inode 4961 * allocation. 4962 * 4963 * The ufs filesystem maintains a number of free block counts (e.g., per 4964 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 4965 * in addition to the bitmaps. These counts are used to improve efficiency 4966 * during allocation and therefore must be consistent with the bitmaps. 4967 * There is no convenient way to guarantee post-crash consistency of these 4968 * counts with simple update ordering, for two main reasons: (1) The counts 4969 * and bitmaps for a single cylinder group block are not in the same disk 4970 * sector. If a disk write is interrupted (e.g., by power failure), one may 4971 * be written and the other not. (2) Some of the counts are located in the 4972 * superblock rather than the cylinder group block. So, we focus our soft 4973 * updates implementation on protecting the bitmaps. When mounting a 4974 * filesystem, we recompute the auxiliary counts from the bitmaps. 4975 */ 4976 4977/* 4978 * Called just after updating the cylinder group block to allocate an inode. 4979 */ 4980void 4981softdep_setup_inomapdep(bp, ip, newinum, mode) 4982 struct buf *bp; /* buffer for cylgroup block with inode map */ 4983 struct inode *ip; /* inode related to allocation */ 4984 ino_t newinum; /* new inode number being allocated */ 4985 int mode; 4986{ 4987 struct inodedep *inodedep; 4988 struct bmsafemap *bmsafemap; 4989 struct jaddref *jaddref; 4990 struct mount *mp; 4991 struct fs *fs; 4992 4993 mp = UFSTOVFS(ip->i_ump); 4994 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 4995 ("softdep_setup_inomapdep called on non-softdep filesystem")); 4996 fs = ip->i_ump->um_fs; 4997 jaddref = NULL; 4998 4999 /* 5000 * Allocate the journal reference add structure so that the bitmap 5001 * can be dependent on it. 5002 */ 5003 if (MOUNTEDSUJ(mp)) { 5004 jaddref = newjaddref(ip, newinum, 0, 0, mode); 5005 jaddref->ja_state |= NEWBLOCK; 5006 } 5007 5008 /* 5009 * Create a dependency for the newly allocated inode. 5010 * Panic if it already exists as something is seriously wrong. 5011 * Otherwise add it to the dependency list for the buffer holding 5012 * the cylinder group map from which it was allocated. 5013 * 5014 * We have to preallocate a bmsafemap entry in case it is needed 5015 * in bmsafemap_lookup since once we allocate the inodedep, we 5016 * have to finish initializing it before we can FREE_LOCK(). 5017 * By preallocating, we avoid FREE_LOCK() while doing a malloc 5018 * in bmsafemap_lookup. We cannot call bmsafemap_lookup before 5019 * creating the inodedep as it can be freed during the time 5020 * that we FREE_LOCK() while allocating the inodedep. We must 5021 * call workitem_alloc() before entering the locked section as 5022 * it also acquires the lock and we must avoid trying doing so 5023 * recursively. 5024 */ 5025 bmsafemap = malloc(sizeof(struct bmsafemap), 5026 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 5027 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); 5028 ACQUIRE_LOCK(ip->i_ump); 5029 if ((inodedep_lookup(mp, newinum, DEPALLOC | NODELAY, &inodedep))) 5030 panic("softdep_setup_inomapdep: dependency %p for new" 5031 "inode already exists", inodedep); 5032 bmsafemap = bmsafemap_lookup(mp, bp, ino_to_cg(fs, newinum), bmsafemap); 5033 if (jaddref) { 5034 LIST_INSERT_HEAD(&bmsafemap->sm_jaddrefhd, jaddref, ja_bmdeps); 5035 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 5036 if_deps); 5037 } else { 5038 inodedep->id_state |= ONDEPLIST; 5039 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 5040 } 5041 inodedep->id_bmsafemap = bmsafemap; 5042 inodedep->id_state &= ~DEPCOMPLETE; 5043 FREE_LOCK(ip->i_ump); 5044} 5045 5046/* 5047 * Called just after updating the cylinder group block to 5048 * allocate block or fragment. 5049 */ 5050void 5051softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags) 5052 struct buf *bp; /* buffer for cylgroup block with block map */ 5053 struct mount *mp; /* filesystem doing allocation */ 5054 ufs2_daddr_t newblkno; /* number of newly allocated block */ 5055 int frags; /* Number of fragments. */ 5056 int oldfrags; /* Previous number of fragments for extend. */ 5057{ 5058 struct newblk *newblk; 5059 struct bmsafemap *bmsafemap; 5060 struct jnewblk *jnewblk; 5061 struct ufsmount *ump; 5062 struct fs *fs; 5063 5064 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5065 ("softdep_setup_blkmapdep called on non-softdep filesystem")); 5066 ump = VFSTOUFS(mp); 5067 fs = ump->um_fs; 5068 jnewblk = NULL; 5069 /* 5070 * Create a dependency for the newly allocated block. 5071 * Add it to the dependency list for the buffer holding 5072 * the cylinder group map from which it was allocated. 5073 */ 5074 if (MOUNTEDSUJ(mp)) { 5075 jnewblk = malloc(sizeof(*jnewblk), M_JNEWBLK, M_SOFTDEP_FLAGS); 5076 workitem_alloc(&jnewblk->jn_list, D_JNEWBLK, mp); 5077 jnewblk->jn_jsegdep = newjsegdep(&jnewblk->jn_list); 5078 jnewblk->jn_state = ATTACHED; 5079 jnewblk->jn_blkno = newblkno; 5080 jnewblk->jn_frags = frags; 5081 jnewblk->jn_oldfrags = oldfrags; 5082#ifdef SUJ_DEBUG 5083 { 5084 struct cg *cgp; 5085 uint8_t *blksfree; 5086 long bno; 5087 int i; 5088 5089 cgp = (struct cg *)bp->b_data; 5090 blksfree = cg_blksfree(cgp); 5091 bno = dtogd(fs, jnewblk->jn_blkno); 5092 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; 5093 i++) { 5094 if (isset(blksfree, bno + i)) 5095 panic("softdep_setup_blkmapdep: " 5096 "free fragment %d from %d-%d " 5097 "state 0x%X dep %p", i, 5098 jnewblk->jn_oldfrags, 5099 jnewblk->jn_frags, 5100 jnewblk->jn_state, 5101 jnewblk->jn_dep); 5102 } 5103 } 5104#endif 5105 } 5106 5107 CTR3(KTR_SUJ, 5108 "softdep_setup_blkmapdep: blkno %jd frags %d oldfrags %d", 5109 newblkno, frags, oldfrags); 5110 ACQUIRE_LOCK(ump); 5111 if (newblk_lookup(mp, newblkno, DEPALLOC, &newblk) != 0) 5112 panic("softdep_setup_blkmapdep: found block"); 5113 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp, 5114 dtog(fs, newblkno), NULL); 5115 if (jnewblk) { 5116 jnewblk->jn_dep = (struct worklist *)newblk; 5117 LIST_INSERT_HEAD(&bmsafemap->sm_jnewblkhd, jnewblk, jn_deps); 5118 } else { 5119 newblk->nb_state |= ONDEPLIST; 5120 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 5121 } 5122 newblk->nb_bmsafemap = bmsafemap; 5123 newblk->nb_jnewblk = jnewblk; 5124 FREE_LOCK(ump); 5125} 5126 5127#define BMSAFEMAP_HASH(ump, cg) \ 5128 (&(ump)->bmsafemap_hashtbl[(cg) & (ump)->bmsafemap_hash_size]) 5129 5130static int 5131bmsafemap_find(bmsafemaphd, cg, bmsafemapp) 5132 struct bmsafemap_hashhead *bmsafemaphd; 5133 int cg; 5134 struct bmsafemap **bmsafemapp; 5135{ 5136 struct bmsafemap *bmsafemap; 5137 5138 LIST_FOREACH(bmsafemap, bmsafemaphd, sm_hash) 5139 if (bmsafemap->sm_cg == cg) 5140 break; 5141 if (bmsafemap) { 5142 *bmsafemapp = bmsafemap; 5143 return (1); 5144 } 5145 *bmsafemapp = NULL; 5146 5147 return (0); 5148} 5149 5150/* 5151 * Find the bmsafemap associated with a cylinder group buffer. 5152 * If none exists, create one. The buffer must be locked when 5153 * this routine is called and this routine must be called with 5154 * the softdep lock held. To avoid giving up the lock while 5155 * allocating a new bmsafemap, a preallocated bmsafemap may be 5156 * provided. If it is provided but not needed, it is freed. 5157 */ 5158static struct bmsafemap * 5159bmsafemap_lookup(mp, bp, cg, newbmsafemap) 5160 struct mount *mp; 5161 struct buf *bp; 5162 int cg; 5163 struct bmsafemap *newbmsafemap; 5164{ 5165 struct bmsafemap_hashhead *bmsafemaphd; 5166 struct bmsafemap *bmsafemap, *collision; 5167 struct worklist *wk; 5168 struct ufsmount *ump; 5169 5170 ump = VFSTOUFS(mp); 5171 LOCK_OWNED(ump); 5172 KASSERT(bp != NULL, ("bmsafemap_lookup: missing buffer")); 5173 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5174 if (wk->wk_type == D_BMSAFEMAP) { 5175 if (newbmsafemap) 5176 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP); 5177 return (WK_BMSAFEMAP(wk)); 5178 } 5179 } 5180 bmsafemaphd = BMSAFEMAP_HASH(ump, cg); 5181 if (bmsafemap_find(bmsafemaphd, cg, &bmsafemap) == 1) { 5182 if (newbmsafemap) 5183 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP); 5184 return (bmsafemap); 5185 } 5186 if (newbmsafemap) { 5187 bmsafemap = newbmsafemap; 5188 } else { 5189 FREE_LOCK(ump); 5190 bmsafemap = malloc(sizeof(struct bmsafemap), 5191 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 5192 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); 5193 ACQUIRE_LOCK(ump); 5194 } 5195 bmsafemap->sm_buf = bp; 5196 LIST_INIT(&bmsafemap->sm_inodedephd); 5197 LIST_INIT(&bmsafemap->sm_inodedepwr); 5198 LIST_INIT(&bmsafemap->sm_newblkhd); 5199 LIST_INIT(&bmsafemap->sm_newblkwr); 5200 LIST_INIT(&bmsafemap->sm_jaddrefhd); 5201 LIST_INIT(&bmsafemap->sm_jnewblkhd); 5202 LIST_INIT(&bmsafemap->sm_freehd); 5203 LIST_INIT(&bmsafemap->sm_freewr); 5204 if (bmsafemap_find(bmsafemaphd, cg, &collision) == 1) { 5205 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 5206 return (collision); 5207 } 5208 bmsafemap->sm_cg = cg; 5209 LIST_INSERT_HEAD(bmsafemaphd, bmsafemap, sm_hash); 5210 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next); 5211 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 5212 return (bmsafemap); 5213} 5214 5215/* 5216 * Direct block allocation dependencies. 5217 * 5218 * When a new block is allocated, the corresponding disk locations must be 5219 * initialized (with zeros or new data) before the on-disk inode points to 5220 * them. Also, the freemap from which the block was allocated must be 5221 * updated (on disk) before the inode's pointer. These two dependencies are 5222 * independent of each other and are needed for all file blocks and indirect 5223 * blocks that are pointed to directly by the inode. Just before the 5224 * "in-core" version of the inode is updated with a newly allocated block 5225 * number, a procedure (below) is called to setup allocation dependency 5226 * structures. These structures are removed when the corresponding 5227 * dependencies are satisfied or when the block allocation becomes obsolete 5228 * (i.e., the file is deleted, the block is de-allocated, or the block is a 5229 * fragment that gets upgraded). All of these cases are handled in 5230 * procedures described later. 5231 * 5232 * When a file extension causes a fragment to be upgraded, either to a larger 5233 * fragment or to a full block, the on-disk location may change (if the 5234 * previous fragment could not simply be extended). In this case, the old 5235 * fragment must be de-allocated, but not until after the inode's pointer has 5236 * been updated. In most cases, this is handled by later procedures, which 5237 * will construct a "freefrag" structure to be added to the workitem queue 5238 * when the inode update is complete (or obsolete). The main exception to 5239 * this is when an allocation occurs while a pending allocation dependency 5240 * (for the same block pointer) remains. This case is handled in the main 5241 * allocation dependency setup procedure by immediately freeing the 5242 * unreferenced fragments. 5243 */ 5244void 5245softdep_setup_allocdirect(ip, off, newblkno, oldblkno, newsize, oldsize, bp) 5246 struct inode *ip; /* inode to which block is being added */ 5247 ufs_lbn_t off; /* block pointer within inode */ 5248 ufs2_daddr_t newblkno; /* disk block number being added */ 5249 ufs2_daddr_t oldblkno; /* previous block number, 0 unless frag */ 5250 long newsize; /* size of new block */ 5251 long oldsize; /* size of new block */ 5252 struct buf *bp; /* bp for allocated block */ 5253{ 5254 struct allocdirect *adp, *oldadp; 5255 struct allocdirectlst *adphead; 5256 struct freefrag *freefrag; 5257 struct inodedep *inodedep; 5258 struct pagedep *pagedep; 5259 struct jnewblk *jnewblk; 5260 struct newblk *newblk; 5261 struct mount *mp; 5262 ufs_lbn_t lbn; 5263 5264 lbn = bp->b_lblkno; 5265 mp = UFSTOVFS(ip->i_ump); 5266 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5267 ("softdep_setup_allocdirect called on non-softdep filesystem")); 5268 if (oldblkno && oldblkno != newblkno) 5269 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn); 5270 else 5271 freefrag = NULL; 5272 5273 CTR6(KTR_SUJ, 5274 "softdep_setup_allocdirect: ino %d blkno %jd oldblkno %jd " 5275 "off %jd newsize %ld oldsize %d", 5276 ip->i_number, newblkno, oldblkno, off, newsize, oldsize); 5277 ACQUIRE_LOCK(ip->i_ump); 5278 if (off >= NDADDR) { 5279 if (lbn > 0) 5280 panic("softdep_setup_allocdirect: bad lbn %jd, off %jd", 5281 lbn, off); 5282 /* allocating an indirect block */ 5283 if (oldblkno != 0) 5284 panic("softdep_setup_allocdirect: non-zero indir"); 5285 } else { 5286 if (off != lbn) 5287 panic("softdep_setup_allocdirect: lbn %jd != off %jd", 5288 lbn, off); 5289 /* 5290 * Allocating a direct block. 5291 * 5292 * If we are allocating a directory block, then we must 5293 * allocate an associated pagedep to track additions and 5294 * deletions. 5295 */ 5296 if ((ip->i_mode & IFMT) == IFDIR) 5297 pagedep_lookup(mp, bp, ip->i_number, off, DEPALLOC, 5298 &pagedep); 5299 } 5300 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0) 5301 panic("softdep_setup_allocdirect: lost block"); 5302 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5303 ("softdep_setup_allocdirect: newblk already initialized")); 5304 /* 5305 * Convert the newblk to an allocdirect. 5306 */ 5307 WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT); 5308 adp = (struct allocdirect *)newblk; 5309 newblk->nb_freefrag = freefrag; 5310 adp->ad_offset = off; 5311 adp->ad_oldblkno = oldblkno; 5312 adp->ad_newsize = newsize; 5313 adp->ad_oldsize = oldsize; 5314 5315 /* 5316 * Finish initializing the journal. 5317 */ 5318 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5319 jnewblk->jn_ino = ip->i_number; 5320 jnewblk->jn_lbn = lbn; 5321 add_to_journal(&jnewblk->jn_list); 5322 } 5323 if (freefrag && freefrag->ff_jdep != NULL && 5324 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5325 add_to_journal(freefrag->ff_jdep); 5326 inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep); 5327 adp->ad_inodedep = inodedep; 5328 5329 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list); 5330 /* 5331 * The list of allocdirects must be kept in sorted and ascending 5332 * order so that the rollback routines can quickly determine the 5333 * first uncommitted block (the size of the file stored on disk 5334 * ends at the end of the lowest committed fragment, or if there 5335 * are no fragments, at the end of the highest committed block). 5336 * Since files generally grow, the typical case is that the new 5337 * block is to be added at the end of the list. We speed this 5338 * special case by checking against the last allocdirect in the 5339 * list before laboriously traversing the list looking for the 5340 * insertion point. 5341 */ 5342 adphead = &inodedep->id_newinoupdt; 5343 oldadp = TAILQ_LAST(adphead, allocdirectlst); 5344 if (oldadp == NULL || oldadp->ad_offset <= off) { 5345 /* insert at end of list */ 5346 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 5347 if (oldadp != NULL && oldadp->ad_offset == off) 5348 allocdirect_merge(adphead, adp, oldadp); 5349 FREE_LOCK(ip->i_ump); 5350 return; 5351 } 5352 TAILQ_FOREACH(oldadp, adphead, ad_next) { 5353 if (oldadp->ad_offset >= off) 5354 break; 5355 } 5356 if (oldadp == NULL) 5357 panic("softdep_setup_allocdirect: lost entry"); 5358 /* insert in middle of list */ 5359 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 5360 if (oldadp->ad_offset == off) 5361 allocdirect_merge(adphead, adp, oldadp); 5362 5363 FREE_LOCK(ip->i_ump); 5364} 5365 5366/* 5367 * Merge a newer and older journal record to be stored either in a 5368 * newblock or freefrag. This handles aggregating journal records for 5369 * fragment allocation into a second record as well as replacing a 5370 * journal free with an aborted journal allocation. A segment for the 5371 * oldest record will be placed on wkhd if it has been written. If not 5372 * the segment for the newer record will suffice. 5373 */ 5374static struct worklist * 5375jnewblk_merge(new, old, wkhd) 5376 struct worklist *new; 5377 struct worklist *old; 5378 struct workhead *wkhd; 5379{ 5380 struct jnewblk *njnewblk; 5381 struct jnewblk *jnewblk; 5382 5383 /* Handle NULLs to simplify callers. */ 5384 if (new == NULL) 5385 return (old); 5386 if (old == NULL) 5387 return (new); 5388 /* Replace a jfreefrag with a jnewblk. */ 5389 if (new->wk_type == D_JFREEFRAG) { 5390 if (WK_JNEWBLK(old)->jn_blkno != WK_JFREEFRAG(new)->fr_blkno) 5391 panic("jnewblk_merge: blkno mismatch: %p, %p", 5392 old, new); 5393 cancel_jfreefrag(WK_JFREEFRAG(new)); 5394 return (old); 5395 } 5396 if (old->wk_type != D_JNEWBLK || new->wk_type != D_JNEWBLK) 5397 panic("jnewblk_merge: Bad type: old %d new %d\n", 5398 old->wk_type, new->wk_type); 5399 /* 5400 * Handle merging of two jnewblk records that describe 5401 * different sets of fragments in the same block. 5402 */ 5403 jnewblk = WK_JNEWBLK(old); 5404 njnewblk = WK_JNEWBLK(new); 5405 if (jnewblk->jn_blkno != njnewblk->jn_blkno) 5406 panic("jnewblk_merge: Merging disparate blocks."); 5407 /* 5408 * The record may be rolled back in the cg. 5409 */ 5410 if (jnewblk->jn_state & UNDONE) { 5411 jnewblk->jn_state &= ~UNDONE; 5412 njnewblk->jn_state |= UNDONE; 5413 njnewblk->jn_state &= ~ATTACHED; 5414 } 5415 /* 5416 * We modify the newer addref and free the older so that if neither 5417 * has been written the most up-to-date copy will be on disk. If 5418 * both have been written but rolled back we only temporarily need 5419 * one of them to fix the bits when the cg write completes. 5420 */ 5421 jnewblk->jn_state |= ATTACHED | COMPLETE; 5422 njnewblk->jn_oldfrags = jnewblk->jn_oldfrags; 5423 cancel_jnewblk(jnewblk, wkhd); 5424 WORKLIST_REMOVE(&jnewblk->jn_list); 5425 free_jnewblk(jnewblk); 5426 return (new); 5427} 5428 5429/* 5430 * Replace an old allocdirect dependency with a newer one. 5431 * This routine must be called with splbio interrupts blocked. 5432 */ 5433static void 5434allocdirect_merge(adphead, newadp, oldadp) 5435 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 5436 struct allocdirect *newadp; /* allocdirect being added */ 5437 struct allocdirect *oldadp; /* existing allocdirect being checked */ 5438{ 5439 struct worklist *wk; 5440 struct freefrag *freefrag; 5441 5442 freefrag = NULL; 5443 LOCK_OWNED(VFSTOUFS(newadp->ad_list.wk_mp)); 5444 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 5445 newadp->ad_oldsize != oldadp->ad_newsize || 5446 newadp->ad_offset >= NDADDR) 5447 panic("%s %jd != new %jd || old size %ld != new %ld", 5448 "allocdirect_merge: old blkno", 5449 (intmax_t)newadp->ad_oldblkno, 5450 (intmax_t)oldadp->ad_newblkno, 5451 newadp->ad_oldsize, oldadp->ad_newsize); 5452 newadp->ad_oldblkno = oldadp->ad_oldblkno; 5453 newadp->ad_oldsize = oldadp->ad_oldsize; 5454 /* 5455 * If the old dependency had a fragment to free or had never 5456 * previously had a block allocated, then the new dependency 5457 * can immediately post its freefrag and adopt the old freefrag. 5458 * This action is done by swapping the freefrag dependencies. 5459 * The new dependency gains the old one's freefrag, and the 5460 * old one gets the new one and then immediately puts it on 5461 * the worklist when it is freed by free_newblk. It is 5462 * not possible to do this swap when the old dependency had a 5463 * non-zero size but no previous fragment to free. This condition 5464 * arises when the new block is an extension of the old block. 5465 * Here, the first part of the fragment allocated to the new 5466 * dependency is part of the block currently claimed on disk by 5467 * the old dependency, so cannot legitimately be freed until the 5468 * conditions for the new dependency are fulfilled. 5469 */ 5470 freefrag = newadp->ad_freefrag; 5471 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 5472 newadp->ad_freefrag = oldadp->ad_freefrag; 5473 oldadp->ad_freefrag = freefrag; 5474 } 5475 /* 5476 * If we are tracking a new directory-block allocation, 5477 * move it from the old allocdirect to the new allocdirect. 5478 */ 5479 if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) { 5480 WORKLIST_REMOVE(wk); 5481 if (!LIST_EMPTY(&oldadp->ad_newdirblk)) 5482 panic("allocdirect_merge: extra newdirblk"); 5483 WORKLIST_INSERT(&newadp->ad_newdirblk, wk); 5484 } 5485 TAILQ_REMOVE(adphead, oldadp, ad_next); 5486 /* 5487 * We need to move any journal dependencies over to the freefrag 5488 * that releases this block if it exists. Otherwise we are 5489 * extending an existing block and we'll wait until that is 5490 * complete to release the journal space and extend the 5491 * new journal to cover this old space as well. 5492 */ 5493 if (freefrag == NULL) { 5494 if (oldadp->ad_newblkno != newadp->ad_newblkno) 5495 panic("allocdirect_merge: %jd != %jd", 5496 oldadp->ad_newblkno, newadp->ad_newblkno); 5497 newadp->ad_block.nb_jnewblk = (struct jnewblk *) 5498 jnewblk_merge(&newadp->ad_block.nb_jnewblk->jn_list, 5499 &oldadp->ad_block.nb_jnewblk->jn_list, 5500 &newadp->ad_block.nb_jwork); 5501 oldadp->ad_block.nb_jnewblk = NULL; 5502 cancel_newblk(&oldadp->ad_block, NULL, 5503 &newadp->ad_block.nb_jwork); 5504 } else { 5505 wk = (struct worklist *) cancel_newblk(&oldadp->ad_block, 5506 &freefrag->ff_list, &freefrag->ff_jwork); 5507 freefrag->ff_jdep = jnewblk_merge(freefrag->ff_jdep, wk, 5508 &freefrag->ff_jwork); 5509 } 5510 free_newblk(&oldadp->ad_block); 5511} 5512 5513/* 5514 * Allocate a jfreefrag structure to journal a single block free. 5515 */ 5516static struct jfreefrag * 5517newjfreefrag(freefrag, ip, blkno, size, lbn) 5518 struct freefrag *freefrag; 5519 struct inode *ip; 5520 ufs2_daddr_t blkno; 5521 long size; 5522 ufs_lbn_t lbn; 5523{ 5524 struct jfreefrag *jfreefrag; 5525 struct fs *fs; 5526 5527 fs = ip->i_fs; 5528 jfreefrag = malloc(sizeof(struct jfreefrag), M_JFREEFRAG, 5529 M_SOFTDEP_FLAGS); 5530 workitem_alloc(&jfreefrag->fr_list, D_JFREEFRAG, UFSTOVFS(ip->i_ump)); 5531 jfreefrag->fr_jsegdep = newjsegdep(&jfreefrag->fr_list); 5532 jfreefrag->fr_state = ATTACHED | DEPCOMPLETE; 5533 jfreefrag->fr_ino = ip->i_number; 5534 jfreefrag->fr_lbn = lbn; 5535 jfreefrag->fr_blkno = blkno; 5536 jfreefrag->fr_frags = numfrags(fs, size); 5537 jfreefrag->fr_freefrag = freefrag; 5538 5539 return (jfreefrag); 5540} 5541 5542/* 5543 * Allocate a new freefrag structure. 5544 */ 5545static struct freefrag * 5546newfreefrag(ip, blkno, size, lbn) 5547 struct inode *ip; 5548 ufs2_daddr_t blkno; 5549 long size; 5550 ufs_lbn_t lbn; 5551{ 5552 struct freefrag *freefrag; 5553 struct fs *fs; 5554 5555 CTR4(KTR_SUJ, "newfreefrag: ino %d blkno %jd size %ld lbn %jd", 5556 ip->i_number, blkno, size, lbn); 5557 fs = ip->i_fs; 5558 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 5559 panic("newfreefrag: frag size"); 5560 freefrag = malloc(sizeof(struct freefrag), 5561 M_FREEFRAG, M_SOFTDEP_FLAGS); 5562 workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ip->i_ump)); 5563 freefrag->ff_state = ATTACHED; 5564 LIST_INIT(&freefrag->ff_jwork); 5565 freefrag->ff_inum = ip->i_number; 5566 freefrag->ff_vtype = ITOV(ip)->v_type; 5567 freefrag->ff_blkno = blkno; 5568 freefrag->ff_fragsize = size; 5569 5570 if (MOUNTEDSUJ(UFSTOVFS(ip->i_ump))) { 5571 freefrag->ff_jdep = (struct worklist *) 5572 newjfreefrag(freefrag, ip, blkno, size, lbn); 5573 } else { 5574 freefrag->ff_state |= DEPCOMPLETE; 5575 freefrag->ff_jdep = NULL; 5576 } 5577 5578 return (freefrag); 5579} 5580 5581/* 5582 * This workitem de-allocates fragments that were replaced during 5583 * file block allocation. 5584 */ 5585static void 5586handle_workitem_freefrag(freefrag) 5587 struct freefrag *freefrag; 5588{ 5589 struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp); 5590 struct workhead wkhd; 5591 5592 CTR3(KTR_SUJ, 5593 "handle_workitem_freefrag: ino %d blkno %jd size %ld", 5594 freefrag->ff_inum, freefrag->ff_blkno, freefrag->ff_fragsize); 5595 /* 5596 * It would be illegal to add new completion items to the 5597 * freefrag after it was schedule to be done so it must be 5598 * safe to modify the list head here. 5599 */ 5600 LIST_INIT(&wkhd); 5601 ACQUIRE_LOCK(ump); 5602 LIST_SWAP(&freefrag->ff_jwork, &wkhd, worklist, wk_list); 5603 /* 5604 * If the journal has not been written we must cancel it here. 5605 */ 5606 if (freefrag->ff_jdep) { 5607 if (freefrag->ff_jdep->wk_type != D_JNEWBLK) 5608 panic("handle_workitem_freefrag: Unexpected type %d\n", 5609 freefrag->ff_jdep->wk_type); 5610 cancel_jnewblk(WK_JNEWBLK(freefrag->ff_jdep), &wkhd); 5611 } 5612 FREE_LOCK(ump); 5613 ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno, 5614 freefrag->ff_fragsize, freefrag->ff_inum, freefrag->ff_vtype, &wkhd); 5615 ACQUIRE_LOCK(ump); 5616 WORKITEM_FREE(freefrag, D_FREEFRAG); 5617 FREE_LOCK(ump); 5618} 5619 5620/* 5621 * Set up a dependency structure for an external attributes data block. 5622 * This routine follows much of the structure of softdep_setup_allocdirect. 5623 * See the description of softdep_setup_allocdirect above for details. 5624 */ 5625void 5626softdep_setup_allocext(ip, off, newblkno, oldblkno, newsize, oldsize, bp) 5627 struct inode *ip; 5628 ufs_lbn_t off; 5629 ufs2_daddr_t newblkno; 5630 ufs2_daddr_t oldblkno; 5631 long newsize; 5632 long oldsize; 5633 struct buf *bp; 5634{ 5635 struct allocdirect *adp, *oldadp; 5636 struct allocdirectlst *adphead; 5637 struct freefrag *freefrag; 5638 struct inodedep *inodedep; 5639 struct jnewblk *jnewblk; 5640 struct newblk *newblk; 5641 struct mount *mp; 5642 ufs_lbn_t lbn; 5643 5644 mp = UFSTOVFS(ip->i_ump); 5645 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5646 ("softdep_setup_allocext called on non-softdep filesystem")); 5647 KASSERT(off < NXADDR, ("softdep_setup_allocext: lbn %lld > NXADDR", 5648 (long long)off)); 5649 5650 lbn = bp->b_lblkno; 5651 if (oldblkno && oldblkno != newblkno) 5652 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn); 5653 else 5654 freefrag = NULL; 5655 5656 ACQUIRE_LOCK(ip->i_ump); 5657 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0) 5658 panic("softdep_setup_allocext: lost block"); 5659 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5660 ("softdep_setup_allocext: newblk already initialized")); 5661 /* 5662 * Convert the newblk to an allocdirect. 5663 */ 5664 WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT); 5665 adp = (struct allocdirect *)newblk; 5666 newblk->nb_freefrag = freefrag; 5667 adp->ad_offset = off; 5668 adp->ad_oldblkno = oldblkno; 5669 adp->ad_newsize = newsize; 5670 adp->ad_oldsize = oldsize; 5671 adp->ad_state |= EXTDATA; 5672 5673 /* 5674 * Finish initializing the journal. 5675 */ 5676 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5677 jnewblk->jn_ino = ip->i_number; 5678 jnewblk->jn_lbn = lbn; 5679 add_to_journal(&jnewblk->jn_list); 5680 } 5681 if (freefrag && freefrag->ff_jdep != NULL && 5682 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5683 add_to_journal(freefrag->ff_jdep); 5684 inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep); 5685 adp->ad_inodedep = inodedep; 5686 5687 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list); 5688 /* 5689 * The list of allocdirects must be kept in sorted and ascending 5690 * order so that the rollback routines can quickly determine the 5691 * first uncommitted block (the size of the file stored on disk 5692 * ends at the end of the lowest committed fragment, or if there 5693 * are no fragments, at the end of the highest committed block). 5694 * Since files generally grow, the typical case is that the new 5695 * block is to be added at the end of the list. We speed this 5696 * special case by checking against the last allocdirect in the 5697 * list before laboriously traversing the list looking for the 5698 * insertion point. 5699 */ 5700 adphead = &inodedep->id_newextupdt; 5701 oldadp = TAILQ_LAST(adphead, allocdirectlst); 5702 if (oldadp == NULL || oldadp->ad_offset <= off) { 5703 /* insert at end of list */ 5704 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 5705 if (oldadp != NULL && oldadp->ad_offset == off) 5706 allocdirect_merge(adphead, adp, oldadp); 5707 FREE_LOCK(ip->i_ump); 5708 return; 5709 } 5710 TAILQ_FOREACH(oldadp, adphead, ad_next) { 5711 if (oldadp->ad_offset >= off) 5712 break; 5713 } 5714 if (oldadp == NULL) 5715 panic("softdep_setup_allocext: lost entry"); 5716 /* insert in middle of list */ 5717 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 5718 if (oldadp->ad_offset == off) 5719 allocdirect_merge(adphead, adp, oldadp); 5720 FREE_LOCK(ip->i_ump); 5721} 5722 5723/* 5724 * Indirect block allocation dependencies. 5725 * 5726 * The same dependencies that exist for a direct block also exist when 5727 * a new block is allocated and pointed to by an entry in a block of 5728 * indirect pointers. The undo/redo states described above are also 5729 * used here. Because an indirect block contains many pointers that 5730 * may have dependencies, a second copy of the entire in-memory indirect 5731 * block is kept. The buffer cache copy is always completely up-to-date. 5732 * The second copy, which is used only as a source for disk writes, 5733 * contains only the safe pointers (i.e., those that have no remaining 5734 * update dependencies). The second copy is freed when all pointers 5735 * are safe. The cache is not allowed to replace indirect blocks with 5736 * pending update dependencies. If a buffer containing an indirect 5737 * block with dependencies is written, these routines will mark it 5738 * dirty again. It can only be successfully written once all the 5739 * dependencies are removed. The ffs_fsync routine in conjunction with 5740 * softdep_sync_metadata work together to get all the dependencies 5741 * removed so that a file can be successfully written to disk. Three 5742 * procedures are used when setting up indirect block pointer 5743 * dependencies. The division is necessary because of the organization 5744 * of the "balloc" routine and because of the distinction between file 5745 * pages and file metadata blocks. 5746 */ 5747 5748/* 5749 * Allocate a new allocindir structure. 5750 */ 5751static struct allocindir * 5752newallocindir(ip, ptrno, newblkno, oldblkno, lbn) 5753 struct inode *ip; /* inode for file being extended */ 5754 int ptrno; /* offset of pointer in indirect block */ 5755 ufs2_daddr_t newblkno; /* disk block number being added */ 5756 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 5757 ufs_lbn_t lbn; 5758{ 5759 struct newblk *newblk; 5760 struct allocindir *aip; 5761 struct freefrag *freefrag; 5762 struct jnewblk *jnewblk; 5763 5764 if (oldblkno) 5765 freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize, lbn); 5766 else 5767 freefrag = NULL; 5768 ACQUIRE_LOCK(ip->i_ump); 5769 if (newblk_lookup(UFSTOVFS(ip->i_ump), newblkno, 0, &newblk) == 0) 5770 panic("new_allocindir: lost block"); 5771 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5772 ("newallocindir: newblk already initialized")); 5773 WORKITEM_REASSIGN(newblk, D_ALLOCINDIR); 5774 newblk->nb_freefrag = freefrag; 5775 aip = (struct allocindir *)newblk; 5776 aip->ai_offset = ptrno; 5777 aip->ai_oldblkno = oldblkno; 5778 aip->ai_lbn = lbn; 5779 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5780 jnewblk->jn_ino = ip->i_number; 5781 jnewblk->jn_lbn = lbn; 5782 add_to_journal(&jnewblk->jn_list); 5783 } 5784 if (freefrag && freefrag->ff_jdep != NULL && 5785 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5786 add_to_journal(freefrag->ff_jdep); 5787 return (aip); 5788} 5789 5790/* 5791 * Called just before setting an indirect block pointer 5792 * to a newly allocated file page. 5793 */ 5794void 5795softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 5796 struct inode *ip; /* inode for file being extended */ 5797 ufs_lbn_t lbn; /* allocated block number within file */ 5798 struct buf *bp; /* buffer with indirect blk referencing page */ 5799 int ptrno; /* offset of pointer in indirect block */ 5800 ufs2_daddr_t newblkno; /* disk block number being added */ 5801 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 5802 struct buf *nbp; /* buffer holding allocated page */ 5803{ 5804 struct inodedep *inodedep; 5805 struct freefrag *freefrag; 5806 struct allocindir *aip; 5807 struct pagedep *pagedep; 5808 struct mount *mp; 5809 int dflags; 5810 5811 mp = UFSTOVFS(ip->i_ump); 5812 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5813 ("softdep_setup_allocindir_page called on non-softdep filesystem")); 5814 KASSERT(lbn == nbp->b_lblkno, 5815 ("softdep_setup_allocindir_page: lbn %jd != lblkno %jd", 5816 lbn, bp->b_lblkno)); 5817 CTR4(KTR_SUJ, 5818 "softdep_setup_allocindir_page: ino %d blkno %jd oldblkno %jd " 5819 "lbn %jd", ip->i_number, newblkno, oldblkno, lbn); 5820 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page"); 5821 aip = newallocindir(ip, ptrno, newblkno, oldblkno, lbn); 5822 dflags = DEPALLOC; 5823 if (IS_SNAPSHOT(ip)) 5824 dflags |= NODELAY; 5825 (void) inodedep_lookup(mp, ip->i_number, dflags, &inodedep); 5826 /* 5827 * If we are allocating a directory page, then we must 5828 * allocate an associated pagedep to track additions and 5829 * deletions. 5830 */ 5831 if ((ip->i_mode & IFMT) == IFDIR) 5832 pagedep_lookup(mp, nbp, ip->i_number, lbn, DEPALLOC, &pagedep); 5833 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list); 5834 freefrag = setup_allocindir_phase2(bp, ip, inodedep, aip, lbn); 5835 FREE_LOCK(ip->i_ump); 5836 if (freefrag) 5837 handle_workitem_freefrag(freefrag); 5838} 5839 5840/* 5841 * Called just before setting an indirect block pointer to a 5842 * newly allocated indirect block. 5843 */ 5844void 5845softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 5846 struct buf *nbp; /* newly allocated indirect block */ 5847 struct inode *ip; /* inode for file being extended */ 5848 struct buf *bp; /* indirect block referencing allocated block */ 5849 int ptrno; /* offset of pointer in indirect block */ 5850 ufs2_daddr_t newblkno; /* disk block number being added */ 5851{ 5852 struct inodedep *inodedep; 5853 struct allocindir *aip; 5854 ufs_lbn_t lbn; 5855 int dflags; 5856 5857 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0, 5858 ("softdep_setup_allocindir_meta called on non-softdep filesystem")); 5859 CTR3(KTR_SUJ, 5860 "softdep_setup_allocindir_meta: ino %d blkno %jd ptrno %d", 5861 ip->i_number, newblkno, ptrno); 5862 lbn = nbp->b_lblkno; 5863 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta"); 5864 aip = newallocindir(ip, ptrno, newblkno, 0, lbn); 5865 dflags = DEPALLOC; 5866 if (IS_SNAPSHOT(ip)) 5867 dflags |= NODELAY; 5868 inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, dflags, &inodedep); 5869 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list); 5870 if (setup_allocindir_phase2(bp, ip, inodedep, aip, lbn)) 5871 panic("softdep_setup_allocindir_meta: Block already existed"); 5872 FREE_LOCK(ip->i_ump); 5873} 5874 5875static void 5876indirdep_complete(indirdep) 5877 struct indirdep *indirdep; 5878{ 5879 struct allocindir *aip; 5880 5881 LIST_REMOVE(indirdep, ir_next); 5882 indirdep->ir_state |= DEPCOMPLETE; 5883 5884 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) { 5885 LIST_REMOVE(aip, ai_next); 5886 free_newblk(&aip->ai_block); 5887 } 5888 /* 5889 * If this indirdep is not attached to a buf it was simply waiting 5890 * on completion to clear completehd. free_indirdep() asserts 5891 * that nothing is dangling. 5892 */ 5893 if ((indirdep->ir_state & ONWORKLIST) == 0) 5894 free_indirdep(indirdep); 5895} 5896 5897static struct indirdep * 5898indirdep_lookup(mp, ip, bp) 5899 struct mount *mp; 5900 struct inode *ip; 5901 struct buf *bp; 5902{ 5903 struct indirdep *indirdep, *newindirdep; 5904 struct newblk *newblk; 5905 struct ufsmount *ump; 5906 struct worklist *wk; 5907 struct fs *fs; 5908 ufs2_daddr_t blkno; 5909 5910 ump = VFSTOUFS(mp); 5911 LOCK_OWNED(ump); 5912 indirdep = NULL; 5913 newindirdep = NULL; 5914 fs = ip->i_fs; 5915 for (;;) { 5916 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5917 if (wk->wk_type != D_INDIRDEP) 5918 continue; 5919 indirdep = WK_INDIRDEP(wk); 5920 break; 5921 } 5922 /* Found on the buffer worklist, no new structure to free. */ 5923 if (indirdep != NULL && newindirdep == NULL) 5924 return (indirdep); 5925 if (indirdep != NULL && newindirdep != NULL) 5926 panic("indirdep_lookup: simultaneous create"); 5927 /* None found on the buffer and a new structure is ready. */ 5928 if (indirdep == NULL && newindirdep != NULL) 5929 break; 5930 /* None found and no new structure available. */ 5931 FREE_LOCK(ump); 5932 newindirdep = malloc(sizeof(struct indirdep), 5933 M_INDIRDEP, M_SOFTDEP_FLAGS); 5934 workitem_alloc(&newindirdep->ir_list, D_INDIRDEP, mp); 5935 newindirdep->ir_state = ATTACHED; 5936 if (ip->i_ump->um_fstype == UFS1) 5937 newindirdep->ir_state |= UFS1FMT; 5938 TAILQ_INIT(&newindirdep->ir_trunc); 5939 newindirdep->ir_saveddata = NULL; 5940 LIST_INIT(&newindirdep->ir_deplisthd); 5941 LIST_INIT(&newindirdep->ir_donehd); 5942 LIST_INIT(&newindirdep->ir_writehd); 5943 LIST_INIT(&newindirdep->ir_completehd); 5944 if (bp->b_blkno == bp->b_lblkno) { 5945 ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp, 5946 NULL, NULL); 5947 bp->b_blkno = blkno; 5948 } 5949 newindirdep->ir_freeblks = NULL; 5950 newindirdep->ir_savebp = 5951 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0); 5952 newindirdep->ir_bp = bp; 5953 BUF_KERNPROC(newindirdep->ir_savebp); 5954 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 5955 ACQUIRE_LOCK(ump); 5956 } 5957 indirdep = newindirdep; 5958 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 5959 /* 5960 * If the block is not yet allocated we don't set DEPCOMPLETE so 5961 * that we don't free dependencies until the pointers are valid. 5962 * This could search b_dep for D_ALLOCDIRECT/D_ALLOCINDIR rather 5963 * than using the hash. 5964 */ 5965 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk)) 5966 LIST_INSERT_HEAD(&newblk->nb_indirdeps, indirdep, ir_next); 5967 else 5968 indirdep->ir_state |= DEPCOMPLETE; 5969 return (indirdep); 5970} 5971 5972/* 5973 * Called to finish the allocation of the "aip" allocated 5974 * by one of the two routines above. 5975 */ 5976static struct freefrag * 5977setup_allocindir_phase2(bp, ip, inodedep, aip, lbn) 5978 struct buf *bp; /* in-memory copy of the indirect block */ 5979 struct inode *ip; /* inode for file being extended */ 5980 struct inodedep *inodedep; /* Inodedep for ip */ 5981 struct allocindir *aip; /* allocindir allocated by the above routines */ 5982 ufs_lbn_t lbn; /* Logical block number for this block. */ 5983{ 5984 struct fs *fs; 5985 struct indirdep *indirdep; 5986 struct allocindir *oldaip; 5987 struct freefrag *freefrag; 5988 struct mount *mp; 5989 5990 LOCK_OWNED(ip->i_ump); 5991 mp = UFSTOVFS(ip->i_ump); 5992 fs = ip->i_fs; 5993 if (bp->b_lblkno >= 0) 5994 panic("setup_allocindir_phase2: not indir blk"); 5995 KASSERT(aip->ai_offset >= 0 && aip->ai_offset < NINDIR(fs), 5996 ("setup_allocindir_phase2: Bad offset %d", aip->ai_offset)); 5997 indirdep = indirdep_lookup(mp, ip, bp); 5998 KASSERT(indirdep->ir_savebp != NULL, 5999 ("setup_allocindir_phase2 NULL ir_savebp")); 6000 aip->ai_indirdep = indirdep; 6001 /* 6002 * Check for an unwritten dependency for this indirect offset. If 6003 * there is, merge the old dependency into the new one. This happens 6004 * as a result of reallocblk only. 6005 */ 6006 freefrag = NULL; 6007 if (aip->ai_oldblkno != 0) { 6008 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) { 6009 if (oldaip->ai_offset == aip->ai_offset) { 6010 freefrag = allocindir_merge(aip, oldaip); 6011 goto done; 6012 } 6013 } 6014 LIST_FOREACH(oldaip, &indirdep->ir_donehd, ai_next) { 6015 if (oldaip->ai_offset == aip->ai_offset) { 6016 freefrag = allocindir_merge(aip, oldaip); 6017 goto done; 6018 } 6019 } 6020 } 6021done: 6022 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 6023 return (freefrag); 6024} 6025 6026/* 6027 * Merge two allocindirs which refer to the same block. Move newblock 6028 * dependencies and setup the freefrags appropriately. 6029 */ 6030static struct freefrag * 6031allocindir_merge(aip, oldaip) 6032 struct allocindir *aip; 6033 struct allocindir *oldaip; 6034{ 6035 struct freefrag *freefrag; 6036 struct worklist *wk; 6037 6038 if (oldaip->ai_newblkno != aip->ai_oldblkno) 6039 panic("allocindir_merge: blkno"); 6040 aip->ai_oldblkno = oldaip->ai_oldblkno; 6041 freefrag = aip->ai_freefrag; 6042 aip->ai_freefrag = oldaip->ai_freefrag; 6043 oldaip->ai_freefrag = NULL; 6044 KASSERT(freefrag != NULL, ("setup_allocindir_phase2: No freefrag")); 6045 /* 6046 * If we are tracking a new directory-block allocation, 6047 * move it from the old allocindir to the new allocindir. 6048 */ 6049 if ((wk = LIST_FIRST(&oldaip->ai_newdirblk)) != NULL) { 6050 WORKLIST_REMOVE(wk); 6051 if (!LIST_EMPTY(&oldaip->ai_newdirblk)) 6052 panic("allocindir_merge: extra newdirblk"); 6053 WORKLIST_INSERT(&aip->ai_newdirblk, wk); 6054 } 6055 /* 6056 * We can skip journaling for this freefrag and just complete 6057 * any pending journal work for the allocindir that is being 6058 * removed after the freefrag completes. 6059 */ 6060 if (freefrag->ff_jdep) 6061 cancel_jfreefrag(WK_JFREEFRAG(freefrag->ff_jdep)); 6062 LIST_REMOVE(oldaip, ai_next); 6063 freefrag->ff_jdep = (struct worklist *)cancel_newblk(&oldaip->ai_block, 6064 &freefrag->ff_list, &freefrag->ff_jwork); 6065 free_newblk(&oldaip->ai_block); 6066 6067 return (freefrag); 6068} 6069 6070static inline void 6071setup_freedirect(freeblks, ip, i, needj) 6072 struct freeblks *freeblks; 6073 struct inode *ip; 6074 int i; 6075 int needj; 6076{ 6077 ufs2_daddr_t blkno; 6078 int frags; 6079 6080 blkno = DIP(ip, i_db[i]); 6081 if (blkno == 0) 6082 return; 6083 DIP_SET(ip, i_db[i], 0); 6084 frags = sblksize(ip->i_fs, ip->i_size, i); 6085 frags = numfrags(ip->i_fs, frags); 6086 newfreework(ip->i_ump, freeblks, NULL, i, blkno, frags, 0, needj); 6087} 6088 6089static inline void 6090setup_freeext(freeblks, ip, i, needj) 6091 struct freeblks *freeblks; 6092 struct inode *ip; 6093 int i; 6094 int needj; 6095{ 6096 ufs2_daddr_t blkno; 6097 int frags; 6098 6099 blkno = ip->i_din2->di_extb[i]; 6100 if (blkno == 0) 6101 return; 6102 ip->i_din2->di_extb[i] = 0; 6103 frags = sblksize(ip->i_fs, ip->i_din2->di_extsize, i); 6104 frags = numfrags(ip->i_fs, frags); 6105 newfreework(ip->i_ump, freeblks, NULL, -1 - i, blkno, frags, 0, needj); 6106} 6107 6108static inline void 6109setup_freeindir(freeblks, ip, i, lbn, needj) 6110 struct freeblks *freeblks; 6111 struct inode *ip; 6112 int i; 6113 ufs_lbn_t lbn; 6114 int needj; 6115{ 6116 ufs2_daddr_t blkno; 6117 6118 blkno = DIP(ip, i_ib[i]); 6119 if (blkno == 0) 6120 return; 6121 DIP_SET(ip, i_ib[i], 0); 6122 newfreework(ip->i_ump, freeblks, NULL, lbn, blkno, ip->i_fs->fs_frag, 6123 0, needj); 6124} 6125 6126static inline struct freeblks * 6127newfreeblks(mp, ip) 6128 struct mount *mp; 6129 struct inode *ip; 6130{ 6131 struct freeblks *freeblks; 6132 6133 freeblks = malloc(sizeof(struct freeblks), 6134 M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO); 6135 workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp); 6136 LIST_INIT(&freeblks->fb_jblkdephd); 6137 LIST_INIT(&freeblks->fb_jwork); 6138 freeblks->fb_ref = 0; 6139 freeblks->fb_cgwait = 0; 6140 freeblks->fb_state = ATTACHED; 6141 freeblks->fb_uid = ip->i_uid; 6142 freeblks->fb_inum = ip->i_number; 6143 freeblks->fb_vtype = ITOV(ip)->v_type; 6144 freeblks->fb_modrev = DIP(ip, i_modrev); 6145 freeblks->fb_devvp = ip->i_devvp; 6146 freeblks->fb_chkcnt = 0; 6147 freeblks->fb_len = 0; 6148 6149 return (freeblks); 6150} 6151 6152static void 6153trunc_indirdep(indirdep, freeblks, bp, off) 6154 struct indirdep *indirdep; 6155 struct freeblks *freeblks; 6156 struct buf *bp; 6157 int off; 6158{ 6159 struct allocindir *aip, *aipn; 6160 6161 /* 6162 * The first set of allocindirs won't be in savedbp. 6163 */ 6164 LIST_FOREACH_SAFE(aip, &indirdep->ir_deplisthd, ai_next, aipn) 6165 if (aip->ai_offset > off) 6166 cancel_allocindir(aip, bp, freeblks, 1); 6167 LIST_FOREACH_SAFE(aip, &indirdep->ir_donehd, ai_next, aipn) 6168 if (aip->ai_offset > off) 6169 cancel_allocindir(aip, bp, freeblks, 1); 6170 /* 6171 * These will exist in savedbp. 6172 */ 6173 LIST_FOREACH_SAFE(aip, &indirdep->ir_writehd, ai_next, aipn) 6174 if (aip->ai_offset > off) 6175 cancel_allocindir(aip, NULL, freeblks, 0); 6176 LIST_FOREACH_SAFE(aip, &indirdep->ir_completehd, ai_next, aipn) 6177 if (aip->ai_offset > off) 6178 cancel_allocindir(aip, NULL, freeblks, 0); 6179} 6180 6181/* 6182 * Follow the chain of indirects down to lastlbn creating a freework 6183 * structure for each. This will be used to start indir_trunc() at 6184 * the right offset and create the journal records for the parrtial 6185 * truncation. A second step will handle the truncated dependencies. 6186 */ 6187static int 6188setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno) 6189 struct freeblks *freeblks; 6190 struct inode *ip; 6191 ufs_lbn_t lbn; 6192 ufs_lbn_t lastlbn; 6193 ufs2_daddr_t blkno; 6194{ 6195 struct indirdep *indirdep; 6196 struct indirdep *indirn; 6197 struct freework *freework; 6198 struct newblk *newblk; 6199 struct mount *mp; 6200 struct buf *bp; 6201 uint8_t *start; 6202 uint8_t *end; 6203 ufs_lbn_t lbnadd; 6204 int level; 6205 int error; 6206 int off; 6207 6208 6209 freework = NULL; 6210 if (blkno == 0) 6211 return (0); 6212 mp = freeblks->fb_list.wk_mp; 6213 bp = getblk(ITOV(ip), lbn, mp->mnt_stat.f_iosize, 0, 0, 0); 6214 if ((bp->b_flags & B_CACHE) == 0) { 6215 bp->b_blkno = blkptrtodb(VFSTOUFS(mp), blkno); 6216 bp->b_iocmd = BIO_READ; 6217 bp->b_flags &= ~B_INVAL; 6218 bp->b_ioflags &= ~BIO_ERROR; 6219 vfs_busy_pages(bp, 0); 6220 bp->b_iooffset = dbtob(bp->b_blkno); 6221 bstrategy(bp); 6222 curthread->td_ru.ru_inblock++; 6223 error = bufwait(bp); 6224 if (error) { 6225 brelse(bp); 6226 return (error); 6227 } 6228 } 6229 level = lbn_level(lbn); 6230 lbnadd = lbn_offset(ip->i_fs, level); 6231 /* 6232 * Compute the offset of the last block we want to keep. Store 6233 * in the freework the first block we want to completely free. 6234 */ 6235 off = (lastlbn - -(lbn + level)) / lbnadd; 6236 if (off + 1 == NINDIR(ip->i_fs)) 6237 goto nowork; 6238 freework = newfreework(ip->i_ump, freeblks, NULL, lbn, blkno, 0, off+1, 6239 0); 6240 /* 6241 * Link the freework into the indirdep. This will prevent any new 6242 * allocations from proceeding until we are finished with the 6243 * truncate and the block is written. 6244 */ 6245 ACQUIRE_LOCK(ip->i_ump); 6246 indirdep = indirdep_lookup(mp, ip, bp); 6247 if (indirdep->ir_freeblks) 6248 panic("setup_trunc_indir: indirdep already truncated."); 6249 TAILQ_INSERT_TAIL(&indirdep->ir_trunc, freework, fw_next); 6250 freework->fw_indir = indirdep; 6251 /* 6252 * Cancel any allocindirs that will not make it to disk. 6253 * We have to do this for all copies of the indirdep that 6254 * live on this newblk. 6255 */ 6256 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 6257 newblk_lookup(mp, dbtofsb(ip->i_fs, bp->b_blkno), 0, &newblk); 6258 LIST_FOREACH(indirn, &newblk->nb_indirdeps, ir_next) 6259 trunc_indirdep(indirn, freeblks, bp, off); 6260 } else 6261 trunc_indirdep(indirdep, freeblks, bp, off); 6262 FREE_LOCK(ip->i_ump); 6263 /* 6264 * Creation is protected by the buf lock. The saveddata is only 6265 * needed if a full truncation follows a partial truncation but it 6266 * is difficult to allocate in that case so we fetch it anyway. 6267 */ 6268 if (indirdep->ir_saveddata == NULL) 6269 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP, 6270 M_SOFTDEP_FLAGS); 6271nowork: 6272 /* Fetch the blkno of the child and the zero start offset. */ 6273 if (ip->i_ump->um_fstype == UFS1) { 6274 blkno = ((ufs1_daddr_t *)bp->b_data)[off]; 6275 start = (uint8_t *)&((ufs1_daddr_t *)bp->b_data)[off+1]; 6276 } else { 6277 blkno = ((ufs2_daddr_t *)bp->b_data)[off]; 6278 start = (uint8_t *)&((ufs2_daddr_t *)bp->b_data)[off+1]; 6279 } 6280 if (freework) { 6281 /* Zero the truncated pointers. */ 6282 end = bp->b_data + bp->b_bcount; 6283 bzero(start, end - start); 6284 bdwrite(bp); 6285 } else 6286 bqrelse(bp); 6287 if (level == 0) 6288 return (0); 6289 lbn++; /* adjust level */ 6290 lbn -= (off * lbnadd); 6291 return setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno); 6292} 6293 6294/* 6295 * Complete the partial truncation of an indirect block setup by 6296 * setup_trunc_indir(). This zeros the truncated pointers in the saved 6297 * copy and writes them to disk before the freeblks is allowed to complete. 6298 */ 6299static void 6300complete_trunc_indir(freework) 6301 struct freework *freework; 6302{ 6303 struct freework *fwn; 6304 struct indirdep *indirdep; 6305 struct ufsmount *ump; 6306 struct buf *bp; 6307 uintptr_t start; 6308 int count; 6309 6310 ump = VFSTOUFS(freework->fw_list.wk_mp); 6311 LOCK_OWNED(ump); 6312 indirdep = freework->fw_indir; 6313 for (;;) { 6314 bp = indirdep->ir_bp; 6315 /* See if the block was discarded. */ 6316 if (bp == NULL) 6317 break; 6318 /* Inline part of getdirtybuf(). We dont want bremfree. */ 6319 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0) 6320 break; 6321 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 6322 LOCK_PTR(ump)) == 0) 6323 BUF_UNLOCK(bp); 6324 ACQUIRE_LOCK(ump); 6325 } 6326 freework->fw_state |= DEPCOMPLETE; 6327 TAILQ_REMOVE(&indirdep->ir_trunc, freework, fw_next); 6328 /* 6329 * Zero the pointers in the saved copy. 6330 */ 6331 if (indirdep->ir_state & UFS1FMT) 6332 start = sizeof(ufs1_daddr_t); 6333 else 6334 start = sizeof(ufs2_daddr_t); 6335 start *= freework->fw_start; 6336 count = indirdep->ir_savebp->b_bcount - start; 6337 start += (uintptr_t)indirdep->ir_savebp->b_data; 6338 bzero((char *)start, count); 6339 /* 6340 * We need to start the next truncation in the list if it has not 6341 * been started yet. 6342 */ 6343 fwn = TAILQ_FIRST(&indirdep->ir_trunc); 6344 if (fwn != NULL) { 6345 if (fwn->fw_freeblks == indirdep->ir_freeblks) 6346 TAILQ_REMOVE(&indirdep->ir_trunc, fwn, fw_next); 6347 if ((fwn->fw_state & ONWORKLIST) == 0) 6348 freework_enqueue(fwn); 6349 } 6350 /* 6351 * If bp is NULL the block was fully truncated, restore 6352 * the saved block list otherwise free it if it is no 6353 * longer needed. 6354 */ 6355 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 6356 if (bp == NULL) 6357 bcopy(indirdep->ir_saveddata, 6358 indirdep->ir_savebp->b_data, 6359 indirdep->ir_savebp->b_bcount); 6360 free(indirdep->ir_saveddata, M_INDIRDEP); 6361 indirdep->ir_saveddata = NULL; 6362 } 6363 /* 6364 * When bp is NULL there is a full truncation pending. We 6365 * must wait for this full truncation to be journaled before 6366 * we can release this freework because the disk pointers will 6367 * never be written as zero. 6368 */ 6369 if (bp == NULL) { 6370 if (LIST_EMPTY(&indirdep->ir_freeblks->fb_jblkdephd)) 6371 handle_written_freework(freework); 6372 else 6373 WORKLIST_INSERT(&indirdep->ir_freeblks->fb_freeworkhd, 6374 &freework->fw_list); 6375 } else { 6376 /* Complete when the real copy is written. */ 6377 WORKLIST_INSERT(&bp->b_dep, &freework->fw_list); 6378 BUF_UNLOCK(bp); 6379 } 6380} 6381 6382/* 6383 * Calculate the number of blocks we are going to release where datablocks 6384 * is the current total and length is the new file size. 6385 */ 6386static ufs2_daddr_t 6387blkcount(fs, datablocks, length) 6388 struct fs *fs; 6389 ufs2_daddr_t datablocks; 6390 off_t length; 6391{ 6392 off_t totblks, numblks; 6393 6394 totblks = 0; 6395 numblks = howmany(length, fs->fs_bsize); 6396 if (numblks <= NDADDR) { 6397 totblks = howmany(length, fs->fs_fsize); 6398 goto out; 6399 } 6400 totblks = blkstofrags(fs, numblks); 6401 numblks -= NDADDR; 6402 /* 6403 * Count all single, then double, then triple indirects required. 6404 * Subtracting one indirects worth of blocks for each pass 6405 * acknowledges one of each pointed to by the inode. 6406 */ 6407 for (;;) { 6408 totblks += blkstofrags(fs, howmany(numblks, NINDIR(fs))); 6409 numblks -= NINDIR(fs); 6410 if (numblks <= 0) 6411 break; 6412 numblks = howmany(numblks, NINDIR(fs)); 6413 } 6414out: 6415 totblks = fsbtodb(fs, totblks); 6416 /* 6417 * Handle sparse files. We can't reclaim more blocks than the inode 6418 * references. We will correct it later in handle_complete_freeblks() 6419 * when we know the real count. 6420 */ 6421 if (totblks > datablocks) 6422 return (0); 6423 return (datablocks - totblks); 6424} 6425 6426/* 6427 * Handle freeblocks for journaled softupdate filesystems. 6428 * 6429 * Contrary to normal softupdates, we must preserve the block pointers in 6430 * indirects until their subordinates are free. This is to avoid journaling 6431 * every block that is freed which may consume more space than the journal 6432 * itself. The recovery program will see the free block journals at the 6433 * base of the truncated area and traverse them to reclaim space. The 6434 * pointers in the inode may be cleared immediately after the journal 6435 * records are written because each direct and indirect pointer in the 6436 * inode is recorded in a journal. This permits full truncation to proceed 6437 * asynchronously. The write order is journal -> inode -> cgs -> indirects. 6438 * 6439 * The algorithm is as follows: 6440 * 1) Traverse the in-memory state and create journal entries to release 6441 * the relevant blocks and full indirect trees. 6442 * 2) Traverse the indirect block chain adding partial truncation freework 6443 * records to indirects in the path to lastlbn. The freework will 6444 * prevent new allocation dependencies from being satisfied in this 6445 * indirect until the truncation completes. 6446 * 3) Read and lock the inode block, performing an update with the new size 6447 * and pointers. This prevents truncated data from becoming valid on 6448 * disk through step 4. 6449 * 4) Reap unsatisfied dependencies that are beyond the truncated area, 6450 * eliminate journal work for those records that do not require it. 6451 * 5) Schedule the journal records to be written followed by the inode block. 6452 * 6) Allocate any necessary frags for the end of file. 6453 * 7) Zero any partially truncated blocks. 6454 * 6455 * From this truncation proceeds asynchronously using the freework and 6456 * indir_trunc machinery. The file will not be extended again into a 6457 * partially truncated indirect block until all work is completed but 6458 * the normal dependency mechanism ensures that it is rolled back/forward 6459 * as appropriate. Further truncation may occur without delay and is 6460 * serialized in indir_trunc(). 6461 */ 6462void 6463softdep_journal_freeblocks(ip, cred, length, flags) 6464 struct inode *ip; /* The inode whose length is to be reduced */ 6465 struct ucred *cred; 6466 off_t length; /* The new length for the file */ 6467 int flags; /* IO_EXT and/or IO_NORMAL */ 6468{ 6469 struct freeblks *freeblks, *fbn; 6470 struct worklist *wk, *wkn; 6471 struct inodedep *inodedep; 6472 struct jblkdep *jblkdep; 6473 struct allocdirect *adp, *adpn; 6474 struct ufsmount *ump; 6475 struct fs *fs; 6476 struct buf *bp; 6477 struct vnode *vp; 6478 struct mount *mp; 6479 ufs2_daddr_t extblocks, datablocks; 6480 ufs_lbn_t tmpval, lbn, lastlbn; 6481 int frags, lastoff, iboff, allocblock, needj, dflags, error, i; 6482 6483 fs = ip->i_fs; 6484 ump = ip->i_ump; 6485 mp = UFSTOVFS(ump); 6486 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 6487 ("softdep_journal_freeblocks called on non-softdep filesystem")); 6488 vp = ITOV(ip); 6489 needj = 1; 6490 iboff = -1; 6491 allocblock = 0; 6492 extblocks = 0; 6493 datablocks = 0; 6494 frags = 0; 6495 freeblks = newfreeblks(mp, ip); 6496 ACQUIRE_LOCK(ump); 6497 /* 6498 * If we're truncating a removed file that will never be written 6499 * we don't need to journal the block frees. The canceled journals 6500 * for the allocations will suffice. 6501 */ 6502 dflags = DEPALLOC; 6503 if (IS_SNAPSHOT(ip)) 6504 dflags |= NODELAY; 6505 inodedep_lookup(mp, ip->i_number, dflags, &inodedep); 6506 if ((inodedep->id_state & (UNLINKED | DEPCOMPLETE)) == UNLINKED && 6507 length == 0) 6508 needj = 0; 6509 CTR3(KTR_SUJ, "softdep_journal_freeblks: ip %d length %ld needj %d", 6510 ip->i_number, length, needj); 6511 FREE_LOCK(ump); 6512 /* 6513 * Calculate the lbn that we are truncating to. This results in -1 6514 * if we're truncating the 0 bytes. So it is the last lbn we want 6515 * to keep, not the first lbn we want to truncate. 6516 */ 6517 lastlbn = lblkno(fs, length + fs->fs_bsize - 1) - 1; 6518 lastoff = blkoff(fs, length); 6519 /* 6520 * Compute frags we are keeping in lastlbn. 0 means all. 6521 */ 6522 if (lastlbn >= 0 && lastlbn < NDADDR) { 6523 frags = fragroundup(fs, lastoff); 6524 /* adp offset of last valid allocdirect. */ 6525 iboff = lastlbn; 6526 } else if (lastlbn > 0) 6527 iboff = NDADDR; 6528 if (fs->fs_magic == FS_UFS2_MAGIC) 6529 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 6530 /* 6531 * Handle normal data blocks and indirects. This section saves 6532 * values used after the inode update to complete frag and indirect 6533 * truncation. 6534 */ 6535 if ((flags & IO_NORMAL) != 0) { 6536 /* 6537 * Handle truncation of whole direct and indirect blocks. 6538 */ 6539 for (i = iboff + 1; i < NDADDR; i++) 6540 setup_freedirect(freeblks, ip, i, needj); 6541 for (i = 0, tmpval = NINDIR(fs), lbn = NDADDR; i < NIADDR; 6542 i++, lbn += tmpval, tmpval *= NINDIR(fs)) { 6543 /* Release a whole indirect tree. */ 6544 if (lbn > lastlbn) { 6545 setup_freeindir(freeblks, ip, i, -lbn -i, 6546 needj); 6547 continue; 6548 } 6549 iboff = i + NDADDR; 6550 /* 6551 * Traverse partially truncated indirect tree. 6552 */ 6553 if (lbn <= lastlbn && lbn + tmpval - 1 > lastlbn) 6554 setup_trunc_indir(freeblks, ip, -lbn - i, 6555 lastlbn, DIP(ip, i_ib[i])); 6556 } 6557 /* 6558 * Handle partial truncation to a frag boundary. 6559 */ 6560 if (frags) { 6561 ufs2_daddr_t blkno; 6562 long oldfrags; 6563 6564 oldfrags = blksize(fs, ip, lastlbn); 6565 blkno = DIP(ip, i_db[lastlbn]); 6566 if (blkno && oldfrags != frags) { 6567 oldfrags -= frags; 6568 oldfrags = numfrags(ip->i_fs, oldfrags); 6569 blkno += numfrags(ip->i_fs, frags); 6570 newfreework(ump, freeblks, NULL, lastlbn, 6571 blkno, oldfrags, 0, needj); 6572 if (needj) 6573 adjust_newfreework(freeblks, 6574 numfrags(ip->i_fs, frags)); 6575 } else if (blkno == 0) 6576 allocblock = 1; 6577 } 6578 /* 6579 * Add a journal record for partial truncate if we are 6580 * handling indirect blocks. Non-indirects need no extra 6581 * journaling. 6582 */ 6583 if (length != 0 && lastlbn >= NDADDR) { 6584 ip->i_flag |= IN_TRUNCATED; 6585 newjtrunc(freeblks, length, 0); 6586 } 6587 ip->i_size = length; 6588 DIP_SET(ip, i_size, ip->i_size); 6589 datablocks = DIP(ip, i_blocks) - extblocks; 6590 if (length != 0) 6591 datablocks = blkcount(ip->i_fs, datablocks, length); 6592 freeblks->fb_len = length; 6593 } 6594 if ((flags & IO_EXT) != 0) { 6595 for (i = 0; i < NXADDR; i++) 6596 setup_freeext(freeblks, ip, i, needj); 6597 ip->i_din2->di_extsize = 0; 6598 datablocks += extblocks; 6599 } 6600#ifdef QUOTA 6601 /* Reference the quotas in case the block count is wrong in the end. */ 6602 quotaref(vp, freeblks->fb_quota); 6603 (void) chkdq(ip, -datablocks, NOCRED, 0); 6604#endif 6605 freeblks->fb_chkcnt = -datablocks; 6606 UFS_LOCK(ump); 6607 fs->fs_pendingblocks += datablocks; 6608 UFS_UNLOCK(ump); 6609 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks); 6610 /* 6611 * Handle truncation of incomplete alloc direct dependencies. We 6612 * hold the inode block locked to prevent incomplete dependencies 6613 * from reaching the disk while we are eliminating those that 6614 * have been truncated. This is a partially inlined ffs_update(). 6615 */ 6616 ufs_itimes(vp); 6617 ip->i_flag &= ~(IN_LAZYACCESS | IN_LAZYMOD | IN_MODIFIED); 6618 error = bread(ip->i_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 6619 (int)fs->fs_bsize, cred, &bp); 6620 if (error) { 6621 brelse(bp); 6622 softdep_error("softdep_journal_freeblocks", error); 6623 return; 6624 } 6625 if (bp->b_bufsize == fs->fs_bsize) 6626 bp->b_flags |= B_CLUSTEROK; 6627 softdep_update_inodeblock(ip, bp, 0); 6628 if (ump->um_fstype == UFS1) 6629 *((struct ufs1_dinode *)bp->b_data + 6630 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1; 6631 else 6632 *((struct ufs2_dinode *)bp->b_data + 6633 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2; 6634 ACQUIRE_LOCK(ump); 6635 (void) inodedep_lookup(mp, ip->i_number, dflags, &inodedep); 6636 if ((inodedep->id_state & IOSTARTED) != 0) 6637 panic("softdep_setup_freeblocks: inode busy"); 6638 /* 6639 * Add the freeblks structure to the list of operations that 6640 * must await the zero'ed inode being written to disk. If we 6641 * still have a bitmap dependency (needj), then the inode 6642 * has never been written to disk, so we can process the 6643 * freeblks below once we have deleted the dependencies. 6644 */ 6645 if (needj) 6646 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list); 6647 else 6648 freeblks->fb_state |= COMPLETE; 6649 if ((flags & IO_NORMAL) != 0) { 6650 TAILQ_FOREACH_SAFE(adp, &inodedep->id_inoupdt, ad_next, adpn) { 6651 if (adp->ad_offset > iboff) 6652 cancel_allocdirect(&inodedep->id_inoupdt, adp, 6653 freeblks); 6654 /* 6655 * Truncate the allocdirect. We could eliminate 6656 * or modify journal records as well. 6657 */ 6658 else if (adp->ad_offset == iboff && frags) 6659 adp->ad_newsize = frags; 6660 } 6661 } 6662 if ((flags & IO_EXT) != 0) 6663 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0) 6664 cancel_allocdirect(&inodedep->id_extupdt, adp, 6665 freeblks); 6666 /* 6667 * Scan the bufwait list for newblock dependencies that will never 6668 * make it to disk. 6669 */ 6670 LIST_FOREACH_SAFE(wk, &inodedep->id_bufwait, wk_list, wkn) { 6671 if (wk->wk_type != D_ALLOCDIRECT) 6672 continue; 6673 adp = WK_ALLOCDIRECT(wk); 6674 if (((flags & IO_NORMAL) != 0 && (adp->ad_offset > iboff)) || 6675 ((flags & IO_EXT) != 0 && (adp->ad_state & EXTDATA))) { 6676 cancel_jfreeblk(freeblks, adp->ad_newblkno); 6677 cancel_newblk(WK_NEWBLK(wk), NULL, &freeblks->fb_jwork); 6678 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk); 6679 } 6680 } 6681 /* 6682 * Add journal work. 6683 */ 6684 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) 6685 add_to_journal(&jblkdep->jb_list); 6686 FREE_LOCK(ump); 6687 bdwrite(bp); 6688 /* 6689 * Truncate dependency structures beyond length. 6690 */ 6691 trunc_dependencies(ip, freeblks, lastlbn, frags, flags); 6692 /* 6693 * This is only set when we need to allocate a fragment because 6694 * none existed at the end of a frag-sized file. It handles only 6695 * allocating a new, zero filled block. 6696 */ 6697 if (allocblock) { 6698 ip->i_size = length - lastoff; 6699 DIP_SET(ip, i_size, ip->i_size); 6700 error = UFS_BALLOC(vp, length - 1, 1, cred, BA_CLRBUF, &bp); 6701 if (error != 0) { 6702 softdep_error("softdep_journal_freeblks", error); 6703 return; 6704 } 6705 ip->i_size = length; 6706 DIP_SET(ip, i_size, length); 6707 ip->i_flag |= IN_CHANGE | IN_UPDATE; 6708 allocbuf(bp, frags); 6709 ffs_update(vp, 0); 6710 bawrite(bp); 6711 } else if (lastoff != 0 && vp->v_type != VDIR) { 6712 int size; 6713 6714 /* 6715 * Zero the end of a truncated frag or block. 6716 */ 6717 size = sblksize(fs, length, lastlbn); 6718 error = bread(vp, lastlbn, size, cred, &bp); 6719 if (error) { 6720 softdep_error("softdep_journal_freeblks", error); 6721 return; 6722 } 6723 bzero((char *)bp->b_data + lastoff, size - lastoff); 6724 bawrite(bp); 6725 6726 } 6727 ACQUIRE_LOCK(ump); 6728 inodedep_lookup(mp, ip->i_number, dflags, &inodedep); 6729 TAILQ_INSERT_TAIL(&inodedep->id_freeblklst, freeblks, fb_next); 6730 freeblks->fb_state |= DEPCOMPLETE | ONDEPLIST; 6731 /* 6732 * We zero earlier truncations so they don't erroneously 6733 * update i_blocks. 6734 */ 6735 if (freeblks->fb_len == 0 && (flags & IO_NORMAL) != 0) 6736 TAILQ_FOREACH(fbn, &inodedep->id_freeblklst, fb_next) 6737 fbn->fb_len = 0; 6738 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE && 6739 LIST_EMPTY(&freeblks->fb_jblkdephd)) 6740 freeblks->fb_state |= INPROGRESS; 6741 else 6742 freeblks = NULL; 6743 FREE_LOCK(ump); 6744 if (freeblks) 6745 handle_workitem_freeblocks(freeblks, 0); 6746 trunc_pages(ip, length, extblocks, flags); 6747 6748} 6749 6750/* 6751 * Flush a JOP_SYNC to the journal. 6752 */ 6753void 6754softdep_journal_fsync(ip) 6755 struct inode *ip; 6756{ 6757 struct jfsync *jfsync; 6758 6759 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0, 6760 ("softdep_journal_fsync called on non-softdep filesystem")); 6761 if ((ip->i_flag & IN_TRUNCATED) == 0) 6762 return; 6763 ip->i_flag &= ~IN_TRUNCATED; 6764 jfsync = malloc(sizeof(*jfsync), M_JFSYNC, M_SOFTDEP_FLAGS | M_ZERO); 6765 workitem_alloc(&jfsync->jfs_list, D_JFSYNC, UFSTOVFS(ip->i_ump)); 6766 jfsync->jfs_size = ip->i_size; 6767 jfsync->jfs_ino = ip->i_number; 6768 ACQUIRE_LOCK(ip->i_ump); 6769 add_to_journal(&jfsync->jfs_list); 6770 jwait(&jfsync->jfs_list, MNT_WAIT); 6771 FREE_LOCK(ip->i_ump); 6772} 6773 6774/* 6775 * Block de-allocation dependencies. 6776 * 6777 * When blocks are de-allocated, the on-disk pointers must be nullified before 6778 * the blocks are made available for use by other files. (The true 6779 * requirement is that old pointers must be nullified before new on-disk 6780 * pointers are set. We chose this slightly more stringent requirement to 6781 * reduce complexity.) Our implementation handles this dependency by updating 6782 * the inode (or indirect block) appropriately but delaying the actual block 6783 * de-allocation (i.e., freemap and free space count manipulation) until 6784 * after the updated versions reach stable storage. After the disk is 6785 * updated, the blocks can be safely de-allocated whenever it is convenient. 6786 * This implementation handles only the common case of reducing a file's 6787 * length to zero. Other cases are handled by the conventional synchronous 6788 * write approach. 6789 * 6790 * The ffs implementation with which we worked double-checks 6791 * the state of the block pointers and file size as it reduces 6792 * a file's length. Some of this code is replicated here in our 6793 * soft updates implementation. The freeblks->fb_chkcnt field is 6794 * used to transfer a part of this information to the procedure 6795 * that eventually de-allocates the blocks. 6796 * 6797 * This routine should be called from the routine that shortens 6798 * a file's length, before the inode's size or block pointers 6799 * are modified. It will save the block pointer information for 6800 * later release and zero the inode so that the calling routine 6801 * can release it. 6802 */ 6803void 6804softdep_setup_freeblocks(ip, length, flags) 6805 struct inode *ip; /* The inode whose length is to be reduced */ 6806 off_t length; /* The new length for the file */ 6807 int flags; /* IO_EXT and/or IO_NORMAL */ 6808{ 6809 struct ufs1_dinode *dp1; 6810 struct ufs2_dinode *dp2; 6811 struct freeblks *freeblks; 6812 struct inodedep *inodedep; 6813 struct allocdirect *adp; 6814 struct ufsmount *ump; 6815 struct buf *bp; 6816 struct fs *fs; 6817 ufs2_daddr_t extblocks, datablocks; 6818 struct mount *mp; 6819 int i, delay, error, dflags; 6820 ufs_lbn_t tmpval; 6821 ufs_lbn_t lbn; 6822 6823 ump = ip->i_ump; 6824 mp = UFSTOVFS(ump); 6825 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 6826 ("softdep_setup_freeblocks called on non-softdep filesystem")); 6827 CTR2(KTR_SUJ, "softdep_setup_freeblks: ip %d length %ld", 6828 ip->i_number, length); 6829 KASSERT(length == 0, ("softdep_setup_freeblocks: non-zero length")); 6830 fs = ip->i_fs; 6831 freeblks = newfreeblks(mp, ip); 6832 extblocks = 0; 6833 datablocks = 0; 6834 if (fs->fs_magic == FS_UFS2_MAGIC) 6835 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 6836 if ((flags & IO_NORMAL) != 0) { 6837 for (i = 0; i < NDADDR; i++) 6838 setup_freedirect(freeblks, ip, i, 0); 6839 for (i = 0, tmpval = NINDIR(fs), lbn = NDADDR; i < NIADDR; 6840 i++, lbn += tmpval, tmpval *= NINDIR(fs)) 6841 setup_freeindir(freeblks, ip, i, -lbn -i, 0); 6842 ip->i_size = 0; 6843 DIP_SET(ip, i_size, 0); 6844 datablocks = DIP(ip, i_blocks) - extblocks; 6845 } 6846 if ((flags & IO_EXT) != 0) { 6847 for (i = 0; i < NXADDR; i++) 6848 setup_freeext(freeblks, ip, i, 0); 6849 ip->i_din2->di_extsize = 0; 6850 datablocks += extblocks; 6851 } 6852#ifdef QUOTA 6853 /* Reference the quotas in case the block count is wrong in the end. */ 6854 quotaref(ITOV(ip), freeblks->fb_quota); 6855 (void) chkdq(ip, -datablocks, NOCRED, 0); 6856#endif 6857 freeblks->fb_chkcnt = -datablocks; 6858 UFS_LOCK(ump); 6859 fs->fs_pendingblocks += datablocks; 6860 UFS_UNLOCK(ump); 6861 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks); 6862 /* 6863 * Push the zero'ed inode to to its disk buffer so that we are free 6864 * to delete its dependencies below. Once the dependencies are gone 6865 * the buffer can be safely released. 6866 */ 6867 if ((error = bread(ip->i_devvp, 6868 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 6869 (int)fs->fs_bsize, NOCRED, &bp)) != 0) { 6870 brelse(bp); 6871 softdep_error("softdep_setup_freeblocks", error); 6872 } 6873 if (ump->um_fstype == UFS1) { 6874 dp1 = ((struct ufs1_dinode *)bp->b_data + 6875 ino_to_fsbo(fs, ip->i_number)); 6876 ip->i_din1->di_freelink = dp1->di_freelink; 6877 *dp1 = *ip->i_din1; 6878 } else { 6879 dp2 = ((struct ufs2_dinode *)bp->b_data + 6880 ino_to_fsbo(fs, ip->i_number)); 6881 ip->i_din2->di_freelink = dp2->di_freelink; 6882 *dp2 = *ip->i_din2; 6883 } 6884 /* 6885 * Find and eliminate any inode dependencies. 6886 */ 6887 ACQUIRE_LOCK(ump); 6888 dflags = DEPALLOC; 6889 if (IS_SNAPSHOT(ip)) 6890 dflags |= NODELAY; 6891 (void) inodedep_lookup(mp, ip->i_number, dflags, &inodedep); 6892 if ((inodedep->id_state & IOSTARTED) != 0) 6893 panic("softdep_setup_freeblocks: inode busy"); 6894 /* 6895 * Add the freeblks structure to the list of operations that 6896 * must await the zero'ed inode being written to disk. If we 6897 * still have a bitmap dependency (delay == 0), then the inode 6898 * has never been written to disk, so we can process the 6899 * freeblks below once we have deleted the dependencies. 6900 */ 6901 delay = (inodedep->id_state & DEPCOMPLETE); 6902 if (delay) 6903 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list); 6904 else 6905 freeblks->fb_state |= COMPLETE; 6906 /* 6907 * Because the file length has been truncated to zero, any 6908 * pending block allocation dependency structures associated 6909 * with this inode are obsolete and can simply be de-allocated. 6910 * We must first merge the two dependency lists to get rid of 6911 * any duplicate freefrag structures, then purge the merged list. 6912 * If we still have a bitmap dependency, then the inode has never 6913 * been written to disk, so we can free any fragments without delay. 6914 */ 6915 if (flags & IO_NORMAL) { 6916 merge_inode_lists(&inodedep->id_newinoupdt, 6917 &inodedep->id_inoupdt); 6918 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 6919 cancel_allocdirect(&inodedep->id_inoupdt, adp, 6920 freeblks); 6921 } 6922 if (flags & IO_EXT) { 6923 merge_inode_lists(&inodedep->id_newextupdt, 6924 &inodedep->id_extupdt); 6925 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0) 6926 cancel_allocdirect(&inodedep->id_extupdt, adp, 6927 freeblks); 6928 } 6929 FREE_LOCK(ump); 6930 bdwrite(bp); 6931 trunc_dependencies(ip, freeblks, -1, 0, flags); 6932 ACQUIRE_LOCK(ump); 6933 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 6934 (void) free_inodedep(inodedep); 6935 freeblks->fb_state |= DEPCOMPLETE; 6936 /* 6937 * If the inode with zeroed block pointers is now on disk 6938 * we can start freeing blocks. 6939 */ 6940 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 6941 freeblks->fb_state |= INPROGRESS; 6942 else 6943 freeblks = NULL; 6944 FREE_LOCK(ump); 6945 if (freeblks) 6946 handle_workitem_freeblocks(freeblks, 0); 6947 trunc_pages(ip, length, extblocks, flags); 6948} 6949 6950/* 6951 * Eliminate pages from the page cache that back parts of this inode and 6952 * adjust the vnode pager's idea of our size. This prevents stale data 6953 * from hanging around in the page cache. 6954 */ 6955static void 6956trunc_pages(ip, length, extblocks, flags) 6957 struct inode *ip; 6958 off_t length; 6959 ufs2_daddr_t extblocks; 6960 int flags; 6961{ 6962 struct vnode *vp; 6963 struct fs *fs; 6964 ufs_lbn_t lbn; 6965 off_t end, extend; 6966 6967 vp = ITOV(ip); 6968 fs = ip->i_fs; 6969 extend = OFF_TO_IDX(lblktosize(fs, -extblocks)); 6970 if ((flags & IO_EXT) != 0) 6971 vn_pages_remove(vp, extend, 0); 6972 if ((flags & IO_NORMAL) == 0) 6973 return; 6974 BO_LOCK(&vp->v_bufobj); 6975 drain_output(vp); 6976 BO_UNLOCK(&vp->v_bufobj); 6977 /* 6978 * The vnode pager eliminates file pages we eliminate indirects 6979 * below. 6980 */ 6981 vnode_pager_setsize(vp, length); 6982 /* 6983 * Calculate the end based on the last indirect we want to keep. If 6984 * the block extends into indirects we can just use the negative of 6985 * its lbn. Doubles and triples exist at lower numbers so we must 6986 * be careful not to remove those, if they exist. double and triple 6987 * indirect lbns do not overlap with others so it is not important 6988 * to verify how many levels are required. 6989 */ 6990 lbn = lblkno(fs, length); 6991 if (lbn >= NDADDR) { 6992 /* Calculate the virtual lbn of the triple indirect. */ 6993 lbn = -lbn - (NIADDR - 1); 6994 end = OFF_TO_IDX(lblktosize(fs, lbn)); 6995 } else 6996 end = extend; 6997 vn_pages_remove(vp, OFF_TO_IDX(OFF_MAX), end); 6998} 6999 7000/* 7001 * See if the buf bp is in the range eliminated by truncation. 7002 */ 7003static int 7004trunc_check_buf(bp, blkoffp, lastlbn, lastoff, flags) 7005 struct buf *bp; 7006 int *blkoffp; 7007 ufs_lbn_t lastlbn; 7008 int lastoff; 7009 int flags; 7010{ 7011 ufs_lbn_t lbn; 7012 7013 *blkoffp = 0; 7014 /* Only match ext/normal blocks as appropriate. */ 7015 if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) || 7016 ((flags & IO_NORMAL) == 0 && (bp->b_xflags & BX_ALTDATA) == 0)) 7017 return (0); 7018 /* ALTDATA is always a full truncation. */ 7019 if ((bp->b_xflags & BX_ALTDATA) != 0) 7020 return (1); 7021 /* -1 is full truncation. */ 7022 if (lastlbn == -1) 7023 return (1); 7024 /* 7025 * If this is a partial truncate we only want those 7026 * blocks and indirect blocks that cover the range 7027 * we're after. 7028 */ 7029 lbn = bp->b_lblkno; 7030 if (lbn < 0) 7031 lbn = -(lbn + lbn_level(lbn)); 7032 if (lbn < lastlbn) 7033 return (0); 7034 /* Here we only truncate lblkno if it's partial. */ 7035 if (lbn == lastlbn) { 7036 if (lastoff == 0) 7037 return (0); 7038 *blkoffp = lastoff; 7039 } 7040 return (1); 7041} 7042 7043/* 7044 * Eliminate any dependencies that exist in memory beyond lblkno:off 7045 */ 7046static void 7047trunc_dependencies(ip, freeblks, lastlbn, lastoff, flags) 7048 struct inode *ip; 7049 struct freeblks *freeblks; 7050 ufs_lbn_t lastlbn; 7051 int lastoff; 7052 int flags; 7053{ 7054 struct bufobj *bo; 7055 struct vnode *vp; 7056 struct buf *bp; 7057 struct fs *fs; 7058 int blkoff; 7059 7060 /* 7061 * We must wait for any I/O in progress to finish so that 7062 * all potential buffers on the dirty list will be visible. 7063 * Once they are all there, walk the list and get rid of 7064 * any dependencies. 7065 */ 7066 fs = ip->i_fs; 7067 vp = ITOV(ip); 7068 bo = &vp->v_bufobj; 7069 BO_LOCK(bo); 7070 drain_output(vp); 7071 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) 7072 bp->b_vflags &= ~BV_SCANNED; 7073restart: 7074 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) { 7075 if (bp->b_vflags & BV_SCANNED) 7076 continue; 7077 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) { 7078 bp->b_vflags |= BV_SCANNED; 7079 continue; 7080 } 7081 KASSERT(bp->b_bufobj == bo, ("Wrong object in buffer")); 7082 if ((bp = getdirtybuf(bp, BO_LOCKPTR(bo), MNT_WAIT)) == NULL) 7083 goto restart; 7084 BO_UNLOCK(bo); 7085 if (deallocate_dependencies(bp, freeblks, blkoff)) 7086 bqrelse(bp); 7087 else 7088 brelse(bp); 7089 BO_LOCK(bo); 7090 goto restart; 7091 } 7092 /* 7093 * Now do the work of vtruncbuf while also matching indirect blocks. 7094 */ 7095 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) 7096 bp->b_vflags &= ~BV_SCANNED; 7097cleanrestart: 7098 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) { 7099 if (bp->b_vflags & BV_SCANNED) 7100 continue; 7101 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) { 7102 bp->b_vflags |= BV_SCANNED; 7103 continue; 7104 } 7105 if (BUF_LOCK(bp, 7106 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 7107 BO_LOCKPTR(bo)) == ENOLCK) { 7108 BO_LOCK(bo); 7109 goto cleanrestart; 7110 } 7111 bp->b_vflags |= BV_SCANNED; 7112 bremfree(bp); 7113 if (blkoff != 0) { 7114 allocbuf(bp, blkoff); 7115 bqrelse(bp); 7116 } else { 7117 bp->b_flags |= B_INVAL | B_NOCACHE | B_RELBUF; 7118 brelse(bp); 7119 } 7120 BO_LOCK(bo); 7121 goto cleanrestart; 7122 } 7123 drain_output(vp); 7124 BO_UNLOCK(bo); 7125} 7126 7127static int 7128cancel_pagedep(pagedep, freeblks, blkoff) 7129 struct pagedep *pagedep; 7130 struct freeblks *freeblks; 7131 int blkoff; 7132{ 7133 struct jremref *jremref; 7134 struct jmvref *jmvref; 7135 struct dirrem *dirrem, *tmp; 7136 int i; 7137 7138 /* 7139 * Copy any directory remove dependencies to the list 7140 * to be processed after the freeblks proceeds. If 7141 * directory entry never made it to disk they 7142 * can be dumped directly onto the work list. 7143 */ 7144 LIST_FOREACH_SAFE(dirrem, &pagedep->pd_dirremhd, dm_next, tmp) { 7145 /* Skip this directory removal if it is intended to remain. */ 7146 if (dirrem->dm_offset < blkoff) 7147 continue; 7148 /* 7149 * If there are any dirrems we wait for the journal write 7150 * to complete and then restart the buf scan as the lock 7151 * has been dropped. 7152 */ 7153 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) { 7154 jwait(&jremref->jr_list, MNT_WAIT); 7155 return (ERESTART); 7156 } 7157 LIST_REMOVE(dirrem, dm_next); 7158 dirrem->dm_dirinum = pagedep->pd_ino; 7159 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &dirrem->dm_list); 7160 } 7161 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) { 7162 jwait(&jmvref->jm_list, MNT_WAIT); 7163 return (ERESTART); 7164 } 7165 /* 7166 * When we're partially truncating a pagedep we just want to flush 7167 * journal entries and return. There can not be any adds in the 7168 * truncated portion of the directory and newblk must remain if 7169 * part of the block remains. 7170 */ 7171 if (blkoff != 0) { 7172 struct diradd *dap; 7173 7174 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 7175 if (dap->da_offset > blkoff) 7176 panic("cancel_pagedep: diradd %p off %d > %d", 7177 dap, dap->da_offset, blkoff); 7178 for (i = 0; i < DAHASHSZ; i++) 7179 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) 7180 if (dap->da_offset > blkoff) 7181 panic("cancel_pagedep: diradd %p off %d > %d", 7182 dap, dap->da_offset, blkoff); 7183 return (0); 7184 } 7185 /* 7186 * There should be no directory add dependencies present 7187 * as the directory could not be truncated until all 7188 * children were removed. 7189 */ 7190 KASSERT(LIST_FIRST(&pagedep->pd_pendinghd) == NULL, 7191 ("deallocate_dependencies: pendinghd != NULL")); 7192 for (i = 0; i < DAHASHSZ; i++) 7193 KASSERT(LIST_FIRST(&pagedep->pd_diraddhd[i]) == NULL, 7194 ("deallocate_dependencies: diraddhd != NULL")); 7195 if ((pagedep->pd_state & NEWBLOCK) != 0) 7196 free_newdirblk(pagedep->pd_newdirblk); 7197 if (free_pagedep(pagedep) == 0) 7198 panic("Failed to free pagedep %p", pagedep); 7199 return (0); 7200} 7201 7202/* 7203 * Reclaim any dependency structures from a buffer that is about to 7204 * be reallocated to a new vnode. The buffer must be locked, thus, 7205 * no I/O completion operations can occur while we are manipulating 7206 * its associated dependencies. The mutex is held so that other I/O's 7207 * associated with related dependencies do not occur. 7208 */ 7209static int 7210deallocate_dependencies(bp, freeblks, off) 7211 struct buf *bp; 7212 struct freeblks *freeblks; 7213 int off; 7214{ 7215 struct indirdep *indirdep; 7216 struct pagedep *pagedep; 7217 struct allocdirect *adp; 7218 struct worklist *wk, *wkn; 7219 struct ufsmount *ump; 7220 7221 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL) 7222 goto done; 7223 ump = VFSTOUFS(wk->wk_mp); 7224 ACQUIRE_LOCK(ump); 7225 LIST_FOREACH_SAFE(wk, &bp->b_dep, wk_list, wkn) { 7226 switch (wk->wk_type) { 7227 case D_INDIRDEP: 7228 indirdep = WK_INDIRDEP(wk); 7229 if (bp->b_lblkno >= 0 || 7230 bp->b_blkno != indirdep->ir_savebp->b_lblkno) 7231 panic("deallocate_dependencies: not indir"); 7232 cancel_indirdep(indirdep, bp, freeblks); 7233 continue; 7234 7235 case D_PAGEDEP: 7236 pagedep = WK_PAGEDEP(wk); 7237 if (cancel_pagedep(pagedep, freeblks, off)) { 7238 FREE_LOCK(ump); 7239 return (ERESTART); 7240 } 7241 continue; 7242 7243 case D_ALLOCINDIR: 7244 /* 7245 * Simply remove the allocindir, we'll find it via 7246 * the indirdep where we can clear pointers if 7247 * needed. 7248 */ 7249 WORKLIST_REMOVE(wk); 7250 continue; 7251 7252 case D_FREEWORK: 7253 /* 7254 * A truncation is waiting for the zero'd pointers 7255 * to be written. It can be freed when the freeblks 7256 * is journaled. 7257 */ 7258 WORKLIST_REMOVE(wk); 7259 wk->wk_state |= ONDEPLIST; 7260 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk); 7261 break; 7262 7263 case D_ALLOCDIRECT: 7264 adp = WK_ALLOCDIRECT(wk); 7265 if (off != 0) 7266 continue; 7267 /* FALLTHROUGH */ 7268 default: 7269 panic("deallocate_dependencies: Unexpected type %s", 7270 TYPENAME(wk->wk_type)); 7271 /* NOTREACHED */ 7272 } 7273 } 7274 FREE_LOCK(ump); 7275done: 7276 /* 7277 * Don't throw away this buf, we were partially truncating and 7278 * some deps may always remain. 7279 */ 7280 if (off) { 7281 allocbuf(bp, off); 7282 bp->b_vflags |= BV_SCANNED; 7283 return (EBUSY); 7284 } 7285 bp->b_flags |= B_INVAL | B_NOCACHE; 7286 7287 return (0); 7288} 7289 7290/* 7291 * An allocdirect is being canceled due to a truncate. We must make sure 7292 * the journal entry is released in concert with the blkfree that releases 7293 * the storage. Completed journal entries must not be released until the 7294 * space is no longer pointed to by the inode or in the bitmap. 7295 */ 7296static void 7297cancel_allocdirect(adphead, adp, freeblks) 7298 struct allocdirectlst *adphead; 7299 struct allocdirect *adp; 7300 struct freeblks *freeblks; 7301{ 7302 struct freework *freework; 7303 struct newblk *newblk; 7304 struct worklist *wk; 7305 7306 TAILQ_REMOVE(adphead, adp, ad_next); 7307 newblk = (struct newblk *)adp; 7308 freework = NULL; 7309 /* 7310 * Find the correct freework structure. 7311 */ 7312 LIST_FOREACH(wk, &freeblks->fb_freeworkhd, wk_list) { 7313 if (wk->wk_type != D_FREEWORK) 7314 continue; 7315 freework = WK_FREEWORK(wk); 7316 if (freework->fw_blkno == newblk->nb_newblkno) 7317 break; 7318 } 7319 if (freework == NULL) 7320 panic("cancel_allocdirect: Freework not found"); 7321 /* 7322 * If a newblk exists at all we still have the journal entry that 7323 * initiated the allocation so we do not need to journal the free. 7324 */ 7325 cancel_jfreeblk(freeblks, freework->fw_blkno); 7326 /* 7327 * If the journal hasn't been written the jnewblk must be passed 7328 * to the call to ffs_blkfree that reclaims the space. We accomplish 7329 * this by linking the journal dependency into the freework to be 7330 * freed when freework_freeblock() is called. If the journal has 7331 * been written we can simply reclaim the journal space when the 7332 * freeblks work is complete. 7333 */ 7334 freework->fw_jnewblk = cancel_newblk(newblk, &freework->fw_list, 7335 &freeblks->fb_jwork); 7336 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list); 7337} 7338 7339 7340/* 7341 * Cancel a new block allocation. May be an indirect or direct block. We 7342 * remove it from various lists and return any journal record that needs to 7343 * be resolved by the caller. 7344 * 7345 * A special consideration is made for indirects which were never pointed 7346 * at on disk and will never be found once this block is released. 7347 */ 7348static struct jnewblk * 7349cancel_newblk(newblk, wk, wkhd) 7350 struct newblk *newblk; 7351 struct worklist *wk; 7352 struct workhead *wkhd; 7353{ 7354 struct jnewblk *jnewblk; 7355 7356 CTR1(KTR_SUJ, "cancel_newblk: blkno %jd", newblk->nb_newblkno); 7357 7358 newblk->nb_state |= GOINGAWAY; 7359 /* 7360 * Previously we traversed the completedhd on each indirdep 7361 * attached to this newblk to cancel them and gather journal 7362 * work. Since we need only the oldest journal segment and 7363 * the lowest point on the tree will always have the oldest 7364 * journal segment we are free to release the segments 7365 * of any subordinates and may leave the indirdep list to 7366 * indirdep_complete() when this newblk is freed. 7367 */ 7368 if (newblk->nb_state & ONDEPLIST) { 7369 newblk->nb_state &= ~ONDEPLIST; 7370 LIST_REMOVE(newblk, nb_deps); 7371 } 7372 if (newblk->nb_state & ONWORKLIST) 7373 WORKLIST_REMOVE(&newblk->nb_list); 7374 /* 7375 * If the journal entry hasn't been written we save a pointer to 7376 * the dependency that frees it until it is written or the 7377 * superseding operation completes. 7378 */ 7379 jnewblk = newblk->nb_jnewblk; 7380 if (jnewblk != NULL && wk != NULL) { 7381 newblk->nb_jnewblk = NULL; 7382 jnewblk->jn_dep = wk; 7383 } 7384 if (!LIST_EMPTY(&newblk->nb_jwork)) 7385 jwork_move(wkhd, &newblk->nb_jwork); 7386 /* 7387 * When truncating we must free the newdirblk early to remove 7388 * the pagedep from the hash before returning. 7389 */ 7390 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL) 7391 free_newdirblk(WK_NEWDIRBLK(wk)); 7392 if (!LIST_EMPTY(&newblk->nb_newdirblk)) 7393 panic("cancel_newblk: extra newdirblk"); 7394 7395 return (jnewblk); 7396} 7397 7398/* 7399 * Schedule the freefrag associated with a newblk to be released once 7400 * the pointers are written and the previous block is no longer needed. 7401 */ 7402static void 7403newblk_freefrag(newblk) 7404 struct newblk *newblk; 7405{ 7406 struct freefrag *freefrag; 7407 7408 if (newblk->nb_freefrag == NULL) 7409 return; 7410 freefrag = newblk->nb_freefrag; 7411 newblk->nb_freefrag = NULL; 7412 freefrag->ff_state |= COMPLETE; 7413 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE) 7414 add_to_worklist(&freefrag->ff_list, 0); 7415} 7416 7417/* 7418 * Free a newblk. Generate a new freefrag work request if appropriate. 7419 * This must be called after the inode pointer and any direct block pointers 7420 * are valid or fully removed via truncate or frag extension. 7421 */ 7422static void 7423free_newblk(newblk) 7424 struct newblk *newblk; 7425{ 7426 struct indirdep *indirdep; 7427 struct worklist *wk; 7428 7429 KASSERT(newblk->nb_jnewblk == NULL, 7430 ("free_newblk: jnewblk %p still attached", newblk->nb_jnewblk)); 7431 KASSERT(newblk->nb_list.wk_type != D_NEWBLK, 7432 ("free_newblk: unclaimed newblk")); 7433 LOCK_OWNED(VFSTOUFS(newblk->nb_list.wk_mp)); 7434 newblk_freefrag(newblk); 7435 if (newblk->nb_state & ONDEPLIST) 7436 LIST_REMOVE(newblk, nb_deps); 7437 if (newblk->nb_state & ONWORKLIST) 7438 WORKLIST_REMOVE(&newblk->nb_list); 7439 LIST_REMOVE(newblk, nb_hash); 7440 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL) 7441 free_newdirblk(WK_NEWDIRBLK(wk)); 7442 if (!LIST_EMPTY(&newblk->nb_newdirblk)) 7443 panic("free_newblk: extra newdirblk"); 7444 while ((indirdep = LIST_FIRST(&newblk->nb_indirdeps)) != NULL) 7445 indirdep_complete(indirdep); 7446 handle_jwork(&newblk->nb_jwork); 7447 WORKITEM_FREE(newblk, D_NEWBLK); 7448} 7449 7450/* 7451 * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep. 7452 * This routine must be called with splbio interrupts blocked. 7453 */ 7454static void 7455free_newdirblk(newdirblk) 7456 struct newdirblk *newdirblk; 7457{ 7458 struct pagedep *pagedep; 7459 struct diradd *dap; 7460 struct worklist *wk; 7461 7462 LOCK_OWNED(VFSTOUFS(newdirblk->db_list.wk_mp)); 7463 WORKLIST_REMOVE(&newdirblk->db_list); 7464 /* 7465 * If the pagedep is still linked onto the directory buffer 7466 * dependency chain, then some of the entries on the 7467 * pd_pendinghd list may not be committed to disk yet. In 7468 * this case, we will simply clear the NEWBLOCK flag and 7469 * let the pd_pendinghd list be processed when the pagedep 7470 * is next written. If the pagedep is no longer on the buffer 7471 * dependency chain, then all the entries on the pd_pending 7472 * list are committed to disk and we can free them here. 7473 */ 7474 pagedep = newdirblk->db_pagedep; 7475 pagedep->pd_state &= ~NEWBLOCK; 7476 if ((pagedep->pd_state & ONWORKLIST) == 0) { 7477 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 7478 free_diradd(dap, NULL); 7479 /* 7480 * If no dependencies remain, the pagedep will be freed. 7481 */ 7482 free_pagedep(pagedep); 7483 } 7484 /* Should only ever be one item in the list. */ 7485 while ((wk = LIST_FIRST(&newdirblk->db_mkdir)) != NULL) { 7486 WORKLIST_REMOVE(wk); 7487 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 7488 } 7489 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 7490} 7491 7492/* 7493 * Prepare an inode to be freed. The actual free operation is not 7494 * done until the zero'ed inode has been written to disk. 7495 */ 7496void 7497softdep_freefile(pvp, ino, mode) 7498 struct vnode *pvp; 7499 ino_t ino; 7500 int mode; 7501{ 7502 struct inode *ip = VTOI(pvp); 7503 struct inodedep *inodedep; 7504 struct freefile *freefile; 7505 struct freeblks *freeblks; 7506 struct ufsmount *ump; 7507 7508 ump = ip->i_ump; 7509 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 7510 ("softdep_freefile called on non-softdep filesystem")); 7511 /* 7512 * This sets up the inode de-allocation dependency. 7513 */ 7514 freefile = malloc(sizeof(struct freefile), 7515 M_FREEFILE, M_SOFTDEP_FLAGS); 7516 workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount); 7517 freefile->fx_mode = mode; 7518 freefile->fx_oldinum = ino; 7519 freefile->fx_devvp = ip->i_devvp; 7520 LIST_INIT(&freefile->fx_jwork); 7521 UFS_LOCK(ump); 7522 ip->i_fs->fs_pendinginodes += 1; 7523 UFS_UNLOCK(ump); 7524 7525 /* 7526 * If the inodedep does not exist, then the zero'ed inode has 7527 * been written to disk. If the allocated inode has never been 7528 * written to disk, then the on-disk inode is zero'ed. In either 7529 * case we can free the file immediately. If the journal was 7530 * canceled before being written the inode will never make it to 7531 * disk and we must send the canceled journal entrys to 7532 * ffs_freefile() to be cleared in conjunction with the bitmap. 7533 * Any blocks waiting on the inode to write can be safely freed 7534 * here as it will never been written. 7535 */ 7536 ACQUIRE_LOCK(ump); 7537 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep); 7538 if (inodedep) { 7539 /* 7540 * Clear out freeblks that no longer need to reference 7541 * this inode. 7542 */ 7543 while ((freeblks = 7544 TAILQ_FIRST(&inodedep->id_freeblklst)) != NULL) { 7545 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, 7546 fb_next); 7547 freeblks->fb_state &= ~ONDEPLIST; 7548 } 7549 /* 7550 * Remove this inode from the unlinked list. 7551 */ 7552 if (inodedep->id_state & UNLINKED) { 7553 /* 7554 * Save the journal work to be freed with the bitmap 7555 * before we clear UNLINKED. Otherwise it can be lost 7556 * if the inode block is written. 7557 */ 7558 handle_bufwait(inodedep, &freefile->fx_jwork); 7559 clear_unlinked_inodedep(inodedep); 7560 /* 7561 * Re-acquire inodedep as we've dropped the 7562 * per-filesystem lock in clear_unlinked_inodedep(). 7563 */ 7564 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep); 7565 } 7566 } 7567 if (inodedep == NULL || check_inode_unwritten(inodedep)) { 7568 FREE_LOCK(ump); 7569 handle_workitem_freefile(freefile); 7570 return; 7571 } 7572 if ((inodedep->id_state & DEPCOMPLETE) == 0) 7573 inodedep->id_state |= GOINGAWAY; 7574 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 7575 FREE_LOCK(ump); 7576 if (ip->i_number == ino) 7577 ip->i_flag |= IN_MODIFIED; 7578} 7579 7580/* 7581 * Check to see if an inode has never been written to disk. If 7582 * so free the inodedep and return success, otherwise return failure. 7583 * This routine must be called with splbio interrupts blocked. 7584 * 7585 * If we still have a bitmap dependency, then the inode has never 7586 * been written to disk. Drop the dependency as it is no longer 7587 * necessary since the inode is being deallocated. We set the 7588 * ALLCOMPLETE flags since the bitmap now properly shows that the 7589 * inode is not allocated. Even if the inode is actively being 7590 * written, it has been rolled back to its zero'ed state, so we 7591 * are ensured that a zero inode is what is on the disk. For short 7592 * lived files, this change will usually result in removing all the 7593 * dependencies from the inode so that it can be freed immediately. 7594 */ 7595static int 7596check_inode_unwritten(inodedep) 7597 struct inodedep *inodedep; 7598{ 7599 7600 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp)); 7601 7602 if ((inodedep->id_state & (DEPCOMPLETE | UNLINKED)) != 0 || 7603 !LIST_EMPTY(&inodedep->id_dirremhd) || 7604 !LIST_EMPTY(&inodedep->id_pendinghd) || 7605 !LIST_EMPTY(&inodedep->id_bufwait) || 7606 !LIST_EMPTY(&inodedep->id_inowait) || 7607 !TAILQ_EMPTY(&inodedep->id_inoreflst) || 7608 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 7609 !TAILQ_EMPTY(&inodedep->id_newinoupdt) || 7610 !TAILQ_EMPTY(&inodedep->id_extupdt) || 7611 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 7612 !TAILQ_EMPTY(&inodedep->id_freeblklst) || 7613 inodedep->id_mkdiradd != NULL || 7614 inodedep->id_nlinkdelta != 0) 7615 return (0); 7616 /* 7617 * Another process might be in initiate_write_inodeblock_ufs[12] 7618 * trying to allocate memory without holding "Softdep Lock". 7619 */ 7620 if ((inodedep->id_state & IOSTARTED) != 0 && 7621 inodedep->id_savedino1 == NULL) 7622 return (0); 7623 7624 if (inodedep->id_state & ONDEPLIST) 7625 LIST_REMOVE(inodedep, id_deps); 7626 inodedep->id_state &= ~ONDEPLIST; 7627 inodedep->id_state |= ALLCOMPLETE; 7628 inodedep->id_bmsafemap = NULL; 7629 if (inodedep->id_state & ONWORKLIST) 7630 WORKLIST_REMOVE(&inodedep->id_list); 7631 if (inodedep->id_savedino1 != NULL) { 7632 free(inodedep->id_savedino1, M_SAVEDINO); 7633 inodedep->id_savedino1 = NULL; 7634 } 7635 if (free_inodedep(inodedep) == 0) 7636 panic("check_inode_unwritten: busy inode"); 7637 return (1); 7638} 7639 7640/* 7641 * Try to free an inodedep structure. Return 1 if it could be freed. 7642 */ 7643static int 7644free_inodedep(inodedep) 7645 struct inodedep *inodedep; 7646{ 7647 7648 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp)); 7649 if ((inodedep->id_state & (ONWORKLIST | UNLINKED)) != 0 || 7650 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 7651 !LIST_EMPTY(&inodedep->id_dirremhd) || 7652 !LIST_EMPTY(&inodedep->id_pendinghd) || 7653 !LIST_EMPTY(&inodedep->id_bufwait) || 7654 !LIST_EMPTY(&inodedep->id_inowait) || 7655 !TAILQ_EMPTY(&inodedep->id_inoreflst) || 7656 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 7657 !TAILQ_EMPTY(&inodedep->id_newinoupdt) || 7658 !TAILQ_EMPTY(&inodedep->id_extupdt) || 7659 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 7660 !TAILQ_EMPTY(&inodedep->id_freeblklst) || 7661 inodedep->id_mkdiradd != NULL || 7662 inodedep->id_nlinkdelta != 0 || 7663 inodedep->id_savedino1 != NULL) 7664 return (0); 7665 if (inodedep->id_state & ONDEPLIST) 7666 LIST_REMOVE(inodedep, id_deps); 7667 LIST_REMOVE(inodedep, id_hash); 7668 WORKITEM_FREE(inodedep, D_INODEDEP); 7669 return (1); 7670} 7671 7672/* 7673 * Free the block referenced by a freework structure. The parent freeblks 7674 * structure is released and completed when the final cg bitmap reaches 7675 * the disk. This routine may be freeing a jnewblk which never made it to 7676 * disk in which case we do not have to wait as the operation is undone 7677 * in memory immediately. 7678 */ 7679static void 7680freework_freeblock(freework) 7681 struct freework *freework; 7682{ 7683 struct freeblks *freeblks; 7684 struct jnewblk *jnewblk; 7685 struct ufsmount *ump; 7686 struct workhead wkhd; 7687 struct fs *fs; 7688 int bsize; 7689 int needj; 7690 7691 ump = VFSTOUFS(freework->fw_list.wk_mp); 7692 LOCK_OWNED(ump); 7693 /* 7694 * Handle partial truncate separately. 7695 */ 7696 if (freework->fw_indir) { 7697 complete_trunc_indir(freework); 7698 return; 7699 } 7700 freeblks = freework->fw_freeblks; 7701 fs = ump->um_fs; 7702 needj = MOUNTEDSUJ(freeblks->fb_list.wk_mp) != 0; 7703 bsize = lfragtosize(fs, freework->fw_frags); 7704 LIST_INIT(&wkhd); 7705 /* 7706 * DEPCOMPLETE is cleared in indirblk_insert() if the block lives 7707 * on the indirblk hashtable and prevents premature freeing. 7708 */ 7709 freework->fw_state |= DEPCOMPLETE; 7710 /* 7711 * SUJ needs to wait for the segment referencing freed indirect 7712 * blocks to expire so that we know the checker will not confuse 7713 * a re-allocated indirect block with its old contents. 7714 */ 7715 if (needj && freework->fw_lbn <= -NDADDR) 7716 indirblk_insert(freework); 7717 /* 7718 * If we are canceling an existing jnewblk pass it to the free 7719 * routine, otherwise pass the freeblk which will ultimately 7720 * release the freeblks. If we're not journaling, we can just 7721 * free the freeblks immediately. 7722 */ 7723 jnewblk = freework->fw_jnewblk; 7724 if (jnewblk != NULL) { 7725 cancel_jnewblk(jnewblk, &wkhd); 7726 needj = 0; 7727 } else if (needj) { 7728 freework->fw_state |= DELAYEDFREE; 7729 freeblks->fb_cgwait++; 7730 WORKLIST_INSERT(&wkhd, &freework->fw_list); 7731 } 7732 FREE_LOCK(ump); 7733 freeblks_free(ump, freeblks, btodb(bsize)); 7734 CTR4(KTR_SUJ, 7735 "freework_freeblock: ino %d blkno %jd lbn %jd size %ld", 7736 freeblks->fb_inum, freework->fw_blkno, freework->fw_lbn, bsize); 7737 ffs_blkfree(ump, fs, freeblks->fb_devvp, freework->fw_blkno, bsize, 7738 freeblks->fb_inum, freeblks->fb_vtype, &wkhd); 7739 ACQUIRE_LOCK(ump); 7740 /* 7741 * The jnewblk will be discarded and the bits in the map never 7742 * made it to disk. We can immediately free the freeblk. 7743 */ 7744 if (needj == 0) 7745 handle_written_freework(freework); 7746} 7747 7748/* 7749 * We enqueue freework items that need processing back on the freeblks and 7750 * add the freeblks to the worklist. This makes it easier to find all work 7751 * required to flush a truncation in process_truncates(). 7752 */ 7753static void 7754freework_enqueue(freework) 7755 struct freework *freework; 7756{ 7757 struct freeblks *freeblks; 7758 7759 freeblks = freework->fw_freeblks; 7760 if ((freework->fw_state & INPROGRESS) == 0) 7761 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list); 7762 if ((freeblks->fb_state & 7763 (ONWORKLIST | INPROGRESS | ALLCOMPLETE)) == ALLCOMPLETE && 7764 LIST_EMPTY(&freeblks->fb_jblkdephd)) 7765 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 7766} 7767 7768/* 7769 * Start, continue, or finish the process of freeing an indirect block tree. 7770 * The free operation may be paused at any point with fw_off containing the 7771 * offset to restart from. This enables us to implement some flow control 7772 * for large truncates which may fan out and generate a huge number of 7773 * dependencies. 7774 */ 7775static void 7776handle_workitem_indirblk(freework) 7777 struct freework *freework; 7778{ 7779 struct freeblks *freeblks; 7780 struct ufsmount *ump; 7781 struct fs *fs; 7782 7783 freeblks = freework->fw_freeblks; 7784 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7785 fs = ump->um_fs; 7786 if (freework->fw_state & DEPCOMPLETE) { 7787 handle_written_freework(freework); 7788 return; 7789 } 7790 if (freework->fw_off == NINDIR(fs)) { 7791 freework_freeblock(freework); 7792 return; 7793 } 7794 freework->fw_state |= INPROGRESS; 7795 FREE_LOCK(ump); 7796 indir_trunc(freework, fsbtodb(fs, freework->fw_blkno), 7797 freework->fw_lbn); 7798 ACQUIRE_LOCK(ump); 7799} 7800 7801/* 7802 * Called when a freework structure attached to a cg buf is written. The 7803 * ref on either the parent or the freeblks structure is released and 7804 * the freeblks is added back to the worklist if there is more work to do. 7805 */ 7806static void 7807handle_written_freework(freework) 7808 struct freework *freework; 7809{ 7810 struct freeblks *freeblks; 7811 struct freework *parent; 7812 7813 freeblks = freework->fw_freeblks; 7814 parent = freework->fw_parent; 7815 if (freework->fw_state & DELAYEDFREE) 7816 freeblks->fb_cgwait--; 7817 freework->fw_state |= COMPLETE; 7818 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE) 7819 WORKITEM_FREE(freework, D_FREEWORK); 7820 if (parent) { 7821 if (--parent->fw_ref == 0) 7822 freework_enqueue(parent); 7823 return; 7824 } 7825 if (--freeblks->fb_ref != 0) 7826 return; 7827 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST | INPROGRESS)) == 7828 ALLCOMPLETE && LIST_EMPTY(&freeblks->fb_jblkdephd)) 7829 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 7830} 7831 7832/* 7833 * This workitem routine performs the block de-allocation. 7834 * The workitem is added to the pending list after the updated 7835 * inode block has been written to disk. As mentioned above, 7836 * checks regarding the number of blocks de-allocated (compared 7837 * to the number of blocks allocated for the file) are also 7838 * performed in this function. 7839 */ 7840static int 7841handle_workitem_freeblocks(freeblks, flags) 7842 struct freeblks *freeblks; 7843 int flags; 7844{ 7845 struct freework *freework; 7846 struct newblk *newblk; 7847 struct allocindir *aip; 7848 struct ufsmount *ump; 7849 struct worklist *wk; 7850 7851 KASSERT(LIST_EMPTY(&freeblks->fb_jblkdephd), 7852 ("handle_workitem_freeblocks: Journal entries not written.")); 7853 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7854 ACQUIRE_LOCK(ump); 7855 while ((wk = LIST_FIRST(&freeblks->fb_freeworkhd)) != NULL) { 7856 WORKLIST_REMOVE(wk); 7857 switch (wk->wk_type) { 7858 case D_DIRREM: 7859 wk->wk_state |= COMPLETE; 7860 add_to_worklist(wk, 0); 7861 continue; 7862 7863 case D_ALLOCDIRECT: 7864 free_newblk(WK_NEWBLK(wk)); 7865 continue; 7866 7867 case D_ALLOCINDIR: 7868 aip = WK_ALLOCINDIR(wk); 7869 freework = NULL; 7870 if (aip->ai_state & DELAYEDFREE) { 7871 FREE_LOCK(ump); 7872 freework = newfreework(ump, freeblks, NULL, 7873 aip->ai_lbn, aip->ai_newblkno, 7874 ump->um_fs->fs_frag, 0, 0); 7875 ACQUIRE_LOCK(ump); 7876 } 7877 newblk = WK_NEWBLK(wk); 7878 if (newblk->nb_jnewblk) { 7879 freework->fw_jnewblk = newblk->nb_jnewblk; 7880 newblk->nb_jnewblk->jn_dep = &freework->fw_list; 7881 newblk->nb_jnewblk = NULL; 7882 } 7883 free_newblk(newblk); 7884 continue; 7885 7886 case D_FREEWORK: 7887 freework = WK_FREEWORK(wk); 7888 if (freework->fw_lbn <= -NDADDR) 7889 handle_workitem_indirblk(freework); 7890 else 7891 freework_freeblock(freework); 7892 continue; 7893 default: 7894 panic("handle_workitem_freeblocks: Unknown type %s", 7895 TYPENAME(wk->wk_type)); 7896 } 7897 } 7898 if (freeblks->fb_ref != 0) { 7899 freeblks->fb_state &= ~INPROGRESS; 7900 wake_worklist(&freeblks->fb_list); 7901 freeblks = NULL; 7902 } 7903 FREE_LOCK(ump); 7904 if (freeblks) 7905 return handle_complete_freeblocks(freeblks, flags); 7906 return (0); 7907} 7908 7909/* 7910 * Handle completion of block free via truncate. This allows fs_pending 7911 * to track the actual free block count more closely than if we only updated 7912 * it at the end. We must be careful to handle cases where the block count 7913 * on free was incorrect. 7914 */ 7915static void 7916freeblks_free(ump, freeblks, blocks) 7917 struct ufsmount *ump; 7918 struct freeblks *freeblks; 7919 int blocks; 7920{ 7921 struct fs *fs; 7922 ufs2_daddr_t remain; 7923 7924 UFS_LOCK(ump); 7925 remain = -freeblks->fb_chkcnt; 7926 freeblks->fb_chkcnt += blocks; 7927 if (remain > 0) { 7928 if (remain < blocks) 7929 blocks = remain; 7930 fs = ump->um_fs; 7931 fs->fs_pendingblocks -= blocks; 7932 } 7933 UFS_UNLOCK(ump); 7934} 7935 7936/* 7937 * Once all of the freework workitems are complete we can retire the 7938 * freeblocks dependency and any journal work awaiting completion. This 7939 * can not be called until all other dependencies are stable on disk. 7940 */ 7941static int 7942handle_complete_freeblocks(freeblks, flags) 7943 struct freeblks *freeblks; 7944 int flags; 7945{ 7946 struct inodedep *inodedep; 7947 struct inode *ip; 7948 struct vnode *vp; 7949 struct fs *fs; 7950 struct ufsmount *ump; 7951 ufs2_daddr_t spare; 7952 7953 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7954 fs = ump->um_fs; 7955 flags = LK_EXCLUSIVE | flags; 7956 spare = freeblks->fb_chkcnt; 7957 7958 /* 7959 * If we did not release the expected number of blocks we may have 7960 * to adjust the inode block count here. Only do so if it wasn't 7961 * a truncation to zero and the modrev still matches. 7962 */ 7963 if (spare && freeblks->fb_len != 0) { 7964 if (ffs_vgetf(freeblks->fb_list.wk_mp, freeblks->fb_inum, 7965 flags, &vp, FFSV_FORCEINSMQ) != 0) 7966 return (EBUSY); 7967 ip = VTOI(vp); 7968 if (DIP(ip, i_modrev) == freeblks->fb_modrev) { 7969 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - spare); 7970 ip->i_flag |= IN_CHANGE; 7971 /* 7972 * We must wait so this happens before the 7973 * journal is reclaimed. 7974 */ 7975 ffs_update(vp, 1); 7976 } 7977 vput(vp); 7978 } 7979 if (spare < 0) { 7980 UFS_LOCK(ump); 7981 fs->fs_pendingblocks += spare; 7982 UFS_UNLOCK(ump); 7983 } 7984#ifdef QUOTA 7985 /* Handle spare. */ 7986 if (spare) 7987 quotaadj(freeblks->fb_quota, ump, -spare); 7988 quotarele(freeblks->fb_quota); 7989#endif 7990 ACQUIRE_LOCK(ump); 7991 if (freeblks->fb_state & ONDEPLIST) { 7992 inodedep_lookup(freeblks->fb_list.wk_mp, freeblks->fb_inum, 7993 0, &inodedep); 7994 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, fb_next); 7995 freeblks->fb_state &= ~ONDEPLIST; 7996 if (TAILQ_EMPTY(&inodedep->id_freeblklst)) 7997 free_inodedep(inodedep); 7998 } 7999 /* 8000 * All of the freeblock deps must be complete prior to this call 8001 * so it's now safe to complete earlier outstanding journal entries. 8002 */ 8003 handle_jwork(&freeblks->fb_jwork); 8004 WORKITEM_FREE(freeblks, D_FREEBLKS); 8005 FREE_LOCK(ump); 8006 return (0); 8007} 8008 8009/* 8010 * Release blocks associated with the freeblks and stored in the indirect 8011 * block dbn. If level is greater than SINGLE, the block is an indirect block 8012 * and recursive calls to indirtrunc must be used to cleanse other indirect 8013 * blocks. 8014 * 8015 * This handles partial and complete truncation of blocks. Partial is noted 8016 * with goingaway == 0. In this case the freework is completed after the 8017 * zero'd indirects are written to disk. For full truncation the freework 8018 * is completed after the block is freed. 8019 */ 8020static void 8021indir_trunc(freework, dbn, lbn) 8022 struct freework *freework; 8023 ufs2_daddr_t dbn; 8024 ufs_lbn_t lbn; 8025{ 8026 struct freework *nfreework; 8027 struct workhead wkhd; 8028 struct freeblks *freeblks; 8029 struct buf *bp; 8030 struct fs *fs; 8031 struct indirdep *indirdep; 8032 struct ufsmount *ump; 8033 ufs1_daddr_t *bap1 = 0; 8034 ufs2_daddr_t nb, nnb, *bap2 = 0; 8035 ufs_lbn_t lbnadd, nlbn; 8036 int i, nblocks, ufs1fmt; 8037 int freedblocks; 8038 int goingaway; 8039 int freedeps; 8040 int needj; 8041 int level; 8042 int cnt; 8043 8044 freeblks = freework->fw_freeblks; 8045 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 8046 fs = ump->um_fs; 8047 /* 8048 * Get buffer of block pointers to be freed. There are three cases: 8049 * 8050 * 1) Partial truncate caches the indirdep pointer in the freework 8051 * which provides us a back copy to the save bp which holds the 8052 * pointers we want to clear. When this completes the zero 8053 * pointers are written to the real copy. 8054 * 2) The indirect is being completely truncated, cancel_indirdep() 8055 * eliminated the real copy and placed the indirdep on the saved 8056 * copy. The indirdep and buf are discarded when this completes. 8057 * 3) The indirect was not in memory, we read a copy off of the disk 8058 * using the devvp and drop and invalidate the buffer when we're 8059 * done. 8060 */ 8061 goingaway = 1; 8062 indirdep = NULL; 8063 if (freework->fw_indir != NULL) { 8064 goingaway = 0; 8065 indirdep = freework->fw_indir; 8066 bp = indirdep->ir_savebp; 8067 if (bp == NULL || bp->b_blkno != dbn) 8068 panic("indir_trunc: Bad saved buf %p blkno %jd", 8069 bp, (intmax_t)dbn); 8070 } else if ((bp = incore(&freeblks->fb_devvp->v_bufobj, dbn)) != NULL) { 8071 /* 8072 * The lock prevents the buf dep list from changing and 8073 * indirects on devvp should only ever have one dependency. 8074 */ 8075 indirdep = WK_INDIRDEP(LIST_FIRST(&bp->b_dep)); 8076 if (indirdep == NULL || (indirdep->ir_state & GOINGAWAY) == 0) 8077 panic("indir_trunc: Bad indirdep %p from buf %p", 8078 indirdep, bp); 8079 } else if (bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 8080 NOCRED, &bp) != 0) { 8081 brelse(bp); 8082 return; 8083 } 8084 ACQUIRE_LOCK(ump); 8085 /* Protects against a race with complete_trunc_indir(). */ 8086 freework->fw_state &= ~INPROGRESS; 8087 /* 8088 * If we have an indirdep we need to enforce the truncation order 8089 * and discard it when it is complete. 8090 */ 8091 if (indirdep) { 8092 if (freework != TAILQ_FIRST(&indirdep->ir_trunc) && 8093 !TAILQ_EMPTY(&indirdep->ir_trunc)) { 8094 /* 8095 * Add the complete truncate to the list on the 8096 * indirdep to enforce in-order processing. 8097 */ 8098 if (freework->fw_indir == NULL) 8099 TAILQ_INSERT_TAIL(&indirdep->ir_trunc, 8100 freework, fw_next); 8101 FREE_LOCK(ump); 8102 return; 8103 } 8104 /* 8105 * If we're goingaway, free the indirdep. Otherwise it will 8106 * linger until the write completes. 8107 */ 8108 if (goingaway) 8109 free_indirdep(indirdep); 8110 } 8111 FREE_LOCK(ump); 8112 /* Initialize pointers depending on block size. */ 8113 if (ump->um_fstype == UFS1) { 8114 bap1 = (ufs1_daddr_t *)bp->b_data; 8115 nb = bap1[freework->fw_off]; 8116 ufs1fmt = 1; 8117 } else { 8118 bap2 = (ufs2_daddr_t *)bp->b_data; 8119 nb = bap2[freework->fw_off]; 8120 ufs1fmt = 0; 8121 } 8122 level = lbn_level(lbn); 8123 needj = MOUNTEDSUJ(UFSTOVFS(ump)) != 0; 8124 lbnadd = lbn_offset(fs, level); 8125 nblocks = btodb(fs->fs_bsize); 8126 nfreework = freework; 8127 freedeps = 0; 8128 cnt = 0; 8129 /* 8130 * Reclaim blocks. Traverses into nested indirect levels and 8131 * arranges for the current level to be freed when subordinates 8132 * are free when journaling. 8133 */ 8134 for (i = freework->fw_off; i < NINDIR(fs); i++, nb = nnb) { 8135 if (i != NINDIR(fs) - 1) { 8136 if (ufs1fmt) 8137 nnb = bap1[i+1]; 8138 else 8139 nnb = bap2[i+1]; 8140 } else 8141 nnb = 0; 8142 if (nb == 0) 8143 continue; 8144 cnt++; 8145 if (level != 0) { 8146 nlbn = (lbn + 1) - (i * lbnadd); 8147 if (needj != 0) { 8148 nfreework = newfreework(ump, freeblks, freework, 8149 nlbn, nb, fs->fs_frag, 0, 0); 8150 freedeps++; 8151 } 8152 indir_trunc(nfreework, fsbtodb(fs, nb), nlbn); 8153 } else { 8154 struct freedep *freedep; 8155 8156 /* 8157 * Attempt to aggregate freedep dependencies for 8158 * all blocks being released to the same CG. 8159 */ 8160 LIST_INIT(&wkhd); 8161 if (needj != 0 && 8162 (nnb == 0 || (dtog(fs, nb) != dtog(fs, nnb)))) { 8163 freedep = newfreedep(freework); 8164 WORKLIST_INSERT_UNLOCKED(&wkhd, 8165 &freedep->fd_list); 8166 freedeps++; 8167 } 8168 CTR3(KTR_SUJ, 8169 "indir_trunc: ino %d blkno %jd size %ld", 8170 freeblks->fb_inum, nb, fs->fs_bsize); 8171 ffs_blkfree(ump, fs, freeblks->fb_devvp, nb, 8172 fs->fs_bsize, freeblks->fb_inum, 8173 freeblks->fb_vtype, &wkhd); 8174 } 8175 } 8176 if (goingaway) { 8177 bp->b_flags |= B_INVAL | B_NOCACHE; 8178 brelse(bp); 8179 } 8180 freedblocks = 0; 8181 if (level == 0) 8182 freedblocks = (nblocks * cnt); 8183 if (needj == 0) 8184 freedblocks += nblocks; 8185 freeblks_free(ump, freeblks, freedblocks); 8186 /* 8187 * If we are journaling set up the ref counts and offset so this 8188 * indirect can be completed when its children are free. 8189 */ 8190 if (needj) { 8191 ACQUIRE_LOCK(ump); 8192 freework->fw_off = i; 8193 freework->fw_ref += freedeps; 8194 freework->fw_ref -= NINDIR(fs) + 1; 8195 if (level == 0) 8196 freeblks->fb_cgwait += freedeps; 8197 if (freework->fw_ref == 0) 8198 freework_freeblock(freework); 8199 FREE_LOCK(ump); 8200 return; 8201 } 8202 /* 8203 * If we're not journaling we can free the indirect now. 8204 */ 8205 dbn = dbtofsb(fs, dbn); 8206 CTR3(KTR_SUJ, 8207 "indir_trunc 2: ino %d blkno %jd size %ld", 8208 freeblks->fb_inum, dbn, fs->fs_bsize); 8209 ffs_blkfree(ump, fs, freeblks->fb_devvp, dbn, fs->fs_bsize, 8210 freeblks->fb_inum, freeblks->fb_vtype, NULL); 8211 /* Non SUJ softdep does single-threaded truncations. */ 8212 if (freework->fw_blkno == dbn) { 8213 freework->fw_state |= ALLCOMPLETE; 8214 ACQUIRE_LOCK(ump); 8215 handle_written_freework(freework); 8216 FREE_LOCK(ump); 8217 } 8218 return; 8219} 8220 8221/* 8222 * Cancel an allocindir when it is removed via truncation. When bp is not 8223 * NULL the indirect never appeared on disk and is scheduled to be freed 8224 * independently of the indir so we can more easily track journal work. 8225 */ 8226static void 8227cancel_allocindir(aip, bp, freeblks, trunc) 8228 struct allocindir *aip; 8229 struct buf *bp; 8230 struct freeblks *freeblks; 8231 int trunc; 8232{ 8233 struct indirdep *indirdep; 8234 struct freefrag *freefrag; 8235 struct newblk *newblk; 8236 8237 newblk = (struct newblk *)aip; 8238 LIST_REMOVE(aip, ai_next); 8239 /* 8240 * We must eliminate the pointer in bp if it must be freed on its 8241 * own due to partial truncate or pending journal work. 8242 */ 8243 if (bp && (trunc || newblk->nb_jnewblk)) { 8244 /* 8245 * Clear the pointer and mark the aip to be freed 8246 * directly if it never existed on disk. 8247 */ 8248 aip->ai_state |= DELAYEDFREE; 8249 indirdep = aip->ai_indirdep; 8250 if (indirdep->ir_state & UFS1FMT) 8251 ((ufs1_daddr_t *)bp->b_data)[aip->ai_offset] = 0; 8252 else 8253 ((ufs2_daddr_t *)bp->b_data)[aip->ai_offset] = 0; 8254 } 8255 /* 8256 * When truncating the previous pointer will be freed via 8257 * savedbp. Eliminate the freefrag which would dup free. 8258 */ 8259 if (trunc && (freefrag = newblk->nb_freefrag) != NULL) { 8260 newblk->nb_freefrag = NULL; 8261 if (freefrag->ff_jdep) 8262 cancel_jfreefrag( 8263 WK_JFREEFRAG(freefrag->ff_jdep)); 8264 jwork_move(&freeblks->fb_jwork, &freefrag->ff_jwork); 8265 WORKITEM_FREE(freefrag, D_FREEFRAG); 8266 } 8267 /* 8268 * If the journal hasn't been written the jnewblk must be passed 8269 * to the call to ffs_blkfree that reclaims the space. We accomplish 8270 * this by leaving the journal dependency on the newblk to be freed 8271 * when a freework is created in handle_workitem_freeblocks(). 8272 */ 8273 cancel_newblk(newblk, NULL, &freeblks->fb_jwork); 8274 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list); 8275} 8276 8277/* 8278 * Create the mkdir dependencies for . and .. in a new directory. Link them 8279 * in to a newdirblk so any subsequent additions are tracked properly. The 8280 * caller is responsible for adding the mkdir1 dependency to the journal 8281 * and updating id_mkdiradd. This function returns with the per-filesystem 8282 * lock held. 8283 */ 8284static struct mkdir * 8285setup_newdir(dap, newinum, dinum, newdirbp, mkdirp) 8286 struct diradd *dap; 8287 ino_t newinum; 8288 ino_t dinum; 8289 struct buf *newdirbp; 8290 struct mkdir **mkdirp; 8291{ 8292 struct newblk *newblk; 8293 struct pagedep *pagedep; 8294 struct inodedep *inodedep; 8295 struct newdirblk *newdirblk = 0; 8296 struct mkdir *mkdir1, *mkdir2; 8297 struct worklist *wk; 8298 struct jaddref *jaddref; 8299 struct ufsmount *ump; 8300 struct mount *mp; 8301 8302 mp = dap->da_list.wk_mp; 8303 ump = VFSTOUFS(mp); 8304 newdirblk = malloc(sizeof(struct newdirblk), M_NEWDIRBLK, 8305 M_SOFTDEP_FLAGS); 8306 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 8307 LIST_INIT(&newdirblk->db_mkdir); 8308 mkdir1 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS); 8309 workitem_alloc(&mkdir1->md_list, D_MKDIR, mp); 8310 mkdir1->md_state = ATTACHED | MKDIR_BODY; 8311 mkdir1->md_diradd = dap; 8312 mkdir1->md_jaddref = NULL; 8313 mkdir2 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS); 8314 workitem_alloc(&mkdir2->md_list, D_MKDIR, mp); 8315 mkdir2->md_state = ATTACHED | MKDIR_PARENT; 8316 mkdir2->md_diradd = dap; 8317 mkdir2->md_jaddref = NULL; 8318 if (MOUNTEDSUJ(mp) == 0) { 8319 mkdir1->md_state |= DEPCOMPLETE; 8320 mkdir2->md_state |= DEPCOMPLETE; 8321 } 8322 /* 8323 * Dependency on "." and ".." being written to disk. 8324 */ 8325 mkdir1->md_buf = newdirbp; 8326 ACQUIRE_LOCK(VFSTOUFS(mp)); 8327 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir1, md_mkdirs); 8328 /* 8329 * We must link the pagedep, allocdirect, and newdirblk for 8330 * the initial file page so the pointer to the new directory 8331 * is not written until the directory contents are live and 8332 * any subsequent additions are not marked live until the 8333 * block is reachable via the inode. 8334 */ 8335 if (pagedep_lookup(mp, newdirbp, newinum, 0, 0, &pagedep) == 0) 8336 panic("setup_newdir: lost pagedep"); 8337 LIST_FOREACH(wk, &newdirbp->b_dep, wk_list) 8338 if (wk->wk_type == D_ALLOCDIRECT) 8339 break; 8340 if (wk == NULL) 8341 panic("setup_newdir: lost allocdirect"); 8342 if (pagedep->pd_state & NEWBLOCK) 8343 panic("setup_newdir: NEWBLOCK already set"); 8344 newblk = WK_NEWBLK(wk); 8345 pagedep->pd_state |= NEWBLOCK; 8346 pagedep->pd_newdirblk = newdirblk; 8347 newdirblk->db_pagedep = pagedep; 8348 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list); 8349 WORKLIST_INSERT(&newdirblk->db_mkdir, &mkdir1->md_list); 8350 /* 8351 * Look up the inodedep for the parent directory so that we 8352 * can link mkdir2 into the pending dotdot jaddref or 8353 * the inode write if there is none. If the inode is 8354 * ALLCOMPLETE and no jaddref is present all dependencies have 8355 * been satisfied and mkdir2 can be freed. 8356 */ 8357 inodedep_lookup(mp, dinum, 0, &inodedep); 8358 if (MOUNTEDSUJ(mp)) { 8359 if (inodedep == NULL) 8360 panic("setup_newdir: Lost parent."); 8361 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 8362 inoreflst); 8363 KASSERT(jaddref != NULL && jaddref->ja_parent == newinum && 8364 (jaddref->ja_state & MKDIR_PARENT), 8365 ("setup_newdir: bad dotdot jaddref %p", jaddref)); 8366 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir2, md_mkdirs); 8367 mkdir2->md_jaddref = jaddref; 8368 jaddref->ja_mkdir = mkdir2; 8369 } else if (inodedep == NULL || 8370 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 8371 dap->da_state &= ~MKDIR_PARENT; 8372 WORKITEM_FREE(mkdir2, D_MKDIR); 8373 mkdir2 = NULL; 8374 } else { 8375 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir2, md_mkdirs); 8376 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir2->md_list); 8377 } 8378 *mkdirp = mkdir2; 8379 8380 return (mkdir1); 8381} 8382 8383/* 8384 * Directory entry addition dependencies. 8385 * 8386 * When adding a new directory entry, the inode (with its incremented link 8387 * count) must be written to disk before the directory entry's pointer to it. 8388 * Also, if the inode is newly allocated, the corresponding freemap must be 8389 * updated (on disk) before the directory entry's pointer. These requirements 8390 * are met via undo/redo on the directory entry's pointer, which consists 8391 * simply of the inode number. 8392 * 8393 * As directory entries are added and deleted, the free space within a 8394 * directory block can become fragmented. The ufs filesystem will compact 8395 * a fragmented directory block to make space for a new entry. When this 8396 * occurs, the offsets of previously added entries change. Any "diradd" 8397 * dependency structures corresponding to these entries must be updated with 8398 * the new offsets. 8399 */ 8400 8401/* 8402 * This routine is called after the in-memory inode's link 8403 * count has been incremented, but before the directory entry's 8404 * pointer to the inode has been set. 8405 */ 8406int 8407softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 8408 struct buf *bp; /* buffer containing directory block */ 8409 struct inode *dp; /* inode for directory */ 8410 off_t diroffset; /* offset of new entry in directory */ 8411 ino_t newinum; /* inode referenced by new directory entry */ 8412 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 8413 int isnewblk; /* entry is in a newly allocated block */ 8414{ 8415 int offset; /* offset of new entry within directory block */ 8416 ufs_lbn_t lbn; /* block in directory containing new entry */ 8417 struct fs *fs; 8418 struct diradd *dap; 8419 struct newblk *newblk; 8420 struct pagedep *pagedep; 8421 struct inodedep *inodedep; 8422 struct newdirblk *newdirblk = 0; 8423 struct mkdir *mkdir1, *mkdir2; 8424 struct jaddref *jaddref; 8425 struct ufsmount *ump; 8426 struct mount *mp; 8427 int isindir; 8428 8429 ump = dp->i_ump; 8430 mp = UFSTOVFS(ump); 8431 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 8432 ("softdep_setup_directory_add called on non-softdep filesystem")); 8433 /* 8434 * Whiteouts have no dependencies. 8435 */ 8436 if (newinum == WINO) { 8437 if (newdirbp != NULL) 8438 bdwrite(newdirbp); 8439 return (0); 8440 } 8441 jaddref = NULL; 8442 mkdir1 = mkdir2 = NULL; 8443 fs = dp->i_fs; 8444 lbn = lblkno(fs, diroffset); 8445 offset = blkoff(fs, diroffset); 8446 dap = malloc(sizeof(struct diradd), M_DIRADD, 8447 M_SOFTDEP_FLAGS|M_ZERO); 8448 workitem_alloc(&dap->da_list, D_DIRADD, mp); 8449 dap->da_offset = offset; 8450 dap->da_newinum = newinum; 8451 dap->da_state = ATTACHED; 8452 LIST_INIT(&dap->da_jwork); 8453 isindir = bp->b_lblkno >= NDADDR; 8454 if (isnewblk && 8455 (isindir ? blkoff(fs, diroffset) : fragoff(fs, diroffset)) == 0) { 8456 newdirblk = malloc(sizeof(struct newdirblk), 8457 M_NEWDIRBLK, M_SOFTDEP_FLAGS); 8458 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 8459 LIST_INIT(&newdirblk->db_mkdir); 8460 } 8461 /* 8462 * If we're creating a new directory setup the dependencies and set 8463 * the dap state to wait for them. Otherwise it's COMPLETE and 8464 * we can move on. 8465 */ 8466 if (newdirbp == NULL) { 8467 dap->da_state |= DEPCOMPLETE; 8468 ACQUIRE_LOCK(ump); 8469 } else { 8470 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 8471 mkdir1 = setup_newdir(dap, newinum, dp->i_number, newdirbp, 8472 &mkdir2); 8473 } 8474 /* 8475 * Link into parent directory pagedep to await its being written. 8476 */ 8477 pagedep_lookup(mp, bp, dp->i_number, lbn, DEPALLOC, &pagedep); 8478#ifdef DEBUG 8479 if (diradd_lookup(pagedep, offset) != NULL) 8480 panic("softdep_setup_directory_add: %p already at off %d\n", 8481 diradd_lookup(pagedep, offset), offset); 8482#endif 8483 dap->da_pagedep = pagedep; 8484 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 8485 da_pdlist); 8486 inodedep_lookup(mp, newinum, DEPALLOC | NODELAY, &inodedep); 8487 /* 8488 * If we're journaling, link the diradd into the jaddref so it 8489 * may be completed after the journal entry is written. Otherwise, 8490 * link the diradd into its inodedep. If the inode is not yet 8491 * written place it on the bufwait list, otherwise do the post-inode 8492 * write processing to put it on the id_pendinghd list. 8493 */ 8494 if (MOUNTEDSUJ(mp)) { 8495 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 8496 inoreflst); 8497 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 8498 ("softdep_setup_directory_add: bad jaddref %p", jaddref)); 8499 jaddref->ja_diroff = diroffset; 8500 jaddref->ja_diradd = dap; 8501 add_to_journal(&jaddref->ja_list); 8502 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 8503 diradd_inode_written(dap, inodedep); 8504 else 8505 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 8506 /* 8507 * Add the journal entries for . and .. links now that the primary 8508 * link is written. 8509 */ 8510 if (mkdir1 != NULL && MOUNTEDSUJ(mp)) { 8511 jaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref, 8512 inoreflst, if_deps); 8513 KASSERT(jaddref != NULL && 8514 jaddref->ja_ino == jaddref->ja_parent && 8515 (jaddref->ja_state & MKDIR_BODY), 8516 ("softdep_setup_directory_add: bad dot jaddref %p", 8517 jaddref)); 8518 mkdir1->md_jaddref = jaddref; 8519 jaddref->ja_mkdir = mkdir1; 8520 /* 8521 * It is important that the dotdot journal entry 8522 * is added prior to the dot entry since dot writes 8523 * both the dot and dotdot links. These both must 8524 * be added after the primary link for the journal 8525 * to remain consistent. 8526 */ 8527 add_to_journal(&mkdir2->md_jaddref->ja_list); 8528 add_to_journal(&jaddref->ja_list); 8529 } 8530 /* 8531 * If we are adding a new directory remember this diradd so that if 8532 * we rename it we can keep the dot and dotdot dependencies. If 8533 * we are adding a new name for an inode that has a mkdiradd we 8534 * must be in rename and we have to move the dot and dotdot 8535 * dependencies to this new name. The old name is being orphaned 8536 * soon. 8537 */ 8538 if (mkdir1 != NULL) { 8539 if (inodedep->id_mkdiradd != NULL) 8540 panic("softdep_setup_directory_add: Existing mkdir"); 8541 inodedep->id_mkdiradd = dap; 8542 } else if (inodedep->id_mkdiradd) 8543 merge_diradd(inodedep, dap); 8544 if (newdirblk) { 8545 /* 8546 * There is nothing to do if we are already tracking 8547 * this block. 8548 */ 8549 if ((pagedep->pd_state & NEWBLOCK) != 0) { 8550 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 8551 FREE_LOCK(ump); 8552 return (0); 8553 } 8554 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk) 8555 == 0) 8556 panic("softdep_setup_directory_add: lost entry"); 8557 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list); 8558 pagedep->pd_state |= NEWBLOCK; 8559 pagedep->pd_newdirblk = newdirblk; 8560 newdirblk->db_pagedep = pagedep; 8561 FREE_LOCK(ump); 8562 /* 8563 * If we extended into an indirect signal direnter to sync. 8564 */ 8565 if (isindir) 8566 return (1); 8567 return (0); 8568 } 8569 FREE_LOCK(ump); 8570 return (0); 8571} 8572 8573/* 8574 * This procedure is called to change the offset of a directory 8575 * entry when compacting a directory block which must be owned 8576 * exclusively by the caller. Note that the actual entry movement 8577 * must be done in this procedure to ensure that no I/O completions 8578 * occur while the move is in progress. 8579 */ 8580void 8581softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize) 8582 struct buf *bp; /* Buffer holding directory block. */ 8583 struct inode *dp; /* inode for directory */ 8584 caddr_t base; /* address of dp->i_offset */ 8585 caddr_t oldloc; /* address of old directory location */ 8586 caddr_t newloc; /* address of new directory location */ 8587 int entrysize; /* size of directory entry */ 8588{ 8589 int offset, oldoffset, newoffset; 8590 struct pagedep *pagedep; 8591 struct jmvref *jmvref; 8592 struct diradd *dap; 8593 struct direct *de; 8594 struct mount *mp; 8595 ufs_lbn_t lbn; 8596 int flags; 8597 8598 mp = UFSTOVFS(dp->i_ump); 8599 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 8600 ("softdep_change_directoryentry_offset called on " 8601 "non-softdep filesystem")); 8602 de = (struct direct *)oldloc; 8603 jmvref = NULL; 8604 flags = 0; 8605 /* 8606 * Moves are always journaled as it would be too complex to 8607 * determine if any affected adds or removes are present in the 8608 * journal. 8609 */ 8610 if (MOUNTEDSUJ(mp)) { 8611 flags = DEPALLOC; 8612 jmvref = newjmvref(dp, de->d_ino, 8613 dp->i_offset + (oldloc - base), 8614 dp->i_offset + (newloc - base)); 8615 } 8616 lbn = lblkno(dp->i_fs, dp->i_offset); 8617 offset = blkoff(dp->i_fs, dp->i_offset); 8618 oldoffset = offset + (oldloc - base); 8619 newoffset = offset + (newloc - base); 8620 ACQUIRE_LOCK(dp->i_ump); 8621 if (pagedep_lookup(mp, bp, dp->i_number, lbn, flags, &pagedep) == 0) 8622 goto done; 8623 dap = diradd_lookup(pagedep, oldoffset); 8624 if (dap) { 8625 dap->da_offset = newoffset; 8626 newoffset = DIRADDHASH(newoffset); 8627 oldoffset = DIRADDHASH(oldoffset); 8628 if ((dap->da_state & ALLCOMPLETE) != ALLCOMPLETE && 8629 newoffset != oldoffset) { 8630 LIST_REMOVE(dap, da_pdlist); 8631 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[newoffset], 8632 dap, da_pdlist); 8633 } 8634 } 8635done: 8636 if (jmvref) { 8637 jmvref->jm_pagedep = pagedep; 8638 LIST_INSERT_HEAD(&pagedep->pd_jmvrefhd, jmvref, jm_deps); 8639 add_to_journal(&jmvref->jm_list); 8640 } 8641 bcopy(oldloc, newloc, entrysize); 8642 FREE_LOCK(dp->i_ump); 8643} 8644 8645/* 8646 * Move the mkdir dependencies and journal work from one diradd to another 8647 * when renaming a directory. The new name must depend on the mkdir deps 8648 * completing as the old name did. Directories can only have one valid link 8649 * at a time so one must be canonical. 8650 */ 8651static void 8652merge_diradd(inodedep, newdap) 8653 struct inodedep *inodedep; 8654 struct diradd *newdap; 8655{ 8656 struct diradd *olddap; 8657 struct mkdir *mkdir, *nextmd; 8658 struct ufsmount *ump; 8659 short state; 8660 8661 olddap = inodedep->id_mkdiradd; 8662 inodedep->id_mkdiradd = newdap; 8663 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8664 newdap->da_state &= ~DEPCOMPLETE; 8665 ump = VFSTOUFS(inodedep->id_list.wk_mp); 8666 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir; 8667 mkdir = nextmd) { 8668 nextmd = LIST_NEXT(mkdir, md_mkdirs); 8669 if (mkdir->md_diradd != olddap) 8670 continue; 8671 mkdir->md_diradd = newdap; 8672 state = mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY); 8673 newdap->da_state |= state; 8674 olddap->da_state &= ~state; 8675 if ((olddap->da_state & 8676 (MKDIR_PARENT | MKDIR_BODY)) == 0) 8677 break; 8678 } 8679 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 8680 panic("merge_diradd: unfound ref"); 8681 } 8682 /* 8683 * Any mkdir related journal items are not safe to be freed until 8684 * the new name is stable. 8685 */ 8686 jwork_move(&newdap->da_jwork, &olddap->da_jwork); 8687 olddap->da_state |= DEPCOMPLETE; 8688 complete_diradd(olddap); 8689} 8690 8691/* 8692 * Move the diradd to the pending list when all diradd dependencies are 8693 * complete. 8694 */ 8695static void 8696complete_diradd(dap) 8697 struct diradd *dap; 8698{ 8699 struct pagedep *pagedep; 8700 8701 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 8702 if (dap->da_state & DIRCHG) 8703 pagedep = dap->da_previous->dm_pagedep; 8704 else 8705 pagedep = dap->da_pagedep; 8706 LIST_REMOVE(dap, da_pdlist); 8707 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 8708 } 8709} 8710 8711/* 8712 * Cancel a diradd when a dirrem overlaps with it. We must cancel the journal 8713 * add entries and conditonally journal the remove. 8714 */ 8715static void 8716cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref) 8717 struct diradd *dap; 8718 struct dirrem *dirrem; 8719 struct jremref *jremref; 8720 struct jremref *dotremref; 8721 struct jremref *dotdotremref; 8722{ 8723 struct inodedep *inodedep; 8724 struct jaddref *jaddref; 8725 struct inoref *inoref; 8726 struct ufsmount *ump; 8727 struct mkdir *mkdir; 8728 8729 /* 8730 * If no remove references were allocated we're on a non-journaled 8731 * filesystem and can skip the cancel step. 8732 */ 8733 if (jremref == NULL) { 8734 free_diradd(dap, NULL); 8735 return; 8736 } 8737 /* 8738 * Cancel the primary name an free it if it does not require 8739 * journaling. 8740 */ 8741 if (inodedep_lookup(dap->da_list.wk_mp, dap->da_newinum, 8742 0, &inodedep) != 0) { 8743 /* Abort the addref that reference this diradd. */ 8744 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 8745 if (inoref->if_list.wk_type != D_JADDREF) 8746 continue; 8747 jaddref = (struct jaddref *)inoref; 8748 if (jaddref->ja_diradd != dap) 8749 continue; 8750 if (cancel_jaddref(jaddref, inodedep, 8751 &dirrem->dm_jwork) == 0) { 8752 free_jremref(jremref); 8753 jremref = NULL; 8754 } 8755 break; 8756 } 8757 } 8758 /* 8759 * Cancel subordinate names and free them if they do not require 8760 * journaling. 8761 */ 8762 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8763 ump = VFSTOUFS(dap->da_list.wk_mp); 8764 LIST_FOREACH(mkdir, &ump->softdep_mkdirlisthd, md_mkdirs) { 8765 if (mkdir->md_diradd != dap) 8766 continue; 8767 if ((jaddref = mkdir->md_jaddref) == NULL) 8768 continue; 8769 mkdir->md_jaddref = NULL; 8770 if (mkdir->md_state & MKDIR_PARENT) { 8771 if (cancel_jaddref(jaddref, NULL, 8772 &dirrem->dm_jwork) == 0) { 8773 free_jremref(dotdotremref); 8774 dotdotremref = NULL; 8775 } 8776 } else { 8777 if (cancel_jaddref(jaddref, inodedep, 8778 &dirrem->dm_jwork) == 0) { 8779 free_jremref(dotremref); 8780 dotremref = NULL; 8781 } 8782 } 8783 } 8784 } 8785 8786 if (jremref) 8787 journal_jremref(dirrem, jremref, inodedep); 8788 if (dotremref) 8789 journal_jremref(dirrem, dotremref, inodedep); 8790 if (dotdotremref) 8791 journal_jremref(dirrem, dotdotremref, NULL); 8792 jwork_move(&dirrem->dm_jwork, &dap->da_jwork); 8793 free_diradd(dap, &dirrem->dm_jwork); 8794} 8795 8796/* 8797 * Free a diradd dependency structure. This routine must be called 8798 * with splbio interrupts blocked. 8799 */ 8800static void 8801free_diradd(dap, wkhd) 8802 struct diradd *dap; 8803 struct workhead *wkhd; 8804{ 8805 struct dirrem *dirrem; 8806 struct pagedep *pagedep; 8807 struct inodedep *inodedep; 8808 struct mkdir *mkdir, *nextmd; 8809 struct ufsmount *ump; 8810 8811 ump = VFSTOUFS(dap->da_list.wk_mp); 8812 LOCK_OWNED(ump); 8813 LIST_REMOVE(dap, da_pdlist); 8814 if (dap->da_state & ONWORKLIST) 8815 WORKLIST_REMOVE(&dap->da_list); 8816 if ((dap->da_state & DIRCHG) == 0) { 8817 pagedep = dap->da_pagedep; 8818 } else { 8819 dirrem = dap->da_previous; 8820 pagedep = dirrem->dm_pagedep; 8821 dirrem->dm_dirinum = pagedep->pd_ino; 8822 dirrem->dm_state |= COMPLETE; 8823 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 8824 add_to_worklist(&dirrem->dm_list, 0); 8825 } 8826 if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum, 8827 0, &inodedep) != 0) 8828 if (inodedep->id_mkdiradd == dap) 8829 inodedep->id_mkdiradd = NULL; 8830 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8831 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir; 8832 mkdir = nextmd) { 8833 nextmd = LIST_NEXT(mkdir, md_mkdirs); 8834 if (mkdir->md_diradd != dap) 8835 continue; 8836 dap->da_state &= 8837 ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)); 8838 LIST_REMOVE(mkdir, md_mkdirs); 8839 if (mkdir->md_state & ONWORKLIST) 8840 WORKLIST_REMOVE(&mkdir->md_list); 8841 if (mkdir->md_jaddref != NULL) 8842 panic("free_diradd: Unexpected jaddref"); 8843 WORKITEM_FREE(mkdir, D_MKDIR); 8844 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 8845 break; 8846 } 8847 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 8848 panic("free_diradd: unfound ref"); 8849 } 8850 if (inodedep) 8851 free_inodedep(inodedep); 8852 /* 8853 * Free any journal segments waiting for the directory write. 8854 */ 8855 handle_jwork(&dap->da_jwork); 8856 WORKITEM_FREE(dap, D_DIRADD); 8857} 8858 8859/* 8860 * Directory entry removal dependencies. 8861 * 8862 * When removing a directory entry, the entry's inode pointer must be 8863 * zero'ed on disk before the corresponding inode's link count is decremented 8864 * (possibly freeing the inode for re-use). This dependency is handled by 8865 * updating the directory entry but delaying the inode count reduction until 8866 * after the directory block has been written to disk. After this point, the 8867 * inode count can be decremented whenever it is convenient. 8868 */ 8869 8870/* 8871 * This routine should be called immediately after removing 8872 * a directory entry. The inode's link count should not be 8873 * decremented by the calling procedure -- the soft updates 8874 * code will do this task when it is safe. 8875 */ 8876void 8877softdep_setup_remove(bp, dp, ip, isrmdir) 8878 struct buf *bp; /* buffer containing directory block */ 8879 struct inode *dp; /* inode for the directory being modified */ 8880 struct inode *ip; /* inode for directory entry being removed */ 8881 int isrmdir; /* indicates if doing RMDIR */ 8882{ 8883 struct dirrem *dirrem, *prevdirrem; 8884 struct inodedep *inodedep; 8885 int direct; 8886 8887 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0, 8888 ("softdep_setup_remove called on non-softdep filesystem")); 8889 /* 8890 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. We want 8891 * newdirrem() to setup the full directory remove which requires 8892 * isrmdir > 1. 8893 */ 8894 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 8895 /* 8896 * Add the dirrem to the inodedep's pending remove list for quick 8897 * discovery later. 8898 */ 8899 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, 8900 &inodedep) == 0) 8901 panic("softdep_setup_remove: Lost inodedep."); 8902 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked")); 8903 dirrem->dm_state |= ONDEPLIST; 8904 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 8905 8906 /* 8907 * If the COMPLETE flag is clear, then there were no active 8908 * entries and we want to roll back to a zeroed entry until 8909 * the new inode is committed to disk. If the COMPLETE flag is 8910 * set then we have deleted an entry that never made it to 8911 * disk. If the entry we deleted resulted from a name change, 8912 * then the old name still resides on disk. We cannot delete 8913 * its inode (returned to us in prevdirrem) until the zeroed 8914 * directory entry gets to disk. The new inode has never been 8915 * referenced on the disk, so can be deleted immediately. 8916 */ 8917 if ((dirrem->dm_state & COMPLETE) == 0) { 8918 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 8919 dm_next); 8920 FREE_LOCK(ip->i_ump); 8921 } else { 8922 if (prevdirrem != NULL) 8923 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 8924 prevdirrem, dm_next); 8925 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 8926 direct = LIST_EMPTY(&dirrem->dm_jremrefhd); 8927 FREE_LOCK(ip->i_ump); 8928 if (direct) 8929 handle_workitem_remove(dirrem, 0); 8930 } 8931} 8932 8933/* 8934 * Check for an entry matching 'offset' on both the pd_dirraddhd list and the 8935 * pd_pendinghd list of a pagedep. 8936 */ 8937static struct diradd * 8938diradd_lookup(pagedep, offset) 8939 struct pagedep *pagedep; 8940 int offset; 8941{ 8942 struct diradd *dap; 8943 8944 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 8945 if (dap->da_offset == offset) 8946 return (dap); 8947 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 8948 if (dap->da_offset == offset) 8949 return (dap); 8950 return (NULL); 8951} 8952 8953/* 8954 * Search for a .. diradd dependency in a directory that is being removed. 8955 * If the directory was renamed to a new parent we have a diradd rather 8956 * than a mkdir for the .. entry. We need to cancel it now before 8957 * it is found in truncate(). 8958 */ 8959static struct jremref * 8960cancel_diradd_dotdot(ip, dirrem, jremref) 8961 struct inode *ip; 8962 struct dirrem *dirrem; 8963 struct jremref *jremref; 8964{ 8965 struct pagedep *pagedep; 8966 struct diradd *dap; 8967 struct worklist *wk; 8968 8969 if (pagedep_lookup(UFSTOVFS(ip->i_ump), NULL, ip->i_number, 0, 0, 8970 &pagedep) == 0) 8971 return (jremref); 8972 dap = diradd_lookup(pagedep, DOTDOT_OFFSET); 8973 if (dap == NULL) 8974 return (jremref); 8975 cancel_diradd(dap, dirrem, jremref, NULL, NULL); 8976 /* 8977 * Mark any journal work as belonging to the parent so it is freed 8978 * with the .. reference. 8979 */ 8980 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list) 8981 wk->wk_state |= MKDIR_PARENT; 8982 return (NULL); 8983} 8984 8985/* 8986 * Cancel the MKDIR_PARENT mkdir component of a diradd when we're going to 8987 * replace it with a dirrem/diradd pair as a result of re-parenting a 8988 * directory. This ensures that we don't simultaneously have a mkdir and 8989 * a diradd for the same .. entry. 8990 */ 8991static struct jremref * 8992cancel_mkdir_dotdot(ip, dirrem, jremref) 8993 struct inode *ip; 8994 struct dirrem *dirrem; 8995 struct jremref *jremref; 8996{ 8997 struct inodedep *inodedep; 8998 struct jaddref *jaddref; 8999 struct ufsmount *ump; 9000 struct mkdir *mkdir; 9001 struct diradd *dap; 9002 9003 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, 9004 &inodedep) == 0) 9005 return (jremref); 9006 dap = inodedep->id_mkdiradd; 9007 if (dap == NULL || (dap->da_state & MKDIR_PARENT) == 0) 9008 return (jremref); 9009 ump = VFSTOUFS(inodedep->id_list.wk_mp); 9010 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir; 9011 mkdir = LIST_NEXT(mkdir, md_mkdirs)) 9012 if (mkdir->md_diradd == dap && mkdir->md_state & MKDIR_PARENT) 9013 break; 9014 if (mkdir == NULL) 9015 panic("cancel_mkdir_dotdot: Unable to find mkdir\n"); 9016 if ((jaddref = mkdir->md_jaddref) != NULL) { 9017 mkdir->md_jaddref = NULL; 9018 jaddref->ja_state &= ~MKDIR_PARENT; 9019 if (inodedep_lookup(UFSTOVFS(ip->i_ump), jaddref->ja_ino, 0, 9020 &inodedep) == 0) 9021 panic("cancel_mkdir_dotdot: Lost parent inodedep"); 9022 if (cancel_jaddref(jaddref, inodedep, &dirrem->dm_jwork)) { 9023 journal_jremref(dirrem, jremref, inodedep); 9024 jremref = NULL; 9025 } 9026 } 9027 if (mkdir->md_state & ONWORKLIST) 9028 WORKLIST_REMOVE(&mkdir->md_list); 9029 mkdir->md_state |= ALLCOMPLETE; 9030 complete_mkdir(mkdir); 9031 return (jremref); 9032} 9033 9034static void 9035journal_jremref(dirrem, jremref, inodedep) 9036 struct dirrem *dirrem; 9037 struct jremref *jremref; 9038 struct inodedep *inodedep; 9039{ 9040 9041 if (inodedep == NULL) 9042 if (inodedep_lookup(jremref->jr_list.wk_mp, 9043 jremref->jr_ref.if_ino, 0, &inodedep) == 0) 9044 panic("journal_jremref: Lost inodedep"); 9045 LIST_INSERT_HEAD(&dirrem->dm_jremrefhd, jremref, jr_deps); 9046 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps); 9047 add_to_journal(&jremref->jr_list); 9048} 9049 9050static void 9051dirrem_journal(dirrem, jremref, dotremref, dotdotremref) 9052 struct dirrem *dirrem; 9053 struct jremref *jremref; 9054 struct jremref *dotremref; 9055 struct jremref *dotdotremref; 9056{ 9057 struct inodedep *inodedep; 9058 9059 9060 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 0, 9061 &inodedep) == 0) 9062 panic("dirrem_journal: Lost inodedep"); 9063 journal_jremref(dirrem, jremref, inodedep); 9064 if (dotremref) 9065 journal_jremref(dirrem, dotremref, inodedep); 9066 if (dotdotremref) 9067 journal_jremref(dirrem, dotdotremref, NULL); 9068} 9069 9070/* 9071 * Allocate a new dirrem if appropriate and return it along with 9072 * its associated pagedep. Called without a lock, returns with lock. 9073 */ 9074static struct dirrem * 9075newdirrem(bp, dp, ip, isrmdir, prevdirremp) 9076 struct buf *bp; /* buffer containing directory block */ 9077 struct inode *dp; /* inode for the directory being modified */ 9078 struct inode *ip; /* inode for directory entry being removed */ 9079 int isrmdir; /* indicates if doing RMDIR */ 9080 struct dirrem **prevdirremp; /* previously referenced inode, if any */ 9081{ 9082 int offset; 9083 ufs_lbn_t lbn; 9084 struct diradd *dap; 9085 struct dirrem *dirrem; 9086 struct pagedep *pagedep; 9087 struct jremref *jremref; 9088 struct jremref *dotremref; 9089 struct jremref *dotdotremref; 9090 struct vnode *dvp; 9091 9092 /* 9093 * Whiteouts have no deletion dependencies. 9094 */ 9095 if (ip == NULL) 9096 panic("newdirrem: whiteout"); 9097 dvp = ITOV(dp); 9098 /* 9099 * If the system is over its limit and our filesystem is 9100 * responsible for more than our share of that usage and 9101 * we are not a snapshot, request some inodedep cleanup. 9102 * Limiting the number of dirrem structures will also limit 9103 * the number of freefile and freeblks structures. 9104 */ 9105 ACQUIRE_LOCK(ip->i_ump); 9106 while (!IS_SNAPSHOT(ip) && dep_current[D_DIRREM] > max_softdeps / 2 && 9107 ip->i_ump->softdep_curdeps[D_DIRREM] > 9108 (max_softdeps / 2) / stat_flush_threads) 9109 (void) request_cleanup(ITOV(dp)->v_mount, FLUSH_BLOCKS); 9110 FREE_LOCK(ip->i_ump); 9111 dirrem = malloc(sizeof(struct dirrem), 9112 M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO); 9113 workitem_alloc(&dirrem->dm_list, D_DIRREM, dvp->v_mount); 9114 LIST_INIT(&dirrem->dm_jremrefhd); 9115 LIST_INIT(&dirrem->dm_jwork); 9116 dirrem->dm_state = isrmdir ? RMDIR : 0; 9117 dirrem->dm_oldinum = ip->i_number; 9118 *prevdirremp = NULL; 9119 /* 9120 * Allocate remove reference structures to track journal write 9121 * dependencies. We will always have one for the link and 9122 * when doing directories we will always have one more for dot. 9123 * When renaming a directory we skip the dotdot link change so 9124 * this is not needed. 9125 */ 9126 jremref = dotremref = dotdotremref = NULL; 9127 if (DOINGSUJ(dvp)) { 9128 if (isrmdir) { 9129 jremref = newjremref(dirrem, dp, ip, dp->i_offset, 9130 ip->i_effnlink + 2); 9131 dotremref = newjremref(dirrem, ip, ip, DOT_OFFSET, 9132 ip->i_effnlink + 1); 9133 dotdotremref = newjremref(dirrem, ip, dp, DOTDOT_OFFSET, 9134 dp->i_effnlink + 1); 9135 dotdotremref->jr_state |= MKDIR_PARENT; 9136 } else 9137 jremref = newjremref(dirrem, dp, ip, dp->i_offset, 9138 ip->i_effnlink + 1); 9139 } 9140 ACQUIRE_LOCK(ip->i_ump); 9141 lbn = lblkno(dp->i_fs, dp->i_offset); 9142 offset = blkoff(dp->i_fs, dp->i_offset); 9143 pagedep_lookup(UFSTOVFS(dp->i_ump), bp, dp->i_number, lbn, DEPALLOC, 9144 &pagedep); 9145 dirrem->dm_pagedep = pagedep; 9146 dirrem->dm_offset = offset; 9147 /* 9148 * If we're renaming a .. link to a new directory, cancel any 9149 * existing MKDIR_PARENT mkdir. If it has already been canceled 9150 * the jremref is preserved for any potential diradd in this 9151 * location. This can not coincide with a rmdir. 9152 */ 9153 if (dp->i_offset == DOTDOT_OFFSET) { 9154 if (isrmdir) 9155 panic("newdirrem: .. directory change during remove?"); 9156 jremref = cancel_mkdir_dotdot(dp, dirrem, jremref); 9157 } 9158 /* 9159 * If we're removing a directory search for the .. dependency now and 9160 * cancel it. Any pending journal work will be added to the dirrem 9161 * to be completed when the workitem remove completes. 9162 */ 9163 if (isrmdir) 9164 dotdotremref = cancel_diradd_dotdot(ip, dirrem, dotdotremref); 9165 /* 9166 * Check for a diradd dependency for the same directory entry. 9167 * If present, then both dependencies become obsolete and can 9168 * be de-allocated. 9169 */ 9170 dap = diradd_lookup(pagedep, offset); 9171 if (dap == NULL) { 9172 /* 9173 * Link the jremref structures into the dirrem so they are 9174 * written prior to the pagedep. 9175 */ 9176 if (jremref) 9177 dirrem_journal(dirrem, jremref, dotremref, 9178 dotdotremref); 9179 return (dirrem); 9180 } 9181 /* 9182 * Must be ATTACHED at this point. 9183 */ 9184 if ((dap->da_state & ATTACHED) == 0) 9185 panic("newdirrem: not ATTACHED"); 9186 if (dap->da_newinum != ip->i_number) 9187 panic("newdirrem: inum %ju should be %ju", 9188 (uintmax_t)ip->i_number, (uintmax_t)dap->da_newinum); 9189 /* 9190 * If we are deleting a changed name that never made it to disk, 9191 * then return the dirrem describing the previous inode (which 9192 * represents the inode currently referenced from this entry on disk). 9193 */ 9194 if ((dap->da_state & DIRCHG) != 0) { 9195 *prevdirremp = dap->da_previous; 9196 dap->da_state &= ~DIRCHG; 9197 dap->da_pagedep = pagedep; 9198 } 9199 /* 9200 * We are deleting an entry that never made it to disk. 9201 * Mark it COMPLETE so we can delete its inode immediately. 9202 */ 9203 dirrem->dm_state |= COMPLETE; 9204 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref); 9205#ifdef SUJ_DEBUG 9206 if (isrmdir == 0) { 9207 struct worklist *wk; 9208 9209 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list) 9210 if (wk->wk_state & (MKDIR_BODY | MKDIR_PARENT)) 9211 panic("bad wk %p (0x%X)\n", wk, wk->wk_state); 9212 } 9213#endif 9214 9215 return (dirrem); 9216} 9217 9218/* 9219 * Directory entry change dependencies. 9220 * 9221 * Changing an existing directory entry requires that an add operation 9222 * be completed first followed by a deletion. The semantics for the addition 9223 * are identical to the description of adding a new entry above except 9224 * that the rollback is to the old inode number rather than zero. Once 9225 * the addition dependency is completed, the removal is done as described 9226 * in the removal routine above. 9227 */ 9228 9229/* 9230 * This routine should be called immediately after changing 9231 * a directory entry. The inode's link count should not be 9232 * decremented by the calling procedure -- the soft updates 9233 * code will perform this task when it is safe. 9234 */ 9235void 9236softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 9237 struct buf *bp; /* buffer containing directory block */ 9238 struct inode *dp; /* inode for the directory being modified */ 9239 struct inode *ip; /* inode for directory entry being removed */ 9240 ino_t newinum; /* new inode number for changed entry */ 9241 int isrmdir; /* indicates if doing RMDIR */ 9242{ 9243 int offset; 9244 struct diradd *dap = NULL; 9245 struct dirrem *dirrem, *prevdirrem; 9246 struct pagedep *pagedep; 9247 struct inodedep *inodedep; 9248 struct jaddref *jaddref; 9249 struct mount *mp; 9250 9251 offset = blkoff(dp->i_fs, dp->i_offset); 9252 mp = UFSTOVFS(dp->i_ump); 9253 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 9254 ("softdep_setup_directory_change called on non-softdep filesystem")); 9255 9256 /* 9257 * Whiteouts do not need diradd dependencies. 9258 */ 9259 if (newinum != WINO) { 9260 dap = malloc(sizeof(struct diradd), 9261 M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO); 9262 workitem_alloc(&dap->da_list, D_DIRADD, mp); 9263 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 9264 dap->da_offset = offset; 9265 dap->da_newinum = newinum; 9266 LIST_INIT(&dap->da_jwork); 9267 } 9268 9269 /* 9270 * Allocate a new dirrem and ACQUIRE_LOCK. 9271 */ 9272 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 9273 pagedep = dirrem->dm_pagedep; 9274 /* 9275 * The possible values for isrmdir: 9276 * 0 - non-directory file rename 9277 * 1 - directory rename within same directory 9278 * inum - directory rename to new directory of given inode number 9279 * When renaming to a new directory, we are both deleting and 9280 * creating a new directory entry, so the link count on the new 9281 * directory should not change. Thus we do not need the followup 9282 * dirrem which is usually done in handle_workitem_remove. We set 9283 * the DIRCHG flag to tell handle_workitem_remove to skip the 9284 * followup dirrem. 9285 */ 9286 if (isrmdir > 1) 9287 dirrem->dm_state |= DIRCHG; 9288 9289 /* 9290 * Whiteouts have no additional dependencies, 9291 * so just put the dirrem on the correct list. 9292 */ 9293 if (newinum == WINO) { 9294 if ((dirrem->dm_state & COMPLETE) == 0) { 9295 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 9296 dm_next); 9297 } else { 9298 dirrem->dm_dirinum = pagedep->pd_ino; 9299 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 9300 add_to_worklist(&dirrem->dm_list, 0); 9301 } 9302 FREE_LOCK(dp->i_ump); 9303 return; 9304 } 9305 /* 9306 * Add the dirrem to the inodedep's pending remove list for quick 9307 * discovery later. A valid nlinkdelta ensures that this lookup 9308 * will not fail. 9309 */ 9310 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 9311 panic("softdep_setup_directory_change: Lost inodedep."); 9312 dirrem->dm_state |= ONDEPLIST; 9313 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 9314 9315 /* 9316 * If the COMPLETE flag is clear, then there were no active 9317 * entries and we want to roll back to the previous inode until 9318 * the new inode is committed to disk. If the COMPLETE flag is 9319 * set, then we have deleted an entry that never made it to disk. 9320 * If the entry we deleted resulted from a name change, then the old 9321 * inode reference still resides on disk. Any rollback that we do 9322 * needs to be to that old inode (returned to us in prevdirrem). If 9323 * the entry we deleted resulted from a create, then there is 9324 * no entry on the disk, so we want to roll back to zero rather 9325 * than the uncommitted inode. In either of the COMPLETE cases we 9326 * want to immediately free the unwritten and unreferenced inode. 9327 */ 9328 if ((dirrem->dm_state & COMPLETE) == 0) { 9329 dap->da_previous = dirrem; 9330 } else { 9331 if (prevdirrem != NULL) { 9332 dap->da_previous = prevdirrem; 9333 } else { 9334 dap->da_state &= ~DIRCHG; 9335 dap->da_pagedep = pagedep; 9336 } 9337 dirrem->dm_dirinum = pagedep->pd_ino; 9338 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 9339 add_to_worklist(&dirrem->dm_list, 0); 9340 } 9341 /* 9342 * Lookup the jaddref for this journal entry. We must finish 9343 * initializing it and make the diradd write dependent on it. 9344 * If we're not journaling, put it on the id_bufwait list if the 9345 * inode is not yet written. If it is written, do the post-inode 9346 * write processing to put it on the id_pendinghd list. 9347 */ 9348 inodedep_lookup(mp, newinum, DEPALLOC | NODELAY, &inodedep); 9349 if (MOUNTEDSUJ(mp)) { 9350 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 9351 inoreflst); 9352 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 9353 ("softdep_setup_directory_change: bad jaddref %p", 9354 jaddref)); 9355 jaddref->ja_diroff = dp->i_offset; 9356 jaddref->ja_diradd = dap; 9357 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 9358 dap, da_pdlist); 9359 add_to_journal(&jaddref->ja_list); 9360 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 9361 dap->da_state |= COMPLETE; 9362 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 9363 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 9364 } else { 9365 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 9366 dap, da_pdlist); 9367 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 9368 } 9369 /* 9370 * If we're making a new name for a directory that has not been 9371 * committed when need to move the dot and dotdot references to 9372 * this new name. 9373 */ 9374 if (inodedep->id_mkdiradd && dp->i_offset != DOTDOT_OFFSET) 9375 merge_diradd(inodedep, dap); 9376 FREE_LOCK(dp->i_ump); 9377} 9378 9379/* 9380 * Called whenever the link count on an inode is changed. 9381 * It creates an inode dependency so that the new reference(s) 9382 * to the inode cannot be committed to disk until the updated 9383 * inode has been written. 9384 */ 9385void 9386softdep_change_linkcnt(ip) 9387 struct inode *ip; /* the inode with the increased link count */ 9388{ 9389 struct inodedep *inodedep; 9390 int dflags; 9391 9392 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0, 9393 ("softdep_change_linkcnt called on non-softdep filesystem")); 9394 ACQUIRE_LOCK(ip->i_ump); 9395 dflags = DEPALLOC; 9396 if (IS_SNAPSHOT(ip)) 9397 dflags |= NODELAY; 9398 inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, dflags, &inodedep); 9399 if (ip->i_nlink < ip->i_effnlink) 9400 panic("softdep_change_linkcnt: bad delta"); 9401 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9402 FREE_LOCK(ip->i_ump); 9403} 9404 9405/* 9406 * Attach a sbdep dependency to the superblock buf so that we can keep 9407 * track of the head of the linked list of referenced but unlinked inodes. 9408 */ 9409void 9410softdep_setup_sbupdate(ump, fs, bp) 9411 struct ufsmount *ump; 9412 struct fs *fs; 9413 struct buf *bp; 9414{ 9415 struct sbdep *sbdep; 9416 struct worklist *wk; 9417 9418 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 9419 ("softdep_setup_sbupdate called on non-softdep filesystem")); 9420 LIST_FOREACH(wk, &bp->b_dep, wk_list) 9421 if (wk->wk_type == D_SBDEP) 9422 break; 9423 if (wk != NULL) 9424 return; 9425 sbdep = malloc(sizeof(struct sbdep), M_SBDEP, M_SOFTDEP_FLAGS); 9426 workitem_alloc(&sbdep->sb_list, D_SBDEP, UFSTOVFS(ump)); 9427 sbdep->sb_fs = fs; 9428 sbdep->sb_ump = ump; 9429 ACQUIRE_LOCK(ump); 9430 WORKLIST_INSERT(&bp->b_dep, &sbdep->sb_list); 9431 FREE_LOCK(ump); 9432} 9433 9434/* 9435 * Return the first unlinked inodedep which is ready to be the head of the 9436 * list. The inodedep and all those after it must have valid next pointers. 9437 */ 9438static struct inodedep * 9439first_unlinked_inodedep(ump) 9440 struct ufsmount *ump; 9441{ 9442 struct inodedep *inodedep; 9443 struct inodedep *idp; 9444 9445 LOCK_OWNED(ump); 9446 for (inodedep = TAILQ_LAST(&ump->softdep_unlinked, inodedeplst); 9447 inodedep; inodedep = idp) { 9448 if ((inodedep->id_state & UNLINKNEXT) == 0) 9449 return (NULL); 9450 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9451 if (idp == NULL || (idp->id_state & UNLINKNEXT) == 0) 9452 break; 9453 if ((inodedep->id_state & UNLINKPREV) == 0) 9454 break; 9455 } 9456 return (inodedep); 9457} 9458 9459/* 9460 * Set the sujfree unlinked head pointer prior to writing a superblock. 9461 */ 9462static void 9463initiate_write_sbdep(sbdep) 9464 struct sbdep *sbdep; 9465{ 9466 struct inodedep *inodedep; 9467 struct fs *bpfs; 9468 struct fs *fs; 9469 9470 bpfs = sbdep->sb_fs; 9471 fs = sbdep->sb_ump->um_fs; 9472 inodedep = first_unlinked_inodedep(sbdep->sb_ump); 9473 if (inodedep) { 9474 fs->fs_sujfree = inodedep->id_ino; 9475 inodedep->id_state |= UNLINKPREV; 9476 } else 9477 fs->fs_sujfree = 0; 9478 bpfs->fs_sujfree = fs->fs_sujfree; 9479} 9480 9481/* 9482 * After a superblock is written determine whether it must be written again 9483 * due to a changing unlinked list head. 9484 */ 9485static int 9486handle_written_sbdep(sbdep, bp) 9487 struct sbdep *sbdep; 9488 struct buf *bp; 9489{ 9490 struct inodedep *inodedep; 9491 struct mount *mp; 9492 struct fs *fs; 9493 9494 LOCK_OWNED(sbdep->sb_ump); 9495 fs = sbdep->sb_fs; 9496 mp = UFSTOVFS(sbdep->sb_ump); 9497 /* 9498 * If the superblock doesn't match the in-memory list start over. 9499 */ 9500 inodedep = first_unlinked_inodedep(sbdep->sb_ump); 9501 if ((inodedep && fs->fs_sujfree != inodedep->id_ino) || 9502 (inodedep == NULL && fs->fs_sujfree != 0)) { 9503 bdirty(bp); 9504 return (1); 9505 } 9506 WORKITEM_FREE(sbdep, D_SBDEP); 9507 if (fs->fs_sujfree == 0) 9508 return (0); 9509 /* 9510 * Now that we have a record of this inode in stable store allow it 9511 * to be written to free up pending work. Inodes may see a lot of 9512 * write activity after they are unlinked which we must not hold up. 9513 */ 9514 for (; inodedep != NULL; inodedep = TAILQ_NEXT(inodedep, id_unlinked)) { 9515 if ((inodedep->id_state & UNLINKLINKS) != UNLINKLINKS) 9516 panic("handle_written_sbdep: Bad inodedep %p (0x%X)", 9517 inodedep, inodedep->id_state); 9518 if (inodedep->id_state & UNLINKONLIST) 9519 break; 9520 inodedep->id_state |= DEPCOMPLETE | UNLINKONLIST; 9521 } 9522 9523 return (0); 9524} 9525 9526/* 9527 * Mark an inodedep as unlinked and insert it into the in-memory unlinked list. 9528 */ 9529static void 9530unlinked_inodedep(mp, inodedep) 9531 struct mount *mp; 9532 struct inodedep *inodedep; 9533{ 9534 struct ufsmount *ump; 9535 9536 ump = VFSTOUFS(mp); 9537 LOCK_OWNED(ump); 9538 if (MOUNTEDSUJ(mp) == 0) 9539 return; 9540 ump->um_fs->fs_fmod = 1; 9541 if (inodedep->id_state & UNLINKED) 9542 panic("unlinked_inodedep: %p already unlinked\n", inodedep); 9543 inodedep->id_state |= UNLINKED; 9544 TAILQ_INSERT_HEAD(&ump->softdep_unlinked, inodedep, id_unlinked); 9545} 9546 9547/* 9548 * Remove an inodedep from the unlinked inodedep list. This may require 9549 * disk writes if the inode has made it that far. 9550 */ 9551static void 9552clear_unlinked_inodedep(inodedep) 9553 struct inodedep *inodedep; 9554{ 9555 struct ufsmount *ump; 9556 struct inodedep *idp; 9557 struct inodedep *idn; 9558 struct fs *fs; 9559 struct buf *bp; 9560 ino_t ino; 9561 ino_t nino; 9562 ino_t pino; 9563 int error; 9564 9565 ump = VFSTOUFS(inodedep->id_list.wk_mp); 9566 fs = ump->um_fs; 9567 ino = inodedep->id_ino; 9568 error = 0; 9569 for (;;) { 9570 LOCK_OWNED(ump); 9571 KASSERT((inodedep->id_state & UNLINKED) != 0, 9572 ("clear_unlinked_inodedep: inodedep %p not unlinked", 9573 inodedep)); 9574 /* 9575 * If nothing has yet been written simply remove us from 9576 * the in memory list and return. This is the most common 9577 * case where handle_workitem_remove() loses the final 9578 * reference. 9579 */ 9580 if ((inodedep->id_state & UNLINKLINKS) == 0) 9581 break; 9582 /* 9583 * If we have a NEXT pointer and no PREV pointer we can simply 9584 * clear NEXT's PREV and remove ourselves from the list. Be 9585 * careful not to clear PREV if the superblock points at 9586 * next as well. 9587 */ 9588 idn = TAILQ_NEXT(inodedep, id_unlinked); 9589 if ((inodedep->id_state & UNLINKLINKS) == UNLINKNEXT) { 9590 if (idn && fs->fs_sujfree != idn->id_ino) 9591 idn->id_state &= ~UNLINKPREV; 9592 break; 9593 } 9594 /* 9595 * Here we have an inodedep which is actually linked into 9596 * the list. We must remove it by forcing a write to the 9597 * link before us, whether it be the superblock or an inode. 9598 * Unfortunately the list may change while we're waiting 9599 * on the buf lock for either resource so we must loop until 9600 * we lock the right one. If both the superblock and an 9601 * inode point to this inode we must clear the inode first 9602 * followed by the superblock. 9603 */ 9604 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9605 pino = 0; 9606 if (idp && (idp->id_state & UNLINKNEXT)) 9607 pino = idp->id_ino; 9608 FREE_LOCK(ump); 9609 if (pino == 0) { 9610 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc), 9611 (int)fs->fs_sbsize, 0, 0, 0); 9612 } else { 9613 error = bread(ump->um_devvp, 9614 fsbtodb(fs, ino_to_fsba(fs, pino)), 9615 (int)fs->fs_bsize, NOCRED, &bp); 9616 if (error) 9617 brelse(bp); 9618 } 9619 ACQUIRE_LOCK(ump); 9620 if (error) 9621 break; 9622 /* If the list has changed restart the loop. */ 9623 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9624 nino = 0; 9625 if (idp && (idp->id_state & UNLINKNEXT)) 9626 nino = idp->id_ino; 9627 if (nino != pino || 9628 (inodedep->id_state & UNLINKPREV) != UNLINKPREV) { 9629 FREE_LOCK(ump); 9630 brelse(bp); 9631 ACQUIRE_LOCK(ump); 9632 continue; 9633 } 9634 nino = 0; 9635 idn = TAILQ_NEXT(inodedep, id_unlinked); 9636 if (idn) 9637 nino = idn->id_ino; 9638 /* 9639 * Remove us from the in memory list. After this we cannot 9640 * access the inodedep. 9641 */ 9642 KASSERT((inodedep->id_state & UNLINKED) != 0, 9643 ("clear_unlinked_inodedep: inodedep %p not unlinked", 9644 inodedep)); 9645 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST); 9646 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked); 9647 FREE_LOCK(ump); 9648 /* 9649 * The predecessor's next pointer is manually updated here 9650 * so that the NEXT flag is never cleared for an element 9651 * that is in the list. 9652 */ 9653 if (pino == 0) { 9654 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize); 9655 ffs_oldfscompat_write((struct fs *)bp->b_data, ump); 9656 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data, 9657 bp); 9658 } else if (fs->fs_magic == FS_UFS1_MAGIC) 9659 ((struct ufs1_dinode *)bp->b_data + 9660 ino_to_fsbo(fs, pino))->di_freelink = nino; 9661 else 9662 ((struct ufs2_dinode *)bp->b_data + 9663 ino_to_fsbo(fs, pino))->di_freelink = nino; 9664 /* 9665 * If the bwrite fails we have no recourse to recover. The 9666 * filesystem is corrupted already. 9667 */ 9668 bwrite(bp); 9669 ACQUIRE_LOCK(ump); 9670 /* 9671 * If the superblock pointer still needs to be cleared force 9672 * a write here. 9673 */ 9674 if (fs->fs_sujfree == ino) { 9675 FREE_LOCK(ump); 9676 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc), 9677 (int)fs->fs_sbsize, 0, 0, 0); 9678 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize); 9679 ffs_oldfscompat_write((struct fs *)bp->b_data, ump); 9680 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data, 9681 bp); 9682 bwrite(bp); 9683 ACQUIRE_LOCK(ump); 9684 } 9685 9686 if (fs->fs_sujfree != ino) 9687 return; 9688 panic("clear_unlinked_inodedep: Failed to clear free head"); 9689 } 9690 if (inodedep->id_ino == fs->fs_sujfree) 9691 panic("clear_unlinked_inodedep: Freeing head of free list"); 9692 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST); 9693 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked); 9694 return; 9695} 9696 9697/* 9698 * This workitem decrements the inode's link count. 9699 * If the link count reaches zero, the file is removed. 9700 */ 9701static int 9702handle_workitem_remove(dirrem, flags) 9703 struct dirrem *dirrem; 9704 int flags; 9705{ 9706 struct inodedep *inodedep; 9707 struct workhead dotdotwk; 9708 struct worklist *wk; 9709 struct ufsmount *ump; 9710 struct mount *mp; 9711 struct vnode *vp; 9712 struct inode *ip; 9713 ino_t oldinum; 9714 9715 if (dirrem->dm_state & ONWORKLIST) 9716 panic("handle_workitem_remove: dirrem %p still on worklist", 9717 dirrem); 9718 oldinum = dirrem->dm_oldinum; 9719 mp = dirrem->dm_list.wk_mp; 9720 ump = VFSTOUFS(mp); 9721 flags |= LK_EXCLUSIVE; 9722 if (ffs_vgetf(mp, oldinum, flags, &vp, FFSV_FORCEINSMQ) != 0) 9723 return (EBUSY); 9724 ip = VTOI(vp); 9725 ACQUIRE_LOCK(ump); 9726 if ((inodedep_lookup(mp, oldinum, 0, &inodedep)) == 0) 9727 panic("handle_workitem_remove: lost inodedep"); 9728 if (dirrem->dm_state & ONDEPLIST) 9729 LIST_REMOVE(dirrem, dm_inonext); 9730 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd), 9731 ("handle_workitem_remove: Journal entries not written.")); 9732 9733 /* 9734 * Move all dependencies waiting on the remove to complete 9735 * from the dirrem to the inode inowait list to be completed 9736 * after the inode has been updated and written to disk. Any 9737 * marked MKDIR_PARENT are saved to be completed when the .. ref 9738 * is removed. 9739 */ 9740 LIST_INIT(&dotdotwk); 9741 while ((wk = LIST_FIRST(&dirrem->dm_jwork)) != NULL) { 9742 WORKLIST_REMOVE(wk); 9743 if (wk->wk_state & MKDIR_PARENT) { 9744 wk->wk_state &= ~MKDIR_PARENT; 9745 WORKLIST_INSERT(&dotdotwk, wk); 9746 continue; 9747 } 9748 WORKLIST_INSERT(&inodedep->id_inowait, wk); 9749 } 9750 LIST_SWAP(&dirrem->dm_jwork, &dotdotwk, worklist, wk_list); 9751 /* 9752 * Normal file deletion. 9753 */ 9754 if ((dirrem->dm_state & RMDIR) == 0) { 9755 ip->i_nlink--; 9756 DIP_SET(ip, i_nlink, ip->i_nlink); 9757 ip->i_flag |= IN_CHANGE; 9758 if (ip->i_nlink < ip->i_effnlink) 9759 panic("handle_workitem_remove: bad file delta"); 9760 if (ip->i_nlink == 0) 9761 unlinked_inodedep(mp, inodedep); 9762 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9763 KASSERT(LIST_EMPTY(&dirrem->dm_jwork), 9764 ("handle_workitem_remove: worklist not empty. %s", 9765 TYPENAME(LIST_FIRST(&dirrem->dm_jwork)->wk_type))); 9766 WORKITEM_FREE(dirrem, D_DIRREM); 9767 FREE_LOCK(ump); 9768 goto out; 9769 } 9770 /* 9771 * Directory deletion. Decrement reference count for both the 9772 * just deleted parent directory entry and the reference for ".". 9773 * Arrange to have the reference count on the parent decremented 9774 * to account for the loss of "..". 9775 */ 9776 ip->i_nlink -= 2; 9777 DIP_SET(ip, i_nlink, ip->i_nlink); 9778 ip->i_flag |= IN_CHANGE; 9779 if (ip->i_nlink < ip->i_effnlink) 9780 panic("handle_workitem_remove: bad dir delta"); 9781 if (ip->i_nlink == 0) 9782 unlinked_inodedep(mp, inodedep); 9783 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9784 /* 9785 * Rename a directory to a new parent. Since, we are both deleting 9786 * and creating a new directory entry, the link count on the new 9787 * directory should not change. Thus we skip the followup dirrem. 9788 */ 9789 if (dirrem->dm_state & DIRCHG) { 9790 KASSERT(LIST_EMPTY(&dirrem->dm_jwork), 9791 ("handle_workitem_remove: DIRCHG and worklist not empty.")); 9792 WORKITEM_FREE(dirrem, D_DIRREM); 9793 FREE_LOCK(ump); 9794 goto out; 9795 } 9796 dirrem->dm_state = ONDEPLIST; 9797 dirrem->dm_oldinum = dirrem->dm_dirinum; 9798 /* 9799 * Place the dirrem on the parent's diremhd list. 9800 */ 9801 if (inodedep_lookup(mp, dirrem->dm_oldinum, 0, &inodedep) == 0) 9802 panic("handle_workitem_remove: lost dir inodedep"); 9803 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 9804 /* 9805 * If the allocated inode has never been written to disk, then 9806 * the on-disk inode is zero'ed and we can remove the file 9807 * immediately. When journaling if the inode has been marked 9808 * unlinked and not DEPCOMPLETE we know it can never be written. 9809 */ 9810 inodedep_lookup(mp, oldinum, 0, &inodedep); 9811 if (inodedep == NULL || 9812 (inodedep->id_state & (DEPCOMPLETE | UNLINKED)) == UNLINKED || 9813 check_inode_unwritten(inodedep)) { 9814 FREE_LOCK(ump); 9815 vput(vp); 9816 return handle_workitem_remove(dirrem, flags); 9817 } 9818 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 9819 FREE_LOCK(ump); 9820 ip->i_flag |= IN_CHANGE; 9821out: 9822 ffs_update(vp, 0); 9823 vput(vp); 9824 return (0); 9825} 9826 9827/* 9828 * Inode de-allocation dependencies. 9829 * 9830 * When an inode's link count is reduced to zero, it can be de-allocated. We 9831 * found it convenient to postpone de-allocation until after the inode is 9832 * written to disk with its new link count (zero). At this point, all of the 9833 * on-disk inode's block pointers are nullified and, with careful dependency 9834 * list ordering, all dependencies related to the inode will be satisfied and 9835 * the corresponding dependency structures de-allocated. So, if/when the 9836 * inode is reused, there will be no mixing of old dependencies with new 9837 * ones. This artificial dependency is set up by the block de-allocation 9838 * procedure above (softdep_setup_freeblocks) and completed by the 9839 * following procedure. 9840 */ 9841static void 9842handle_workitem_freefile(freefile) 9843 struct freefile *freefile; 9844{ 9845 struct workhead wkhd; 9846 struct fs *fs; 9847 struct inodedep *idp; 9848 struct ufsmount *ump; 9849 int error; 9850 9851 ump = VFSTOUFS(freefile->fx_list.wk_mp); 9852 fs = ump->um_fs; 9853#ifdef DEBUG 9854 ACQUIRE_LOCK(ump); 9855 error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp); 9856 FREE_LOCK(ump); 9857 if (error) 9858 panic("handle_workitem_freefile: inodedep %p survived", idp); 9859#endif 9860 UFS_LOCK(ump); 9861 fs->fs_pendinginodes -= 1; 9862 UFS_UNLOCK(ump); 9863 LIST_INIT(&wkhd); 9864 LIST_SWAP(&freefile->fx_jwork, &wkhd, worklist, wk_list); 9865 if ((error = ffs_freefile(ump, fs, freefile->fx_devvp, 9866 freefile->fx_oldinum, freefile->fx_mode, &wkhd)) != 0) 9867 softdep_error("handle_workitem_freefile", error); 9868 ACQUIRE_LOCK(ump); 9869 WORKITEM_FREE(freefile, D_FREEFILE); 9870 FREE_LOCK(ump); 9871} 9872 9873 9874/* 9875 * Helper function which unlinks marker element from work list and returns 9876 * the next element on the list. 9877 */ 9878static __inline struct worklist * 9879markernext(struct worklist *marker) 9880{ 9881 struct worklist *next; 9882 9883 next = LIST_NEXT(marker, wk_list); 9884 LIST_REMOVE(marker, wk_list); 9885 return next; 9886} 9887 9888/* 9889 * Disk writes. 9890 * 9891 * The dependency structures constructed above are most actively used when file 9892 * system blocks are written to disk. No constraints are placed on when a 9893 * block can be written, but unsatisfied update dependencies are made safe by 9894 * modifying (or replacing) the source memory for the duration of the disk 9895 * write. When the disk write completes, the memory block is again brought 9896 * up-to-date. 9897 * 9898 * In-core inode structure reclamation. 9899 * 9900 * Because there are a finite number of "in-core" inode structures, they are 9901 * reused regularly. By transferring all inode-related dependencies to the 9902 * in-memory inode block and indexing them separately (via "inodedep"s), we 9903 * can allow "in-core" inode structures to be reused at any time and avoid 9904 * any increase in contention. 9905 * 9906 * Called just before entering the device driver to initiate a new disk I/O. 9907 * The buffer must be locked, thus, no I/O completion operations can occur 9908 * while we are manipulating its associated dependencies. 9909 */ 9910static void 9911softdep_disk_io_initiation(bp) 9912 struct buf *bp; /* structure describing disk write to occur */ 9913{ 9914 struct worklist *wk; 9915 struct worklist marker; 9916 struct inodedep *inodedep; 9917 struct freeblks *freeblks; 9918 struct jblkdep *jblkdep; 9919 struct newblk *newblk; 9920 struct ufsmount *ump; 9921 9922 /* 9923 * We only care about write operations. There should never 9924 * be dependencies for reads. 9925 */ 9926 if (bp->b_iocmd != BIO_WRITE) 9927 panic("softdep_disk_io_initiation: not write"); 9928 9929 if (bp->b_vflags & BV_BKGRDINPROG) 9930 panic("softdep_disk_io_initiation: Writing buffer with " 9931 "background write in progress: %p", bp); 9932 9933 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL) 9934 return; 9935 ump = VFSTOUFS(wk->wk_mp); 9936 9937 marker.wk_type = D_LAST + 1; /* Not a normal workitem */ 9938 PHOLD(curproc); /* Don't swap out kernel stack */ 9939 ACQUIRE_LOCK(ump); 9940 /* 9941 * Do any necessary pre-I/O processing. 9942 */ 9943 for (wk = LIST_FIRST(&bp->b_dep); wk != NULL; 9944 wk = markernext(&marker)) { 9945 LIST_INSERT_AFTER(wk, &marker, wk_list); 9946 switch (wk->wk_type) { 9947 9948 case D_PAGEDEP: 9949 initiate_write_filepage(WK_PAGEDEP(wk), bp); 9950 continue; 9951 9952 case D_INODEDEP: 9953 inodedep = WK_INODEDEP(wk); 9954 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) 9955 initiate_write_inodeblock_ufs1(inodedep, bp); 9956 else 9957 initiate_write_inodeblock_ufs2(inodedep, bp); 9958 continue; 9959 9960 case D_INDIRDEP: 9961 initiate_write_indirdep(WK_INDIRDEP(wk), bp); 9962 continue; 9963 9964 case D_BMSAFEMAP: 9965 initiate_write_bmsafemap(WK_BMSAFEMAP(wk), bp); 9966 continue; 9967 9968 case D_JSEG: 9969 WK_JSEG(wk)->js_buf = NULL; 9970 continue; 9971 9972 case D_FREEBLKS: 9973 freeblks = WK_FREEBLKS(wk); 9974 jblkdep = LIST_FIRST(&freeblks->fb_jblkdephd); 9975 /* 9976 * We have to wait for the freeblks to be journaled 9977 * before we can write an inodeblock with updated 9978 * pointers. Be careful to arrange the marker so 9979 * we revisit the freeblks if it's not removed by 9980 * the first jwait(). 9981 */ 9982 if (jblkdep != NULL) { 9983 LIST_REMOVE(&marker, wk_list); 9984 LIST_INSERT_BEFORE(wk, &marker, wk_list); 9985 jwait(&jblkdep->jb_list, MNT_WAIT); 9986 } 9987 continue; 9988 case D_ALLOCDIRECT: 9989 case D_ALLOCINDIR: 9990 /* 9991 * We have to wait for the jnewblk to be journaled 9992 * before we can write to a block if the contents 9993 * may be confused with an earlier file's indirect 9994 * at recovery time. Handle the marker as described 9995 * above. 9996 */ 9997 newblk = WK_NEWBLK(wk); 9998 if (newblk->nb_jnewblk != NULL && 9999 indirblk_lookup(newblk->nb_list.wk_mp, 10000 newblk->nb_newblkno)) { 10001 LIST_REMOVE(&marker, wk_list); 10002 LIST_INSERT_BEFORE(wk, &marker, wk_list); 10003 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 10004 } 10005 continue; 10006 10007 case D_SBDEP: 10008 initiate_write_sbdep(WK_SBDEP(wk)); 10009 continue; 10010 10011 case D_MKDIR: 10012 case D_FREEWORK: 10013 case D_FREEDEP: 10014 case D_JSEGDEP: 10015 continue; 10016 10017 default: 10018 panic("handle_disk_io_initiation: Unexpected type %s", 10019 TYPENAME(wk->wk_type)); 10020 /* NOTREACHED */ 10021 } 10022 } 10023 FREE_LOCK(ump); 10024 PRELE(curproc); /* Allow swapout of kernel stack */ 10025} 10026 10027/* 10028 * Called from within the procedure above to deal with unsatisfied 10029 * allocation dependencies in a directory. The buffer must be locked, 10030 * thus, no I/O completion operations can occur while we are 10031 * manipulating its associated dependencies. 10032 */ 10033static void 10034initiate_write_filepage(pagedep, bp) 10035 struct pagedep *pagedep; 10036 struct buf *bp; 10037{ 10038 struct jremref *jremref; 10039 struct jmvref *jmvref; 10040 struct dirrem *dirrem; 10041 struct diradd *dap; 10042 struct direct *ep; 10043 int i; 10044 10045 if (pagedep->pd_state & IOSTARTED) { 10046 /* 10047 * This can only happen if there is a driver that does not 10048 * understand chaining. Here biodone will reissue the call 10049 * to strategy for the incomplete buffers. 10050 */ 10051 printf("initiate_write_filepage: already started\n"); 10052 return; 10053 } 10054 pagedep->pd_state |= IOSTARTED; 10055 /* 10056 * Wait for all journal remove dependencies to hit the disk. 10057 * We can not allow any potentially conflicting directory adds 10058 * to be visible before removes and rollback is too difficult. 10059 * The per-filesystem lock may be dropped and re-acquired, however 10060 * we hold the buf locked so the dependency can not go away. 10061 */ 10062 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) 10063 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) 10064 jwait(&jremref->jr_list, MNT_WAIT); 10065 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) 10066 jwait(&jmvref->jm_list, MNT_WAIT); 10067 for (i = 0; i < DAHASHSZ; i++) { 10068 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 10069 ep = (struct direct *) 10070 ((char *)bp->b_data + dap->da_offset); 10071 if (ep->d_ino != dap->da_newinum) 10072 panic("%s: dir inum %ju != new %ju", 10073 "initiate_write_filepage", 10074 (uintmax_t)ep->d_ino, 10075 (uintmax_t)dap->da_newinum); 10076 if (dap->da_state & DIRCHG) 10077 ep->d_ino = dap->da_previous->dm_oldinum; 10078 else 10079 ep->d_ino = 0; 10080 dap->da_state &= ~ATTACHED; 10081 dap->da_state |= UNDONE; 10082 } 10083 } 10084} 10085 10086/* 10087 * Version of initiate_write_inodeblock that handles UFS1 dinodes. 10088 * Note that any bug fixes made to this routine must be done in the 10089 * version found below. 10090 * 10091 * Called from within the procedure above to deal with unsatisfied 10092 * allocation dependencies in an inodeblock. The buffer must be 10093 * locked, thus, no I/O completion operations can occur while we 10094 * are manipulating its associated dependencies. 10095 */ 10096static void 10097initiate_write_inodeblock_ufs1(inodedep, bp) 10098 struct inodedep *inodedep; 10099 struct buf *bp; /* The inode block */ 10100{ 10101 struct allocdirect *adp, *lastadp; 10102 struct ufs1_dinode *dp; 10103 struct ufs1_dinode *sip; 10104 struct inoref *inoref; 10105 struct ufsmount *ump; 10106 struct fs *fs; 10107 ufs_lbn_t i; 10108#ifdef INVARIANTS 10109 ufs_lbn_t prevlbn = 0; 10110#endif 10111 int deplist; 10112 10113 if (inodedep->id_state & IOSTARTED) 10114 panic("initiate_write_inodeblock_ufs1: already started"); 10115 inodedep->id_state |= IOSTARTED; 10116 fs = inodedep->id_fs; 10117 ump = VFSTOUFS(inodedep->id_list.wk_mp); 10118 LOCK_OWNED(ump); 10119 dp = (struct ufs1_dinode *)bp->b_data + 10120 ino_to_fsbo(fs, inodedep->id_ino); 10121 10122 /* 10123 * If we're on the unlinked list but have not yet written our 10124 * next pointer initialize it here. 10125 */ 10126 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) { 10127 struct inodedep *inon; 10128 10129 inon = TAILQ_NEXT(inodedep, id_unlinked); 10130 dp->di_freelink = inon ? inon->id_ino : 0; 10131 } 10132 /* 10133 * If the bitmap is not yet written, then the allocated 10134 * inode cannot be written to disk. 10135 */ 10136 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 10137 if (inodedep->id_savedino1 != NULL) 10138 panic("initiate_write_inodeblock_ufs1: I/O underway"); 10139 FREE_LOCK(ump); 10140 sip = malloc(sizeof(struct ufs1_dinode), 10141 M_SAVEDINO, M_SOFTDEP_FLAGS); 10142 ACQUIRE_LOCK(ump); 10143 inodedep->id_savedino1 = sip; 10144 *inodedep->id_savedino1 = *dp; 10145 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 10146 dp->di_gen = inodedep->id_savedino1->di_gen; 10147 dp->di_freelink = inodedep->id_savedino1->di_freelink; 10148 return; 10149 } 10150 /* 10151 * If no dependencies, then there is nothing to roll back. 10152 */ 10153 inodedep->id_savedsize = dp->di_size; 10154 inodedep->id_savedextsize = 0; 10155 inodedep->id_savednlink = dp->di_nlink; 10156 if (TAILQ_EMPTY(&inodedep->id_inoupdt) && 10157 TAILQ_EMPTY(&inodedep->id_inoreflst)) 10158 return; 10159 /* 10160 * Revert the link count to that of the first unwritten journal entry. 10161 */ 10162 inoref = TAILQ_FIRST(&inodedep->id_inoreflst); 10163 if (inoref) 10164 dp->di_nlink = inoref->if_nlink; 10165 /* 10166 * Set the dependencies to busy. 10167 */ 10168 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10169 adp = TAILQ_NEXT(adp, ad_next)) { 10170#ifdef INVARIANTS 10171 if (deplist != 0 && prevlbn >= adp->ad_offset) 10172 panic("softdep_write_inodeblock: lbn order"); 10173 prevlbn = adp->ad_offset; 10174 if (adp->ad_offset < NDADDR && 10175 dp->di_db[adp->ad_offset] != adp->ad_newblkno) 10176 panic("%s: direct pointer #%jd mismatch %d != %jd", 10177 "softdep_write_inodeblock", 10178 (intmax_t)adp->ad_offset, 10179 dp->di_db[adp->ad_offset], 10180 (intmax_t)adp->ad_newblkno); 10181 if (adp->ad_offset >= NDADDR && 10182 dp->di_ib[adp->ad_offset - NDADDR] != adp->ad_newblkno) 10183 panic("%s: indirect pointer #%jd mismatch %d != %jd", 10184 "softdep_write_inodeblock", 10185 (intmax_t)adp->ad_offset - NDADDR, 10186 dp->di_ib[adp->ad_offset - NDADDR], 10187 (intmax_t)adp->ad_newblkno); 10188 deplist |= 1 << adp->ad_offset; 10189 if ((adp->ad_state & ATTACHED) == 0) 10190 panic("softdep_write_inodeblock: Unknown state 0x%x", 10191 adp->ad_state); 10192#endif /* INVARIANTS */ 10193 adp->ad_state &= ~ATTACHED; 10194 adp->ad_state |= UNDONE; 10195 } 10196 /* 10197 * The on-disk inode cannot claim to be any larger than the last 10198 * fragment that has been written. Otherwise, the on-disk inode 10199 * might have fragments that were not the last block in the file 10200 * which would corrupt the filesystem. 10201 */ 10202 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10203 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10204 if (adp->ad_offset >= NDADDR) 10205 break; 10206 dp->di_db[adp->ad_offset] = adp->ad_oldblkno; 10207 /* keep going until hitting a rollback to a frag */ 10208 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10209 continue; 10210 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10211 for (i = adp->ad_offset + 1; i < NDADDR; i++) { 10212#ifdef INVARIANTS 10213 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 10214 panic("softdep_write_inodeblock: lost dep1"); 10215#endif /* INVARIANTS */ 10216 dp->di_db[i] = 0; 10217 } 10218 for (i = 0; i < NIADDR; i++) { 10219#ifdef INVARIANTS 10220 if (dp->di_ib[i] != 0 && 10221 (deplist & ((1 << NDADDR) << i)) == 0) 10222 panic("softdep_write_inodeblock: lost dep2"); 10223#endif /* INVARIANTS */ 10224 dp->di_ib[i] = 0; 10225 } 10226 return; 10227 } 10228 /* 10229 * If we have zero'ed out the last allocated block of the file, 10230 * roll back the size to the last currently allocated block. 10231 * We know that this last allocated block is a full-sized as 10232 * we already checked for fragments in the loop above. 10233 */ 10234 if (lastadp != NULL && 10235 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10236 for (i = lastadp->ad_offset; i >= 0; i--) 10237 if (dp->di_db[i] != 0) 10238 break; 10239 dp->di_size = (i + 1) * fs->fs_bsize; 10240 } 10241 /* 10242 * The only dependencies are for indirect blocks. 10243 * 10244 * The file size for indirect block additions is not guaranteed. 10245 * Such a guarantee would be non-trivial to achieve. The conventional 10246 * synchronous write implementation also does not make this guarantee. 10247 * Fsck should catch and fix discrepancies. Arguably, the file size 10248 * can be over-estimated without destroying integrity when the file 10249 * moves into the indirect blocks (i.e., is large). If we want to 10250 * postpone fsck, we are stuck with this argument. 10251 */ 10252 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 10253 dp->di_ib[adp->ad_offset - NDADDR] = 0; 10254} 10255 10256/* 10257 * Version of initiate_write_inodeblock that handles UFS2 dinodes. 10258 * Note that any bug fixes made to this routine must be done in the 10259 * version found above. 10260 * 10261 * Called from within the procedure above to deal with unsatisfied 10262 * allocation dependencies in an inodeblock. The buffer must be 10263 * locked, thus, no I/O completion operations can occur while we 10264 * are manipulating its associated dependencies. 10265 */ 10266static void 10267initiate_write_inodeblock_ufs2(inodedep, bp) 10268 struct inodedep *inodedep; 10269 struct buf *bp; /* The inode block */ 10270{ 10271 struct allocdirect *adp, *lastadp; 10272 struct ufs2_dinode *dp; 10273 struct ufs2_dinode *sip; 10274 struct inoref *inoref; 10275 struct ufsmount *ump; 10276 struct fs *fs; 10277 ufs_lbn_t i; 10278#ifdef INVARIANTS 10279 ufs_lbn_t prevlbn = 0; 10280#endif 10281 int deplist; 10282 10283 if (inodedep->id_state & IOSTARTED) 10284 panic("initiate_write_inodeblock_ufs2: already started"); 10285 inodedep->id_state |= IOSTARTED; 10286 fs = inodedep->id_fs; 10287 ump = VFSTOUFS(inodedep->id_list.wk_mp); 10288 LOCK_OWNED(ump); 10289 dp = (struct ufs2_dinode *)bp->b_data + 10290 ino_to_fsbo(fs, inodedep->id_ino); 10291 10292 /* 10293 * If we're on the unlinked list but have not yet written our 10294 * next pointer initialize it here. 10295 */ 10296 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) { 10297 struct inodedep *inon; 10298 10299 inon = TAILQ_NEXT(inodedep, id_unlinked); 10300 dp->di_freelink = inon ? inon->id_ino : 0; 10301 } 10302 /* 10303 * If the bitmap is not yet written, then the allocated 10304 * inode cannot be written to disk. 10305 */ 10306 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 10307 if (inodedep->id_savedino2 != NULL) 10308 panic("initiate_write_inodeblock_ufs2: I/O underway"); 10309 FREE_LOCK(ump); 10310 sip = malloc(sizeof(struct ufs2_dinode), 10311 M_SAVEDINO, M_SOFTDEP_FLAGS); 10312 ACQUIRE_LOCK(ump); 10313 inodedep->id_savedino2 = sip; 10314 *inodedep->id_savedino2 = *dp; 10315 bzero((caddr_t)dp, sizeof(struct ufs2_dinode)); 10316 dp->di_gen = inodedep->id_savedino2->di_gen; 10317 dp->di_freelink = inodedep->id_savedino2->di_freelink; 10318 return; 10319 } 10320 /* 10321 * If no dependencies, then there is nothing to roll back. 10322 */ 10323 inodedep->id_savedsize = dp->di_size; 10324 inodedep->id_savedextsize = dp->di_extsize; 10325 inodedep->id_savednlink = dp->di_nlink; 10326 if (TAILQ_EMPTY(&inodedep->id_inoupdt) && 10327 TAILQ_EMPTY(&inodedep->id_extupdt) && 10328 TAILQ_EMPTY(&inodedep->id_inoreflst)) 10329 return; 10330 /* 10331 * Revert the link count to that of the first unwritten journal entry. 10332 */ 10333 inoref = TAILQ_FIRST(&inodedep->id_inoreflst); 10334 if (inoref) 10335 dp->di_nlink = inoref->if_nlink; 10336 10337 /* 10338 * Set the ext data dependencies to busy. 10339 */ 10340 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 10341 adp = TAILQ_NEXT(adp, ad_next)) { 10342#ifdef INVARIANTS 10343 if (deplist != 0 && prevlbn >= adp->ad_offset) 10344 panic("softdep_write_inodeblock: lbn order"); 10345 prevlbn = adp->ad_offset; 10346 if (dp->di_extb[adp->ad_offset] != adp->ad_newblkno) 10347 panic("%s: direct pointer #%jd mismatch %jd != %jd", 10348 "softdep_write_inodeblock", 10349 (intmax_t)adp->ad_offset, 10350 (intmax_t)dp->di_extb[adp->ad_offset], 10351 (intmax_t)adp->ad_newblkno); 10352 deplist |= 1 << adp->ad_offset; 10353 if ((adp->ad_state & ATTACHED) == 0) 10354 panic("softdep_write_inodeblock: Unknown state 0x%x", 10355 adp->ad_state); 10356#endif /* INVARIANTS */ 10357 adp->ad_state &= ~ATTACHED; 10358 adp->ad_state |= UNDONE; 10359 } 10360 /* 10361 * The on-disk inode cannot claim to be any larger than the last 10362 * fragment that has been written. Otherwise, the on-disk inode 10363 * might have fragments that were not the last block in the ext 10364 * data which would corrupt the filesystem. 10365 */ 10366 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 10367 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10368 dp->di_extb[adp->ad_offset] = adp->ad_oldblkno; 10369 /* keep going until hitting a rollback to a frag */ 10370 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10371 continue; 10372 dp->di_extsize = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10373 for (i = adp->ad_offset + 1; i < NXADDR; i++) { 10374#ifdef INVARIANTS 10375 if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0) 10376 panic("softdep_write_inodeblock: lost dep1"); 10377#endif /* INVARIANTS */ 10378 dp->di_extb[i] = 0; 10379 } 10380 lastadp = NULL; 10381 break; 10382 } 10383 /* 10384 * If we have zero'ed out the last allocated block of the ext 10385 * data, roll back the size to the last currently allocated block. 10386 * We know that this last allocated block is a full-sized as 10387 * we already checked for fragments in the loop above. 10388 */ 10389 if (lastadp != NULL && 10390 dp->di_extsize <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10391 for (i = lastadp->ad_offset; i >= 0; i--) 10392 if (dp->di_extb[i] != 0) 10393 break; 10394 dp->di_extsize = (i + 1) * fs->fs_bsize; 10395 } 10396 /* 10397 * Set the file data dependencies to busy. 10398 */ 10399 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10400 adp = TAILQ_NEXT(adp, ad_next)) { 10401#ifdef INVARIANTS 10402 if (deplist != 0 && prevlbn >= adp->ad_offset) 10403 panic("softdep_write_inodeblock: lbn order"); 10404 if ((adp->ad_state & ATTACHED) == 0) 10405 panic("inodedep %p and adp %p not attached", inodedep, adp); 10406 prevlbn = adp->ad_offset; 10407 if (adp->ad_offset < NDADDR && 10408 dp->di_db[adp->ad_offset] != adp->ad_newblkno) 10409 panic("%s: direct pointer #%jd mismatch %jd != %jd", 10410 "softdep_write_inodeblock", 10411 (intmax_t)adp->ad_offset, 10412 (intmax_t)dp->di_db[adp->ad_offset], 10413 (intmax_t)adp->ad_newblkno); 10414 if (adp->ad_offset >= NDADDR && 10415 dp->di_ib[adp->ad_offset - NDADDR] != adp->ad_newblkno) 10416 panic("%s indirect pointer #%jd mismatch %jd != %jd", 10417 "softdep_write_inodeblock:", 10418 (intmax_t)adp->ad_offset - NDADDR, 10419 (intmax_t)dp->di_ib[adp->ad_offset - NDADDR], 10420 (intmax_t)adp->ad_newblkno); 10421 deplist |= 1 << adp->ad_offset; 10422 if ((adp->ad_state & ATTACHED) == 0) 10423 panic("softdep_write_inodeblock: Unknown state 0x%x", 10424 adp->ad_state); 10425#endif /* INVARIANTS */ 10426 adp->ad_state &= ~ATTACHED; 10427 adp->ad_state |= UNDONE; 10428 } 10429 /* 10430 * The on-disk inode cannot claim to be any larger than the last 10431 * fragment that has been written. Otherwise, the on-disk inode 10432 * might have fragments that were not the last block in the file 10433 * which would corrupt the filesystem. 10434 */ 10435 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10436 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10437 if (adp->ad_offset >= NDADDR) 10438 break; 10439 dp->di_db[adp->ad_offset] = adp->ad_oldblkno; 10440 /* keep going until hitting a rollback to a frag */ 10441 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10442 continue; 10443 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10444 for (i = adp->ad_offset + 1; i < NDADDR; i++) { 10445#ifdef INVARIANTS 10446 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 10447 panic("softdep_write_inodeblock: lost dep2"); 10448#endif /* INVARIANTS */ 10449 dp->di_db[i] = 0; 10450 } 10451 for (i = 0; i < NIADDR; i++) { 10452#ifdef INVARIANTS 10453 if (dp->di_ib[i] != 0 && 10454 (deplist & ((1 << NDADDR) << i)) == 0) 10455 panic("softdep_write_inodeblock: lost dep3"); 10456#endif /* INVARIANTS */ 10457 dp->di_ib[i] = 0; 10458 } 10459 return; 10460 } 10461 /* 10462 * If we have zero'ed out the last allocated block of the file, 10463 * roll back the size to the last currently allocated block. 10464 * We know that this last allocated block is a full-sized as 10465 * we already checked for fragments in the loop above. 10466 */ 10467 if (lastadp != NULL && 10468 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10469 for (i = lastadp->ad_offset; i >= 0; i--) 10470 if (dp->di_db[i] != 0) 10471 break; 10472 dp->di_size = (i + 1) * fs->fs_bsize; 10473 } 10474 /* 10475 * The only dependencies are for indirect blocks. 10476 * 10477 * The file size for indirect block additions is not guaranteed. 10478 * Such a guarantee would be non-trivial to achieve. The conventional 10479 * synchronous write implementation also does not make this guarantee. 10480 * Fsck should catch and fix discrepancies. Arguably, the file size 10481 * can be over-estimated without destroying integrity when the file 10482 * moves into the indirect blocks (i.e., is large). If we want to 10483 * postpone fsck, we are stuck with this argument. 10484 */ 10485 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 10486 dp->di_ib[adp->ad_offset - NDADDR] = 0; 10487} 10488 10489/* 10490 * Cancel an indirdep as a result of truncation. Release all of the 10491 * children allocindirs and place their journal work on the appropriate 10492 * list. 10493 */ 10494static void 10495cancel_indirdep(indirdep, bp, freeblks) 10496 struct indirdep *indirdep; 10497 struct buf *bp; 10498 struct freeblks *freeblks; 10499{ 10500 struct allocindir *aip; 10501 10502 /* 10503 * None of the indirect pointers will ever be visible, 10504 * so they can simply be tossed. GOINGAWAY ensures 10505 * that allocated pointers will be saved in the buffer 10506 * cache until they are freed. Note that they will 10507 * only be able to be found by their physical address 10508 * since the inode mapping the logical address will 10509 * be gone. The save buffer used for the safe copy 10510 * was allocated in setup_allocindir_phase2 using 10511 * the physical address so it could be used for this 10512 * purpose. Hence we swap the safe copy with the real 10513 * copy, allowing the safe copy to be freed and holding 10514 * on to the real copy for later use in indir_trunc. 10515 */ 10516 if (indirdep->ir_state & GOINGAWAY) 10517 panic("cancel_indirdep: already gone"); 10518 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 10519 indirdep->ir_state |= DEPCOMPLETE; 10520 LIST_REMOVE(indirdep, ir_next); 10521 } 10522 indirdep->ir_state |= GOINGAWAY; 10523 /* 10524 * Pass in bp for blocks still have journal writes 10525 * pending so we can cancel them on their own. 10526 */ 10527 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 10528 cancel_allocindir(aip, bp, freeblks, 0); 10529 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) 10530 cancel_allocindir(aip, NULL, freeblks, 0); 10531 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != 0) 10532 cancel_allocindir(aip, NULL, freeblks, 0); 10533 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != 0) 10534 cancel_allocindir(aip, NULL, freeblks, 0); 10535 /* 10536 * If there are pending partial truncations we need to keep the 10537 * old block copy around until they complete. This is because 10538 * the current b_data is not a perfect superset of the available 10539 * blocks. 10540 */ 10541 if (TAILQ_EMPTY(&indirdep->ir_trunc)) 10542 bcopy(bp->b_data, indirdep->ir_savebp->b_data, bp->b_bcount); 10543 else 10544 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 10545 WORKLIST_REMOVE(&indirdep->ir_list); 10546 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, &indirdep->ir_list); 10547 indirdep->ir_bp = NULL; 10548 indirdep->ir_freeblks = freeblks; 10549} 10550 10551/* 10552 * Free an indirdep once it no longer has new pointers to track. 10553 */ 10554static void 10555free_indirdep(indirdep) 10556 struct indirdep *indirdep; 10557{ 10558 10559 KASSERT(TAILQ_EMPTY(&indirdep->ir_trunc), 10560 ("free_indirdep: Indir trunc list not empty.")); 10561 KASSERT(LIST_EMPTY(&indirdep->ir_completehd), 10562 ("free_indirdep: Complete head not empty.")); 10563 KASSERT(LIST_EMPTY(&indirdep->ir_writehd), 10564 ("free_indirdep: write head not empty.")); 10565 KASSERT(LIST_EMPTY(&indirdep->ir_donehd), 10566 ("free_indirdep: done head not empty.")); 10567 KASSERT(LIST_EMPTY(&indirdep->ir_deplisthd), 10568 ("free_indirdep: deplist head not empty.")); 10569 KASSERT((indirdep->ir_state & DEPCOMPLETE), 10570 ("free_indirdep: %p still on newblk list.", indirdep)); 10571 KASSERT(indirdep->ir_saveddata == NULL, 10572 ("free_indirdep: %p still has saved data.", indirdep)); 10573 if (indirdep->ir_state & ONWORKLIST) 10574 WORKLIST_REMOVE(&indirdep->ir_list); 10575 WORKITEM_FREE(indirdep, D_INDIRDEP); 10576} 10577 10578/* 10579 * Called before a write to an indirdep. This routine is responsible for 10580 * rolling back pointers to a safe state which includes only those 10581 * allocindirs which have been completed. 10582 */ 10583static void 10584initiate_write_indirdep(indirdep, bp) 10585 struct indirdep *indirdep; 10586 struct buf *bp; 10587{ 10588 struct ufsmount *ump; 10589 10590 indirdep->ir_state |= IOSTARTED; 10591 if (indirdep->ir_state & GOINGAWAY) 10592 panic("disk_io_initiation: indirdep gone"); 10593 /* 10594 * If there are no remaining dependencies, this will be writing 10595 * the real pointers. 10596 */ 10597 if (LIST_EMPTY(&indirdep->ir_deplisthd) && 10598 TAILQ_EMPTY(&indirdep->ir_trunc)) 10599 return; 10600 /* 10601 * Replace up-to-date version with safe version. 10602 */ 10603 if (indirdep->ir_saveddata == NULL) { 10604 ump = VFSTOUFS(indirdep->ir_list.wk_mp); 10605 LOCK_OWNED(ump); 10606 FREE_LOCK(ump); 10607 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP, 10608 M_SOFTDEP_FLAGS); 10609 ACQUIRE_LOCK(ump); 10610 } 10611 indirdep->ir_state &= ~ATTACHED; 10612 indirdep->ir_state |= UNDONE; 10613 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 10614 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 10615 bp->b_bcount); 10616} 10617 10618/* 10619 * Called when an inode has been cleared in a cg bitmap. This finally 10620 * eliminates any canceled jaddrefs 10621 */ 10622void 10623softdep_setup_inofree(mp, bp, ino, wkhd) 10624 struct mount *mp; 10625 struct buf *bp; 10626 ino_t ino; 10627 struct workhead *wkhd; 10628{ 10629 struct worklist *wk, *wkn; 10630 struct inodedep *inodedep; 10631 struct ufsmount *ump; 10632 uint8_t *inosused; 10633 struct cg *cgp; 10634 struct fs *fs; 10635 10636 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 10637 ("softdep_setup_inofree called on non-softdep filesystem")); 10638 ump = VFSTOUFS(mp); 10639 ACQUIRE_LOCK(ump); 10640 fs = ump->um_fs; 10641 cgp = (struct cg *)bp->b_data; 10642 inosused = cg_inosused(cgp); 10643 if (isset(inosused, ino % fs->fs_ipg)) 10644 panic("softdep_setup_inofree: inode %ju not freed.", 10645 (uintmax_t)ino); 10646 if (inodedep_lookup(mp, ino, 0, &inodedep)) 10647 panic("softdep_setup_inofree: ino %ju has existing inodedep %p", 10648 (uintmax_t)ino, inodedep); 10649 if (wkhd) { 10650 LIST_FOREACH_SAFE(wk, wkhd, wk_list, wkn) { 10651 if (wk->wk_type != D_JADDREF) 10652 continue; 10653 WORKLIST_REMOVE(wk); 10654 /* 10655 * We can free immediately even if the jaddref 10656 * isn't attached in a background write as now 10657 * the bitmaps are reconciled. 10658 */ 10659 wk->wk_state |= COMPLETE | ATTACHED; 10660 free_jaddref(WK_JADDREF(wk)); 10661 } 10662 jwork_move(&bp->b_dep, wkhd); 10663 } 10664 FREE_LOCK(ump); 10665} 10666 10667 10668/* 10669 * Called via ffs_blkfree() after a set of frags has been cleared from a cg 10670 * map. Any dependencies waiting for the write to clear are added to the 10671 * buf's list and any jnewblks that are being canceled are discarded 10672 * immediately. 10673 */ 10674void 10675softdep_setup_blkfree(mp, bp, blkno, frags, wkhd) 10676 struct mount *mp; 10677 struct buf *bp; 10678 ufs2_daddr_t blkno; 10679 int frags; 10680 struct workhead *wkhd; 10681{ 10682 struct bmsafemap *bmsafemap; 10683 struct jnewblk *jnewblk; 10684 struct ufsmount *ump; 10685 struct worklist *wk; 10686 struct fs *fs; 10687#ifdef SUJ_DEBUG 10688 uint8_t *blksfree; 10689 struct cg *cgp; 10690 ufs2_daddr_t jstart; 10691 ufs2_daddr_t jend; 10692 ufs2_daddr_t end; 10693 long bno; 10694 int i; 10695#endif 10696 10697 CTR3(KTR_SUJ, 10698 "softdep_setup_blkfree: blkno %jd frags %d wk head %p", 10699 blkno, frags, wkhd); 10700 10701 ump = VFSTOUFS(mp); 10702 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 10703 ("softdep_setup_blkfree called on non-softdep filesystem")); 10704 ACQUIRE_LOCK(ump); 10705 /* Lookup the bmsafemap so we track when it is dirty. */ 10706 fs = ump->um_fs; 10707 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL); 10708 /* 10709 * Detach any jnewblks which have been canceled. They must linger 10710 * until the bitmap is cleared again by ffs_blkfree() to prevent 10711 * an unjournaled allocation from hitting the disk. 10712 */ 10713 if (wkhd) { 10714 while ((wk = LIST_FIRST(wkhd)) != NULL) { 10715 CTR2(KTR_SUJ, 10716 "softdep_setup_blkfree: blkno %jd wk type %d", 10717 blkno, wk->wk_type); 10718 WORKLIST_REMOVE(wk); 10719 if (wk->wk_type != D_JNEWBLK) { 10720 WORKLIST_INSERT(&bmsafemap->sm_freehd, wk); 10721 continue; 10722 } 10723 jnewblk = WK_JNEWBLK(wk); 10724 KASSERT(jnewblk->jn_state & GOINGAWAY, 10725 ("softdep_setup_blkfree: jnewblk not canceled.")); 10726#ifdef SUJ_DEBUG 10727 /* 10728 * Assert that this block is free in the bitmap 10729 * before we discard the jnewblk. 10730 */ 10731 cgp = (struct cg *)bp->b_data; 10732 blksfree = cg_blksfree(cgp); 10733 bno = dtogd(fs, jnewblk->jn_blkno); 10734 for (i = jnewblk->jn_oldfrags; 10735 i < jnewblk->jn_frags; i++) { 10736 if (isset(blksfree, bno + i)) 10737 continue; 10738 panic("softdep_setup_blkfree: not free"); 10739 } 10740#endif 10741 /* 10742 * Even if it's not attached we can free immediately 10743 * as the new bitmap is correct. 10744 */ 10745 wk->wk_state |= COMPLETE | ATTACHED; 10746 free_jnewblk(jnewblk); 10747 } 10748 } 10749 10750#ifdef SUJ_DEBUG 10751 /* 10752 * Assert that we are not freeing a block which has an outstanding 10753 * allocation dependency. 10754 */ 10755 fs = VFSTOUFS(mp)->um_fs; 10756 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL); 10757 end = blkno + frags; 10758 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) { 10759 /* 10760 * Don't match against blocks that will be freed when the 10761 * background write is done. 10762 */ 10763 if ((jnewblk->jn_state & (ATTACHED | COMPLETE | DEPCOMPLETE)) == 10764 (COMPLETE | DEPCOMPLETE)) 10765 continue; 10766 jstart = jnewblk->jn_blkno + jnewblk->jn_oldfrags; 10767 jend = jnewblk->jn_blkno + jnewblk->jn_frags; 10768 if ((blkno >= jstart && blkno < jend) || 10769 (end > jstart && end <= jend)) { 10770 printf("state 0x%X %jd - %d %d dep %p\n", 10771 jnewblk->jn_state, jnewblk->jn_blkno, 10772 jnewblk->jn_oldfrags, jnewblk->jn_frags, 10773 jnewblk->jn_dep); 10774 panic("softdep_setup_blkfree: " 10775 "%jd-%jd(%d) overlaps with %jd-%jd", 10776 blkno, end, frags, jstart, jend); 10777 } 10778 } 10779#endif 10780 FREE_LOCK(ump); 10781} 10782 10783/* 10784 * Revert a block allocation when the journal record that describes it 10785 * is not yet written. 10786 */ 10787static int 10788jnewblk_rollback(jnewblk, fs, cgp, blksfree) 10789 struct jnewblk *jnewblk; 10790 struct fs *fs; 10791 struct cg *cgp; 10792 uint8_t *blksfree; 10793{ 10794 ufs1_daddr_t fragno; 10795 long cgbno, bbase; 10796 int frags, blk; 10797 int i; 10798 10799 frags = 0; 10800 cgbno = dtogd(fs, jnewblk->jn_blkno); 10801 /* 10802 * We have to test which frags need to be rolled back. We may 10803 * be operating on a stale copy when doing background writes. 10804 */ 10805 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) 10806 if (isclr(blksfree, cgbno + i)) 10807 frags++; 10808 if (frags == 0) 10809 return (0); 10810 /* 10811 * This is mostly ffs_blkfree() sans some validation and 10812 * superblock updates. 10813 */ 10814 if (frags == fs->fs_frag) { 10815 fragno = fragstoblks(fs, cgbno); 10816 ffs_setblock(fs, blksfree, fragno); 10817 ffs_clusteracct(fs, cgp, fragno, 1); 10818 cgp->cg_cs.cs_nbfree++; 10819 } else { 10820 cgbno += jnewblk->jn_oldfrags; 10821 bbase = cgbno - fragnum(fs, cgbno); 10822 /* Decrement the old frags. */ 10823 blk = blkmap(fs, blksfree, bbase); 10824 ffs_fragacct(fs, blk, cgp->cg_frsum, -1); 10825 /* Deallocate the fragment */ 10826 for (i = 0; i < frags; i++) 10827 setbit(blksfree, cgbno + i); 10828 cgp->cg_cs.cs_nffree += frags; 10829 /* Add back in counts associated with the new frags */ 10830 blk = blkmap(fs, blksfree, bbase); 10831 ffs_fragacct(fs, blk, cgp->cg_frsum, 1); 10832 /* If a complete block has been reassembled, account for it. */ 10833 fragno = fragstoblks(fs, bbase); 10834 if (ffs_isblock(fs, blksfree, fragno)) { 10835 cgp->cg_cs.cs_nffree -= fs->fs_frag; 10836 ffs_clusteracct(fs, cgp, fragno, 1); 10837 cgp->cg_cs.cs_nbfree++; 10838 } 10839 } 10840 stat_jnewblk++; 10841 jnewblk->jn_state &= ~ATTACHED; 10842 jnewblk->jn_state |= UNDONE; 10843 10844 return (frags); 10845} 10846 10847static void 10848initiate_write_bmsafemap(bmsafemap, bp) 10849 struct bmsafemap *bmsafemap; 10850 struct buf *bp; /* The cg block. */ 10851{ 10852 struct jaddref *jaddref; 10853 struct jnewblk *jnewblk; 10854 uint8_t *inosused; 10855 uint8_t *blksfree; 10856 struct cg *cgp; 10857 struct fs *fs; 10858 ino_t ino; 10859 10860 if (bmsafemap->sm_state & IOSTARTED) 10861 return; 10862 bmsafemap->sm_state |= IOSTARTED; 10863 /* 10864 * Clear any inode allocations which are pending journal writes. 10865 */ 10866 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd) != NULL) { 10867 cgp = (struct cg *)bp->b_data; 10868 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 10869 inosused = cg_inosused(cgp); 10870 LIST_FOREACH(jaddref, &bmsafemap->sm_jaddrefhd, ja_bmdeps) { 10871 ino = jaddref->ja_ino % fs->fs_ipg; 10872 if (isset(inosused, ino)) { 10873 if ((jaddref->ja_mode & IFMT) == IFDIR) 10874 cgp->cg_cs.cs_ndir--; 10875 cgp->cg_cs.cs_nifree++; 10876 clrbit(inosused, ino); 10877 jaddref->ja_state &= ~ATTACHED; 10878 jaddref->ja_state |= UNDONE; 10879 stat_jaddref++; 10880 } else 10881 panic("initiate_write_bmsafemap: inode %ju " 10882 "marked free", (uintmax_t)jaddref->ja_ino); 10883 } 10884 } 10885 /* 10886 * Clear any block allocations which are pending journal writes. 10887 */ 10888 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) { 10889 cgp = (struct cg *)bp->b_data; 10890 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 10891 blksfree = cg_blksfree(cgp); 10892 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) { 10893 if (jnewblk_rollback(jnewblk, fs, cgp, blksfree)) 10894 continue; 10895 panic("initiate_write_bmsafemap: block %jd " 10896 "marked free", jnewblk->jn_blkno); 10897 } 10898 } 10899 /* 10900 * Move allocation lists to the written lists so they can be 10901 * cleared once the block write is complete. 10902 */ 10903 LIST_SWAP(&bmsafemap->sm_inodedephd, &bmsafemap->sm_inodedepwr, 10904 inodedep, id_deps); 10905 LIST_SWAP(&bmsafemap->sm_newblkhd, &bmsafemap->sm_newblkwr, 10906 newblk, nb_deps); 10907 LIST_SWAP(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, worklist, 10908 wk_list); 10909} 10910 10911/* 10912 * This routine is called during the completion interrupt 10913 * service routine for a disk write (from the procedure called 10914 * by the device driver to inform the filesystem caches of 10915 * a request completion). It should be called early in this 10916 * procedure, before the block is made available to other 10917 * processes or other routines are called. 10918 * 10919 */ 10920static void 10921softdep_disk_write_complete(bp) 10922 struct buf *bp; /* describes the completed disk write */ 10923{ 10924 struct worklist *wk; 10925 struct worklist *owk; 10926 struct ufsmount *ump; 10927 struct workhead reattach; 10928 struct freeblks *freeblks; 10929 struct buf *sbp; 10930 10931 /* 10932 * If an error occurred while doing the write, then the data 10933 * has not hit the disk and the dependencies cannot be unrolled. 10934 */ 10935 if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0) 10936 return; 10937 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL) 10938 return; 10939 ump = VFSTOUFS(wk->wk_mp); 10940 LIST_INIT(&reattach); 10941 /* 10942 * This lock must not be released anywhere in this code segment. 10943 */ 10944 sbp = NULL; 10945 owk = NULL; 10946 ACQUIRE_LOCK(ump); 10947 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 10948 WORKLIST_REMOVE(wk); 10949 atomic_add_long(&dep_write[wk->wk_type], 1); 10950 if (wk == owk) 10951 panic("duplicate worklist: %p\n", wk); 10952 owk = wk; 10953 switch (wk->wk_type) { 10954 10955 case D_PAGEDEP: 10956 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 10957 WORKLIST_INSERT(&reattach, wk); 10958 continue; 10959 10960 case D_INODEDEP: 10961 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 10962 WORKLIST_INSERT(&reattach, wk); 10963 continue; 10964 10965 case D_BMSAFEMAP: 10966 if (handle_written_bmsafemap(WK_BMSAFEMAP(wk), bp)) 10967 WORKLIST_INSERT(&reattach, wk); 10968 continue; 10969 10970 case D_MKDIR: 10971 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 10972 continue; 10973 10974 case D_ALLOCDIRECT: 10975 wk->wk_state |= COMPLETE; 10976 handle_allocdirect_partdone(WK_ALLOCDIRECT(wk), NULL); 10977 continue; 10978 10979 case D_ALLOCINDIR: 10980 wk->wk_state |= COMPLETE; 10981 handle_allocindir_partdone(WK_ALLOCINDIR(wk)); 10982 continue; 10983 10984 case D_INDIRDEP: 10985 if (handle_written_indirdep(WK_INDIRDEP(wk), bp, &sbp)) 10986 WORKLIST_INSERT(&reattach, wk); 10987 continue; 10988 10989 case D_FREEBLKS: 10990 wk->wk_state |= COMPLETE; 10991 freeblks = WK_FREEBLKS(wk); 10992 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE && 10993 LIST_EMPTY(&freeblks->fb_jblkdephd)) 10994 add_to_worklist(wk, WK_NODELAY); 10995 continue; 10996 10997 case D_FREEWORK: 10998 handle_written_freework(WK_FREEWORK(wk)); 10999 break; 11000 11001 case D_JSEGDEP: 11002 free_jsegdep(WK_JSEGDEP(wk)); 11003 continue; 11004 11005 case D_JSEG: 11006 handle_written_jseg(WK_JSEG(wk), bp); 11007 continue; 11008 11009 case D_SBDEP: 11010 if (handle_written_sbdep(WK_SBDEP(wk), bp)) 11011 WORKLIST_INSERT(&reattach, wk); 11012 continue; 11013 11014 case D_FREEDEP: 11015 free_freedep(WK_FREEDEP(wk)); 11016 continue; 11017 11018 default: 11019 panic("handle_disk_write_complete: Unknown type %s", 11020 TYPENAME(wk->wk_type)); 11021 /* NOTREACHED */ 11022 } 11023 } 11024 /* 11025 * Reattach any requests that must be redone. 11026 */ 11027 while ((wk = LIST_FIRST(&reattach)) != NULL) { 11028 WORKLIST_REMOVE(wk); 11029 WORKLIST_INSERT(&bp->b_dep, wk); 11030 } 11031 FREE_LOCK(ump); 11032 if (sbp) 11033 brelse(sbp); 11034} 11035 11036/* 11037 * Called from within softdep_disk_write_complete above. Note that 11038 * this routine is always called from interrupt level with further 11039 * splbio interrupts blocked. 11040 */ 11041static void 11042handle_allocdirect_partdone(adp, wkhd) 11043 struct allocdirect *adp; /* the completed allocdirect */ 11044 struct workhead *wkhd; /* Work to do when inode is writtne. */ 11045{ 11046 struct allocdirectlst *listhead; 11047 struct allocdirect *listadp; 11048 struct inodedep *inodedep; 11049 long bsize; 11050 11051 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 11052 return; 11053 /* 11054 * The on-disk inode cannot claim to be any larger than the last 11055 * fragment that has been written. Otherwise, the on-disk inode 11056 * might have fragments that were not the last block in the file 11057 * which would corrupt the filesystem. Thus, we cannot free any 11058 * allocdirects after one whose ad_oldblkno claims a fragment as 11059 * these blocks must be rolled back to zero before writing the inode. 11060 * We check the currently active set of allocdirects in id_inoupdt 11061 * or id_extupdt as appropriate. 11062 */ 11063 inodedep = adp->ad_inodedep; 11064 bsize = inodedep->id_fs->fs_bsize; 11065 if (adp->ad_state & EXTDATA) 11066 listhead = &inodedep->id_extupdt; 11067 else 11068 listhead = &inodedep->id_inoupdt; 11069 TAILQ_FOREACH(listadp, listhead, ad_next) { 11070 /* found our block */ 11071 if (listadp == adp) 11072 break; 11073 /* continue if ad_oldlbn is not a fragment */ 11074 if (listadp->ad_oldsize == 0 || 11075 listadp->ad_oldsize == bsize) 11076 continue; 11077 /* hit a fragment */ 11078 return; 11079 } 11080 /* 11081 * If we have reached the end of the current list without 11082 * finding the just finished dependency, then it must be 11083 * on the future dependency list. Future dependencies cannot 11084 * be freed until they are moved to the current list. 11085 */ 11086 if (listadp == NULL) { 11087#ifdef DEBUG 11088 if (adp->ad_state & EXTDATA) 11089 listhead = &inodedep->id_newextupdt; 11090 else 11091 listhead = &inodedep->id_newinoupdt; 11092 TAILQ_FOREACH(listadp, listhead, ad_next) 11093 /* found our block */ 11094 if (listadp == adp) 11095 break; 11096 if (listadp == NULL) 11097 panic("handle_allocdirect_partdone: lost dep"); 11098#endif /* DEBUG */ 11099 return; 11100 } 11101 /* 11102 * If we have found the just finished dependency, then queue 11103 * it along with anything that follows it that is complete. 11104 * Since the pointer has not yet been written in the inode 11105 * as the dependency prevents it, place the allocdirect on the 11106 * bufwait list where it will be freed once the pointer is 11107 * valid. 11108 */ 11109 if (wkhd == NULL) 11110 wkhd = &inodedep->id_bufwait; 11111 for (; adp; adp = listadp) { 11112 listadp = TAILQ_NEXT(adp, ad_next); 11113 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 11114 return; 11115 TAILQ_REMOVE(listhead, adp, ad_next); 11116 WORKLIST_INSERT(wkhd, &adp->ad_block.nb_list); 11117 } 11118} 11119 11120/* 11121 * Called from within softdep_disk_write_complete above. This routine 11122 * completes successfully written allocindirs. 11123 */ 11124static void 11125handle_allocindir_partdone(aip) 11126 struct allocindir *aip; /* the completed allocindir */ 11127{ 11128 struct indirdep *indirdep; 11129 11130 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 11131 return; 11132 indirdep = aip->ai_indirdep; 11133 LIST_REMOVE(aip, ai_next); 11134 /* 11135 * Don't set a pointer while the buffer is undergoing IO or while 11136 * we have active truncations. 11137 */ 11138 if (indirdep->ir_state & UNDONE || !TAILQ_EMPTY(&indirdep->ir_trunc)) { 11139 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 11140 return; 11141 } 11142 if (indirdep->ir_state & UFS1FMT) 11143 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 11144 aip->ai_newblkno; 11145 else 11146 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 11147 aip->ai_newblkno; 11148 /* 11149 * Await the pointer write before freeing the allocindir. 11150 */ 11151 LIST_INSERT_HEAD(&indirdep->ir_writehd, aip, ai_next); 11152} 11153 11154/* 11155 * Release segments held on a jwork list. 11156 */ 11157static void 11158handle_jwork(wkhd) 11159 struct workhead *wkhd; 11160{ 11161 struct worklist *wk; 11162 11163 while ((wk = LIST_FIRST(wkhd)) != NULL) { 11164 WORKLIST_REMOVE(wk); 11165 switch (wk->wk_type) { 11166 case D_JSEGDEP: 11167 free_jsegdep(WK_JSEGDEP(wk)); 11168 continue; 11169 case D_FREEDEP: 11170 free_freedep(WK_FREEDEP(wk)); 11171 continue; 11172 case D_FREEFRAG: 11173 rele_jseg(WK_JSEG(WK_FREEFRAG(wk)->ff_jdep)); 11174 WORKITEM_FREE(wk, D_FREEFRAG); 11175 continue; 11176 case D_FREEWORK: 11177 handle_written_freework(WK_FREEWORK(wk)); 11178 continue; 11179 default: 11180 panic("handle_jwork: Unknown type %s\n", 11181 TYPENAME(wk->wk_type)); 11182 } 11183 } 11184} 11185 11186/* 11187 * Handle the bufwait list on an inode when it is safe to release items 11188 * held there. This normally happens after an inode block is written but 11189 * may be delayed and handled later if there are pending journal items that 11190 * are not yet safe to be released. 11191 */ 11192static struct freefile * 11193handle_bufwait(inodedep, refhd) 11194 struct inodedep *inodedep; 11195 struct workhead *refhd; 11196{ 11197 struct jaddref *jaddref; 11198 struct freefile *freefile; 11199 struct worklist *wk; 11200 11201 freefile = NULL; 11202 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 11203 WORKLIST_REMOVE(wk); 11204 switch (wk->wk_type) { 11205 case D_FREEFILE: 11206 /* 11207 * We defer adding freefile to the worklist 11208 * until all other additions have been made to 11209 * ensure that it will be done after all the 11210 * old blocks have been freed. 11211 */ 11212 if (freefile != NULL) 11213 panic("handle_bufwait: freefile"); 11214 freefile = WK_FREEFILE(wk); 11215 continue; 11216 11217 case D_MKDIR: 11218 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 11219 continue; 11220 11221 case D_DIRADD: 11222 diradd_inode_written(WK_DIRADD(wk), inodedep); 11223 continue; 11224 11225 case D_FREEFRAG: 11226 wk->wk_state |= COMPLETE; 11227 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE) 11228 add_to_worklist(wk, 0); 11229 continue; 11230 11231 case D_DIRREM: 11232 wk->wk_state |= COMPLETE; 11233 add_to_worklist(wk, 0); 11234 continue; 11235 11236 case D_ALLOCDIRECT: 11237 case D_ALLOCINDIR: 11238 free_newblk(WK_NEWBLK(wk)); 11239 continue; 11240 11241 case D_JNEWBLK: 11242 wk->wk_state |= COMPLETE; 11243 free_jnewblk(WK_JNEWBLK(wk)); 11244 continue; 11245 11246 /* 11247 * Save freed journal segments and add references on 11248 * the supplied list which will delay their release 11249 * until the cg bitmap is cleared on disk. 11250 */ 11251 case D_JSEGDEP: 11252 if (refhd == NULL) 11253 free_jsegdep(WK_JSEGDEP(wk)); 11254 else 11255 WORKLIST_INSERT(refhd, wk); 11256 continue; 11257 11258 case D_JADDREF: 11259 jaddref = WK_JADDREF(wk); 11260 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, 11261 if_deps); 11262 /* 11263 * Transfer any jaddrefs to the list to be freed with 11264 * the bitmap if we're handling a removed file. 11265 */ 11266 if (refhd == NULL) { 11267 wk->wk_state |= COMPLETE; 11268 free_jaddref(jaddref); 11269 } else 11270 WORKLIST_INSERT(refhd, wk); 11271 continue; 11272 11273 default: 11274 panic("handle_bufwait: Unknown type %p(%s)", 11275 wk, TYPENAME(wk->wk_type)); 11276 /* NOTREACHED */ 11277 } 11278 } 11279 return (freefile); 11280} 11281/* 11282 * Called from within softdep_disk_write_complete above to restore 11283 * in-memory inode block contents to their most up-to-date state. Note 11284 * that this routine is always called from interrupt level with further 11285 * splbio interrupts blocked. 11286 */ 11287static int 11288handle_written_inodeblock(inodedep, bp) 11289 struct inodedep *inodedep; 11290 struct buf *bp; /* buffer containing the inode block */ 11291{ 11292 struct freefile *freefile; 11293 struct allocdirect *adp, *nextadp; 11294 struct ufs1_dinode *dp1 = NULL; 11295 struct ufs2_dinode *dp2 = NULL; 11296 struct workhead wkhd; 11297 int hadchanges, fstype; 11298 ino_t freelink; 11299 11300 LIST_INIT(&wkhd); 11301 hadchanges = 0; 11302 freefile = NULL; 11303 if ((inodedep->id_state & IOSTARTED) == 0) 11304 panic("handle_written_inodeblock: not started"); 11305 inodedep->id_state &= ~IOSTARTED; 11306 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) { 11307 fstype = UFS1; 11308 dp1 = (struct ufs1_dinode *)bp->b_data + 11309 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 11310 freelink = dp1->di_freelink; 11311 } else { 11312 fstype = UFS2; 11313 dp2 = (struct ufs2_dinode *)bp->b_data + 11314 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 11315 freelink = dp2->di_freelink; 11316 } 11317 /* 11318 * Leave this inodeblock dirty until it's in the list. 11319 */ 11320 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) == UNLINKED) { 11321 struct inodedep *inon; 11322 11323 inon = TAILQ_NEXT(inodedep, id_unlinked); 11324 if ((inon == NULL && freelink == 0) || 11325 (inon && inon->id_ino == freelink)) { 11326 if (inon) 11327 inon->id_state |= UNLINKPREV; 11328 inodedep->id_state |= UNLINKNEXT; 11329 } 11330 hadchanges = 1; 11331 } 11332 /* 11333 * If we had to rollback the inode allocation because of 11334 * bitmaps being incomplete, then simply restore it. 11335 * Keep the block dirty so that it will not be reclaimed until 11336 * all associated dependencies have been cleared and the 11337 * corresponding updates written to disk. 11338 */ 11339 if (inodedep->id_savedino1 != NULL) { 11340 hadchanges = 1; 11341 if (fstype == UFS1) 11342 *dp1 = *inodedep->id_savedino1; 11343 else 11344 *dp2 = *inodedep->id_savedino2; 11345 free(inodedep->id_savedino1, M_SAVEDINO); 11346 inodedep->id_savedino1 = NULL; 11347 if ((bp->b_flags & B_DELWRI) == 0) 11348 stat_inode_bitmap++; 11349 bdirty(bp); 11350 /* 11351 * If the inode is clear here and GOINGAWAY it will never 11352 * be written. Process the bufwait and clear any pending 11353 * work which may include the freefile. 11354 */ 11355 if (inodedep->id_state & GOINGAWAY) 11356 goto bufwait; 11357 return (1); 11358 } 11359 inodedep->id_state |= COMPLETE; 11360 /* 11361 * Roll forward anything that had to be rolled back before 11362 * the inode could be updated. 11363 */ 11364 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 11365 nextadp = TAILQ_NEXT(adp, ad_next); 11366 if (adp->ad_state & ATTACHED) 11367 panic("handle_written_inodeblock: new entry"); 11368 if (fstype == UFS1) { 11369 if (adp->ad_offset < NDADDR) { 11370 if (dp1->di_db[adp->ad_offset]!=adp->ad_oldblkno) 11371 panic("%s %s #%jd mismatch %d != %jd", 11372 "handle_written_inodeblock:", 11373 "direct pointer", 11374 (intmax_t)adp->ad_offset, 11375 dp1->di_db[adp->ad_offset], 11376 (intmax_t)adp->ad_oldblkno); 11377 dp1->di_db[adp->ad_offset] = adp->ad_newblkno; 11378 } else { 11379 if (dp1->di_ib[adp->ad_offset - NDADDR] != 0) 11380 panic("%s: %s #%jd allocated as %d", 11381 "handle_written_inodeblock", 11382 "indirect pointer", 11383 (intmax_t)adp->ad_offset - NDADDR, 11384 dp1->di_ib[adp->ad_offset - NDADDR]); 11385 dp1->di_ib[adp->ad_offset - NDADDR] = 11386 adp->ad_newblkno; 11387 } 11388 } else { 11389 if (adp->ad_offset < NDADDR) { 11390 if (dp2->di_db[adp->ad_offset]!=adp->ad_oldblkno) 11391 panic("%s: %s #%jd %s %jd != %jd", 11392 "handle_written_inodeblock", 11393 "direct pointer", 11394 (intmax_t)adp->ad_offset, "mismatch", 11395 (intmax_t)dp2->di_db[adp->ad_offset], 11396 (intmax_t)adp->ad_oldblkno); 11397 dp2->di_db[adp->ad_offset] = adp->ad_newblkno; 11398 } else { 11399 if (dp2->di_ib[adp->ad_offset - NDADDR] != 0) 11400 panic("%s: %s #%jd allocated as %jd", 11401 "handle_written_inodeblock", 11402 "indirect pointer", 11403 (intmax_t)adp->ad_offset - NDADDR, 11404 (intmax_t) 11405 dp2->di_ib[adp->ad_offset - NDADDR]); 11406 dp2->di_ib[adp->ad_offset - NDADDR] = 11407 adp->ad_newblkno; 11408 } 11409 } 11410 adp->ad_state &= ~UNDONE; 11411 adp->ad_state |= ATTACHED; 11412 hadchanges = 1; 11413 } 11414 for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) { 11415 nextadp = TAILQ_NEXT(adp, ad_next); 11416 if (adp->ad_state & ATTACHED) 11417 panic("handle_written_inodeblock: new entry"); 11418 if (dp2->di_extb[adp->ad_offset] != adp->ad_oldblkno) 11419 panic("%s: direct pointers #%jd %s %jd != %jd", 11420 "handle_written_inodeblock", 11421 (intmax_t)adp->ad_offset, "mismatch", 11422 (intmax_t)dp2->di_extb[adp->ad_offset], 11423 (intmax_t)adp->ad_oldblkno); 11424 dp2->di_extb[adp->ad_offset] = adp->ad_newblkno; 11425 adp->ad_state &= ~UNDONE; 11426 adp->ad_state |= ATTACHED; 11427 hadchanges = 1; 11428 } 11429 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 11430 stat_direct_blk_ptrs++; 11431 /* 11432 * Reset the file size to its most up-to-date value. 11433 */ 11434 if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1) 11435 panic("handle_written_inodeblock: bad size"); 11436 if (inodedep->id_savednlink > LINK_MAX) 11437 panic("handle_written_inodeblock: Invalid link count " 11438 "%d for inodedep %p", inodedep->id_savednlink, inodedep); 11439 if (fstype == UFS1) { 11440 if (dp1->di_nlink != inodedep->id_savednlink) { 11441 dp1->di_nlink = inodedep->id_savednlink; 11442 hadchanges = 1; 11443 } 11444 if (dp1->di_size != inodedep->id_savedsize) { 11445 dp1->di_size = inodedep->id_savedsize; 11446 hadchanges = 1; 11447 } 11448 } else { 11449 if (dp2->di_nlink != inodedep->id_savednlink) { 11450 dp2->di_nlink = inodedep->id_savednlink; 11451 hadchanges = 1; 11452 } 11453 if (dp2->di_size != inodedep->id_savedsize) { 11454 dp2->di_size = inodedep->id_savedsize; 11455 hadchanges = 1; 11456 } 11457 if (dp2->di_extsize != inodedep->id_savedextsize) { 11458 dp2->di_extsize = inodedep->id_savedextsize; 11459 hadchanges = 1; 11460 } 11461 } 11462 inodedep->id_savedsize = -1; 11463 inodedep->id_savedextsize = -1; 11464 inodedep->id_savednlink = -1; 11465 /* 11466 * If there were any rollbacks in the inode block, then it must be 11467 * marked dirty so that its will eventually get written back in 11468 * its correct form. 11469 */ 11470 if (hadchanges) 11471 bdirty(bp); 11472bufwait: 11473 /* 11474 * Process any allocdirects that completed during the update. 11475 */ 11476 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 11477 handle_allocdirect_partdone(adp, &wkhd); 11478 if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 11479 handle_allocdirect_partdone(adp, &wkhd); 11480 /* 11481 * Process deallocations that were held pending until the 11482 * inode had been written to disk. Freeing of the inode 11483 * is delayed until after all blocks have been freed to 11484 * avoid creation of new <vfsid, inum, lbn> triples 11485 * before the old ones have been deleted. Completely 11486 * unlinked inodes are not processed until the unlinked 11487 * inode list is written or the last reference is removed. 11488 */ 11489 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) != UNLINKED) { 11490 freefile = handle_bufwait(inodedep, NULL); 11491 if (freefile && !LIST_EMPTY(&wkhd)) { 11492 WORKLIST_INSERT(&wkhd, &freefile->fx_list); 11493 freefile = NULL; 11494 } 11495 } 11496 /* 11497 * Move rolled forward dependency completions to the bufwait list 11498 * now that those that were already written have been processed. 11499 */ 11500 if (!LIST_EMPTY(&wkhd) && hadchanges == 0) 11501 panic("handle_written_inodeblock: bufwait but no changes"); 11502 jwork_move(&inodedep->id_bufwait, &wkhd); 11503 11504 if (freefile != NULL) { 11505 /* 11506 * If the inode is goingaway it was never written. Fake up 11507 * the state here so free_inodedep() can succeed. 11508 */ 11509 if (inodedep->id_state & GOINGAWAY) 11510 inodedep->id_state |= COMPLETE | DEPCOMPLETE; 11511 if (free_inodedep(inodedep) == 0) 11512 panic("handle_written_inodeblock: live inodedep %p", 11513 inodedep); 11514 add_to_worklist(&freefile->fx_list, 0); 11515 return (0); 11516 } 11517 11518 /* 11519 * If no outstanding dependencies, free it. 11520 */ 11521 if (free_inodedep(inodedep) || 11522 (TAILQ_FIRST(&inodedep->id_inoreflst) == 0 && 11523 TAILQ_FIRST(&inodedep->id_inoupdt) == 0 && 11524 TAILQ_FIRST(&inodedep->id_extupdt) == 0 && 11525 LIST_FIRST(&inodedep->id_bufwait) == 0)) 11526 return (0); 11527 return (hadchanges); 11528} 11529 11530static int 11531handle_written_indirdep(indirdep, bp, bpp) 11532 struct indirdep *indirdep; 11533 struct buf *bp; 11534 struct buf **bpp; 11535{ 11536 struct allocindir *aip; 11537 struct buf *sbp; 11538 int chgs; 11539 11540 if (indirdep->ir_state & GOINGAWAY) 11541 panic("handle_written_indirdep: indirdep gone"); 11542 if ((indirdep->ir_state & IOSTARTED) == 0) 11543 panic("handle_written_indirdep: IO not started"); 11544 chgs = 0; 11545 /* 11546 * If there were rollbacks revert them here. 11547 */ 11548 if (indirdep->ir_saveddata) { 11549 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 11550 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 11551 free(indirdep->ir_saveddata, M_INDIRDEP); 11552 indirdep->ir_saveddata = NULL; 11553 } 11554 chgs = 1; 11555 } 11556 indirdep->ir_state &= ~(UNDONE | IOSTARTED); 11557 indirdep->ir_state |= ATTACHED; 11558 /* 11559 * Move allocindirs with written pointers to the completehd if 11560 * the indirdep's pointer is not yet written. Otherwise 11561 * free them here. 11562 */ 11563 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != 0) { 11564 LIST_REMOVE(aip, ai_next); 11565 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 11566 LIST_INSERT_HEAD(&indirdep->ir_completehd, aip, 11567 ai_next); 11568 newblk_freefrag(&aip->ai_block); 11569 continue; 11570 } 11571 free_newblk(&aip->ai_block); 11572 } 11573 /* 11574 * Move allocindirs that have finished dependency processing from 11575 * the done list to the write list after updating the pointers. 11576 */ 11577 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 11578 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 11579 handle_allocindir_partdone(aip); 11580 if (aip == LIST_FIRST(&indirdep->ir_donehd)) 11581 panic("disk_write_complete: not gone"); 11582 chgs = 1; 11583 } 11584 } 11585 /* 11586 * Preserve the indirdep if there were any changes or if it is not 11587 * yet valid on disk. 11588 */ 11589 if (chgs) { 11590 stat_indir_blk_ptrs++; 11591 bdirty(bp); 11592 return (1); 11593 } 11594 /* 11595 * If there were no changes we can discard the savedbp and detach 11596 * ourselves from the buf. We are only carrying completed pointers 11597 * in this case. 11598 */ 11599 sbp = indirdep->ir_savebp; 11600 sbp->b_flags |= B_INVAL | B_NOCACHE; 11601 indirdep->ir_savebp = NULL; 11602 indirdep->ir_bp = NULL; 11603 if (*bpp != NULL) 11604 panic("handle_written_indirdep: bp already exists."); 11605 *bpp = sbp; 11606 /* 11607 * The indirdep may not be freed until its parent points at it. 11608 */ 11609 if (indirdep->ir_state & DEPCOMPLETE) 11610 free_indirdep(indirdep); 11611 11612 return (0); 11613} 11614 11615/* 11616 * Process a diradd entry after its dependent inode has been written. 11617 * This routine must be called with splbio interrupts blocked. 11618 */ 11619static void 11620diradd_inode_written(dap, inodedep) 11621 struct diradd *dap; 11622 struct inodedep *inodedep; 11623{ 11624 11625 dap->da_state |= COMPLETE; 11626 complete_diradd(dap); 11627 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 11628} 11629 11630/* 11631 * Returns true if the bmsafemap will have rollbacks when written. Must only 11632 * be called with the per-filesystem lock and the buf lock on the cg held. 11633 */ 11634static int 11635bmsafemap_backgroundwrite(bmsafemap, bp) 11636 struct bmsafemap *bmsafemap; 11637 struct buf *bp; 11638{ 11639 int dirty; 11640 11641 LOCK_OWNED(VFSTOUFS(bmsafemap->sm_list.wk_mp)); 11642 dirty = !LIST_EMPTY(&bmsafemap->sm_jaddrefhd) | 11643 !LIST_EMPTY(&bmsafemap->sm_jnewblkhd); 11644 /* 11645 * If we're initiating a background write we need to process the 11646 * rollbacks as they exist now, not as they exist when IO starts. 11647 * No other consumers will look at the contents of the shadowed 11648 * buf so this is safe to do here. 11649 */ 11650 if (bp->b_xflags & BX_BKGRDMARKER) 11651 initiate_write_bmsafemap(bmsafemap, bp); 11652 11653 return (dirty); 11654} 11655 11656/* 11657 * Re-apply an allocation when a cg write is complete. 11658 */ 11659static int 11660jnewblk_rollforward(jnewblk, fs, cgp, blksfree) 11661 struct jnewblk *jnewblk; 11662 struct fs *fs; 11663 struct cg *cgp; 11664 uint8_t *blksfree; 11665{ 11666 ufs1_daddr_t fragno; 11667 ufs2_daddr_t blkno; 11668 long cgbno, bbase; 11669 int frags, blk; 11670 int i; 11671 11672 frags = 0; 11673 cgbno = dtogd(fs, jnewblk->jn_blkno); 11674 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) { 11675 if (isclr(blksfree, cgbno + i)) 11676 panic("jnewblk_rollforward: re-allocated fragment"); 11677 frags++; 11678 } 11679 if (frags == fs->fs_frag) { 11680 blkno = fragstoblks(fs, cgbno); 11681 ffs_clrblock(fs, blksfree, (long)blkno); 11682 ffs_clusteracct(fs, cgp, blkno, -1); 11683 cgp->cg_cs.cs_nbfree--; 11684 } else { 11685 bbase = cgbno - fragnum(fs, cgbno); 11686 cgbno += jnewblk->jn_oldfrags; 11687 /* If a complete block had been reassembled, account for it. */ 11688 fragno = fragstoblks(fs, bbase); 11689 if (ffs_isblock(fs, blksfree, fragno)) { 11690 cgp->cg_cs.cs_nffree += fs->fs_frag; 11691 ffs_clusteracct(fs, cgp, fragno, -1); 11692 cgp->cg_cs.cs_nbfree--; 11693 } 11694 /* Decrement the old frags. */ 11695 blk = blkmap(fs, blksfree, bbase); 11696 ffs_fragacct(fs, blk, cgp->cg_frsum, -1); 11697 /* Allocate the fragment */ 11698 for (i = 0; i < frags; i++) 11699 clrbit(blksfree, cgbno + i); 11700 cgp->cg_cs.cs_nffree -= frags; 11701 /* Add back in counts associated with the new frags */ 11702 blk = blkmap(fs, blksfree, bbase); 11703 ffs_fragacct(fs, blk, cgp->cg_frsum, 1); 11704 } 11705 return (frags); 11706} 11707 11708/* 11709 * Complete a write to a bmsafemap structure. Roll forward any bitmap 11710 * changes if it's not a background write. Set all written dependencies 11711 * to DEPCOMPLETE and free the structure if possible. 11712 */ 11713static int 11714handle_written_bmsafemap(bmsafemap, bp) 11715 struct bmsafemap *bmsafemap; 11716 struct buf *bp; 11717{ 11718 struct newblk *newblk; 11719 struct inodedep *inodedep; 11720 struct jaddref *jaddref, *jatmp; 11721 struct jnewblk *jnewblk, *jntmp; 11722 struct ufsmount *ump; 11723 uint8_t *inosused; 11724 uint8_t *blksfree; 11725 struct cg *cgp; 11726 struct fs *fs; 11727 ino_t ino; 11728 int foreground; 11729 int chgs; 11730 11731 if ((bmsafemap->sm_state & IOSTARTED) == 0) 11732 panic("initiate_write_bmsafemap: Not started\n"); 11733 ump = VFSTOUFS(bmsafemap->sm_list.wk_mp); 11734 chgs = 0; 11735 bmsafemap->sm_state &= ~IOSTARTED; 11736 foreground = (bp->b_xflags & BX_BKGRDMARKER) == 0; 11737 /* 11738 * Release journal work that was waiting on the write. 11739 */ 11740 handle_jwork(&bmsafemap->sm_freewr); 11741 11742 /* 11743 * Restore unwritten inode allocation pending jaddref writes. 11744 */ 11745 if (!LIST_EMPTY(&bmsafemap->sm_jaddrefhd)) { 11746 cgp = (struct cg *)bp->b_data; 11747 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 11748 inosused = cg_inosused(cgp); 11749 LIST_FOREACH_SAFE(jaddref, &bmsafemap->sm_jaddrefhd, 11750 ja_bmdeps, jatmp) { 11751 if ((jaddref->ja_state & UNDONE) == 0) 11752 continue; 11753 ino = jaddref->ja_ino % fs->fs_ipg; 11754 if (isset(inosused, ino)) 11755 panic("handle_written_bmsafemap: " 11756 "re-allocated inode"); 11757 /* Do the roll-forward only if it's a real copy. */ 11758 if (foreground) { 11759 if ((jaddref->ja_mode & IFMT) == IFDIR) 11760 cgp->cg_cs.cs_ndir++; 11761 cgp->cg_cs.cs_nifree--; 11762 setbit(inosused, ino); 11763 chgs = 1; 11764 } 11765 jaddref->ja_state &= ~UNDONE; 11766 jaddref->ja_state |= ATTACHED; 11767 free_jaddref(jaddref); 11768 } 11769 } 11770 /* 11771 * Restore any block allocations which are pending journal writes. 11772 */ 11773 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) { 11774 cgp = (struct cg *)bp->b_data; 11775 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 11776 blksfree = cg_blksfree(cgp); 11777 LIST_FOREACH_SAFE(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps, 11778 jntmp) { 11779 if ((jnewblk->jn_state & UNDONE) == 0) 11780 continue; 11781 /* Do the roll-forward only if it's a real copy. */ 11782 if (foreground && 11783 jnewblk_rollforward(jnewblk, fs, cgp, blksfree)) 11784 chgs = 1; 11785 jnewblk->jn_state &= ~(UNDONE | NEWBLOCK); 11786 jnewblk->jn_state |= ATTACHED; 11787 free_jnewblk(jnewblk); 11788 } 11789 } 11790 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkwr))) { 11791 newblk->nb_state |= DEPCOMPLETE; 11792 newblk->nb_state &= ~ONDEPLIST; 11793 newblk->nb_bmsafemap = NULL; 11794 LIST_REMOVE(newblk, nb_deps); 11795 if (newblk->nb_list.wk_type == D_ALLOCDIRECT) 11796 handle_allocdirect_partdone( 11797 WK_ALLOCDIRECT(&newblk->nb_list), NULL); 11798 else if (newblk->nb_list.wk_type == D_ALLOCINDIR) 11799 handle_allocindir_partdone( 11800 WK_ALLOCINDIR(&newblk->nb_list)); 11801 else if (newblk->nb_list.wk_type != D_NEWBLK) 11802 panic("handle_written_bmsafemap: Unexpected type: %s", 11803 TYPENAME(newblk->nb_list.wk_type)); 11804 } 11805 while ((inodedep = LIST_FIRST(&bmsafemap->sm_inodedepwr)) != NULL) { 11806 inodedep->id_state |= DEPCOMPLETE; 11807 inodedep->id_state &= ~ONDEPLIST; 11808 LIST_REMOVE(inodedep, id_deps); 11809 inodedep->id_bmsafemap = NULL; 11810 } 11811 LIST_REMOVE(bmsafemap, sm_next); 11812 if (chgs == 0 && LIST_EMPTY(&bmsafemap->sm_jaddrefhd) && 11813 LIST_EMPTY(&bmsafemap->sm_jnewblkhd) && 11814 LIST_EMPTY(&bmsafemap->sm_newblkhd) && 11815 LIST_EMPTY(&bmsafemap->sm_inodedephd) && 11816 LIST_EMPTY(&bmsafemap->sm_freehd)) { 11817 LIST_REMOVE(bmsafemap, sm_hash); 11818 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 11819 return (0); 11820 } 11821 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next); 11822 if (foreground) 11823 bdirty(bp); 11824 return (1); 11825} 11826 11827/* 11828 * Try to free a mkdir dependency. 11829 */ 11830static void 11831complete_mkdir(mkdir) 11832 struct mkdir *mkdir; 11833{ 11834 struct diradd *dap; 11835 11836 if ((mkdir->md_state & ALLCOMPLETE) != ALLCOMPLETE) 11837 return; 11838 LIST_REMOVE(mkdir, md_mkdirs); 11839 dap = mkdir->md_diradd; 11840 dap->da_state &= ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)); 11841 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) { 11842 dap->da_state |= DEPCOMPLETE; 11843 complete_diradd(dap); 11844 } 11845 WORKITEM_FREE(mkdir, D_MKDIR); 11846} 11847 11848/* 11849 * Handle the completion of a mkdir dependency. 11850 */ 11851static void 11852handle_written_mkdir(mkdir, type) 11853 struct mkdir *mkdir; 11854 int type; 11855{ 11856 11857 if ((mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)) != type) 11858 panic("handle_written_mkdir: bad type"); 11859 mkdir->md_state |= COMPLETE; 11860 complete_mkdir(mkdir); 11861} 11862 11863static int 11864free_pagedep(pagedep) 11865 struct pagedep *pagedep; 11866{ 11867 int i; 11868 11869 if (pagedep->pd_state & NEWBLOCK) 11870 return (0); 11871 if (!LIST_EMPTY(&pagedep->pd_dirremhd)) 11872 return (0); 11873 for (i = 0; i < DAHASHSZ; i++) 11874 if (!LIST_EMPTY(&pagedep->pd_diraddhd[i])) 11875 return (0); 11876 if (!LIST_EMPTY(&pagedep->pd_pendinghd)) 11877 return (0); 11878 if (!LIST_EMPTY(&pagedep->pd_jmvrefhd)) 11879 return (0); 11880 if (pagedep->pd_state & ONWORKLIST) 11881 WORKLIST_REMOVE(&pagedep->pd_list); 11882 LIST_REMOVE(pagedep, pd_hash); 11883 WORKITEM_FREE(pagedep, D_PAGEDEP); 11884 11885 return (1); 11886} 11887 11888/* 11889 * Called from within softdep_disk_write_complete above. 11890 * A write operation was just completed. Removed inodes can 11891 * now be freed and associated block pointers may be committed. 11892 * Note that this routine is always called from interrupt level 11893 * with further splbio interrupts blocked. 11894 */ 11895static int 11896handle_written_filepage(pagedep, bp) 11897 struct pagedep *pagedep; 11898 struct buf *bp; /* buffer containing the written page */ 11899{ 11900 struct dirrem *dirrem; 11901 struct diradd *dap, *nextdap; 11902 struct direct *ep; 11903 int i, chgs; 11904 11905 if ((pagedep->pd_state & IOSTARTED) == 0) 11906 panic("handle_written_filepage: not started"); 11907 pagedep->pd_state &= ~IOSTARTED; 11908 /* 11909 * Process any directory removals that have been committed. 11910 */ 11911 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 11912 LIST_REMOVE(dirrem, dm_next); 11913 dirrem->dm_state |= COMPLETE; 11914 dirrem->dm_dirinum = pagedep->pd_ino; 11915 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd), 11916 ("handle_written_filepage: Journal entries not written.")); 11917 add_to_worklist(&dirrem->dm_list, 0); 11918 } 11919 /* 11920 * Free any directory additions that have been committed. 11921 * If it is a newly allocated block, we have to wait until 11922 * the on-disk directory inode claims the new block. 11923 */ 11924 if ((pagedep->pd_state & NEWBLOCK) == 0) 11925 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 11926 free_diradd(dap, NULL); 11927 /* 11928 * Uncommitted directory entries must be restored. 11929 */ 11930 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 11931 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 11932 dap = nextdap) { 11933 nextdap = LIST_NEXT(dap, da_pdlist); 11934 if (dap->da_state & ATTACHED) 11935 panic("handle_written_filepage: attached"); 11936 ep = (struct direct *) 11937 ((char *)bp->b_data + dap->da_offset); 11938 ep->d_ino = dap->da_newinum; 11939 dap->da_state &= ~UNDONE; 11940 dap->da_state |= ATTACHED; 11941 chgs = 1; 11942 /* 11943 * If the inode referenced by the directory has 11944 * been written out, then the dependency can be 11945 * moved to the pending list. 11946 */ 11947 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 11948 LIST_REMOVE(dap, da_pdlist); 11949 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 11950 da_pdlist); 11951 } 11952 } 11953 } 11954 /* 11955 * If there were any rollbacks in the directory, then it must be 11956 * marked dirty so that its will eventually get written back in 11957 * its correct form. 11958 */ 11959 if (chgs) { 11960 if ((bp->b_flags & B_DELWRI) == 0) 11961 stat_dir_entry++; 11962 bdirty(bp); 11963 return (1); 11964 } 11965 /* 11966 * If we are not waiting for a new directory block to be 11967 * claimed by its inode, then the pagedep will be freed. 11968 * Otherwise it will remain to track any new entries on 11969 * the page in case they are fsync'ed. 11970 */ 11971 free_pagedep(pagedep); 11972 return (0); 11973} 11974 11975/* 11976 * Writing back in-core inode structures. 11977 * 11978 * The filesystem only accesses an inode's contents when it occupies an 11979 * "in-core" inode structure. These "in-core" structures are separate from 11980 * the page frames used to cache inode blocks. Only the latter are 11981 * transferred to/from the disk. So, when the updated contents of the 11982 * "in-core" inode structure are copied to the corresponding in-memory inode 11983 * block, the dependencies are also transferred. The following procedure is 11984 * called when copying a dirty "in-core" inode to a cached inode block. 11985 */ 11986 11987/* 11988 * Called when an inode is loaded from disk. If the effective link count 11989 * differed from the actual link count when it was last flushed, then we 11990 * need to ensure that the correct effective link count is put back. 11991 */ 11992void 11993softdep_load_inodeblock(ip) 11994 struct inode *ip; /* the "in_core" copy of the inode */ 11995{ 11996 struct inodedep *inodedep; 11997 11998 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0, 11999 ("softdep_load_inodeblock called on non-softdep filesystem")); 12000 /* 12001 * Check for alternate nlink count. 12002 */ 12003 ip->i_effnlink = ip->i_nlink; 12004 ACQUIRE_LOCK(ip->i_ump); 12005 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, 12006 &inodedep) == 0) { 12007 FREE_LOCK(ip->i_ump); 12008 return; 12009 } 12010 ip->i_effnlink -= inodedep->id_nlinkdelta; 12011 FREE_LOCK(ip->i_ump); 12012} 12013 12014/* 12015 * This routine is called just before the "in-core" inode 12016 * information is to be copied to the in-memory inode block. 12017 * Recall that an inode block contains several inodes. If 12018 * the force flag is set, then the dependencies will be 12019 * cleared so that the update can always be made. Note that 12020 * the buffer is locked when this routine is called, so we 12021 * will never be in the middle of writing the inode block 12022 * to disk. 12023 */ 12024void 12025softdep_update_inodeblock(ip, bp, waitfor) 12026 struct inode *ip; /* the "in_core" copy of the inode */ 12027 struct buf *bp; /* the buffer containing the inode block */ 12028 int waitfor; /* nonzero => update must be allowed */ 12029{ 12030 struct inodedep *inodedep; 12031 struct inoref *inoref; 12032 struct ufsmount *ump; 12033 struct worklist *wk; 12034 struct mount *mp; 12035 struct buf *ibp; 12036 struct fs *fs; 12037 int error; 12038 12039 ump = ip->i_ump; 12040 mp = UFSTOVFS(ump); 12041 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 12042 ("softdep_update_inodeblock called on non-softdep filesystem")); 12043 fs = ip->i_fs; 12044 /* 12045 * Preserve the freelink that is on disk. clear_unlinked_inodedep() 12046 * does not have access to the in-core ip so must write directly into 12047 * the inode block buffer when setting freelink. 12048 */ 12049 if (fs->fs_magic == FS_UFS1_MAGIC) 12050 DIP_SET(ip, i_freelink, ((struct ufs1_dinode *)bp->b_data + 12051 ino_to_fsbo(fs, ip->i_number))->di_freelink); 12052 else 12053 DIP_SET(ip, i_freelink, ((struct ufs2_dinode *)bp->b_data + 12054 ino_to_fsbo(fs, ip->i_number))->di_freelink); 12055 /* 12056 * If the effective link count is not equal to the actual link 12057 * count, then we must track the difference in an inodedep while 12058 * the inode is (potentially) tossed out of the cache. Otherwise, 12059 * if there is no existing inodedep, then there are no dependencies 12060 * to track. 12061 */ 12062 ACQUIRE_LOCK(ump); 12063again: 12064 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 12065 FREE_LOCK(ump); 12066 if (ip->i_effnlink != ip->i_nlink) 12067 panic("softdep_update_inodeblock: bad link count"); 12068 return; 12069 } 12070 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) 12071 panic("softdep_update_inodeblock: bad delta"); 12072 /* 12073 * If we're flushing all dependencies we must also move any waiting 12074 * for journal writes onto the bufwait list prior to I/O. 12075 */ 12076 if (waitfor) { 12077 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12078 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12079 == DEPCOMPLETE) { 12080 jwait(&inoref->if_list, MNT_WAIT); 12081 goto again; 12082 } 12083 } 12084 } 12085 /* 12086 * Changes have been initiated. Anything depending on these 12087 * changes cannot occur until this inode has been written. 12088 */ 12089 inodedep->id_state &= ~COMPLETE; 12090 if ((inodedep->id_state & ONWORKLIST) == 0) 12091 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 12092 /* 12093 * Any new dependencies associated with the incore inode must 12094 * now be moved to the list associated with the buffer holding 12095 * the in-memory copy of the inode. Once merged process any 12096 * allocdirects that are completed by the merger. 12097 */ 12098 merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt); 12099 if (!TAILQ_EMPTY(&inodedep->id_inoupdt)) 12100 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt), 12101 NULL); 12102 merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt); 12103 if (!TAILQ_EMPTY(&inodedep->id_extupdt)) 12104 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt), 12105 NULL); 12106 /* 12107 * Now that the inode has been pushed into the buffer, the 12108 * operations dependent on the inode being written to disk 12109 * can be moved to the id_bufwait so that they will be 12110 * processed when the buffer I/O completes. 12111 */ 12112 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 12113 WORKLIST_REMOVE(wk); 12114 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 12115 } 12116 /* 12117 * Newly allocated inodes cannot be written until the bitmap 12118 * that allocates them have been written (indicated by 12119 * DEPCOMPLETE being set in id_state). If we are doing a 12120 * forced sync (e.g., an fsync on a file), we force the bitmap 12121 * to be written so that the update can be done. 12122 */ 12123 if (waitfor == 0) { 12124 FREE_LOCK(ump); 12125 return; 12126 } 12127retry: 12128 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) != 0) { 12129 FREE_LOCK(ump); 12130 return; 12131 } 12132 ibp = inodedep->id_bmsafemap->sm_buf; 12133 ibp = getdirtybuf(ibp, LOCK_PTR(ump), MNT_WAIT); 12134 if (ibp == NULL) { 12135 /* 12136 * If ibp came back as NULL, the dependency could have been 12137 * freed while we slept. Look it up again, and check to see 12138 * that it has completed. 12139 */ 12140 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 12141 goto retry; 12142 FREE_LOCK(ump); 12143 return; 12144 } 12145 FREE_LOCK(ump); 12146 if ((error = bwrite(ibp)) != 0) 12147 softdep_error("softdep_update_inodeblock: bwrite", error); 12148} 12149 12150/* 12151 * Merge the a new inode dependency list (such as id_newinoupdt) into an 12152 * old inode dependency list (such as id_inoupdt). This routine must be 12153 * called with splbio interrupts blocked. 12154 */ 12155static void 12156merge_inode_lists(newlisthead, oldlisthead) 12157 struct allocdirectlst *newlisthead; 12158 struct allocdirectlst *oldlisthead; 12159{ 12160 struct allocdirect *listadp, *newadp; 12161 12162 newadp = TAILQ_FIRST(newlisthead); 12163 for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) { 12164 if (listadp->ad_offset < newadp->ad_offset) { 12165 listadp = TAILQ_NEXT(listadp, ad_next); 12166 continue; 12167 } 12168 TAILQ_REMOVE(newlisthead, newadp, ad_next); 12169 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 12170 if (listadp->ad_offset == newadp->ad_offset) { 12171 allocdirect_merge(oldlisthead, newadp, 12172 listadp); 12173 listadp = newadp; 12174 } 12175 newadp = TAILQ_FIRST(newlisthead); 12176 } 12177 while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) { 12178 TAILQ_REMOVE(newlisthead, newadp, ad_next); 12179 TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next); 12180 } 12181} 12182 12183/* 12184 * If we are doing an fsync, then we must ensure that any directory 12185 * entries for the inode have been written after the inode gets to disk. 12186 */ 12187int 12188softdep_fsync(vp) 12189 struct vnode *vp; /* the "in_core" copy of the inode */ 12190{ 12191 struct inodedep *inodedep; 12192 struct pagedep *pagedep; 12193 struct inoref *inoref; 12194 struct ufsmount *ump; 12195 struct worklist *wk; 12196 struct diradd *dap; 12197 struct mount *mp; 12198 struct vnode *pvp; 12199 struct inode *ip; 12200 struct buf *bp; 12201 struct fs *fs; 12202 struct thread *td = curthread; 12203 int error, flushparent, pagedep_new_block; 12204 ino_t parentino; 12205 ufs_lbn_t lbn; 12206 12207 ip = VTOI(vp); 12208 fs = ip->i_fs; 12209 ump = ip->i_ump; 12210 mp = vp->v_mount; 12211 if (MOUNTEDSOFTDEP(mp) == 0) 12212 return (0); 12213 ACQUIRE_LOCK(ump); 12214restart: 12215 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 12216 FREE_LOCK(ump); 12217 return (0); 12218 } 12219 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12220 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12221 == DEPCOMPLETE) { 12222 jwait(&inoref->if_list, MNT_WAIT); 12223 goto restart; 12224 } 12225 } 12226 if (!LIST_EMPTY(&inodedep->id_inowait) || 12227 !TAILQ_EMPTY(&inodedep->id_extupdt) || 12228 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 12229 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 12230 !TAILQ_EMPTY(&inodedep->id_newinoupdt)) 12231 panic("softdep_fsync: pending ops %p", inodedep); 12232 for (error = 0, flushparent = 0; ; ) { 12233 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 12234 break; 12235 if (wk->wk_type != D_DIRADD) 12236 panic("softdep_fsync: Unexpected type %s", 12237 TYPENAME(wk->wk_type)); 12238 dap = WK_DIRADD(wk); 12239 /* 12240 * Flush our parent if this directory entry has a MKDIR_PARENT 12241 * dependency or is contained in a newly allocated block. 12242 */ 12243 if (dap->da_state & DIRCHG) 12244 pagedep = dap->da_previous->dm_pagedep; 12245 else 12246 pagedep = dap->da_pagedep; 12247 parentino = pagedep->pd_ino; 12248 lbn = pagedep->pd_lbn; 12249 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) 12250 panic("softdep_fsync: dirty"); 12251 if ((dap->da_state & MKDIR_PARENT) || 12252 (pagedep->pd_state & NEWBLOCK)) 12253 flushparent = 1; 12254 else 12255 flushparent = 0; 12256 /* 12257 * If we are being fsync'ed as part of vgone'ing this vnode, 12258 * then we will not be able to release and recover the 12259 * vnode below, so we just have to give up on writing its 12260 * directory entry out. It will eventually be written, just 12261 * not now, but then the user was not asking to have it 12262 * written, so we are not breaking any promises. 12263 */ 12264 if (vp->v_iflag & VI_DOOMED) 12265 break; 12266 /* 12267 * We prevent deadlock by always fetching inodes from the 12268 * root, moving down the directory tree. Thus, when fetching 12269 * our parent directory, we first try to get the lock. If 12270 * that fails, we must unlock ourselves before requesting 12271 * the lock on our parent. See the comment in ufs_lookup 12272 * for details on possible races. 12273 */ 12274 FREE_LOCK(ump); 12275 if (ffs_vgetf(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp, 12276 FFSV_FORCEINSMQ)) { 12277 error = vfs_busy(mp, MBF_NOWAIT); 12278 if (error != 0) { 12279 vfs_ref(mp); 12280 VOP_UNLOCK(vp, 0); 12281 error = vfs_busy(mp, 0); 12282 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 12283 vfs_rel(mp); 12284 if (error != 0) 12285 return (ENOENT); 12286 if (vp->v_iflag & VI_DOOMED) { 12287 vfs_unbusy(mp); 12288 return (ENOENT); 12289 } 12290 } 12291 VOP_UNLOCK(vp, 0); 12292 error = ffs_vgetf(mp, parentino, LK_EXCLUSIVE, 12293 &pvp, FFSV_FORCEINSMQ); 12294 vfs_unbusy(mp); 12295 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 12296 if (vp->v_iflag & VI_DOOMED) { 12297 if (error == 0) 12298 vput(pvp); 12299 error = ENOENT; 12300 } 12301 if (error != 0) 12302 return (error); 12303 } 12304 /* 12305 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps 12306 * that are contained in direct blocks will be resolved by 12307 * doing a ffs_update. Pagedeps contained in indirect blocks 12308 * may require a complete sync'ing of the directory. So, we 12309 * try the cheap and fast ffs_update first, and if that fails, 12310 * then we do the slower ffs_syncvnode of the directory. 12311 */ 12312 if (flushparent) { 12313 int locked; 12314 12315 if ((error = ffs_update(pvp, 1)) != 0) { 12316 vput(pvp); 12317 return (error); 12318 } 12319 ACQUIRE_LOCK(ump); 12320 locked = 1; 12321 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) { 12322 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) != NULL) { 12323 if (wk->wk_type != D_DIRADD) 12324 panic("softdep_fsync: Unexpected type %s", 12325 TYPENAME(wk->wk_type)); 12326 dap = WK_DIRADD(wk); 12327 if (dap->da_state & DIRCHG) 12328 pagedep = dap->da_previous->dm_pagedep; 12329 else 12330 pagedep = dap->da_pagedep; 12331 pagedep_new_block = pagedep->pd_state & NEWBLOCK; 12332 FREE_LOCK(ump); 12333 locked = 0; 12334 if (pagedep_new_block && (error = 12335 ffs_syncvnode(pvp, MNT_WAIT, 0))) { 12336 vput(pvp); 12337 return (error); 12338 } 12339 } 12340 } 12341 if (locked) 12342 FREE_LOCK(ump); 12343 } 12344 /* 12345 * Flush directory page containing the inode's name. 12346 */ 12347 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred, 12348 &bp); 12349 if (error == 0) 12350 error = bwrite(bp); 12351 else 12352 brelse(bp); 12353 vput(pvp); 12354 if (error != 0) 12355 return (error); 12356 ACQUIRE_LOCK(ump); 12357 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 12358 break; 12359 } 12360 FREE_LOCK(ump); 12361 return (0); 12362} 12363 12364/* 12365 * Flush all the dirty bitmaps associated with the block device 12366 * before flushing the rest of the dirty blocks so as to reduce 12367 * the number of dependencies that will have to be rolled back. 12368 * 12369 * XXX Unused? 12370 */ 12371void 12372softdep_fsync_mountdev(vp) 12373 struct vnode *vp; 12374{ 12375 struct buf *bp, *nbp; 12376 struct worklist *wk; 12377 struct bufobj *bo; 12378 12379 if (!vn_isdisk(vp, NULL)) 12380 panic("softdep_fsync_mountdev: vnode not a disk"); 12381 bo = &vp->v_bufobj; 12382restart: 12383 BO_LOCK(bo); 12384 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 12385 /* 12386 * If it is already scheduled, skip to the next buffer. 12387 */ 12388 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 12389 continue; 12390 12391 if ((bp->b_flags & B_DELWRI) == 0) 12392 panic("softdep_fsync_mountdev: not dirty"); 12393 /* 12394 * We are only interested in bitmaps with outstanding 12395 * dependencies. 12396 */ 12397 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 12398 wk->wk_type != D_BMSAFEMAP || 12399 (bp->b_vflags & BV_BKGRDINPROG)) { 12400 BUF_UNLOCK(bp); 12401 continue; 12402 } 12403 BO_UNLOCK(bo); 12404 bremfree(bp); 12405 (void) bawrite(bp); 12406 goto restart; 12407 } 12408 drain_output(vp); 12409 BO_UNLOCK(bo); 12410} 12411 12412/* 12413 * Sync all cylinder groups that were dirty at the time this function is 12414 * called. Newly dirtied cgs will be inserted before the sentinel. This 12415 * is used to flush freedep activity that may be holding up writes to a 12416 * indirect block. 12417 */ 12418static int 12419sync_cgs(mp, waitfor) 12420 struct mount *mp; 12421 int waitfor; 12422{ 12423 struct bmsafemap *bmsafemap; 12424 struct bmsafemap *sentinel; 12425 struct ufsmount *ump; 12426 struct buf *bp; 12427 int error; 12428 12429 sentinel = malloc(sizeof(*sentinel), M_BMSAFEMAP, M_ZERO | M_WAITOK); 12430 sentinel->sm_cg = -1; 12431 ump = VFSTOUFS(mp); 12432 error = 0; 12433 ACQUIRE_LOCK(ump); 12434 LIST_INSERT_HEAD(&ump->softdep_dirtycg, sentinel, sm_next); 12435 for (bmsafemap = LIST_NEXT(sentinel, sm_next); bmsafemap != NULL; 12436 bmsafemap = LIST_NEXT(sentinel, sm_next)) { 12437 /* Skip sentinels and cgs with no work to release. */ 12438 if (bmsafemap->sm_cg == -1 || 12439 (LIST_EMPTY(&bmsafemap->sm_freehd) && 12440 LIST_EMPTY(&bmsafemap->sm_freewr))) { 12441 LIST_REMOVE(sentinel, sm_next); 12442 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next); 12443 continue; 12444 } 12445 /* 12446 * If we don't get the lock and we're waiting try again, if 12447 * not move on to the next buf and try to sync it. 12448 */ 12449 bp = getdirtybuf(bmsafemap->sm_buf, LOCK_PTR(ump), waitfor); 12450 if (bp == NULL && waitfor == MNT_WAIT) 12451 continue; 12452 LIST_REMOVE(sentinel, sm_next); 12453 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next); 12454 if (bp == NULL) 12455 continue; 12456 FREE_LOCK(ump); 12457 if (waitfor == MNT_NOWAIT) 12458 bawrite(bp); 12459 else 12460 error = bwrite(bp); 12461 ACQUIRE_LOCK(ump); 12462 if (error) 12463 break; 12464 } 12465 LIST_REMOVE(sentinel, sm_next); 12466 FREE_LOCK(ump); 12467 free(sentinel, M_BMSAFEMAP); 12468 return (error); 12469} 12470 12471/* 12472 * This routine is called when we are trying to synchronously flush a 12473 * file. This routine must eliminate any filesystem metadata dependencies 12474 * so that the syncing routine can succeed. 12475 */ 12476int 12477softdep_sync_metadata(struct vnode *vp) 12478{ 12479 struct inode *ip; 12480 int error; 12481 12482 ip = VTOI(vp); 12483 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0, 12484 ("softdep_sync_metadata called on non-softdep filesystem")); 12485 /* 12486 * Ensure that any direct block dependencies have been cleared, 12487 * truncations are started, and inode references are journaled. 12488 */ 12489 ACQUIRE_LOCK(ip->i_ump); 12490 /* 12491 * Write all journal records to prevent rollbacks on devvp. 12492 */ 12493 if (vp->v_type == VCHR) 12494 softdep_flushjournal(vp->v_mount); 12495 error = flush_inodedep_deps(vp, vp->v_mount, ip->i_number); 12496 /* 12497 * Ensure that all truncates are written so we won't find deps on 12498 * indirect blocks. 12499 */ 12500 process_truncates(vp); 12501 FREE_LOCK(ip->i_ump); 12502 12503 return (error); 12504} 12505 12506/* 12507 * This routine is called when we are attempting to sync a buf with 12508 * dependencies. If waitfor is MNT_NOWAIT it attempts to schedule any 12509 * other IO it can but returns EBUSY if the buffer is not yet able to 12510 * be written. Dependencies which will not cause rollbacks will always 12511 * return 0. 12512 */ 12513int 12514softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor) 12515{ 12516 struct indirdep *indirdep; 12517 struct pagedep *pagedep; 12518 struct allocindir *aip; 12519 struct newblk *newblk; 12520 struct ufsmount *ump; 12521 struct buf *nbp; 12522 struct worklist *wk; 12523 int i, error; 12524 12525 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 12526 ("softdep_sync_buf called on non-softdep filesystem")); 12527 /* 12528 * For VCHR we just don't want to force flush any dependencies that 12529 * will cause rollbacks. 12530 */ 12531 if (vp->v_type == VCHR) { 12532 if (waitfor == MNT_NOWAIT && softdep_count_dependencies(bp, 0)) 12533 return (EBUSY); 12534 return (0); 12535 } 12536 ump = VTOI(vp)->i_ump; 12537 ACQUIRE_LOCK(ump); 12538 /* 12539 * As we hold the buffer locked, none of its dependencies 12540 * will disappear. 12541 */ 12542 error = 0; 12543top: 12544 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 12545 switch (wk->wk_type) { 12546 12547 case D_ALLOCDIRECT: 12548 case D_ALLOCINDIR: 12549 newblk = WK_NEWBLK(wk); 12550 if (newblk->nb_jnewblk != NULL) { 12551 if (waitfor == MNT_NOWAIT) { 12552 error = EBUSY; 12553 goto out_unlock; 12554 } 12555 jwait(&newblk->nb_jnewblk->jn_list, waitfor); 12556 goto top; 12557 } 12558 if (newblk->nb_state & DEPCOMPLETE || 12559 waitfor == MNT_NOWAIT) 12560 continue; 12561 nbp = newblk->nb_bmsafemap->sm_buf; 12562 nbp = getdirtybuf(nbp, LOCK_PTR(ump), waitfor); 12563 if (nbp == NULL) 12564 goto top; 12565 FREE_LOCK(ump); 12566 if ((error = bwrite(nbp)) != 0) 12567 goto out; 12568 ACQUIRE_LOCK(ump); 12569 continue; 12570 12571 case D_INDIRDEP: 12572 indirdep = WK_INDIRDEP(wk); 12573 if (waitfor == MNT_NOWAIT) { 12574 if (!TAILQ_EMPTY(&indirdep->ir_trunc) || 12575 !LIST_EMPTY(&indirdep->ir_deplisthd)) { 12576 error = EBUSY; 12577 goto out_unlock; 12578 } 12579 } 12580 if (!TAILQ_EMPTY(&indirdep->ir_trunc)) 12581 panic("softdep_sync_buf: truncation pending."); 12582 restart: 12583 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 12584 newblk = (struct newblk *)aip; 12585 if (newblk->nb_jnewblk != NULL) { 12586 jwait(&newblk->nb_jnewblk->jn_list, 12587 waitfor); 12588 goto restart; 12589 } 12590 if (newblk->nb_state & DEPCOMPLETE) 12591 continue; 12592 nbp = newblk->nb_bmsafemap->sm_buf; 12593 nbp = getdirtybuf(nbp, LOCK_PTR(ump), waitfor); 12594 if (nbp == NULL) 12595 goto restart; 12596 FREE_LOCK(ump); 12597 if ((error = bwrite(nbp)) != 0) 12598 goto out; 12599 ACQUIRE_LOCK(ump); 12600 goto restart; 12601 } 12602 continue; 12603 12604 case D_PAGEDEP: 12605 /* 12606 * Only flush directory entries in synchronous passes. 12607 */ 12608 if (waitfor != MNT_WAIT) { 12609 error = EBUSY; 12610 goto out_unlock; 12611 } 12612 /* 12613 * While syncing snapshots, we must allow recursive 12614 * lookups. 12615 */ 12616 BUF_AREC(bp); 12617 /* 12618 * We are trying to sync a directory that may 12619 * have dependencies on both its own metadata 12620 * and/or dependencies on the inodes of any 12621 * recently allocated files. We walk its diradd 12622 * lists pushing out the associated inode. 12623 */ 12624 pagedep = WK_PAGEDEP(wk); 12625 for (i = 0; i < DAHASHSZ; i++) { 12626 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 12627 continue; 12628 if ((error = flush_pagedep_deps(vp, wk->wk_mp, 12629 &pagedep->pd_diraddhd[i]))) { 12630 BUF_NOREC(bp); 12631 goto out_unlock; 12632 } 12633 } 12634 BUF_NOREC(bp); 12635 continue; 12636 12637 case D_FREEWORK: 12638 case D_FREEDEP: 12639 case D_JSEGDEP: 12640 case D_JNEWBLK: 12641 continue; 12642 12643 default: 12644 panic("softdep_sync_buf: Unknown type %s", 12645 TYPENAME(wk->wk_type)); 12646 /* NOTREACHED */ 12647 } 12648 } 12649out_unlock: 12650 FREE_LOCK(ump); 12651out: 12652 return (error); 12653} 12654 12655/* 12656 * Flush the dependencies associated with an inodedep. 12657 * Called with splbio blocked. 12658 */ 12659static int 12660flush_inodedep_deps(vp, mp, ino) 12661 struct vnode *vp; 12662 struct mount *mp; 12663 ino_t ino; 12664{ 12665 struct inodedep *inodedep; 12666 struct inoref *inoref; 12667 struct ufsmount *ump; 12668 int error, waitfor; 12669 12670 /* 12671 * This work is done in two passes. The first pass grabs most 12672 * of the buffers and begins asynchronously writing them. The 12673 * only way to wait for these asynchronous writes is to sleep 12674 * on the filesystem vnode which may stay busy for a long time 12675 * if the filesystem is active. So, instead, we make a second 12676 * pass over the dependencies blocking on each write. In the 12677 * usual case we will be blocking against a write that we 12678 * initiated, so when it is done the dependency will have been 12679 * resolved. Thus the second pass is expected to end quickly. 12680 * We give a brief window at the top of the loop to allow 12681 * any pending I/O to complete. 12682 */ 12683 ump = VFSTOUFS(mp); 12684 LOCK_OWNED(ump); 12685 for (error = 0, waitfor = MNT_NOWAIT; ; ) { 12686 if (error) 12687 return (error); 12688 FREE_LOCK(ump); 12689 ACQUIRE_LOCK(ump); 12690restart: 12691 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 12692 return (0); 12693 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12694 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12695 == DEPCOMPLETE) { 12696 jwait(&inoref->if_list, MNT_WAIT); 12697 goto restart; 12698 } 12699 } 12700 if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) || 12701 flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) || 12702 flush_deplist(&inodedep->id_extupdt, waitfor, &error) || 12703 flush_deplist(&inodedep->id_newextupdt, waitfor, &error)) 12704 continue; 12705 /* 12706 * If pass2, we are done, otherwise do pass 2. 12707 */ 12708 if (waitfor == MNT_WAIT) 12709 break; 12710 waitfor = MNT_WAIT; 12711 } 12712 /* 12713 * Try freeing inodedep in case all dependencies have been removed. 12714 */ 12715 if (inodedep_lookup(mp, ino, 0, &inodedep) != 0) 12716 (void) free_inodedep(inodedep); 12717 return (0); 12718} 12719 12720/* 12721 * Flush an inode dependency list. 12722 * Called with splbio blocked. 12723 */ 12724static int 12725flush_deplist(listhead, waitfor, errorp) 12726 struct allocdirectlst *listhead; 12727 int waitfor; 12728 int *errorp; 12729{ 12730 struct allocdirect *adp; 12731 struct newblk *newblk; 12732 struct ufsmount *ump; 12733 struct buf *bp; 12734 12735 if ((adp = TAILQ_FIRST(listhead)) == NULL) 12736 return (0); 12737 ump = VFSTOUFS(adp->ad_list.wk_mp); 12738 LOCK_OWNED(ump); 12739 TAILQ_FOREACH(adp, listhead, ad_next) { 12740 newblk = (struct newblk *)adp; 12741 if (newblk->nb_jnewblk != NULL) { 12742 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 12743 return (1); 12744 } 12745 if (newblk->nb_state & DEPCOMPLETE) 12746 continue; 12747 bp = newblk->nb_bmsafemap->sm_buf; 12748 bp = getdirtybuf(bp, LOCK_PTR(ump), waitfor); 12749 if (bp == NULL) { 12750 if (waitfor == MNT_NOWAIT) 12751 continue; 12752 return (1); 12753 } 12754 FREE_LOCK(ump); 12755 if (waitfor == MNT_NOWAIT) 12756 bawrite(bp); 12757 else 12758 *errorp = bwrite(bp); 12759 ACQUIRE_LOCK(ump); 12760 return (1); 12761 } 12762 return (0); 12763} 12764 12765/* 12766 * Flush dependencies associated with an allocdirect block. 12767 */ 12768static int 12769flush_newblk_dep(vp, mp, lbn) 12770 struct vnode *vp; 12771 struct mount *mp; 12772 ufs_lbn_t lbn; 12773{ 12774 struct newblk *newblk; 12775 struct ufsmount *ump; 12776 struct bufobj *bo; 12777 struct inode *ip; 12778 struct buf *bp; 12779 ufs2_daddr_t blkno; 12780 int error; 12781 12782 error = 0; 12783 bo = &vp->v_bufobj; 12784 ip = VTOI(vp); 12785 blkno = DIP(ip, i_db[lbn]); 12786 if (blkno == 0) 12787 panic("flush_newblk_dep: Missing block"); 12788 ump = VFSTOUFS(mp); 12789 ACQUIRE_LOCK(ump); 12790 /* 12791 * Loop until all dependencies related to this block are satisfied. 12792 * We must be careful to restart after each sleep in case a write 12793 * completes some part of this process for us. 12794 */ 12795 for (;;) { 12796 if (newblk_lookup(mp, blkno, 0, &newblk) == 0) { 12797 FREE_LOCK(ump); 12798 break; 12799 } 12800 if (newblk->nb_list.wk_type != D_ALLOCDIRECT) 12801 panic("flush_newblk_deps: Bad newblk %p", newblk); 12802 /* 12803 * Flush the journal. 12804 */ 12805 if (newblk->nb_jnewblk != NULL) { 12806 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 12807 continue; 12808 } 12809 /* 12810 * Write the bitmap dependency. 12811 */ 12812 if ((newblk->nb_state & DEPCOMPLETE) == 0) { 12813 bp = newblk->nb_bmsafemap->sm_buf; 12814 bp = getdirtybuf(bp, LOCK_PTR(ump), MNT_WAIT); 12815 if (bp == NULL) 12816 continue; 12817 FREE_LOCK(ump); 12818 error = bwrite(bp); 12819 if (error) 12820 break; 12821 ACQUIRE_LOCK(ump); 12822 continue; 12823 } 12824 /* 12825 * Write the buffer. 12826 */ 12827 FREE_LOCK(ump); 12828 BO_LOCK(bo); 12829 bp = gbincore(bo, lbn); 12830 if (bp != NULL) { 12831 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | 12832 LK_INTERLOCK, BO_LOCKPTR(bo)); 12833 if (error == ENOLCK) { 12834 ACQUIRE_LOCK(ump); 12835 continue; /* Slept, retry */ 12836 } 12837 if (error != 0) 12838 break; /* Failed */ 12839 if (bp->b_flags & B_DELWRI) { 12840 bremfree(bp); 12841 error = bwrite(bp); 12842 if (error) 12843 break; 12844 } else 12845 BUF_UNLOCK(bp); 12846 } else 12847 BO_UNLOCK(bo); 12848 /* 12849 * We have to wait for the direct pointers to 12850 * point at the newdirblk before the dependency 12851 * will go away. 12852 */ 12853 error = ffs_update(vp, 1); 12854 if (error) 12855 break; 12856 ACQUIRE_LOCK(ump); 12857 } 12858 return (error); 12859} 12860 12861/* 12862 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 12863 * Called with splbio blocked. 12864 */ 12865static int 12866flush_pagedep_deps(pvp, mp, diraddhdp) 12867 struct vnode *pvp; 12868 struct mount *mp; 12869 struct diraddhd *diraddhdp; 12870{ 12871 struct inodedep *inodedep; 12872 struct inoref *inoref; 12873 struct ufsmount *ump; 12874 struct diradd *dap; 12875 struct vnode *vp; 12876 int error = 0; 12877 struct buf *bp; 12878 ino_t inum; 12879 struct diraddhd unfinished; 12880 12881 LIST_INIT(&unfinished); 12882 ump = VFSTOUFS(mp); 12883 LOCK_OWNED(ump); 12884restart: 12885 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 12886 /* 12887 * Flush ourselves if this directory entry 12888 * has a MKDIR_PARENT dependency. 12889 */ 12890 if (dap->da_state & MKDIR_PARENT) { 12891 FREE_LOCK(ump); 12892 if ((error = ffs_update(pvp, 1)) != 0) 12893 break; 12894 ACQUIRE_LOCK(ump); 12895 /* 12896 * If that cleared dependencies, go on to next. 12897 */ 12898 if (dap != LIST_FIRST(diraddhdp)) 12899 continue; 12900 /* 12901 * All MKDIR_PARENT dependencies and all the 12902 * NEWBLOCK pagedeps that are contained in direct 12903 * blocks were resolved by doing above ffs_update. 12904 * Pagedeps contained in indirect blocks may 12905 * require a complete sync'ing of the directory. 12906 * We are in the midst of doing a complete sync, 12907 * so if they are not resolved in this pass we 12908 * defer them for now as they will be sync'ed by 12909 * our caller shortly. 12910 */ 12911 LIST_REMOVE(dap, da_pdlist); 12912 LIST_INSERT_HEAD(&unfinished, dap, da_pdlist); 12913 continue; 12914 } 12915 /* 12916 * A newly allocated directory must have its "." and 12917 * ".." entries written out before its name can be 12918 * committed in its parent. 12919 */ 12920 inum = dap->da_newinum; 12921 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 12922 panic("flush_pagedep_deps: lost inode1"); 12923 /* 12924 * Wait for any pending journal adds to complete so we don't 12925 * cause rollbacks while syncing. 12926 */ 12927 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12928 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12929 == DEPCOMPLETE) { 12930 jwait(&inoref->if_list, MNT_WAIT); 12931 goto restart; 12932 } 12933 } 12934 if (dap->da_state & MKDIR_BODY) { 12935 FREE_LOCK(ump); 12936 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp, 12937 FFSV_FORCEINSMQ))) 12938 break; 12939 error = flush_newblk_dep(vp, mp, 0); 12940 /* 12941 * If we still have the dependency we might need to 12942 * update the vnode to sync the new link count to 12943 * disk. 12944 */ 12945 if (error == 0 && dap == LIST_FIRST(diraddhdp)) 12946 error = ffs_update(vp, 1); 12947 vput(vp); 12948 if (error != 0) 12949 break; 12950 ACQUIRE_LOCK(ump); 12951 /* 12952 * If that cleared dependencies, go on to next. 12953 */ 12954 if (dap != LIST_FIRST(diraddhdp)) 12955 continue; 12956 if (dap->da_state & MKDIR_BODY) { 12957 inodedep_lookup(UFSTOVFS(ump), inum, 0, 12958 &inodedep); 12959 panic("flush_pagedep_deps: MKDIR_BODY " 12960 "inodedep %p dap %p vp %p", 12961 inodedep, dap, vp); 12962 } 12963 } 12964 /* 12965 * Flush the inode on which the directory entry depends. 12966 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 12967 * the only remaining dependency is that the updated inode 12968 * count must get pushed to disk. The inode has already 12969 * been pushed into its inode buffer (via VOP_UPDATE) at 12970 * the time of the reference count change. So we need only 12971 * locate that buffer, ensure that there will be no rollback 12972 * caused by a bitmap dependency, then write the inode buffer. 12973 */ 12974retry: 12975 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 12976 panic("flush_pagedep_deps: lost inode"); 12977 /* 12978 * If the inode still has bitmap dependencies, 12979 * push them to disk. 12980 */ 12981 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) == 0) { 12982 bp = inodedep->id_bmsafemap->sm_buf; 12983 bp = getdirtybuf(bp, LOCK_PTR(ump), MNT_WAIT); 12984 if (bp == NULL) 12985 goto retry; 12986 FREE_LOCK(ump); 12987 if ((error = bwrite(bp)) != 0) 12988 break; 12989 ACQUIRE_LOCK(ump); 12990 if (dap != LIST_FIRST(diraddhdp)) 12991 continue; 12992 } 12993 /* 12994 * If the inode is still sitting in a buffer waiting 12995 * to be written or waiting for the link count to be 12996 * adjusted update it here to flush it to disk. 12997 */ 12998 if (dap == LIST_FIRST(diraddhdp)) { 12999 FREE_LOCK(ump); 13000 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp, 13001 FFSV_FORCEINSMQ))) 13002 break; 13003 error = ffs_update(vp, 1); 13004 vput(vp); 13005 if (error) 13006 break; 13007 ACQUIRE_LOCK(ump); 13008 } 13009 /* 13010 * If we have failed to get rid of all the dependencies 13011 * then something is seriously wrong. 13012 */ 13013 if (dap == LIST_FIRST(diraddhdp)) { 13014 inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep); 13015 panic("flush_pagedep_deps: failed to flush " 13016 "inodedep %p ino %ju dap %p", 13017 inodedep, (uintmax_t)inum, dap); 13018 } 13019 } 13020 if (error) 13021 ACQUIRE_LOCK(ump); 13022 while ((dap = LIST_FIRST(&unfinished)) != NULL) { 13023 LIST_REMOVE(dap, da_pdlist); 13024 LIST_INSERT_HEAD(diraddhdp, dap, da_pdlist); 13025 } 13026 return (error); 13027} 13028 13029/* 13030 * A large burst of file addition or deletion activity can drive the 13031 * memory load excessively high. First attempt to slow things down 13032 * using the techniques below. If that fails, this routine requests 13033 * the offending operations to fall back to running synchronously 13034 * until the memory load returns to a reasonable level. 13035 */ 13036int 13037softdep_slowdown(vp) 13038 struct vnode *vp; 13039{ 13040 struct ufsmount *ump; 13041 int jlow; 13042 int max_softdeps_hard; 13043 13044 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 13045 ("softdep_slowdown called on non-softdep filesystem")); 13046 ump = VFSTOUFS(vp->v_mount); 13047 ACQUIRE_LOCK(ump); 13048 jlow = 0; 13049 /* 13050 * Check for journal space if needed. 13051 */ 13052 if (DOINGSUJ(vp)) { 13053 if (journal_space(ump, 0) == 0) 13054 jlow = 1; 13055 } 13056 /* 13057 * If the system is under its limits and our filesystem is 13058 * not responsible for more than our share of the usage and 13059 * we are not low on journal space, then no need to slow down. 13060 */ 13061 max_softdeps_hard = max_softdeps * 11 / 10; 13062 if (dep_current[D_DIRREM] < max_softdeps_hard / 2 && 13063 dep_current[D_INODEDEP] < max_softdeps_hard && 13064 dep_current[D_INDIRDEP] < max_softdeps_hard / 1000 && 13065 dep_current[D_FREEBLKS] < max_softdeps_hard && jlow == 0 && 13066 ump->softdep_curdeps[D_DIRREM] < 13067 (max_softdeps_hard / 2) / stat_flush_threads && 13068 ump->softdep_curdeps[D_INODEDEP] < 13069 max_softdeps_hard / stat_flush_threads && 13070 ump->softdep_curdeps[D_INDIRDEP] < 13071 (max_softdeps_hard / 1000) / stat_flush_threads && 13072 ump->softdep_curdeps[D_FREEBLKS] < 13073 max_softdeps_hard / stat_flush_threads) { 13074 FREE_LOCK(ump); 13075 return (0); 13076 } 13077 /* 13078 * If the journal is low or our filesystem is over its limit 13079 * then speedup the cleanup. 13080 */ 13081 if (ump->softdep_curdeps[D_INDIRDEP] < 13082 (max_softdeps_hard / 1000) / stat_flush_threads || jlow) 13083 softdep_speedup(ump); 13084 stat_sync_limit_hit += 1; 13085 FREE_LOCK(ump); 13086 /* 13087 * We only slow down the rate at which new dependencies are 13088 * generated if we are not using journaling. With journaling, 13089 * the cleanup should always be sufficient to keep things 13090 * under control. 13091 */ 13092 if (DOINGSUJ(vp)) 13093 return (0); 13094 return (1); 13095} 13096 13097/* 13098 * Called by the allocation routines when they are about to fail 13099 * in the hope that we can free up the requested resource (inodes 13100 * or disk space). 13101 * 13102 * First check to see if the work list has anything on it. If it has, 13103 * clean up entries until we successfully free the requested resource. 13104 * Because this process holds inodes locked, we cannot handle any remove 13105 * requests that might block on a locked inode as that could lead to 13106 * deadlock. If the worklist yields none of the requested resource, 13107 * start syncing out vnodes to free up the needed space. 13108 */ 13109int 13110softdep_request_cleanup(fs, vp, cred, resource) 13111 struct fs *fs; 13112 struct vnode *vp; 13113 struct ucred *cred; 13114 int resource; 13115{ 13116 struct ufsmount *ump; 13117 struct mount *mp; 13118 struct vnode *lvp, *mvp; 13119 long starttime; 13120 ufs2_daddr_t needed; 13121 int error; 13122 13123 /* 13124 * If we are being called because of a process doing a 13125 * copy-on-write, then it is not safe to process any 13126 * worklist items as we will recurse into the copyonwrite 13127 * routine. This will result in an incoherent snapshot. 13128 * If the vnode that we hold is a snapshot, we must avoid 13129 * handling other resources that could cause deadlock. 13130 */ 13131 if ((curthread->td_pflags & TDP_COWINPROGRESS) || IS_SNAPSHOT(VTOI(vp))) 13132 return (0); 13133 13134 if (resource == FLUSH_BLOCKS_WAIT) 13135 stat_cleanup_blkrequests += 1; 13136 else 13137 stat_cleanup_inorequests += 1; 13138 13139 mp = vp->v_mount; 13140 ump = VFSTOUFS(mp); 13141 mtx_assert(UFS_MTX(ump), MA_OWNED); 13142 UFS_UNLOCK(ump); 13143 error = ffs_update(vp, 1); 13144 if (error != 0 || MOUNTEDSOFTDEP(mp) == 0) { 13145 UFS_LOCK(ump); 13146 return (0); 13147 } 13148 /* 13149 * If we are in need of resources, start by cleaning up 13150 * any block removals associated with our inode. 13151 */ 13152 ACQUIRE_LOCK(ump); 13153 process_removes(vp); 13154 process_truncates(vp); 13155 FREE_LOCK(ump); 13156 /* 13157 * Now clean up at least as many resources as we will need. 13158 * 13159 * When requested to clean up inodes, the number that are needed 13160 * is set by the number of simultaneous writers (mnt_writeopcount) 13161 * plus a bit of slop (2) in case some more writers show up while 13162 * we are cleaning. 13163 * 13164 * When requested to free up space, the amount of space that 13165 * we need is enough blocks to allocate a full-sized segment 13166 * (fs_contigsumsize). The number of such segments that will 13167 * be needed is set by the number of simultaneous writers 13168 * (mnt_writeopcount) plus a bit of slop (2) in case some more 13169 * writers show up while we are cleaning. 13170 * 13171 * Additionally, if we are unpriviledged and allocating space, 13172 * we need to ensure that we clean up enough blocks to get the 13173 * needed number of blocks over the threshhold of the minimum 13174 * number of blocks required to be kept free by the filesystem 13175 * (fs_minfree). 13176 */ 13177 if (resource == FLUSH_INODES_WAIT) { 13178 needed = vp->v_mount->mnt_writeopcount + 2; 13179 } else if (resource == FLUSH_BLOCKS_WAIT) { 13180 needed = (vp->v_mount->mnt_writeopcount + 2) * 13181 fs->fs_contigsumsize; 13182 if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE, 0)) 13183 needed += fragstoblks(fs, 13184 roundup((fs->fs_dsize * fs->fs_minfree / 100) - 13185 fs->fs_cstotal.cs_nffree, fs->fs_frag)); 13186 } else { 13187 UFS_LOCK(ump); 13188 printf("softdep_request_cleanup: Unknown resource type %d\n", 13189 resource); 13190 return (0); 13191 } 13192 starttime = time_second; 13193retry: 13194 if ((resource == FLUSH_BLOCKS_WAIT && ump->softdep_on_worklist > 0 && 13195 fs->fs_cstotal.cs_nbfree <= needed) || 13196 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 && 13197 fs->fs_cstotal.cs_nifree <= needed)) { 13198 ACQUIRE_LOCK(ump); 13199 if (ump->softdep_on_worklist > 0 && 13200 process_worklist_item(UFSTOVFS(ump), 13201 ump->softdep_on_worklist, LK_NOWAIT) != 0) 13202 stat_worklist_push += 1; 13203 FREE_LOCK(ump); 13204 } 13205 /* 13206 * If we still need resources and there are no more worklist 13207 * entries to process to obtain them, we have to start flushing 13208 * the dirty vnodes to force the release of additional requests 13209 * to the worklist that we can then process to reap addition 13210 * resources. We walk the vnodes associated with the mount point 13211 * until we get the needed worklist requests that we can reap. 13212 */ 13213 if ((resource == FLUSH_BLOCKS_WAIT && 13214 fs->fs_cstotal.cs_nbfree <= needed) || 13215 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 && 13216 fs->fs_cstotal.cs_nifree <= needed)) { 13217 MNT_VNODE_FOREACH_ALL(lvp, mp, mvp) { 13218 if (TAILQ_FIRST(&lvp->v_bufobj.bo_dirty.bv_hd) == 0) { 13219 VI_UNLOCK(lvp); 13220 continue; 13221 } 13222 if (vget(lvp, LK_EXCLUSIVE | LK_INTERLOCK | LK_NOWAIT, 13223 curthread)) 13224 continue; 13225 if (lvp->v_vflag & VV_NOSYNC) { /* unlinked */ 13226 vput(lvp); 13227 continue; 13228 } 13229 (void) ffs_syncvnode(lvp, MNT_NOWAIT, 0); 13230 vput(lvp); 13231 } 13232 lvp = ump->um_devvp; 13233 if (vn_lock(lvp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 13234 VOP_FSYNC(lvp, MNT_NOWAIT, curthread); 13235 VOP_UNLOCK(lvp, 0); 13236 } 13237 if (ump->softdep_on_worklist > 0) { 13238 stat_cleanup_retries += 1; 13239 goto retry; 13240 } 13241 stat_cleanup_failures += 1; 13242 } 13243 if (time_second - starttime > stat_cleanup_high_delay) 13244 stat_cleanup_high_delay = time_second - starttime; 13245 UFS_LOCK(ump); 13246 return (1); 13247} 13248 13249/* 13250 * If memory utilization has gotten too high, deliberately slow things 13251 * down and speed up the I/O processing. 13252 */ 13253static int 13254request_cleanup(mp, resource) 13255 struct mount *mp; 13256 int resource; 13257{ 13258 struct thread *td = curthread; 13259 struct ufsmount *ump; 13260 13261 ump = VFSTOUFS(mp); 13262 LOCK_OWNED(ump); 13263 /* 13264 * We never hold up the filesystem syncer or buf daemon. 13265 */ 13266 if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF)) 13267 return (0); 13268 /* 13269 * First check to see if the work list has gotten backlogged. 13270 * If it has, co-opt this process to help clean up two entries. 13271 * Because this process may hold inodes locked, we cannot 13272 * handle any remove requests that might block on a locked 13273 * inode as that could lead to deadlock. We set TDP_SOFTDEP 13274 * to avoid recursively processing the worklist. 13275 */ 13276 if (ump->softdep_on_worklist > max_softdeps / 10) { 13277 td->td_pflags |= TDP_SOFTDEP; 13278 process_worklist_item(mp, 2, LK_NOWAIT); 13279 td->td_pflags &= ~TDP_SOFTDEP; 13280 stat_worklist_push += 2; 13281 return(1); 13282 } 13283 /* 13284 * Next, we attempt to speed up the syncer process. If that 13285 * is successful, then we allow the process to continue. 13286 */ 13287 if (softdep_speedup(ump) && 13288 resource != FLUSH_BLOCKS_WAIT && 13289 resource != FLUSH_INODES_WAIT) 13290 return(0); 13291 /* 13292 * If we are resource constrained on inode dependencies, try 13293 * flushing some dirty inodes. Otherwise, we are constrained 13294 * by file deletions, so try accelerating flushes of directories 13295 * with removal dependencies. We would like to do the cleanup 13296 * here, but we probably hold an inode locked at this point and 13297 * that might deadlock against one that we try to clean. So, 13298 * the best that we can do is request the syncer daemon to do 13299 * the cleanup for us. 13300 */ 13301 switch (resource) { 13302 13303 case FLUSH_INODES: 13304 case FLUSH_INODES_WAIT: 13305 ACQUIRE_GBLLOCK(&lk); 13306 stat_ino_limit_push += 1; 13307 req_clear_inodedeps += 1; 13308 FREE_GBLLOCK(&lk); 13309 stat_countp = &stat_ino_limit_hit; 13310 break; 13311 13312 case FLUSH_BLOCKS: 13313 case FLUSH_BLOCKS_WAIT: 13314 ACQUIRE_GBLLOCK(&lk); 13315 stat_blk_limit_push += 1; 13316 req_clear_remove += 1; 13317 FREE_GBLLOCK(&lk); 13318 stat_countp = &stat_blk_limit_hit; 13319 break; 13320 13321 default: 13322 panic("request_cleanup: unknown type"); 13323 } 13324 /* 13325 * Hopefully the syncer daemon will catch up and awaken us. 13326 * We wait at most tickdelay before proceeding in any case. 13327 */ 13328 ACQUIRE_GBLLOCK(&lk); 13329 FREE_LOCK(ump); 13330 proc_waiting += 1; 13331 if (callout_pending(&softdep_callout) == FALSE) 13332 callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2, 13333 pause_timer, 0); 13334 13335 msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0); 13336 proc_waiting -= 1; 13337 FREE_GBLLOCK(&lk); 13338 ACQUIRE_LOCK(ump); 13339 return (1); 13340} 13341 13342/* 13343 * Awaken processes pausing in request_cleanup and clear proc_waiting 13344 * to indicate that there is no longer a timer running. Pause_timer 13345 * will be called with the global softdep mutex (&lk) locked. 13346 */ 13347static void 13348pause_timer(arg) 13349 void *arg; 13350{ 13351 13352 GBLLOCK_OWNED(&lk); 13353 /* 13354 * The callout_ API has acquired mtx and will hold it around this 13355 * function call. 13356 */ 13357 *stat_countp += proc_waiting; 13358 wakeup(&proc_waiting); 13359} 13360 13361/* 13362 * If requested, try removing inode or removal dependencies. 13363 */ 13364static void 13365check_clear_deps(mp) 13366 struct mount *mp; 13367{ 13368 13369 /* 13370 * If we are suspended, it may be because of our using 13371 * too many inodedeps, so help clear them out. 13372 */ 13373 if (MOUNTEDSUJ(mp) && VFSTOUFS(mp)->softdep_jblocks->jb_suspended) 13374 clear_inodedeps(mp); 13375 /* 13376 * General requests for cleanup of backed up dependencies 13377 */ 13378 ACQUIRE_GBLLOCK(&lk); 13379 if (req_clear_inodedeps) { 13380 req_clear_inodedeps -= 1; 13381 FREE_GBLLOCK(&lk); 13382 clear_inodedeps(mp); 13383 ACQUIRE_GBLLOCK(&lk); 13384 wakeup(&proc_waiting); 13385 } 13386 if (req_clear_remove) { 13387 req_clear_remove -= 1; 13388 FREE_GBLLOCK(&lk); 13389 clear_remove(mp); 13390 ACQUIRE_GBLLOCK(&lk); 13391 wakeup(&proc_waiting); 13392 } 13393 FREE_GBLLOCK(&lk); 13394} 13395 13396/* 13397 * Flush out a directory with at least one removal dependency in an effort to 13398 * reduce the number of dirrem, freefile, and freeblks dependency structures. 13399 */ 13400static void 13401clear_remove(mp) 13402 struct mount *mp; 13403{ 13404 struct pagedep_hashhead *pagedephd; 13405 struct pagedep *pagedep; 13406 struct ufsmount *ump; 13407 struct vnode *vp; 13408 struct bufobj *bo; 13409 int error, cnt; 13410 ino_t ino; 13411 13412 ump = VFSTOUFS(mp); 13413 LOCK_OWNED(ump); 13414 13415 for (cnt = 0; cnt <= ump->pagedep_hash_size; cnt++) { 13416 pagedephd = &ump->pagedep_hashtbl[ump->pagedep_nextclean++]; 13417 if (ump->pagedep_nextclean > ump->pagedep_hash_size) 13418 ump->pagedep_nextclean = 0; 13419 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 13420 if (LIST_EMPTY(&pagedep->pd_dirremhd)) 13421 continue; 13422 ino = pagedep->pd_ino; 13423 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 13424 continue; 13425 FREE_LOCK(ump); 13426 13427 /* 13428 * Let unmount clear deps 13429 */ 13430 error = vfs_busy(mp, MBF_NOWAIT); 13431 if (error != 0) 13432 goto finish_write; 13433 error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp, 13434 FFSV_FORCEINSMQ); 13435 vfs_unbusy(mp); 13436 if (error != 0) { 13437 softdep_error("clear_remove: vget", error); 13438 goto finish_write; 13439 } 13440 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0))) 13441 softdep_error("clear_remove: fsync", error); 13442 bo = &vp->v_bufobj; 13443 BO_LOCK(bo); 13444 drain_output(vp); 13445 BO_UNLOCK(bo); 13446 vput(vp); 13447 finish_write: 13448 vn_finished_write(mp); 13449 ACQUIRE_LOCK(ump); 13450 return; 13451 } 13452 } 13453} 13454 13455/* 13456 * Clear out a block of dirty inodes in an effort to reduce 13457 * the number of inodedep dependency structures. 13458 */ 13459static void 13460clear_inodedeps(mp) 13461 struct mount *mp; 13462{ 13463 struct inodedep_hashhead *inodedephd; 13464 struct inodedep *inodedep; 13465 struct ufsmount *ump; 13466 struct vnode *vp; 13467 struct fs *fs; 13468 int error, cnt; 13469 ino_t firstino, lastino, ino; 13470 13471 ump = VFSTOUFS(mp); 13472 fs = ump->um_fs; 13473 LOCK_OWNED(ump); 13474 /* 13475 * Pick a random inode dependency to be cleared. 13476 * We will then gather up all the inodes in its block 13477 * that have dependencies and flush them out. 13478 */ 13479 for (cnt = 0; cnt <= ump->inodedep_hash_size; cnt++) { 13480 inodedephd = &ump->inodedep_hashtbl[ump->inodedep_nextclean++]; 13481 if (ump->inodedep_nextclean > ump->inodedep_hash_size) 13482 ump->inodedep_nextclean = 0; 13483 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 13484 break; 13485 } 13486 if (inodedep == NULL) 13487 return; 13488 /* 13489 * Find the last inode in the block with dependencies. 13490 */ 13491 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 13492 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 13493 if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0) 13494 break; 13495 /* 13496 * Asynchronously push all but the last inode with dependencies. 13497 * Synchronously push the last inode with dependencies to ensure 13498 * that the inode block gets written to free up the inodedeps. 13499 */ 13500 for (ino = firstino; ino <= lastino; ino++) { 13501 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 13502 continue; 13503 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 13504 continue; 13505 FREE_LOCK(ump); 13506 error = vfs_busy(mp, MBF_NOWAIT); /* Let unmount clear deps */ 13507 if (error != 0) { 13508 vn_finished_write(mp); 13509 ACQUIRE_LOCK(ump); 13510 return; 13511 } 13512 if ((error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp, 13513 FFSV_FORCEINSMQ)) != 0) { 13514 softdep_error("clear_inodedeps: vget", error); 13515 vfs_unbusy(mp); 13516 vn_finished_write(mp); 13517 ACQUIRE_LOCK(ump); 13518 return; 13519 } 13520 vfs_unbusy(mp); 13521 if (ino == lastino) { 13522 if ((error = ffs_syncvnode(vp, MNT_WAIT, 0))) 13523 softdep_error("clear_inodedeps: fsync1", error); 13524 } else { 13525 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0))) 13526 softdep_error("clear_inodedeps: fsync2", error); 13527 BO_LOCK(&vp->v_bufobj); 13528 drain_output(vp); 13529 BO_UNLOCK(&vp->v_bufobj); 13530 } 13531 vput(vp); 13532 vn_finished_write(mp); 13533 ACQUIRE_LOCK(ump); 13534 } 13535} 13536 13537void 13538softdep_buf_append(bp, wkhd) 13539 struct buf *bp; 13540 struct workhead *wkhd; 13541{ 13542 struct worklist *wk; 13543 struct ufsmount *ump; 13544 13545 if ((wk = LIST_FIRST(wkhd)) == NULL) 13546 return; 13547 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0, 13548 ("softdep_buf_append called on non-softdep filesystem")); 13549 ump = VFSTOUFS(wk->wk_mp); 13550 ACQUIRE_LOCK(ump); 13551 while ((wk = LIST_FIRST(wkhd)) != NULL) { 13552 WORKLIST_REMOVE(wk); 13553 WORKLIST_INSERT(&bp->b_dep, wk); 13554 } 13555 FREE_LOCK(ump); 13556 13557} 13558 13559void 13560softdep_inode_append(ip, cred, wkhd) 13561 struct inode *ip; 13562 struct ucred *cred; 13563 struct workhead *wkhd; 13564{ 13565 struct buf *bp; 13566 struct fs *fs; 13567 int error; 13568 13569 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0, 13570 ("softdep_inode_append called on non-softdep filesystem")); 13571 fs = ip->i_fs; 13572 error = bread(ip->i_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 13573 (int)fs->fs_bsize, cred, &bp); 13574 if (error) { 13575 bqrelse(bp); 13576 softdep_freework(wkhd); 13577 return; 13578 } 13579 softdep_buf_append(bp, wkhd); 13580 bqrelse(bp); 13581} 13582 13583void 13584softdep_freework(wkhd) 13585 struct workhead *wkhd; 13586{ 13587 struct worklist *wk; 13588 struct ufsmount *ump; 13589 13590 if ((wk = LIST_FIRST(wkhd)) == NULL) 13591 return; 13592 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0, 13593 ("softdep_freework called on non-softdep filesystem")); 13594 ump = VFSTOUFS(wk->wk_mp); 13595 ACQUIRE_LOCK(ump); 13596 handle_jwork(wkhd); 13597 FREE_LOCK(ump); 13598} 13599 13600/* 13601 * Function to determine if the buffer has outstanding dependencies 13602 * that will cause a roll-back if the buffer is written. If wantcount 13603 * is set, return number of dependencies, otherwise just yes or no. 13604 */ 13605static int 13606softdep_count_dependencies(bp, wantcount) 13607 struct buf *bp; 13608 int wantcount; 13609{ 13610 struct worklist *wk; 13611 struct ufsmount *ump; 13612 struct bmsafemap *bmsafemap; 13613 struct freework *freework; 13614 struct inodedep *inodedep; 13615 struct indirdep *indirdep; 13616 struct freeblks *freeblks; 13617 struct allocindir *aip; 13618 struct pagedep *pagedep; 13619 struct dirrem *dirrem; 13620 struct newblk *newblk; 13621 struct mkdir *mkdir; 13622 struct diradd *dap; 13623 int i, retval; 13624 13625 retval = 0; 13626 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL) 13627 return (0); 13628 ump = VFSTOUFS(wk->wk_mp); 13629 ACQUIRE_LOCK(ump); 13630 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 13631 switch (wk->wk_type) { 13632 13633 case D_INODEDEP: 13634 inodedep = WK_INODEDEP(wk); 13635 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 13636 /* bitmap allocation dependency */ 13637 retval += 1; 13638 if (!wantcount) 13639 goto out; 13640 } 13641 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 13642 /* direct block pointer dependency */ 13643 retval += 1; 13644 if (!wantcount) 13645 goto out; 13646 } 13647 if (TAILQ_FIRST(&inodedep->id_extupdt)) { 13648 /* direct block pointer dependency */ 13649 retval += 1; 13650 if (!wantcount) 13651 goto out; 13652 } 13653 if (TAILQ_FIRST(&inodedep->id_inoreflst)) { 13654 /* Add reference dependency. */ 13655 retval += 1; 13656 if (!wantcount) 13657 goto out; 13658 } 13659 continue; 13660 13661 case D_INDIRDEP: 13662 indirdep = WK_INDIRDEP(wk); 13663 13664 TAILQ_FOREACH(freework, &indirdep->ir_trunc, fw_next) { 13665 /* indirect truncation dependency */ 13666 retval += 1; 13667 if (!wantcount) 13668 goto out; 13669 } 13670 13671 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 13672 /* indirect block pointer dependency */ 13673 retval += 1; 13674 if (!wantcount) 13675 goto out; 13676 } 13677 continue; 13678 13679 case D_PAGEDEP: 13680 pagedep = WK_PAGEDEP(wk); 13681 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 13682 if (LIST_FIRST(&dirrem->dm_jremrefhd)) { 13683 /* Journal remove ref dependency. */ 13684 retval += 1; 13685 if (!wantcount) 13686 goto out; 13687 } 13688 } 13689 for (i = 0; i < DAHASHSZ; i++) { 13690 13691 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 13692 /* directory entry dependency */ 13693 retval += 1; 13694 if (!wantcount) 13695 goto out; 13696 } 13697 } 13698 continue; 13699 13700 case D_BMSAFEMAP: 13701 bmsafemap = WK_BMSAFEMAP(wk); 13702 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd)) { 13703 /* Add reference dependency. */ 13704 retval += 1; 13705 if (!wantcount) 13706 goto out; 13707 } 13708 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd)) { 13709 /* Allocate block dependency. */ 13710 retval += 1; 13711 if (!wantcount) 13712 goto out; 13713 } 13714 continue; 13715 13716 case D_FREEBLKS: 13717 freeblks = WK_FREEBLKS(wk); 13718 if (LIST_FIRST(&freeblks->fb_jblkdephd)) { 13719 /* Freeblk journal dependency. */ 13720 retval += 1; 13721 if (!wantcount) 13722 goto out; 13723 } 13724 continue; 13725 13726 case D_ALLOCDIRECT: 13727 case D_ALLOCINDIR: 13728 newblk = WK_NEWBLK(wk); 13729 if (newblk->nb_jnewblk) { 13730 /* Journal allocate dependency. */ 13731 retval += 1; 13732 if (!wantcount) 13733 goto out; 13734 } 13735 continue; 13736 13737 case D_MKDIR: 13738 mkdir = WK_MKDIR(wk); 13739 if (mkdir->md_jaddref) { 13740 /* Journal reference dependency. */ 13741 retval += 1; 13742 if (!wantcount) 13743 goto out; 13744 } 13745 continue; 13746 13747 case D_FREEWORK: 13748 case D_FREEDEP: 13749 case D_JSEGDEP: 13750 case D_JSEG: 13751 case D_SBDEP: 13752 /* never a dependency on these blocks */ 13753 continue; 13754 13755 default: 13756 panic("softdep_count_dependencies: Unexpected type %s", 13757 TYPENAME(wk->wk_type)); 13758 /* NOTREACHED */ 13759 } 13760 } 13761out: 13762 FREE_LOCK(ump); 13763 return retval; 13764} 13765 13766/* 13767 * Acquire exclusive access to a buffer. 13768 * Must be called with a locked mtx parameter. 13769 * Return acquired buffer or NULL on failure. 13770 */ 13771static struct buf * 13772getdirtybuf(bp, lock, waitfor) 13773 struct buf *bp; 13774 struct rwlock *lock; 13775 int waitfor; 13776{ 13777 int error; 13778 13779 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) { 13780 if (waitfor != MNT_WAIT) 13781 return (NULL); 13782 error = BUF_LOCK(bp, 13783 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, lock); 13784 /* 13785 * Even if we sucessfully acquire bp here, we have dropped 13786 * lock, which may violates our guarantee. 13787 */ 13788 if (error == 0) 13789 BUF_UNLOCK(bp); 13790 else if (error != ENOLCK) 13791 panic("getdirtybuf: inconsistent lock: %d", error); 13792 rw_wlock(lock); 13793 return (NULL); 13794 } 13795 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 13796 if (lock != BO_LOCKPTR(bp->b_bufobj) && waitfor == MNT_WAIT) { 13797 rw_wunlock(lock); 13798 BO_LOCK(bp->b_bufobj); 13799 BUF_UNLOCK(bp); 13800 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 13801 bp->b_vflags |= BV_BKGRDWAIT; 13802 msleep(&bp->b_xflags, BO_LOCKPTR(bp->b_bufobj), 13803 PRIBIO | PDROP, "getbuf", 0); 13804 } else 13805 BO_UNLOCK(bp->b_bufobj); 13806 rw_wlock(lock); 13807 return (NULL); 13808 } 13809 BUF_UNLOCK(bp); 13810 if (waitfor != MNT_WAIT) 13811 return (NULL); 13812 /* 13813 * The lock argument must be bp->b_vp's mutex in 13814 * this case. 13815 */ 13816#ifdef DEBUG_VFS_LOCKS 13817 if (bp->b_vp->v_type != VCHR) 13818 ASSERT_BO_WLOCKED(bp->b_bufobj); 13819#endif 13820 bp->b_vflags |= BV_BKGRDWAIT; 13821 rw_sleep(&bp->b_xflags, lock, PRIBIO, "getbuf", 0); 13822 return (NULL); 13823 } 13824 if ((bp->b_flags & B_DELWRI) == 0) { 13825 BUF_UNLOCK(bp); 13826 return (NULL); 13827 } 13828 bremfree(bp); 13829 return (bp); 13830} 13831 13832 13833/* 13834 * Check if it is safe to suspend the file system now. On entry, 13835 * the vnode interlock for devvp should be held. Return 0 with 13836 * the mount interlock held if the file system can be suspended now, 13837 * otherwise return EAGAIN with the mount interlock held. 13838 */ 13839int 13840softdep_check_suspend(struct mount *mp, 13841 struct vnode *devvp, 13842 int softdep_depcnt, 13843 int softdep_accdepcnt, 13844 int secondary_writes, 13845 int secondary_accwrites) 13846{ 13847 struct bufobj *bo; 13848 struct ufsmount *ump; 13849 int error; 13850 13851 bo = &devvp->v_bufobj; 13852 ASSERT_BO_WLOCKED(bo); 13853 13854 /* 13855 * If we are not running with soft updates, then we need only 13856 * deal with secondary writes as we try to suspend. 13857 */ 13858 if (MOUNTEDSOFTDEP(mp) == 0) { 13859 MNT_ILOCK(mp); 13860 while (mp->mnt_secondary_writes != 0) { 13861 BO_UNLOCK(bo); 13862 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp), 13863 (PUSER - 1) | PDROP, "secwr", 0); 13864 BO_LOCK(bo); 13865 MNT_ILOCK(mp); 13866 } 13867 13868 /* 13869 * Reasons for needing more work before suspend: 13870 * - Dirty buffers on devvp. 13871 * - Secondary writes occurred after start of vnode sync loop 13872 */ 13873 error = 0; 13874 if (bo->bo_numoutput > 0 || 13875 bo->bo_dirty.bv_cnt > 0 || 13876 secondary_writes != 0 || 13877 mp->mnt_secondary_writes != 0 || 13878 secondary_accwrites != mp->mnt_secondary_accwrites) 13879 error = EAGAIN; 13880 BO_UNLOCK(bo); 13881 return (error); 13882 } 13883 13884 /* 13885 * If we are running with soft updates, then we need to coordinate 13886 * with them as we try to suspend. 13887 */ 13888 ump = VFSTOUFS(mp); 13889 for (;;) { 13890 if (!TRY_ACQUIRE_LOCK(ump)) { 13891 BO_UNLOCK(bo); 13892 ACQUIRE_LOCK(ump); 13893 FREE_LOCK(ump); 13894 BO_LOCK(bo); 13895 continue; 13896 } 13897 MNT_ILOCK(mp); 13898 if (mp->mnt_secondary_writes != 0) { 13899 FREE_LOCK(ump); 13900 BO_UNLOCK(bo); 13901 msleep(&mp->mnt_secondary_writes, 13902 MNT_MTX(mp), 13903 (PUSER - 1) | PDROP, "secwr", 0); 13904 BO_LOCK(bo); 13905 continue; 13906 } 13907 break; 13908 } 13909 13910 /* 13911 * Reasons for needing more work before suspend: 13912 * - Dirty buffers on devvp. 13913 * - Softdep activity occurred after start of vnode sync loop 13914 * - Secondary writes occurred after start of vnode sync loop 13915 */ 13916 error = 0; 13917 if (bo->bo_numoutput > 0 || 13918 bo->bo_dirty.bv_cnt > 0 || 13919 softdep_depcnt != 0 || 13920 ump->softdep_deps != 0 || 13921 softdep_accdepcnt != ump->softdep_accdeps || 13922 secondary_writes != 0 || 13923 mp->mnt_secondary_writes != 0 || 13924 secondary_accwrites != mp->mnt_secondary_accwrites) 13925 error = EAGAIN; 13926 FREE_LOCK(ump); 13927 BO_UNLOCK(bo); 13928 return (error); 13929} 13930 13931 13932/* 13933 * Get the number of dependency structures for the file system, both 13934 * the current number and the total number allocated. These will 13935 * later be used to detect that softdep processing has occurred. 13936 */ 13937void 13938softdep_get_depcounts(struct mount *mp, 13939 int *softdep_depsp, 13940 int *softdep_accdepsp) 13941{ 13942 struct ufsmount *ump; 13943 13944 if (MOUNTEDSOFTDEP(mp) == 0) { 13945 *softdep_depsp = 0; 13946 *softdep_accdepsp = 0; 13947 return; 13948 } 13949 ump = VFSTOUFS(mp); 13950 ACQUIRE_LOCK(ump); 13951 *softdep_depsp = ump->softdep_deps; 13952 *softdep_accdepsp = ump->softdep_accdeps; 13953 FREE_LOCK(ump); 13954} 13955 13956/* 13957 * Wait for pending output on a vnode to complete. 13958 * Must be called with vnode lock and interlock locked. 13959 * 13960 * XXX: Should just be a call to bufobj_wwait(). 13961 */ 13962static void 13963drain_output(vp) 13964 struct vnode *vp; 13965{ 13966 struct bufobj *bo; 13967 13968 bo = &vp->v_bufobj; 13969 ASSERT_VOP_LOCKED(vp, "drain_output"); 13970 ASSERT_BO_WLOCKED(bo); 13971 13972 while (bo->bo_numoutput) { 13973 bo->bo_flag |= BO_WWAIT; 13974 msleep((caddr_t)&bo->bo_numoutput, 13975 BO_LOCKPTR(bo), PRIBIO + 1, "drainvp", 0); 13976 } 13977} 13978 13979/* 13980 * Called whenever a buffer that is being invalidated or reallocated 13981 * contains dependencies. This should only happen if an I/O error has 13982 * occurred. The routine is called with the buffer locked. 13983 */ 13984static void 13985softdep_deallocate_dependencies(bp) 13986 struct buf *bp; 13987{ 13988 13989 if ((bp->b_ioflags & BIO_ERROR) == 0) 13990 panic("softdep_deallocate_dependencies: dangling deps"); 13991 if (bp->b_vp != NULL && bp->b_vp->v_mount != NULL) 13992 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); 13993 else 13994 printf("softdep_deallocate_dependencies: " 13995 "got error %d while accessing filesystem\n", bp->b_error); 13996 if (bp->b_error != ENXIO) 13997 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 13998} 13999 14000/* 14001 * Function to handle asynchronous write errors in the filesystem. 14002 */ 14003static void 14004softdep_error(func, error) 14005 char *func; 14006 int error; 14007{ 14008 14009 /* XXX should do something better! */ 14010 printf("%s: got error %d while accessing filesystem\n", func, error); 14011} 14012 14013#ifdef DDB 14014 14015static void 14016inodedep_print(struct inodedep *inodedep, int verbose) 14017{ 14018 db_printf("%p fs %p st %x ino %jd inoblk %jd delta %d nlink %d" 14019 " saveino %p\n", 14020 inodedep, inodedep->id_fs, inodedep->id_state, 14021 (intmax_t)inodedep->id_ino, 14022 (intmax_t)fsbtodb(inodedep->id_fs, 14023 ino_to_fsba(inodedep->id_fs, inodedep->id_ino)), 14024 inodedep->id_nlinkdelta, inodedep->id_savednlink, 14025 inodedep->id_savedino1); 14026 14027 if (verbose == 0) 14028 return; 14029 14030 db_printf("\tpendinghd %p, bufwait %p, inowait %p, inoreflst %p, " 14031 "mkdiradd %p\n", 14032 LIST_FIRST(&inodedep->id_pendinghd), 14033 LIST_FIRST(&inodedep->id_bufwait), 14034 LIST_FIRST(&inodedep->id_inowait), 14035 TAILQ_FIRST(&inodedep->id_inoreflst), 14036 inodedep->id_mkdiradd); 14037 db_printf("\tinoupdt %p, newinoupdt %p, extupdt %p, newextupdt %p\n", 14038 TAILQ_FIRST(&inodedep->id_inoupdt), 14039 TAILQ_FIRST(&inodedep->id_newinoupdt), 14040 TAILQ_FIRST(&inodedep->id_extupdt), 14041 TAILQ_FIRST(&inodedep->id_newextupdt)); 14042} 14043 14044DB_SHOW_COMMAND(inodedep, db_show_inodedep) 14045{ 14046 14047 if (have_addr == 0) { 14048 db_printf("Address required\n"); 14049 return; 14050 } 14051 inodedep_print((struct inodedep*)addr, 1); 14052} 14053 14054DB_SHOW_COMMAND(inodedeps, db_show_inodedeps) 14055{ 14056 struct inodedep_hashhead *inodedephd; 14057 struct inodedep *inodedep; 14058 struct ufsmount *ump; 14059 int cnt; 14060 14061 if (have_addr == 0) { 14062 db_printf("Address required\n"); 14063 return; 14064 } 14065 ump = (struct ufsmount *)addr; 14066 for (cnt = 0; cnt < ump->inodedep_hash_size; cnt++) { 14067 inodedephd = &ump->inodedep_hashtbl[cnt]; 14068 LIST_FOREACH(inodedep, inodedephd, id_hash) { 14069 inodedep_print(inodedep, 0); 14070 } 14071 } 14072} 14073 14074DB_SHOW_COMMAND(worklist, db_show_worklist) 14075{ 14076 struct worklist *wk; 14077 14078 if (have_addr == 0) { 14079 db_printf("Address required\n"); 14080 return; 14081 } 14082 wk = (struct worklist *)addr; 14083 printf("worklist: %p type %s state 0x%X\n", 14084 wk, TYPENAME(wk->wk_type), wk->wk_state); 14085} 14086 14087DB_SHOW_COMMAND(workhead, db_show_workhead) 14088{ 14089 struct workhead *wkhd; 14090 struct worklist *wk; 14091 int i; 14092 14093 if (have_addr == 0) { 14094 db_printf("Address required\n"); 14095 return; 14096 } 14097 wkhd = (struct workhead *)addr; 14098 wk = LIST_FIRST(wkhd); 14099 for (i = 0; i < 100 && wk != NULL; i++, wk = LIST_NEXT(wk, wk_list)) 14100 db_printf("worklist: %p type %s state 0x%X", 14101 wk, TYPENAME(wk->wk_type), wk->wk_state); 14102 if (i == 100) 14103 db_printf("workhead overflow"); 14104 printf("\n"); 14105} 14106 14107 14108DB_SHOW_COMMAND(mkdirs, db_show_mkdirs) 14109{ 14110 struct mkdirlist *mkdirlisthd; 14111 struct jaddref *jaddref; 14112 struct diradd *diradd; 14113 struct mkdir *mkdir; 14114 14115 if (have_addr == 0) { 14116 db_printf("Address required\n"); 14117 return; 14118 } 14119 mkdirlisthd = (struct mkdirlist *)addr; 14120 LIST_FOREACH(mkdir, mkdirlisthd, md_mkdirs) { 14121 diradd = mkdir->md_diradd; 14122 db_printf("mkdir: %p state 0x%X dap %p state 0x%X", 14123 mkdir, mkdir->md_state, diradd, diradd->da_state); 14124 if ((jaddref = mkdir->md_jaddref) != NULL) 14125 db_printf(" jaddref %p jaddref state 0x%X", 14126 jaddref, jaddref->ja_state); 14127 db_printf("\n"); 14128 } 14129} 14130 14131/* exported to ffs_vfsops.c */ 14132extern void db_print_ffs(struct ufsmount *ump); 14133void 14134db_print_ffs(struct ufsmount *ump) 14135{ 14136 db_printf("mp %p %s devvp %p fs %p su_wl %d su_deps %d su_req %d\n", 14137 ump->um_mountp, ump->um_mountp->mnt_stat.f_mntonname, 14138 ump->um_devvp, ump->um_fs, ump->softdep_on_worklist, 14139 ump->softdep_deps, ump->softdep_req); 14140} 14141 14142#endif /* DDB */ 14143 14144#endif /* SOFTUPDATES */ 14145