ffs_softdep.c revision 242259
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 242259 2012-10-28 18:53:28Z trasz $"); 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/limits.h> 64#include <sys/lock.h> 65#include <sys/malloc.h> 66#include <sys/mount.h> 67#include <sys/mutex.h> 68#include <sys/namei.h> 69#include <sys/priv.h> 70#include <sys/proc.h> 71#include <sys/stat.h> 72#include <sys/sysctl.h> 73#include <sys/syslog.h> 74#include <sys/vnode.h> 75#include <sys/conf.h> 76 77#include <ufs/ufs/dir.h> 78#include <ufs/ufs/extattr.h> 79#include <ufs/ufs/quota.h> 80#include <ufs/ufs/inode.h> 81#include <ufs/ufs/ufsmount.h> 82#include <ufs/ffs/fs.h> 83#include <ufs/ffs/softdep.h> 84#include <ufs/ffs/ffs_extern.h> 85#include <ufs/ufs/ufs_extern.h> 86 87#include <vm/vm.h> 88#include <vm/vm_extern.h> 89#include <vm/vm_object.h> 90 91#include <ddb/ddb.h> 92 93#ifndef SOFTUPDATES 94 95int 96softdep_flushfiles(oldmnt, flags, td) 97 struct mount *oldmnt; 98 int flags; 99 struct thread *td; 100{ 101 102 panic("softdep_flushfiles called"); 103} 104 105int 106softdep_mount(devvp, mp, fs, cred) 107 struct vnode *devvp; 108 struct mount *mp; 109 struct fs *fs; 110 struct ucred *cred; 111{ 112 113 return (0); 114} 115 116void 117softdep_initialize() 118{ 119 120 return; 121} 122 123void 124softdep_uninitialize() 125{ 126 127 return; 128} 129 130void 131softdep_unmount(mp) 132 struct mount *mp; 133{ 134 135} 136 137void 138softdep_setup_sbupdate(ump, fs, bp) 139 struct ufsmount *ump; 140 struct fs *fs; 141 struct buf *bp; 142{ 143} 144 145void 146softdep_setup_inomapdep(bp, ip, newinum, mode) 147 struct buf *bp; 148 struct inode *ip; 149 ino_t newinum; 150 int mode; 151{ 152 153 panic("softdep_setup_inomapdep called"); 154} 155 156void 157softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags) 158 struct buf *bp; 159 struct mount *mp; 160 ufs2_daddr_t newblkno; 161 int frags; 162 int oldfrags; 163{ 164 165 panic("softdep_setup_blkmapdep called"); 166} 167 168void 169softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 170 struct inode *ip; 171 ufs_lbn_t lbn; 172 ufs2_daddr_t newblkno; 173 ufs2_daddr_t oldblkno; 174 long newsize; 175 long oldsize; 176 struct buf *bp; 177{ 178 179 panic("softdep_setup_allocdirect called"); 180} 181 182void 183softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 184 struct inode *ip; 185 ufs_lbn_t lbn; 186 ufs2_daddr_t newblkno; 187 ufs2_daddr_t oldblkno; 188 long newsize; 189 long oldsize; 190 struct buf *bp; 191{ 192 193 panic("softdep_setup_allocext called"); 194} 195 196void 197softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 198 struct inode *ip; 199 ufs_lbn_t lbn; 200 struct buf *bp; 201 int ptrno; 202 ufs2_daddr_t newblkno; 203 ufs2_daddr_t oldblkno; 204 struct buf *nbp; 205{ 206 207 panic("softdep_setup_allocindir_page called"); 208} 209 210void 211softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 212 struct buf *nbp; 213 struct inode *ip; 214 struct buf *bp; 215 int ptrno; 216 ufs2_daddr_t newblkno; 217{ 218 219 panic("softdep_setup_allocindir_meta called"); 220} 221 222void 223softdep_journal_freeblocks(ip, cred, length, flags) 224 struct inode *ip; 225 struct ucred *cred; 226 off_t length; 227 int flags; 228{ 229 230 panic("softdep_journal_freeblocks called"); 231} 232 233void 234softdep_journal_fsync(ip) 235 struct inode *ip; 236{ 237 238 panic("softdep_journal_fsync called"); 239} 240 241void 242softdep_setup_freeblocks(ip, length, flags) 243 struct inode *ip; 244 off_t length; 245 int flags; 246{ 247 248 panic("softdep_setup_freeblocks called"); 249} 250 251void 252softdep_freefile(pvp, ino, mode) 253 struct vnode *pvp; 254 ino_t ino; 255 int mode; 256{ 257 258 panic("softdep_freefile called"); 259} 260 261int 262softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 263 struct buf *bp; 264 struct inode *dp; 265 off_t diroffset; 266 ino_t newinum; 267 struct buf *newdirbp; 268 int isnewblk; 269{ 270 271 panic("softdep_setup_directory_add called"); 272} 273 274void 275softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize) 276 struct buf *bp; 277 struct inode *dp; 278 caddr_t base; 279 caddr_t oldloc; 280 caddr_t newloc; 281 int entrysize; 282{ 283 284 panic("softdep_change_directoryentry_offset called"); 285} 286 287void 288softdep_setup_remove(bp, dp, ip, isrmdir) 289 struct buf *bp; 290 struct inode *dp; 291 struct inode *ip; 292 int isrmdir; 293{ 294 295 panic("softdep_setup_remove called"); 296} 297 298void 299softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 300 struct buf *bp; 301 struct inode *dp; 302 struct inode *ip; 303 ino_t newinum; 304 int isrmdir; 305{ 306 307 panic("softdep_setup_directory_change called"); 308} 309 310void 311softdep_setup_blkfree(mp, bp, blkno, frags, wkhd) 312 struct mount *mp; 313 struct buf *bp; 314 ufs2_daddr_t blkno; 315 int frags; 316 struct workhead *wkhd; 317{ 318 319 panic("%s called", __FUNCTION__); 320} 321 322void 323softdep_setup_inofree(mp, bp, ino, wkhd) 324 struct mount *mp; 325 struct buf *bp; 326 ino_t ino; 327 struct workhead *wkhd; 328{ 329 330 panic("%s called", __FUNCTION__); 331} 332 333void 334softdep_setup_unlink(dp, ip) 335 struct inode *dp; 336 struct inode *ip; 337{ 338 339 panic("%s called", __FUNCTION__); 340} 341 342void 343softdep_setup_link(dp, ip) 344 struct inode *dp; 345 struct inode *ip; 346{ 347 348 panic("%s called", __FUNCTION__); 349} 350 351void 352softdep_revert_link(dp, ip) 353 struct inode *dp; 354 struct inode *ip; 355{ 356 357 panic("%s called", __FUNCTION__); 358} 359 360void 361softdep_setup_rmdir(dp, ip) 362 struct inode *dp; 363 struct inode *ip; 364{ 365 366 panic("%s called", __FUNCTION__); 367} 368 369void 370softdep_revert_rmdir(dp, ip) 371 struct inode *dp; 372 struct inode *ip; 373{ 374 375 panic("%s called", __FUNCTION__); 376} 377 378void 379softdep_setup_create(dp, ip) 380 struct inode *dp; 381 struct inode *ip; 382{ 383 384 panic("%s called", __FUNCTION__); 385} 386 387void 388softdep_revert_create(dp, ip) 389 struct inode *dp; 390 struct inode *ip; 391{ 392 393 panic("%s called", __FUNCTION__); 394} 395 396void 397softdep_setup_mkdir(dp, ip) 398 struct inode *dp; 399 struct inode *ip; 400{ 401 402 panic("%s called", __FUNCTION__); 403} 404 405void 406softdep_revert_mkdir(dp, ip) 407 struct inode *dp; 408 struct inode *ip; 409{ 410 411 panic("%s called", __FUNCTION__); 412} 413 414void 415softdep_setup_dotdot_link(dp, ip) 416 struct inode *dp; 417 struct inode *ip; 418{ 419 420 panic("%s called", __FUNCTION__); 421} 422 423int 424softdep_prealloc(vp, waitok) 425 struct vnode *vp; 426 int waitok; 427{ 428 429 panic("%s called", __FUNCTION__); 430 431 return (0); 432} 433 434int 435softdep_journal_lookup(mp, vpp) 436 struct mount *mp; 437 struct vnode **vpp; 438{ 439 440 return (ENOENT); 441} 442 443void 444softdep_change_linkcnt(ip) 445 struct inode *ip; 446{ 447 448 panic("softdep_change_linkcnt called"); 449} 450 451void 452softdep_load_inodeblock(ip) 453 struct inode *ip; 454{ 455 456 panic("softdep_load_inodeblock called"); 457} 458 459void 460softdep_update_inodeblock(ip, bp, waitfor) 461 struct inode *ip; 462 struct buf *bp; 463 int waitfor; 464{ 465 466 panic("softdep_update_inodeblock called"); 467} 468 469int 470softdep_fsync(vp) 471 struct vnode *vp; /* the "in_core" copy of the inode */ 472{ 473 474 return (0); 475} 476 477void 478softdep_fsync_mountdev(vp) 479 struct vnode *vp; 480{ 481 482 return; 483} 484 485int 486softdep_flushworklist(oldmnt, countp, td) 487 struct mount *oldmnt; 488 int *countp; 489 struct thread *td; 490{ 491 492 *countp = 0; 493 return (0); 494} 495 496int 497softdep_sync_metadata(struct vnode *vp) 498{ 499 500 return (0); 501} 502 503int 504softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor) 505{ 506 507 return (0); 508} 509 510int 511softdep_slowdown(vp) 512 struct vnode *vp; 513{ 514 515 panic("softdep_slowdown called"); 516} 517 518void 519softdep_releasefile(ip) 520 struct inode *ip; /* inode with the zero effective link count */ 521{ 522 523 panic("softdep_releasefile called"); 524} 525 526int 527softdep_request_cleanup(fs, vp, cred, resource) 528 struct fs *fs; 529 struct vnode *vp; 530 struct ucred *cred; 531 int resource; 532{ 533 534 return (0); 535} 536 537int 538softdep_check_suspend(struct mount *mp, 539 struct vnode *devvp, 540 int softdep_deps, 541 int softdep_accdeps, 542 int secondary_writes, 543 int secondary_accwrites) 544{ 545 struct bufobj *bo; 546 int error; 547 548 (void) softdep_deps, 549 (void) softdep_accdeps; 550 551 bo = &devvp->v_bufobj; 552 ASSERT_BO_LOCKED(bo); 553 554 MNT_ILOCK(mp); 555 while (mp->mnt_secondary_writes != 0) { 556 BO_UNLOCK(bo); 557 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp), 558 (PUSER - 1) | PDROP, "secwr", 0); 559 BO_LOCK(bo); 560 MNT_ILOCK(mp); 561 } 562 563 /* 564 * Reasons for needing more work before suspend: 565 * - Dirty buffers on devvp. 566 * - Secondary writes occurred after start of vnode sync loop 567 */ 568 error = 0; 569 if (bo->bo_numoutput > 0 || 570 bo->bo_dirty.bv_cnt > 0 || 571 secondary_writes != 0 || 572 mp->mnt_secondary_writes != 0 || 573 secondary_accwrites != mp->mnt_secondary_accwrites) 574 error = EAGAIN; 575 BO_UNLOCK(bo); 576 return (error); 577} 578 579void 580softdep_get_depcounts(struct mount *mp, 581 int *softdepactivep, 582 int *softdepactiveaccp) 583{ 584 (void) mp; 585 *softdepactivep = 0; 586 *softdepactiveaccp = 0; 587} 588 589void 590softdep_buf_append(bp, wkhd) 591 struct buf *bp; 592 struct workhead *wkhd; 593{ 594 595 panic("softdep_buf_appendwork called"); 596} 597 598void 599softdep_inode_append(ip, cred, wkhd) 600 struct inode *ip; 601 struct ucred *cred; 602 struct workhead *wkhd; 603{ 604 605 panic("softdep_inode_appendwork called"); 606} 607 608void 609softdep_freework(wkhd) 610 struct workhead *wkhd; 611{ 612 613 panic("softdep_freework called"); 614} 615 616#else 617 618FEATURE(softupdates, "FFS soft-updates support"); 619 620/* 621 * These definitions need to be adapted to the system to which 622 * this file is being ported. 623 */ 624 625#define M_SOFTDEP_FLAGS (M_WAITOK) 626 627#define D_PAGEDEP 0 628#define D_INODEDEP 1 629#define D_BMSAFEMAP 2 630#define D_NEWBLK 3 631#define D_ALLOCDIRECT 4 632#define D_INDIRDEP 5 633#define D_ALLOCINDIR 6 634#define D_FREEFRAG 7 635#define D_FREEBLKS 8 636#define D_FREEFILE 9 637#define D_DIRADD 10 638#define D_MKDIR 11 639#define D_DIRREM 12 640#define D_NEWDIRBLK 13 641#define D_FREEWORK 14 642#define D_FREEDEP 15 643#define D_JADDREF 16 644#define D_JREMREF 17 645#define D_JMVREF 18 646#define D_JNEWBLK 19 647#define D_JFREEBLK 20 648#define D_JFREEFRAG 21 649#define D_JSEG 22 650#define D_JSEGDEP 23 651#define D_SBDEP 24 652#define D_JTRUNC 25 653#define D_JFSYNC 26 654#define D_SENTINAL 27 655#define D_LAST D_SENTINAL 656 657unsigned long dep_current[D_LAST + 1]; 658unsigned long dep_total[D_LAST + 1]; 659unsigned long dep_write[D_LAST + 1]; 660 661 662static SYSCTL_NODE(_debug, OID_AUTO, softdep, CTLFLAG_RW, 0, 663 "soft updates stats"); 664static SYSCTL_NODE(_debug_softdep, OID_AUTO, total, CTLFLAG_RW, 0, 665 "total dependencies allocated"); 666static SYSCTL_NODE(_debug_softdep, OID_AUTO, current, CTLFLAG_RW, 0, 667 "current dependencies allocated"); 668static SYSCTL_NODE(_debug_softdep, OID_AUTO, write, CTLFLAG_RW, 0, 669 "current dependencies written"); 670 671#define SOFTDEP_TYPE(type, str, long) \ 672 static MALLOC_DEFINE(M_ ## type, #str, long); \ 673 SYSCTL_ULONG(_debug_softdep_total, OID_AUTO, str, CTLFLAG_RD, \ 674 &dep_total[D_ ## type], 0, ""); \ 675 SYSCTL_ULONG(_debug_softdep_current, OID_AUTO, str, CTLFLAG_RD, \ 676 &dep_current[D_ ## type], 0, ""); \ 677 SYSCTL_ULONG(_debug_softdep_write, OID_AUTO, str, CTLFLAG_RD, \ 678 &dep_write[D_ ## type], 0, ""); 679 680SOFTDEP_TYPE(PAGEDEP, pagedep, "File page dependencies"); 681SOFTDEP_TYPE(INODEDEP, inodedep, "Inode dependencies"); 682SOFTDEP_TYPE(BMSAFEMAP, bmsafemap, 683 "Block or frag allocated from cyl group map"); 684SOFTDEP_TYPE(NEWBLK, newblk, "New block or frag allocation dependency"); 685SOFTDEP_TYPE(ALLOCDIRECT, allocdirect, "Block or frag dependency for an inode"); 686SOFTDEP_TYPE(INDIRDEP, indirdep, "Indirect block dependencies"); 687SOFTDEP_TYPE(ALLOCINDIR, allocindir, "Block dependency for an indirect block"); 688SOFTDEP_TYPE(FREEFRAG, freefrag, "Previously used frag for an inode"); 689SOFTDEP_TYPE(FREEBLKS, freeblks, "Blocks freed from an inode"); 690SOFTDEP_TYPE(FREEFILE, freefile, "Inode deallocated"); 691SOFTDEP_TYPE(DIRADD, diradd, "New directory entry"); 692SOFTDEP_TYPE(MKDIR, mkdir, "New directory"); 693SOFTDEP_TYPE(DIRREM, dirrem, "Directory entry deleted"); 694SOFTDEP_TYPE(NEWDIRBLK, newdirblk, "Unclaimed new directory block"); 695SOFTDEP_TYPE(FREEWORK, freework, "free an inode block"); 696SOFTDEP_TYPE(FREEDEP, freedep, "track a block free"); 697SOFTDEP_TYPE(JADDREF, jaddref, "Journal inode ref add"); 698SOFTDEP_TYPE(JREMREF, jremref, "Journal inode ref remove"); 699SOFTDEP_TYPE(JMVREF, jmvref, "Journal inode ref move"); 700SOFTDEP_TYPE(JNEWBLK, jnewblk, "Journal new block"); 701SOFTDEP_TYPE(JFREEBLK, jfreeblk, "Journal free block"); 702SOFTDEP_TYPE(JFREEFRAG, jfreefrag, "Journal free frag"); 703SOFTDEP_TYPE(JSEG, jseg, "Journal segment"); 704SOFTDEP_TYPE(JSEGDEP, jsegdep, "Journal segment complete"); 705SOFTDEP_TYPE(SBDEP, sbdep, "Superblock write dependency"); 706SOFTDEP_TYPE(JTRUNC, jtrunc, "Journal inode truncation"); 707SOFTDEP_TYPE(JFSYNC, jfsync, "Journal fsync complete"); 708 709static MALLOC_DEFINE(M_SAVEDINO, "savedino", "Saved inodes"); 710static MALLOC_DEFINE(M_JBLOCKS, "jblocks", "Journal block locations"); 711 712/* 713 * translate from workitem type to memory type 714 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 715 */ 716static struct malloc_type *memtype[] = { 717 M_PAGEDEP, 718 M_INODEDEP, 719 M_BMSAFEMAP, 720 M_NEWBLK, 721 M_ALLOCDIRECT, 722 M_INDIRDEP, 723 M_ALLOCINDIR, 724 M_FREEFRAG, 725 M_FREEBLKS, 726 M_FREEFILE, 727 M_DIRADD, 728 M_MKDIR, 729 M_DIRREM, 730 M_NEWDIRBLK, 731 M_FREEWORK, 732 M_FREEDEP, 733 M_JADDREF, 734 M_JREMREF, 735 M_JMVREF, 736 M_JNEWBLK, 737 M_JFREEBLK, 738 M_JFREEFRAG, 739 M_JSEG, 740 M_JSEGDEP, 741 M_SBDEP, 742 M_JTRUNC, 743 M_JFSYNC 744}; 745 746static LIST_HEAD(mkdirlist, mkdir) mkdirlisthd; 747 748#define DtoM(type) (memtype[type]) 749 750/* 751 * Names of malloc types. 752 */ 753#define TYPENAME(type) \ 754 ((unsigned)(type) <= D_LAST ? memtype[type]->ks_shortdesc : "???") 755/* 756 * End system adaptation definitions. 757 */ 758 759#define DOTDOT_OFFSET offsetof(struct dirtemplate, dotdot_ino) 760#define DOT_OFFSET offsetof(struct dirtemplate, dot_ino) 761 762/* 763 * Forward declarations. 764 */ 765struct inodedep_hashhead; 766struct newblk_hashhead; 767struct pagedep_hashhead; 768struct bmsafemap_hashhead; 769 770/* 771 * Internal function prototypes. 772 */ 773static void softdep_error(char *, int); 774static void drain_output(struct vnode *); 775static struct buf *getdirtybuf(struct buf *, struct mtx *, int); 776static void clear_remove(void); 777static void clear_inodedeps(void); 778static void unlinked_inodedep(struct mount *, struct inodedep *); 779static void clear_unlinked_inodedep(struct inodedep *); 780static struct inodedep *first_unlinked_inodedep(struct ufsmount *); 781static int flush_pagedep_deps(struct vnode *, struct mount *, 782 struct diraddhd *); 783static int free_pagedep(struct pagedep *); 784static int flush_newblk_dep(struct vnode *, struct mount *, ufs_lbn_t); 785static int flush_inodedep_deps(struct vnode *, struct mount *, ino_t); 786static int flush_deplist(struct allocdirectlst *, int, int *); 787static int sync_cgs(struct mount *, int); 788static int handle_written_filepage(struct pagedep *, struct buf *); 789static int handle_written_sbdep(struct sbdep *, struct buf *); 790static void initiate_write_sbdep(struct sbdep *); 791static void diradd_inode_written(struct diradd *, struct inodedep *); 792static int handle_written_indirdep(struct indirdep *, struct buf *, 793 struct buf**); 794static int handle_written_inodeblock(struct inodedep *, struct buf *); 795static int jnewblk_rollforward(struct jnewblk *, struct fs *, struct cg *, 796 uint8_t *); 797static int handle_written_bmsafemap(struct bmsafemap *, struct buf *); 798static void handle_written_jaddref(struct jaddref *); 799static void handle_written_jremref(struct jremref *); 800static void handle_written_jseg(struct jseg *, struct buf *); 801static void handle_written_jnewblk(struct jnewblk *); 802static void handle_written_jblkdep(struct jblkdep *); 803static void handle_written_jfreefrag(struct jfreefrag *); 804static void complete_jseg(struct jseg *); 805static void jseg_write(struct ufsmount *ump, struct jseg *, uint8_t *); 806static void jaddref_write(struct jaddref *, struct jseg *, uint8_t *); 807static void jremref_write(struct jremref *, struct jseg *, uint8_t *); 808static void jmvref_write(struct jmvref *, struct jseg *, uint8_t *); 809static void jtrunc_write(struct jtrunc *, struct jseg *, uint8_t *); 810static void jfsync_write(struct jfsync *, struct jseg *, uint8_t *data); 811static void jnewblk_write(struct jnewblk *, struct jseg *, uint8_t *); 812static void jfreeblk_write(struct jfreeblk *, struct jseg *, uint8_t *); 813static void jfreefrag_write(struct jfreefrag *, struct jseg *, uint8_t *); 814static inline void inoref_write(struct inoref *, struct jseg *, 815 struct jrefrec *); 816static void handle_allocdirect_partdone(struct allocdirect *, 817 struct workhead *); 818static struct jnewblk *cancel_newblk(struct newblk *, struct worklist *, 819 struct workhead *); 820static void indirdep_complete(struct indirdep *); 821static int indirblk_lookup(struct mount *, ufs2_daddr_t); 822static void indirblk_insert(struct freework *); 823static void indirblk_remove(struct freework *); 824static void handle_allocindir_partdone(struct allocindir *); 825static void initiate_write_filepage(struct pagedep *, struct buf *); 826static void initiate_write_indirdep(struct indirdep*, struct buf *); 827static void handle_written_mkdir(struct mkdir *, int); 828static int jnewblk_rollback(struct jnewblk *, struct fs *, struct cg *, 829 uint8_t *); 830static void initiate_write_bmsafemap(struct bmsafemap *, struct buf *); 831static void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *); 832static void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *); 833static void handle_workitem_freefile(struct freefile *); 834static int handle_workitem_remove(struct dirrem *, int); 835static struct dirrem *newdirrem(struct buf *, struct inode *, 836 struct inode *, int, struct dirrem **); 837static struct indirdep *indirdep_lookup(struct mount *, struct inode *, 838 struct buf *); 839static void cancel_indirdep(struct indirdep *, struct buf *, 840 struct freeblks *); 841static void free_indirdep(struct indirdep *); 842static void free_diradd(struct diradd *, struct workhead *); 843static void merge_diradd(struct inodedep *, struct diradd *); 844static void complete_diradd(struct diradd *); 845static struct diradd *diradd_lookup(struct pagedep *, int); 846static struct jremref *cancel_diradd_dotdot(struct inode *, struct dirrem *, 847 struct jremref *); 848static struct jremref *cancel_mkdir_dotdot(struct inode *, struct dirrem *, 849 struct jremref *); 850static void cancel_diradd(struct diradd *, struct dirrem *, struct jremref *, 851 struct jremref *, struct jremref *); 852static void dirrem_journal(struct dirrem *, struct jremref *, struct jremref *, 853 struct jremref *); 854static void cancel_allocindir(struct allocindir *, struct buf *bp, 855 struct freeblks *, int); 856static int setup_trunc_indir(struct freeblks *, struct inode *, 857 ufs_lbn_t, ufs_lbn_t, ufs2_daddr_t); 858static void complete_trunc_indir(struct freework *); 859static void trunc_indirdep(struct indirdep *, struct freeblks *, struct buf *, 860 int); 861static void complete_mkdir(struct mkdir *); 862static void free_newdirblk(struct newdirblk *); 863static void free_jremref(struct jremref *); 864static void free_jaddref(struct jaddref *); 865static void free_jsegdep(struct jsegdep *); 866static void free_jsegs(struct jblocks *); 867static void rele_jseg(struct jseg *); 868static void free_jseg(struct jseg *, struct jblocks *); 869static void free_jnewblk(struct jnewblk *); 870static void free_jblkdep(struct jblkdep *); 871static void free_jfreefrag(struct jfreefrag *); 872static void free_freedep(struct freedep *); 873static void journal_jremref(struct dirrem *, struct jremref *, 874 struct inodedep *); 875static void cancel_jnewblk(struct jnewblk *, struct workhead *); 876static int cancel_jaddref(struct jaddref *, struct inodedep *, 877 struct workhead *); 878static void cancel_jfreefrag(struct jfreefrag *); 879static inline void setup_freedirect(struct freeblks *, struct inode *, 880 int, int); 881static inline void setup_freeext(struct freeblks *, struct inode *, int, int); 882static inline void setup_freeindir(struct freeblks *, struct inode *, int, 883 ufs_lbn_t, int); 884static inline struct freeblks *newfreeblks(struct mount *, struct inode *); 885static void freeblks_free(struct ufsmount *, struct freeblks *, int); 886static void indir_trunc(struct freework *, ufs2_daddr_t, ufs_lbn_t); 887ufs2_daddr_t blkcount(struct fs *, ufs2_daddr_t, off_t); 888static int trunc_check_buf(struct buf *, int *, ufs_lbn_t, int, int); 889static void trunc_dependencies(struct inode *, struct freeblks *, ufs_lbn_t, 890 int, int); 891static void trunc_pages(struct inode *, off_t, ufs2_daddr_t, int); 892static int cancel_pagedep(struct pagedep *, struct freeblks *, int); 893static int deallocate_dependencies(struct buf *, struct freeblks *, int); 894static void newblk_freefrag(struct newblk*); 895static void free_newblk(struct newblk *); 896static void cancel_allocdirect(struct allocdirectlst *, 897 struct allocdirect *, struct freeblks *); 898static int check_inode_unwritten(struct inodedep *); 899static int free_inodedep(struct inodedep *); 900static void freework_freeblock(struct freework *); 901static void freework_enqueue(struct freework *); 902static int handle_workitem_freeblocks(struct freeblks *, int); 903static int handle_complete_freeblocks(struct freeblks *, int); 904static void handle_workitem_indirblk(struct freework *); 905static void handle_written_freework(struct freework *); 906static void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *); 907static struct worklist *jnewblk_merge(struct worklist *, struct worklist *, 908 struct workhead *); 909static struct freefrag *setup_allocindir_phase2(struct buf *, struct inode *, 910 struct inodedep *, struct allocindir *, ufs_lbn_t); 911static struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t, 912 ufs2_daddr_t, ufs_lbn_t); 913static void handle_workitem_freefrag(struct freefrag *); 914static struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long, 915 ufs_lbn_t); 916static void allocdirect_merge(struct allocdirectlst *, 917 struct allocdirect *, struct allocdirect *); 918static struct freefrag *allocindir_merge(struct allocindir *, 919 struct allocindir *); 920static int bmsafemap_find(struct bmsafemap_hashhead *, struct mount *, int, 921 struct bmsafemap **); 922static struct bmsafemap *bmsafemap_lookup(struct mount *, struct buf *, 923 int cg, struct bmsafemap *); 924static int newblk_find(struct newblk_hashhead *, struct mount *, ufs2_daddr_t, 925 int, struct newblk **); 926static int newblk_lookup(struct mount *, ufs2_daddr_t, int, struct newblk **); 927static int inodedep_find(struct inodedep_hashhead *, struct fs *, ino_t, 928 struct inodedep **); 929static int inodedep_lookup(struct mount *, ino_t, int, struct inodedep **); 930static int pagedep_lookup(struct mount *, struct buf *bp, ino_t, ufs_lbn_t, 931 int, struct pagedep **); 932static int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t, 933 struct mount *mp, int, struct pagedep **); 934static void pause_timer(void *); 935static int request_cleanup(struct mount *, int); 936static int process_worklist_item(struct mount *, int, int); 937static void process_removes(struct vnode *); 938static void process_truncates(struct vnode *); 939static void jwork_move(struct workhead *, struct workhead *); 940static void jwork_insert(struct workhead *, struct jsegdep *); 941static void add_to_worklist(struct worklist *, int); 942static void wake_worklist(struct worklist *); 943static void wait_worklist(struct worklist *, char *); 944static void remove_from_worklist(struct worklist *); 945static void softdep_flush(void); 946static void softdep_flushjournal(struct mount *); 947static int softdep_speedup(void); 948static void worklist_speedup(void); 949static int journal_mount(struct mount *, struct fs *, struct ucred *); 950static void journal_unmount(struct mount *); 951static int journal_space(struct ufsmount *, int); 952static void journal_suspend(struct ufsmount *); 953static int journal_unsuspend(struct ufsmount *ump); 954static void softdep_prelink(struct vnode *, struct vnode *); 955static void add_to_journal(struct worklist *); 956static void remove_from_journal(struct worklist *); 957static void softdep_process_journal(struct mount *, struct worklist *, int); 958static struct jremref *newjremref(struct dirrem *, struct inode *, 959 struct inode *ip, off_t, nlink_t); 960static struct jaddref *newjaddref(struct inode *, ino_t, off_t, int16_t, 961 uint16_t); 962static inline void newinoref(struct inoref *, ino_t, ino_t, off_t, nlink_t, 963 uint16_t); 964static inline struct jsegdep *inoref_jseg(struct inoref *); 965static struct jmvref *newjmvref(struct inode *, ino_t, off_t, off_t); 966static struct jfreeblk *newjfreeblk(struct freeblks *, ufs_lbn_t, 967 ufs2_daddr_t, int); 968static struct jtrunc *newjtrunc(struct freeblks *, off_t, int); 969static void move_newblock_dep(struct jaddref *, struct inodedep *); 970static void cancel_jfreeblk(struct freeblks *, ufs2_daddr_t); 971static struct jfreefrag *newjfreefrag(struct freefrag *, struct inode *, 972 ufs2_daddr_t, long, ufs_lbn_t); 973static struct freework *newfreework(struct ufsmount *, struct freeblks *, 974 struct freework *, ufs_lbn_t, ufs2_daddr_t, int, int, int); 975static int jwait(struct worklist *, int); 976static struct inodedep *inodedep_lookup_ip(struct inode *); 977static int bmsafemap_rollbacks(struct bmsafemap *); 978static struct freefile *handle_bufwait(struct inodedep *, struct workhead *); 979static void handle_jwork(struct workhead *); 980static struct mkdir *setup_newdir(struct diradd *, ino_t, ino_t, struct buf *, 981 struct mkdir **); 982static struct jblocks *jblocks_create(void); 983static ufs2_daddr_t jblocks_alloc(struct jblocks *, int, int *); 984static void jblocks_free(struct jblocks *, struct mount *, int); 985static void jblocks_destroy(struct jblocks *); 986static void jblocks_add(struct jblocks *, ufs2_daddr_t, int); 987 988/* 989 * Exported softdep operations. 990 */ 991static void softdep_disk_io_initiation(struct buf *); 992static void softdep_disk_write_complete(struct buf *); 993static void softdep_deallocate_dependencies(struct buf *); 994static int softdep_count_dependencies(struct buf *bp, int); 995 996static struct mtx lk; 997MTX_SYSINIT(softdep_lock, &lk, "Softdep Lock", MTX_DEF); 998 999#define TRY_ACQUIRE_LOCK(lk) mtx_trylock(lk) 1000#define ACQUIRE_LOCK(lk) mtx_lock(lk) 1001#define FREE_LOCK(lk) mtx_unlock(lk) 1002 1003#define BUF_AREC(bp) lockallowrecurse(&(bp)->b_lock) 1004#define BUF_NOREC(bp) lockdisablerecurse(&(bp)->b_lock) 1005 1006/* 1007 * Worklist queue management. 1008 * These routines require that the lock be held. 1009 */ 1010#ifndef /* NOT */ DEBUG 1011#define WORKLIST_INSERT(head, item) do { \ 1012 (item)->wk_state |= ONWORKLIST; \ 1013 LIST_INSERT_HEAD(head, item, wk_list); \ 1014} while (0) 1015#define WORKLIST_REMOVE(item) do { \ 1016 (item)->wk_state &= ~ONWORKLIST; \ 1017 LIST_REMOVE(item, wk_list); \ 1018} while (0) 1019#define WORKLIST_INSERT_UNLOCKED WORKLIST_INSERT 1020#define WORKLIST_REMOVE_UNLOCKED WORKLIST_REMOVE 1021 1022#else /* DEBUG */ 1023static void worklist_insert(struct workhead *, struct worklist *, int); 1024static void worklist_remove(struct worklist *, int); 1025 1026#define WORKLIST_INSERT(head, item) worklist_insert(head, item, 1) 1027#define WORKLIST_INSERT_UNLOCKED(head, item) worklist_insert(head, item, 0) 1028#define WORKLIST_REMOVE(item) worklist_remove(item, 1) 1029#define WORKLIST_REMOVE_UNLOCKED(item) worklist_remove(item, 0) 1030 1031static void 1032worklist_insert(head, item, locked) 1033 struct workhead *head; 1034 struct worklist *item; 1035 int locked; 1036{ 1037 1038 if (locked) 1039 mtx_assert(&lk, MA_OWNED); 1040 if (item->wk_state & ONWORKLIST) 1041 panic("worklist_insert: %p %s(0x%X) already on list", 1042 item, TYPENAME(item->wk_type), item->wk_state); 1043 item->wk_state |= ONWORKLIST; 1044 LIST_INSERT_HEAD(head, item, wk_list); 1045} 1046 1047static void 1048worklist_remove(item, locked) 1049 struct worklist *item; 1050 int locked; 1051{ 1052 1053 if (locked) 1054 mtx_assert(&lk, MA_OWNED); 1055 if ((item->wk_state & ONWORKLIST) == 0) 1056 panic("worklist_remove: %p %s(0x%X) not on list", 1057 item, TYPENAME(item->wk_type), item->wk_state); 1058 item->wk_state &= ~ONWORKLIST; 1059 LIST_REMOVE(item, wk_list); 1060} 1061#endif /* DEBUG */ 1062 1063/* 1064 * Merge two jsegdeps keeping only the oldest one as newer references 1065 * can't be discarded until after older references. 1066 */ 1067static inline struct jsegdep * 1068jsegdep_merge(struct jsegdep *one, struct jsegdep *two) 1069{ 1070 struct jsegdep *swp; 1071 1072 if (two == NULL) 1073 return (one); 1074 1075 if (one->jd_seg->js_seq > two->jd_seg->js_seq) { 1076 swp = one; 1077 one = two; 1078 two = swp; 1079 } 1080 WORKLIST_REMOVE(&two->jd_list); 1081 free_jsegdep(two); 1082 1083 return (one); 1084} 1085 1086/* 1087 * If two freedeps are compatible free one to reduce list size. 1088 */ 1089static inline struct freedep * 1090freedep_merge(struct freedep *one, struct freedep *two) 1091{ 1092 if (two == NULL) 1093 return (one); 1094 1095 if (one->fd_freework == two->fd_freework) { 1096 WORKLIST_REMOVE(&two->fd_list); 1097 free_freedep(two); 1098 } 1099 return (one); 1100} 1101 1102/* 1103 * Move journal work from one list to another. Duplicate freedeps and 1104 * jsegdeps are coalesced to keep the lists as small as possible. 1105 */ 1106static void 1107jwork_move(dst, src) 1108 struct workhead *dst; 1109 struct workhead *src; 1110{ 1111 struct freedep *freedep; 1112 struct jsegdep *jsegdep; 1113 struct worklist *wkn; 1114 struct worklist *wk; 1115 1116 KASSERT(dst != src, 1117 ("jwork_move: dst == src")); 1118 freedep = NULL; 1119 jsegdep = NULL; 1120 LIST_FOREACH_SAFE(wk, dst, wk_list, wkn) { 1121 if (wk->wk_type == D_JSEGDEP) 1122 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep); 1123 if (wk->wk_type == D_FREEDEP) 1124 freedep = freedep_merge(WK_FREEDEP(wk), freedep); 1125 } 1126 1127 mtx_assert(&lk, MA_OWNED); 1128 while ((wk = LIST_FIRST(src)) != NULL) { 1129 WORKLIST_REMOVE(wk); 1130 WORKLIST_INSERT(dst, wk); 1131 if (wk->wk_type == D_JSEGDEP) { 1132 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep); 1133 continue; 1134 } 1135 if (wk->wk_type == D_FREEDEP) 1136 freedep = freedep_merge(WK_FREEDEP(wk), freedep); 1137 } 1138} 1139 1140static void 1141jwork_insert(dst, jsegdep) 1142 struct workhead *dst; 1143 struct jsegdep *jsegdep; 1144{ 1145 struct jsegdep *jsegdepn; 1146 struct worklist *wk; 1147 1148 LIST_FOREACH(wk, dst, wk_list) 1149 if (wk->wk_type == D_JSEGDEP) 1150 break; 1151 if (wk == NULL) { 1152 WORKLIST_INSERT(dst, &jsegdep->jd_list); 1153 return; 1154 } 1155 jsegdepn = WK_JSEGDEP(wk); 1156 if (jsegdep->jd_seg->js_seq < jsegdepn->jd_seg->js_seq) { 1157 WORKLIST_REMOVE(wk); 1158 free_jsegdep(jsegdepn); 1159 WORKLIST_INSERT(dst, &jsegdep->jd_list); 1160 } else 1161 free_jsegdep(jsegdep); 1162} 1163 1164/* 1165 * Routines for tracking and managing workitems. 1166 */ 1167static void workitem_free(struct worklist *, int); 1168static void workitem_alloc(struct worklist *, int, struct mount *); 1169 1170#define WORKITEM_FREE(item, type) workitem_free((struct worklist *)(item), (type)) 1171 1172static void 1173workitem_free(item, type) 1174 struct worklist *item; 1175 int type; 1176{ 1177 struct ufsmount *ump; 1178 mtx_assert(&lk, MA_OWNED); 1179 1180#ifdef DEBUG 1181 if (item->wk_state & ONWORKLIST) 1182 panic("workitem_free: %s(0x%X) still on list", 1183 TYPENAME(item->wk_type), item->wk_state); 1184 if (item->wk_type != type) 1185 panic("workitem_free: type mismatch %s != %s", 1186 TYPENAME(item->wk_type), TYPENAME(type)); 1187#endif 1188 if (item->wk_state & IOWAITING) 1189 wakeup(item); 1190 ump = VFSTOUFS(item->wk_mp); 1191 if (--ump->softdep_deps == 0 && ump->softdep_req) 1192 wakeup(&ump->softdep_deps); 1193 dep_current[type]--; 1194 free(item, DtoM(type)); 1195} 1196 1197static void 1198workitem_alloc(item, type, mp) 1199 struct worklist *item; 1200 int type; 1201 struct mount *mp; 1202{ 1203 struct ufsmount *ump; 1204 1205 item->wk_type = type; 1206 item->wk_mp = mp; 1207 item->wk_state = 0; 1208 1209 ump = VFSTOUFS(mp); 1210 ACQUIRE_LOCK(&lk); 1211 dep_current[type]++; 1212 dep_total[type]++; 1213 ump->softdep_deps++; 1214 ump->softdep_accdeps++; 1215 FREE_LOCK(&lk); 1216} 1217 1218/* 1219 * Workitem queue management 1220 */ 1221static int max_softdeps; /* maximum number of structs before slowdown */ 1222static int maxindirdeps = 50; /* max number of indirdeps before slowdown */ 1223static int tickdelay = 2; /* number of ticks to pause during slowdown */ 1224static int proc_waiting; /* tracks whether we have a timeout posted */ 1225static int *stat_countp; /* statistic to count in proc_waiting timeout */ 1226static struct callout softdep_callout; 1227static int req_pending; 1228static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 1229static int req_clear_remove; /* syncer process flush some freeblks */ 1230 1231/* 1232 * runtime statistics 1233 */ 1234static int stat_worklist_push; /* number of worklist cleanups */ 1235static int stat_blk_limit_push; /* number of times block limit neared */ 1236static int stat_ino_limit_push; /* number of times inode limit neared */ 1237static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 1238static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 1239static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 1240static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 1241static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 1242static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 1243static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 1244static int stat_jaddref; /* bufs redirtied as ino bitmap can not write */ 1245static int stat_jnewblk; /* bufs redirtied as blk bitmap can not write */ 1246static int stat_journal_min; /* Times hit journal min threshold */ 1247static int stat_journal_low; /* Times hit journal low threshold */ 1248static int stat_journal_wait; /* Times blocked in jwait(). */ 1249static int stat_jwait_filepage; /* Times blocked in jwait() for filepage. */ 1250static int stat_jwait_freeblks; /* Times blocked in jwait() for freeblks. */ 1251static int stat_jwait_inode; /* Times blocked in jwait() for inodes. */ 1252static int stat_jwait_newblk; /* Times blocked in jwait() for newblks. */ 1253static int stat_cleanup_high_delay; /* Maximum cleanup delay (in ticks) */ 1254static int stat_cleanup_blkrequests; /* Number of block cleanup requests */ 1255static int stat_cleanup_inorequests; /* Number of inode cleanup requests */ 1256static int stat_cleanup_retries; /* Number of cleanups that needed to flush */ 1257static int stat_cleanup_failures; /* Number of cleanup requests that failed */ 1258 1259SYSCTL_INT(_debug_softdep, OID_AUTO, max_softdeps, CTLFLAG_RW, 1260 &max_softdeps, 0, ""); 1261SYSCTL_INT(_debug_softdep, OID_AUTO, tickdelay, CTLFLAG_RW, 1262 &tickdelay, 0, ""); 1263SYSCTL_INT(_debug_softdep, OID_AUTO, maxindirdeps, CTLFLAG_RW, 1264 &maxindirdeps, 0, ""); 1265SYSCTL_INT(_debug_softdep, OID_AUTO, worklist_push, CTLFLAG_RW, 1266 &stat_worklist_push, 0,""); 1267SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_push, CTLFLAG_RW, 1268 &stat_blk_limit_push, 0,""); 1269SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_push, CTLFLAG_RW, 1270 &stat_ino_limit_push, 0,""); 1271SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_hit, CTLFLAG_RW, 1272 &stat_blk_limit_hit, 0, ""); 1273SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_hit, CTLFLAG_RW, 1274 &stat_ino_limit_hit, 0, ""); 1275SYSCTL_INT(_debug_softdep, OID_AUTO, sync_limit_hit, CTLFLAG_RW, 1276 &stat_sync_limit_hit, 0, ""); 1277SYSCTL_INT(_debug_softdep, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, 1278 &stat_indir_blk_ptrs, 0, ""); 1279SYSCTL_INT(_debug_softdep, OID_AUTO, inode_bitmap, CTLFLAG_RW, 1280 &stat_inode_bitmap, 0, ""); 1281SYSCTL_INT(_debug_softdep, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, 1282 &stat_direct_blk_ptrs, 0, ""); 1283SYSCTL_INT(_debug_softdep, OID_AUTO, dir_entry, CTLFLAG_RW, 1284 &stat_dir_entry, 0, ""); 1285SYSCTL_INT(_debug_softdep, OID_AUTO, jaddref_rollback, CTLFLAG_RW, 1286 &stat_jaddref, 0, ""); 1287SYSCTL_INT(_debug_softdep, OID_AUTO, jnewblk_rollback, CTLFLAG_RW, 1288 &stat_jnewblk, 0, ""); 1289SYSCTL_INT(_debug_softdep, OID_AUTO, journal_low, CTLFLAG_RW, 1290 &stat_journal_low, 0, ""); 1291SYSCTL_INT(_debug_softdep, OID_AUTO, journal_min, CTLFLAG_RW, 1292 &stat_journal_min, 0, ""); 1293SYSCTL_INT(_debug_softdep, OID_AUTO, journal_wait, CTLFLAG_RW, 1294 &stat_journal_wait, 0, ""); 1295SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_filepage, CTLFLAG_RW, 1296 &stat_jwait_filepage, 0, ""); 1297SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_freeblks, CTLFLAG_RW, 1298 &stat_jwait_freeblks, 0, ""); 1299SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_inode, CTLFLAG_RW, 1300 &stat_jwait_inode, 0, ""); 1301SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_newblk, CTLFLAG_RW, 1302 &stat_jwait_newblk, 0, ""); 1303SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_blkrequests, CTLFLAG_RW, 1304 &stat_cleanup_blkrequests, 0, ""); 1305SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_inorequests, CTLFLAG_RW, 1306 &stat_cleanup_inorequests, 0, ""); 1307SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_high_delay, CTLFLAG_RW, 1308 &stat_cleanup_high_delay, 0, ""); 1309SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_retries, CTLFLAG_RW, 1310 &stat_cleanup_retries, 0, ""); 1311SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_failures, CTLFLAG_RW, 1312 &stat_cleanup_failures, 0, ""); 1313 1314SYSCTL_DECL(_vfs_ffs); 1315 1316LIST_HEAD(bmsafemap_hashhead, bmsafemap) *bmsafemap_hashtbl; 1317static u_long bmsafemap_hash; /* size of hash table - 1 */ 1318 1319static int compute_summary_at_mount = 0; /* Whether to recompute the summary at mount time */ 1320SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW, 1321 &compute_summary_at_mount, 0, "Recompute summary at mount"); 1322 1323static struct proc *softdepproc; 1324static struct kproc_desc softdep_kp = { 1325 "softdepflush", 1326 softdep_flush, 1327 &softdepproc 1328}; 1329SYSINIT(sdproc, SI_SUB_KTHREAD_UPDATE, SI_ORDER_ANY, kproc_start, 1330 &softdep_kp); 1331 1332static void 1333softdep_flush(void) 1334{ 1335 struct mount *nmp; 1336 struct mount *mp; 1337 struct ufsmount *ump; 1338 struct thread *td; 1339 int remaining; 1340 int progress; 1341 1342 td = curthread; 1343 td->td_pflags |= TDP_NORUNNINGBUF; 1344 1345 for (;;) { 1346 kproc_suspend_check(softdepproc); 1347 ACQUIRE_LOCK(&lk); 1348 /* 1349 * If requested, try removing inode or removal dependencies. 1350 */ 1351 if (req_clear_inodedeps) { 1352 clear_inodedeps(); 1353 req_clear_inodedeps -= 1; 1354 wakeup_one(&proc_waiting); 1355 } 1356 if (req_clear_remove) { 1357 clear_remove(); 1358 req_clear_remove -= 1; 1359 wakeup_one(&proc_waiting); 1360 } 1361 FREE_LOCK(&lk); 1362 remaining = progress = 0; 1363 mtx_lock(&mountlist_mtx); 1364 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 1365 nmp = TAILQ_NEXT(mp, mnt_list); 1366 if (MOUNTEDSOFTDEP(mp) == 0) 1367 continue; 1368 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) 1369 continue; 1370 progress += softdep_process_worklist(mp, 0); 1371 ump = VFSTOUFS(mp); 1372 remaining += ump->softdep_on_worklist; 1373 mtx_lock(&mountlist_mtx); 1374 nmp = TAILQ_NEXT(mp, mnt_list); 1375 vfs_unbusy(mp); 1376 } 1377 mtx_unlock(&mountlist_mtx); 1378 if (remaining && progress) 1379 continue; 1380 ACQUIRE_LOCK(&lk); 1381 if (!req_pending) 1382 msleep(&req_pending, &lk, PVM, "sdflush", hz); 1383 req_pending = 0; 1384 FREE_LOCK(&lk); 1385 } 1386} 1387 1388static void 1389worklist_speedup(void) 1390{ 1391 mtx_assert(&lk, MA_OWNED); 1392 if (req_pending == 0) { 1393 req_pending = 1; 1394 wakeup(&req_pending); 1395 } 1396} 1397 1398static int 1399softdep_speedup(void) 1400{ 1401 1402 worklist_speedup(); 1403 bd_speedup(); 1404 return speedup_syncer(); 1405} 1406 1407/* 1408 * Add an item to the end of the work queue. 1409 * This routine requires that the lock be held. 1410 * This is the only routine that adds items to the list. 1411 * The following routine is the only one that removes items 1412 * and does so in order from first to last. 1413 */ 1414 1415#define WK_HEAD 0x0001 /* Add to HEAD. */ 1416#define WK_NODELAY 0x0002 /* Process immediately. */ 1417 1418static void 1419add_to_worklist(wk, flags) 1420 struct worklist *wk; 1421 int flags; 1422{ 1423 struct ufsmount *ump; 1424 1425 mtx_assert(&lk, MA_OWNED); 1426 ump = VFSTOUFS(wk->wk_mp); 1427 if (wk->wk_state & ONWORKLIST) 1428 panic("add_to_worklist: %s(0x%X) already on list", 1429 TYPENAME(wk->wk_type), wk->wk_state); 1430 wk->wk_state |= ONWORKLIST; 1431 if (ump->softdep_on_worklist == 0) { 1432 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); 1433 ump->softdep_worklist_tail = wk; 1434 } else if (flags & WK_HEAD) { 1435 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); 1436 } else { 1437 LIST_INSERT_AFTER(ump->softdep_worklist_tail, wk, wk_list); 1438 ump->softdep_worklist_tail = wk; 1439 } 1440 ump->softdep_on_worklist += 1; 1441 if (flags & WK_NODELAY) 1442 worklist_speedup(); 1443} 1444 1445/* 1446 * Remove the item to be processed. If we are removing the last 1447 * item on the list, we need to recalculate the tail pointer. 1448 */ 1449static void 1450remove_from_worklist(wk) 1451 struct worklist *wk; 1452{ 1453 struct ufsmount *ump; 1454 1455 ump = VFSTOUFS(wk->wk_mp); 1456 WORKLIST_REMOVE(wk); 1457 if (ump->softdep_worklist_tail == wk) 1458 ump->softdep_worklist_tail = 1459 (struct worklist *)wk->wk_list.le_prev; 1460 ump->softdep_on_worklist -= 1; 1461} 1462 1463static void 1464wake_worklist(wk) 1465 struct worklist *wk; 1466{ 1467 if (wk->wk_state & IOWAITING) { 1468 wk->wk_state &= ~IOWAITING; 1469 wakeup(wk); 1470 } 1471} 1472 1473static void 1474wait_worklist(wk, wmesg) 1475 struct worklist *wk; 1476 char *wmesg; 1477{ 1478 1479 wk->wk_state |= IOWAITING; 1480 msleep(wk, &lk, PVM, wmesg, 0); 1481} 1482 1483/* 1484 * Process that runs once per second to handle items in the background queue. 1485 * 1486 * Note that we ensure that everything is done in the order in which they 1487 * appear in the queue. The code below depends on this property to ensure 1488 * that blocks of a file are freed before the inode itself is freed. This 1489 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 1490 * until all the old ones have been purged from the dependency lists. 1491 */ 1492int 1493softdep_process_worklist(mp, full) 1494 struct mount *mp; 1495 int full; 1496{ 1497 int cnt, matchcnt; 1498 struct ufsmount *ump; 1499 long starttime; 1500 1501 KASSERT(mp != NULL, ("softdep_process_worklist: NULL mp")); 1502 /* 1503 * Record the process identifier of our caller so that we can give 1504 * this process preferential treatment in request_cleanup below. 1505 */ 1506 matchcnt = 0; 1507 ump = VFSTOUFS(mp); 1508 ACQUIRE_LOCK(&lk); 1509 starttime = time_second; 1510 softdep_process_journal(mp, NULL, full?MNT_WAIT:0); 1511 while (ump->softdep_on_worklist > 0) { 1512 if ((cnt = process_worklist_item(mp, 10, LK_NOWAIT)) == 0) 1513 break; 1514 else 1515 matchcnt += cnt; 1516 /* 1517 * If requested, try removing inode or removal dependencies. 1518 */ 1519 if (req_clear_inodedeps) { 1520 clear_inodedeps(); 1521 req_clear_inodedeps -= 1; 1522 wakeup_one(&proc_waiting); 1523 } 1524 if (req_clear_remove) { 1525 clear_remove(); 1526 req_clear_remove -= 1; 1527 wakeup_one(&proc_waiting); 1528 } 1529 /* 1530 * We do not generally want to stop for buffer space, but if 1531 * we are really being a buffer hog, we will stop and wait. 1532 */ 1533 if (should_yield()) { 1534 FREE_LOCK(&lk); 1535 kern_yield(PRI_UNCHANGED); 1536 bwillwrite(); 1537 ACQUIRE_LOCK(&lk); 1538 } 1539 /* 1540 * Never allow processing to run for more than one 1541 * second. Otherwise the other mountpoints may get 1542 * excessively backlogged. 1543 */ 1544 if (!full && starttime != time_second) 1545 break; 1546 } 1547 if (full == 0) 1548 journal_unsuspend(ump); 1549 FREE_LOCK(&lk); 1550 return (matchcnt); 1551} 1552 1553/* 1554 * Process all removes associated with a vnode if we are running out of 1555 * journal space. Any other process which attempts to flush these will 1556 * be unable as we have the vnodes locked. 1557 */ 1558static void 1559process_removes(vp) 1560 struct vnode *vp; 1561{ 1562 struct inodedep *inodedep; 1563 struct dirrem *dirrem; 1564 struct mount *mp; 1565 ino_t inum; 1566 1567 mtx_assert(&lk, MA_OWNED); 1568 1569 mp = vp->v_mount; 1570 inum = VTOI(vp)->i_number; 1571 for (;;) { 1572top: 1573 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0) 1574 return; 1575 LIST_FOREACH(dirrem, &inodedep->id_dirremhd, dm_inonext) { 1576 /* 1577 * If another thread is trying to lock this vnode 1578 * it will fail but we must wait for it to do so 1579 * before we can proceed. 1580 */ 1581 if (dirrem->dm_state & INPROGRESS) { 1582 wait_worklist(&dirrem->dm_list, "pwrwait"); 1583 goto top; 1584 } 1585 if ((dirrem->dm_state & (COMPLETE | ONWORKLIST)) == 1586 (COMPLETE | ONWORKLIST)) 1587 break; 1588 } 1589 if (dirrem == NULL) 1590 return; 1591 remove_from_worklist(&dirrem->dm_list); 1592 FREE_LOCK(&lk); 1593 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT)) 1594 panic("process_removes: suspended filesystem"); 1595 handle_workitem_remove(dirrem, 0); 1596 vn_finished_secondary_write(mp); 1597 ACQUIRE_LOCK(&lk); 1598 } 1599} 1600 1601/* 1602 * Process all truncations associated with a vnode if we are running out 1603 * of journal space. This is called when the vnode lock is already held 1604 * and no other process can clear the truncation. This function returns 1605 * a value greater than zero if it did any work. 1606 */ 1607static void 1608process_truncates(vp) 1609 struct vnode *vp; 1610{ 1611 struct inodedep *inodedep; 1612 struct freeblks *freeblks; 1613 struct mount *mp; 1614 ino_t inum; 1615 int cgwait; 1616 1617 mtx_assert(&lk, MA_OWNED); 1618 1619 mp = vp->v_mount; 1620 inum = VTOI(vp)->i_number; 1621 for (;;) { 1622 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0) 1623 return; 1624 cgwait = 0; 1625 TAILQ_FOREACH(freeblks, &inodedep->id_freeblklst, fb_next) { 1626 /* Journal entries not yet written. */ 1627 if (!LIST_EMPTY(&freeblks->fb_jblkdephd)) { 1628 jwait(&LIST_FIRST( 1629 &freeblks->fb_jblkdephd)->jb_list, 1630 MNT_WAIT); 1631 break; 1632 } 1633 /* Another thread is executing this item. */ 1634 if (freeblks->fb_state & INPROGRESS) { 1635 wait_worklist(&freeblks->fb_list, "ptrwait"); 1636 break; 1637 } 1638 /* Freeblks is waiting on a inode write. */ 1639 if ((freeblks->fb_state & COMPLETE) == 0) { 1640 FREE_LOCK(&lk); 1641 ffs_update(vp, 1); 1642 ACQUIRE_LOCK(&lk); 1643 break; 1644 } 1645 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST)) == 1646 (ALLCOMPLETE | ONWORKLIST)) { 1647 remove_from_worklist(&freeblks->fb_list); 1648 freeblks->fb_state |= INPROGRESS; 1649 FREE_LOCK(&lk); 1650 if (vn_start_secondary_write(NULL, &mp, 1651 V_NOWAIT)) 1652 panic("process_truncates: " 1653 "suspended filesystem"); 1654 handle_workitem_freeblocks(freeblks, 0); 1655 vn_finished_secondary_write(mp); 1656 ACQUIRE_LOCK(&lk); 1657 break; 1658 } 1659 if (freeblks->fb_cgwait) 1660 cgwait++; 1661 } 1662 if (cgwait) { 1663 FREE_LOCK(&lk); 1664 sync_cgs(mp, MNT_WAIT); 1665 ffs_sync_snap(mp, MNT_WAIT); 1666 ACQUIRE_LOCK(&lk); 1667 continue; 1668 } 1669 if (freeblks == NULL) 1670 break; 1671 } 1672 return; 1673} 1674 1675/* 1676 * Process one item on the worklist. 1677 */ 1678static int 1679process_worklist_item(mp, target, flags) 1680 struct mount *mp; 1681 int target; 1682 int flags; 1683{ 1684 struct worklist sintenel; 1685 struct worklist *wk; 1686 struct ufsmount *ump; 1687 int matchcnt; 1688 int error; 1689 1690 mtx_assert(&lk, MA_OWNED); 1691 KASSERT(mp != NULL, ("process_worklist_item: NULL mp")); 1692 /* 1693 * If we are being called because of a process doing a 1694 * copy-on-write, then it is not safe to write as we may 1695 * recurse into the copy-on-write routine. 1696 */ 1697 if (curthread->td_pflags & TDP_COWINPROGRESS) 1698 return (-1); 1699 PHOLD(curproc); /* Don't let the stack go away. */ 1700 ump = VFSTOUFS(mp); 1701 matchcnt = 0; 1702 sintenel.wk_mp = NULL; 1703 sintenel.wk_type = D_SENTINAL; 1704 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, &sintenel, wk_list); 1705 for (wk = LIST_NEXT(&sintenel, wk_list); wk != NULL; 1706 wk = LIST_NEXT(&sintenel, wk_list)) { 1707 if (wk->wk_type == D_SENTINAL) { 1708 LIST_REMOVE(&sintenel, wk_list); 1709 LIST_INSERT_AFTER(wk, &sintenel, wk_list); 1710 continue; 1711 } 1712 if (wk->wk_state & INPROGRESS) 1713 panic("process_worklist_item: %p already in progress.", 1714 wk); 1715 wk->wk_state |= INPROGRESS; 1716 remove_from_worklist(wk); 1717 FREE_LOCK(&lk); 1718 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT)) 1719 panic("process_worklist_item: suspended filesystem"); 1720 switch (wk->wk_type) { 1721 case D_DIRREM: 1722 /* removal of a directory entry */ 1723 error = handle_workitem_remove(WK_DIRREM(wk), flags); 1724 break; 1725 1726 case D_FREEBLKS: 1727 /* releasing blocks and/or fragments from a file */ 1728 error = handle_workitem_freeblocks(WK_FREEBLKS(wk), 1729 flags); 1730 break; 1731 1732 case D_FREEFRAG: 1733 /* releasing a fragment when replaced as a file grows */ 1734 handle_workitem_freefrag(WK_FREEFRAG(wk)); 1735 error = 0; 1736 break; 1737 1738 case D_FREEFILE: 1739 /* releasing an inode when its link count drops to 0 */ 1740 handle_workitem_freefile(WK_FREEFILE(wk)); 1741 error = 0; 1742 break; 1743 1744 default: 1745 panic("%s_process_worklist: Unknown type %s", 1746 "softdep", TYPENAME(wk->wk_type)); 1747 /* NOTREACHED */ 1748 } 1749 vn_finished_secondary_write(mp); 1750 ACQUIRE_LOCK(&lk); 1751 if (error == 0) { 1752 if (++matchcnt == target) 1753 break; 1754 continue; 1755 } 1756 /* 1757 * We have to retry the worklist item later. Wake up any 1758 * waiters who may be able to complete it immediately and 1759 * add the item back to the head so we don't try to execute 1760 * it again. 1761 */ 1762 wk->wk_state &= ~INPROGRESS; 1763 wake_worklist(wk); 1764 add_to_worklist(wk, WK_HEAD); 1765 } 1766 LIST_REMOVE(&sintenel, wk_list); 1767 /* Sentinal could've become the tail from remove_from_worklist. */ 1768 if (ump->softdep_worklist_tail == &sintenel) 1769 ump->softdep_worklist_tail = 1770 (struct worklist *)sintenel.wk_list.le_prev; 1771 PRELE(curproc); 1772 return (matchcnt); 1773} 1774 1775/* 1776 * Move dependencies from one buffer to another. 1777 */ 1778int 1779softdep_move_dependencies(oldbp, newbp) 1780 struct buf *oldbp; 1781 struct buf *newbp; 1782{ 1783 struct worklist *wk, *wktail; 1784 int dirty; 1785 1786 dirty = 0; 1787 wktail = NULL; 1788 ACQUIRE_LOCK(&lk); 1789 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 1790 LIST_REMOVE(wk, wk_list); 1791 if (wk->wk_type == D_BMSAFEMAP && 1792 bmsafemap_rollbacks(WK_BMSAFEMAP(wk))) 1793 dirty = 1; 1794 if (wktail == 0) 1795 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 1796 else 1797 LIST_INSERT_AFTER(wktail, wk, wk_list); 1798 wktail = wk; 1799 } 1800 FREE_LOCK(&lk); 1801 1802 return (dirty); 1803} 1804 1805/* 1806 * Purge the work list of all items associated with a particular mount point. 1807 */ 1808int 1809softdep_flushworklist(oldmnt, countp, td) 1810 struct mount *oldmnt; 1811 int *countp; 1812 struct thread *td; 1813{ 1814 struct vnode *devvp; 1815 int count, error = 0; 1816 struct ufsmount *ump; 1817 1818 /* 1819 * Alternately flush the block device associated with the mount 1820 * point and process any dependencies that the flushing 1821 * creates. We continue until no more worklist dependencies 1822 * are found. 1823 */ 1824 *countp = 0; 1825 ump = VFSTOUFS(oldmnt); 1826 devvp = ump->um_devvp; 1827 while ((count = softdep_process_worklist(oldmnt, 1)) > 0) { 1828 *countp += count; 1829 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1830 error = VOP_FSYNC(devvp, MNT_WAIT, td); 1831 VOP_UNLOCK(devvp, 0); 1832 if (error) 1833 break; 1834 } 1835 return (error); 1836} 1837 1838int 1839softdep_waitidle(struct mount *mp) 1840{ 1841 struct ufsmount *ump; 1842 int error; 1843 int i; 1844 1845 ump = VFSTOUFS(mp); 1846 ACQUIRE_LOCK(&lk); 1847 for (i = 0; i < 10 && ump->softdep_deps; i++) { 1848 ump->softdep_req = 1; 1849 if (ump->softdep_on_worklist) 1850 panic("softdep_waitidle: work added after flush."); 1851 msleep(&ump->softdep_deps, &lk, PVM, "softdeps", 1); 1852 } 1853 ump->softdep_req = 0; 1854 FREE_LOCK(&lk); 1855 error = 0; 1856 if (i == 10) { 1857 error = EBUSY; 1858 printf("softdep_waitidle: Failed to flush worklist for %p\n", 1859 mp); 1860 } 1861 1862 return (error); 1863} 1864 1865/* 1866 * Flush all vnodes and worklist items associated with a specified mount point. 1867 */ 1868int 1869softdep_flushfiles(oldmnt, flags, td) 1870 struct mount *oldmnt; 1871 int flags; 1872 struct thread *td; 1873{ 1874 int error, depcount, loopcnt, retry_flush_count, retry; 1875 1876 loopcnt = 10; 1877 retry_flush_count = 3; 1878retry_flush: 1879 error = 0; 1880 1881 /* 1882 * Alternately flush the vnodes associated with the mount 1883 * point and process any dependencies that the flushing 1884 * creates. In theory, this loop can happen at most twice, 1885 * but we give it a few extra just to be sure. 1886 */ 1887 for (; loopcnt > 0; loopcnt--) { 1888 /* 1889 * Do another flush in case any vnodes were brought in 1890 * as part of the cleanup operations. 1891 */ 1892 if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0) 1893 break; 1894 if ((error = softdep_flushworklist(oldmnt, &depcount, td)) != 0 || 1895 depcount == 0) 1896 break; 1897 } 1898 /* 1899 * If we are unmounting then it is an error to fail. If we 1900 * are simply trying to downgrade to read-only, then filesystem 1901 * activity can keep us busy forever, so we just fail with EBUSY. 1902 */ 1903 if (loopcnt == 0) { 1904 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 1905 panic("softdep_flushfiles: looping"); 1906 error = EBUSY; 1907 } 1908 if (!error) 1909 error = softdep_waitidle(oldmnt); 1910 if (!error) { 1911 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) { 1912 retry = 0; 1913 MNT_ILOCK(oldmnt); 1914 KASSERT((oldmnt->mnt_kern_flag & MNTK_NOINSMNTQ) != 0, 1915 ("softdep_flushfiles: !MNTK_NOINSMNTQ")); 1916 if (oldmnt->mnt_nvnodelistsize > 0) { 1917 if (--retry_flush_count > 0) { 1918 retry = 1; 1919 loopcnt = 3; 1920 } else 1921 error = EBUSY; 1922 } 1923 MNT_IUNLOCK(oldmnt); 1924 if (retry) 1925 goto retry_flush; 1926 } 1927 } 1928 return (error); 1929} 1930 1931/* 1932 * Structure hashing. 1933 * 1934 * There are three types of structures that can be looked up: 1935 * 1) pagedep structures identified by mount point, inode number, 1936 * and logical block. 1937 * 2) inodedep structures identified by mount point and inode number. 1938 * 3) newblk structures identified by mount point and 1939 * physical block number. 1940 * 1941 * The "pagedep" and "inodedep" dependency structures are hashed 1942 * separately from the file blocks and inodes to which they correspond. 1943 * This separation helps when the in-memory copy of an inode or 1944 * file block must be replaced. It also obviates the need to access 1945 * an inode or file page when simply updating (or de-allocating) 1946 * dependency structures. Lookup of newblk structures is needed to 1947 * find newly allocated blocks when trying to associate them with 1948 * their allocdirect or allocindir structure. 1949 * 1950 * The lookup routines optionally create and hash a new instance when 1951 * an existing entry is not found. 1952 */ 1953#define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 1954#define NODELAY 0x0002 /* cannot do background work */ 1955 1956/* 1957 * Structures and routines associated with pagedep caching. 1958 */ 1959LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl; 1960u_long pagedep_hash; /* size of hash table - 1 */ 1961#define PAGEDEP_HASH(mp, inum, lbn) \ 1962 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \ 1963 pagedep_hash]) 1964 1965static int 1966pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp) 1967 struct pagedep_hashhead *pagedephd; 1968 ino_t ino; 1969 ufs_lbn_t lbn; 1970 struct mount *mp; 1971 int flags; 1972 struct pagedep **pagedeppp; 1973{ 1974 struct pagedep *pagedep; 1975 1976 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 1977 if (ino == pagedep->pd_ino && lbn == pagedep->pd_lbn && 1978 mp == pagedep->pd_list.wk_mp) { 1979 *pagedeppp = pagedep; 1980 return (1); 1981 } 1982 } 1983 *pagedeppp = NULL; 1984 return (0); 1985} 1986/* 1987 * Look up a pagedep. Return 1 if found, 0 otherwise. 1988 * If not found, allocate if DEPALLOC flag is passed. 1989 * Found or allocated entry is returned in pagedeppp. 1990 * This routine must be called with splbio interrupts blocked. 1991 */ 1992static int 1993pagedep_lookup(mp, bp, ino, lbn, flags, pagedeppp) 1994 struct mount *mp; 1995 struct buf *bp; 1996 ino_t ino; 1997 ufs_lbn_t lbn; 1998 int flags; 1999 struct pagedep **pagedeppp; 2000{ 2001 struct pagedep *pagedep; 2002 struct pagedep_hashhead *pagedephd; 2003 struct worklist *wk; 2004 int ret; 2005 int i; 2006 2007 mtx_assert(&lk, MA_OWNED); 2008 if (bp) { 2009 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 2010 if (wk->wk_type == D_PAGEDEP) { 2011 *pagedeppp = WK_PAGEDEP(wk); 2012 return (1); 2013 } 2014 } 2015 } 2016 pagedephd = PAGEDEP_HASH(mp, ino, lbn); 2017 ret = pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp); 2018 if (ret) { 2019 if (((*pagedeppp)->pd_state & ONWORKLIST) == 0 && bp) 2020 WORKLIST_INSERT(&bp->b_dep, &(*pagedeppp)->pd_list); 2021 return (1); 2022 } 2023 if ((flags & DEPALLOC) == 0) 2024 return (0); 2025 FREE_LOCK(&lk); 2026 pagedep = malloc(sizeof(struct pagedep), 2027 M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO); 2028 workitem_alloc(&pagedep->pd_list, D_PAGEDEP, mp); 2029 ACQUIRE_LOCK(&lk); 2030 ret = pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp); 2031 if (*pagedeppp) { 2032 /* 2033 * This should never happen since we only create pagedeps 2034 * with the vnode lock held. Could be an assert. 2035 */ 2036 WORKITEM_FREE(pagedep, D_PAGEDEP); 2037 return (ret); 2038 } 2039 pagedep->pd_ino = ino; 2040 pagedep->pd_lbn = lbn; 2041 LIST_INIT(&pagedep->pd_dirremhd); 2042 LIST_INIT(&pagedep->pd_pendinghd); 2043 for (i = 0; i < DAHASHSZ; i++) 2044 LIST_INIT(&pagedep->pd_diraddhd[i]); 2045 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 2046 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2047 *pagedeppp = pagedep; 2048 return (0); 2049} 2050 2051/* 2052 * Structures and routines associated with inodedep caching. 2053 */ 2054LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl; 2055static u_long inodedep_hash; /* size of hash table - 1 */ 2056#define INODEDEP_HASH(fs, inum) \ 2057 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash]) 2058 2059static int 2060inodedep_find(inodedephd, fs, inum, inodedeppp) 2061 struct inodedep_hashhead *inodedephd; 2062 struct fs *fs; 2063 ino_t inum; 2064 struct inodedep **inodedeppp; 2065{ 2066 struct inodedep *inodedep; 2067 2068 LIST_FOREACH(inodedep, inodedephd, id_hash) 2069 if (inum == inodedep->id_ino && fs == inodedep->id_fs) 2070 break; 2071 if (inodedep) { 2072 *inodedeppp = inodedep; 2073 return (1); 2074 } 2075 *inodedeppp = NULL; 2076 2077 return (0); 2078} 2079/* 2080 * Look up an inodedep. Return 1 if found, 0 if not found. 2081 * If not found, allocate if DEPALLOC flag is passed. 2082 * Found or allocated entry is returned in inodedeppp. 2083 * This routine must be called with splbio interrupts blocked. 2084 */ 2085static int 2086inodedep_lookup(mp, inum, flags, inodedeppp) 2087 struct mount *mp; 2088 ino_t inum; 2089 int flags; 2090 struct inodedep **inodedeppp; 2091{ 2092 struct inodedep *inodedep; 2093 struct inodedep_hashhead *inodedephd; 2094 struct fs *fs; 2095 2096 mtx_assert(&lk, MA_OWNED); 2097 fs = VFSTOUFS(mp)->um_fs; 2098 inodedephd = INODEDEP_HASH(fs, inum); 2099 2100 if (inodedep_find(inodedephd, fs, inum, inodedeppp)) 2101 return (1); 2102 if ((flags & DEPALLOC) == 0) 2103 return (0); 2104 /* 2105 * If we are over our limit, try to improve the situation. 2106 */ 2107 if (dep_current[D_INODEDEP] > max_softdeps && (flags & NODELAY) == 0) 2108 request_cleanup(mp, FLUSH_INODES); 2109 FREE_LOCK(&lk); 2110 inodedep = malloc(sizeof(struct inodedep), 2111 M_INODEDEP, M_SOFTDEP_FLAGS); 2112 workitem_alloc(&inodedep->id_list, D_INODEDEP, mp); 2113 ACQUIRE_LOCK(&lk); 2114 if (inodedep_find(inodedephd, fs, inum, inodedeppp)) { 2115 WORKITEM_FREE(inodedep, D_INODEDEP); 2116 return (1); 2117 } 2118 inodedep->id_fs = fs; 2119 inodedep->id_ino = inum; 2120 inodedep->id_state = ALLCOMPLETE; 2121 inodedep->id_nlinkdelta = 0; 2122 inodedep->id_savedino1 = NULL; 2123 inodedep->id_savedsize = -1; 2124 inodedep->id_savedextsize = -1; 2125 inodedep->id_savednlink = -1; 2126 inodedep->id_bmsafemap = NULL; 2127 inodedep->id_mkdiradd = NULL; 2128 LIST_INIT(&inodedep->id_dirremhd); 2129 LIST_INIT(&inodedep->id_pendinghd); 2130 LIST_INIT(&inodedep->id_inowait); 2131 LIST_INIT(&inodedep->id_bufwait); 2132 TAILQ_INIT(&inodedep->id_inoreflst); 2133 TAILQ_INIT(&inodedep->id_inoupdt); 2134 TAILQ_INIT(&inodedep->id_newinoupdt); 2135 TAILQ_INIT(&inodedep->id_extupdt); 2136 TAILQ_INIT(&inodedep->id_newextupdt); 2137 TAILQ_INIT(&inodedep->id_freeblklst); 2138 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 2139 *inodedeppp = inodedep; 2140 return (0); 2141} 2142 2143/* 2144 * Structures and routines associated with newblk caching. 2145 */ 2146LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl; 2147u_long newblk_hash; /* size of hash table - 1 */ 2148#define NEWBLK_HASH(fs, inum) \ 2149 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash]) 2150 2151static int 2152newblk_find(newblkhd, mp, newblkno, flags, newblkpp) 2153 struct newblk_hashhead *newblkhd; 2154 struct mount *mp; 2155 ufs2_daddr_t newblkno; 2156 int flags; 2157 struct newblk **newblkpp; 2158{ 2159 struct newblk *newblk; 2160 2161 LIST_FOREACH(newblk, newblkhd, nb_hash) { 2162 if (newblkno != newblk->nb_newblkno) 2163 continue; 2164 if (mp != newblk->nb_list.wk_mp) 2165 continue; 2166 /* 2167 * If we're creating a new dependency don't match those that 2168 * have already been converted to allocdirects. This is for 2169 * a frag extend. 2170 */ 2171 if ((flags & DEPALLOC) && newblk->nb_list.wk_type != D_NEWBLK) 2172 continue; 2173 break; 2174 } 2175 if (newblk) { 2176 *newblkpp = newblk; 2177 return (1); 2178 } 2179 *newblkpp = NULL; 2180 return (0); 2181} 2182 2183/* 2184 * Look up a newblk. Return 1 if found, 0 if not found. 2185 * If not found, allocate if DEPALLOC flag is passed. 2186 * Found or allocated entry is returned in newblkpp. 2187 */ 2188static int 2189newblk_lookup(mp, newblkno, flags, newblkpp) 2190 struct mount *mp; 2191 ufs2_daddr_t newblkno; 2192 int flags; 2193 struct newblk **newblkpp; 2194{ 2195 struct newblk *newblk; 2196 struct newblk_hashhead *newblkhd; 2197 2198 newblkhd = NEWBLK_HASH(VFSTOUFS(mp)->um_fs, newblkno); 2199 if (newblk_find(newblkhd, mp, newblkno, flags, newblkpp)) 2200 return (1); 2201 if ((flags & DEPALLOC) == 0) 2202 return (0); 2203 FREE_LOCK(&lk); 2204 newblk = malloc(sizeof(union allblk), M_NEWBLK, 2205 M_SOFTDEP_FLAGS | M_ZERO); 2206 workitem_alloc(&newblk->nb_list, D_NEWBLK, mp); 2207 ACQUIRE_LOCK(&lk); 2208 if (newblk_find(newblkhd, mp, newblkno, flags, newblkpp)) { 2209 WORKITEM_FREE(newblk, D_NEWBLK); 2210 return (1); 2211 } 2212 newblk->nb_freefrag = NULL; 2213 LIST_INIT(&newblk->nb_indirdeps); 2214 LIST_INIT(&newblk->nb_newdirblk); 2215 LIST_INIT(&newblk->nb_jwork); 2216 newblk->nb_state = ATTACHED; 2217 newblk->nb_newblkno = newblkno; 2218 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 2219 *newblkpp = newblk; 2220 return (0); 2221} 2222 2223/* 2224 * Structures and routines associated with freed indirect block caching. 2225 */ 2226struct freeworklst *indir_hashtbl; 2227u_long indir_hash; /* size of hash table - 1 */ 2228#define INDIR_HASH(mp, blkno) \ 2229 (&indir_hashtbl[((((register_t)(mp)) >> 13) + (blkno)) & indir_hash]) 2230 2231/* 2232 * Lookup an indirect block in the indir hash table. The freework is 2233 * removed and potentially freed. The caller must do a blocking journal 2234 * write before writing to the blkno. 2235 */ 2236static int 2237indirblk_lookup(mp, blkno) 2238 struct mount *mp; 2239 ufs2_daddr_t blkno; 2240{ 2241 struct freework *freework; 2242 struct freeworklst *wkhd; 2243 2244 wkhd = INDIR_HASH(mp, blkno); 2245 TAILQ_FOREACH(freework, wkhd, fw_next) { 2246 if (freework->fw_blkno != blkno) 2247 continue; 2248 if (freework->fw_list.wk_mp != mp) 2249 continue; 2250 indirblk_remove(freework); 2251 return (1); 2252 } 2253 return (0); 2254} 2255 2256/* 2257 * Insert an indirect block represented by freework into the indirblk 2258 * hash table so that it may prevent the block from being re-used prior 2259 * to the journal being written. 2260 */ 2261static void 2262indirblk_insert(freework) 2263 struct freework *freework; 2264{ 2265 struct freeblks *freeblks; 2266 struct jsegdep *jsegdep; 2267 struct worklist *wk; 2268 2269 freeblks = freework->fw_freeblks; 2270 LIST_FOREACH(wk, &freeblks->fb_jwork, wk_list) 2271 if (wk->wk_type == D_JSEGDEP) 2272 break; 2273 if (wk == NULL) 2274 return; 2275 2276 jsegdep = WK_JSEGDEP(wk); 2277 LIST_INSERT_HEAD(&jsegdep->jd_seg->js_indirs, freework, fw_segs); 2278 TAILQ_INSERT_HEAD(INDIR_HASH(freework->fw_list.wk_mp, 2279 freework->fw_blkno), freework, fw_next); 2280 freework->fw_state &= ~DEPCOMPLETE; 2281} 2282 2283static void 2284indirblk_remove(freework) 2285 struct freework *freework; 2286{ 2287 2288 LIST_REMOVE(freework, fw_segs); 2289 TAILQ_REMOVE(INDIR_HASH(freework->fw_list.wk_mp, 2290 freework->fw_blkno), freework, fw_next); 2291 freework->fw_state |= DEPCOMPLETE; 2292 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE) 2293 WORKITEM_FREE(freework, D_FREEWORK); 2294} 2295 2296/* 2297 * Executed during filesystem system initialization before 2298 * mounting any filesystems. 2299 */ 2300void 2301softdep_initialize() 2302{ 2303 int i; 2304 2305 LIST_INIT(&mkdirlisthd); 2306 max_softdeps = desiredvnodes * 4; 2307 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, &pagedep_hash); 2308 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash); 2309 newblk_hashtbl = hashinit(desiredvnodes / 5, M_NEWBLK, &newblk_hash); 2310 bmsafemap_hashtbl = hashinit(1024, M_BMSAFEMAP, &bmsafemap_hash); 2311 i = 1 << (ffs(desiredvnodes / 10) - 1); 2312 indir_hashtbl = malloc(i * sizeof(indir_hashtbl[0]), M_FREEWORK, 2313 M_WAITOK); 2314 indir_hash = i - 1; 2315 for (i = 0; i <= indir_hash; i++) 2316 TAILQ_INIT(&indir_hashtbl[i]); 2317 2318 /* initialise bioops hack */ 2319 bioops.io_start = softdep_disk_io_initiation; 2320 bioops.io_complete = softdep_disk_write_complete; 2321 bioops.io_deallocate = softdep_deallocate_dependencies; 2322 bioops.io_countdeps = softdep_count_dependencies; 2323 2324 /* Initialize the callout with an mtx. */ 2325 callout_init_mtx(&softdep_callout, &lk, 0); 2326} 2327 2328/* 2329 * Executed after all filesystems have been unmounted during 2330 * filesystem module unload. 2331 */ 2332void 2333softdep_uninitialize() 2334{ 2335 2336 callout_drain(&softdep_callout); 2337 hashdestroy(pagedep_hashtbl, M_PAGEDEP, pagedep_hash); 2338 hashdestroy(inodedep_hashtbl, M_INODEDEP, inodedep_hash); 2339 hashdestroy(newblk_hashtbl, M_NEWBLK, newblk_hash); 2340 hashdestroy(bmsafemap_hashtbl, M_BMSAFEMAP, bmsafemap_hash); 2341 free(indir_hashtbl, M_FREEWORK); 2342} 2343 2344/* 2345 * Called at mount time to notify the dependency code that a 2346 * filesystem wishes to use it. 2347 */ 2348int 2349softdep_mount(devvp, mp, fs, cred) 2350 struct vnode *devvp; 2351 struct mount *mp; 2352 struct fs *fs; 2353 struct ucred *cred; 2354{ 2355 struct csum_total cstotal; 2356 struct ufsmount *ump; 2357 struct cg *cgp; 2358 struct buf *bp; 2359 int error, cyl; 2360 2361 MNT_ILOCK(mp); 2362 mp->mnt_flag = (mp->mnt_flag & ~MNT_ASYNC) | MNT_SOFTDEP; 2363 if ((mp->mnt_kern_flag & MNTK_SOFTDEP) == 0) { 2364 mp->mnt_kern_flag = (mp->mnt_kern_flag & ~MNTK_ASYNC) | 2365 MNTK_SOFTDEP | MNTK_NOASYNC; 2366 } 2367 MNT_IUNLOCK(mp); 2368 ump = VFSTOUFS(mp); 2369 LIST_INIT(&ump->softdep_workitem_pending); 2370 LIST_INIT(&ump->softdep_journal_pending); 2371 TAILQ_INIT(&ump->softdep_unlinked); 2372 LIST_INIT(&ump->softdep_dirtycg); 2373 ump->softdep_worklist_tail = NULL; 2374 ump->softdep_on_worklist = 0; 2375 ump->softdep_deps = 0; 2376 if ((fs->fs_flags & FS_SUJ) && 2377 (error = journal_mount(mp, fs, cred)) != 0) { 2378 printf("Failed to start journal: %d\n", error); 2379 return (error); 2380 } 2381 /* 2382 * When doing soft updates, the counters in the 2383 * superblock may have gotten out of sync. Recomputation 2384 * can take a long time and can be deferred for background 2385 * fsck. However, the old behavior of scanning the cylinder 2386 * groups and recalculating them at mount time is available 2387 * by setting vfs.ffs.compute_summary_at_mount to one. 2388 */ 2389 if (compute_summary_at_mount == 0 || fs->fs_clean != 0) 2390 return (0); 2391 bzero(&cstotal, sizeof cstotal); 2392 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 2393 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 2394 fs->fs_cgsize, cred, &bp)) != 0) { 2395 brelse(bp); 2396 return (error); 2397 } 2398 cgp = (struct cg *)bp->b_data; 2399 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 2400 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 2401 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 2402 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 2403 fs->fs_cs(fs, cyl) = cgp->cg_cs; 2404 brelse(bp); 2405 } 2406#ifdef DEBUG 2407 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 2408 printf("%s: superblock summary recomputed\n", fs->fs_fsmnt); 2409#endif 2410 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 2411 return (0); 2412} 2413 2414void 2415softdep_unmount(mp) 2416 struct mount *mp; 2417{ 2418 2419 MNT_ILOCK(mp); 2420 mp->mnt_flag &= ~MNT_SOFTDEP; 2421 if (MOUNTEDSUJ(mp) == 0) { 2422 MNT_IUNLOCK(mp); 2423 return; 2424 } 2425 mp->mnt_flag &= ~MNT_SUJ; 2426 MNT_IUNLOCK(mp); 2427 journal_unmount(mp); 2428} 2429 2430struct jblocks { 2431 struct jseglst jb_segs; /* TAILQ of current segments. */ 2432 struct jseg *jb_writeseg; /* Next write to complete. */ 2433 struct jseg *jb_oldestseg; /* Oldest segment with valid entries. */ 2434 struct jextent *jb_extent; /* Extent array. */ 2435 uint64_t jb_nextseq; /* Next sequence number. */ 2436 uint64_t jb_oldestwrseq; /* Oldest written sequence number. */ 2437 uint8_t jb_needseg; /* Need a forced segment. */ 2438 uint8_t jb_suspended; /* Did journal suspend writes? */ 2439 int jb_avail; /* Available extents. */ 2440 int jb_used; /* Last used extent. */ 2441 int jb_head; /* Allocator head. */ 2442 int jb_off; /* Allocator extent offset. */ 2443 int jb_blocks; /* Total disk blocks covered. */ 2444 int jb_free; /* Total disk blocks free. */ 2445 int jb_min; /* Minimum free space. */ 2446 int jb_low; /* Low on space. */ 2447 int jb_age; /* Insertion time of oldest rec. */ 2448}; 2449 2450struct jextent { 2451 ufs2_daddr_t je_daddr; /* Disk block address. */ 2452 int je_blocks; /* Disk block count. */ 2453}; 2454 2455static struct jblocks * 2456jblocks_create(void) 2457{ 2458 struct jblocks *jblocks; 2459 2460 jblocks = malloc(sizeof(*jblocks), M_JBLOCKS, M_WAITOK | M_ZERO); 2461 TAILQ_INIT(&jblocks->jb_segs); 2462 jblocks->jb_avail = 10; 2463 jblocks->jb_extent = malloc(sizeof(struct jextent) * jblocks->jb_avail, 2464 M_JBLOCKS, M_WAITOK | M_ZERO); 2465 2466 return (jblocks); 2467} 2468 2469static ufs2_daddr_t 2470jblocks_alloc(jblocks, bytes, actual) 2471 struct jblocks *jblocks; 2472 int bytes; 2473 int *actual; 2474{ 2475 ufs2_daddr_t daddr; 2476 struct jextent *jext; 2477 int freecnt; 2478 int blocks; 2479 2480 blocks = bytes / DEV_BSIZE; 2481 jext = &jblocks->jb_extent[jblocks->jb_head]; 2482 freecnt = jext->je_blocks - jblocks->jb_off; 2483 if (freecnt == 0) { 2484 jblocks->jb_off = 0; 2485 if (++jblocks->jb_head > jblocks->jb_used) 2486 jblocks->jb_head = 0; 2487 jext = &jblocks->jb_extent[jblocks->jb_head]; 2488 freecnt = jext->je_blocks; 2489 } 2490 if (freecnt > blocks) 2491 freecnt = blocks; 2492 *actual = freecnt * DEV_BSIZE; 2493 daddr = jext->je_daddr + jblocks->jb_off; 2494 jblocks->jb_off += freecnt; 2495 jblocks->jb_free -= freecnt; 2496 2497 return (daddr); 2498} 2499 2500static void 2501jblocks_free(jblocks, mp, bytes) 2502 struct jblocks *jblocks; 2503 struct mount *mp; 2504 int bytes; 2505{ 2506 2507 jblocks->jb_free += bytes / DEV_BSIZE; 2508 if (jblocks->jb_suspended) 2509 worklist_speedup(); 2510 wakeup(jblocks); 2511} 2512 2513static void 2514jblocks_destroy(jblocks) 2515 struct jblocks *jblocks; 2516{ 2517 2518 if (jblocks->jb_extent) 2519 free(jblocks->jb_extent, M_JBLOCKS); 2520 free(jblocks, M_JBLOCKS); 2521} 2522 2523static void 2524jblocks_add(jblocks, daddr, blocks) 2525 struct jblocks *jblocks; 2526 ufs2_daddr_t daddr; 2527 int blocks; 2528{ 2529 struct jextent *jext; 2530 2531 jblocks->jb_blocks += blocks; 2532 jblocks->jb_free += blocks; 2533 jext = &jblocks->jb_extent[jblocks->jb_used]; 2534 /* Adding the first block. */ 2535 if (jext->je_daddr == 0) { 2536 jext->je_daddr = daddr; 2537 jext->je_blocks = blocks; 2538 return; 2539 } 2540 /* Extending the last extent. */ 2541 if (jext->je_daddr + jext->je_blocks == daddr) { 2542 jext->je_blocks += blocks; 2543 return; 2544 } 2545 /* Adding a new extent. */ 2546 if (++jblocks->jb_used == jblocks->jb_avail) { 2547 jblocks->jb_avail *= 2; 2548 jext = malloc(sizeof(struct jextent) * jblocks->jb_avail, 2549 M_JBLOCKS, M_WAITOK | M_ZERO); 2550 memcpy(jext, jblocks->jb_extent, 2551 sizeof(struct jextent) * jblocks->jb_used); 2552 free(jblocks->jb_extent, M_JBLOCKS); 2553 jblocks->jb_extent = jext; 2554 } 2555 jext = &jblocks->jb_extent[jblocks->jb_used]; 2556 jext->je_daddr = daddr; 2557 jext->je_blocks = blocks; 2558 return; 2559} 2560 2561int 2562softdep_journal_lookup(mp, vpp) 2563 struct mount *mp; 2564 struct vnode **vpp; 2565{ 2566 struct componentname cnp; 2567 struct vnode *dvp; 2568 ino_t sujournal; 2569 int error; 2570 2571 error = VFS_VGET(mp, ROOTINO, LK_EXCLUSIVE, &dvp); 2572 if (error) 2573 return (error); 2574 bzero(&cnp, sizeof(cnp)); 2575 cnp.cn_nameiop = LOOKUP; 2576 cnp.cn_flags = ISLASTCN; 2577 cnp.cn_thread = curthread; 2578 cnp.cn_cred = curthread->td_ucred; 2579 cnp.cn_pnbuf = SUJ_FILE; 2580 cnp.cn_nameptr = SUJ_FILE; 2581 cnp.cn_namelen = strlen(SUJ_FILE); 2582 error = ufs_lookup_ino(dvp, NULL, &cnp, &sujournal); 2583 vput(dvp); 2584 if (error != 0) 2585 return (error); 2586 error = VFS_VGET(mp, sujournal, LK_EXCLUSIVE, vpp); 2587 return (error); 2588} 2589 2590/* 2591 * Open and verify the journal file. 2592 */ 2593static int 2594journal_mount(mp, fs, cred) 2595 struct mount *mp; 2596 struct fs *fs; 2597 struct ucred *cred; 2598{ 2599 struct jblocks *jblocks; 2600 struct vnode *vp; 2601 struct inode *ip; 2602 ufs2_daddr_t blkno; 2603 int bcount; 2604 int error; 2605 int i; 2606 2607 error = softdep_journal_lookup(mp, &vp); 2608 if (error != 0) { 2609 printf("Failed to find journal. Use tunefs to create one\n"); 2610 return (error); 2611 } 2612 ip = VTOI(vp); 2613 if (ip->i_size < SUJ_MIN) { 2614 error = ENOSPC; 2615 goto out; 2616 } 2617 bcount = lblkno(fs, ip->i_size); /* Only use whole blocks. */ 2618 jblocks = jblocks_create(); 2619 for (i = 0; i < bcount; i++) { 2620 error = ufs_bmaparray(vp, i, &blkno, NULL, NULL, NULL); 2621 if (error) 2622 break; 2623 jblocks_add(jblocks, blkno, fsbtodb(fs, fs->fs_frag)); 2624 } 2625 if (error) { 2626 jblocks_destroy(jblocks); 2627 goto out; 2628 } 2629 jblocks->jb_low = jblocks->jb_free / 3; /* Reserve 33%. */ 2630 jblocks->jb_min = jblocks->jb_free / 10; /* Suspend at 10%. */ 2631 VFSTOUFS(mp)->softdep_jblocks = jblocks; 2632out: 2633 if (error == 0) { 2634 MNT_ILOCK(mp); 2635 mp->mnt_flag |= MNT_SUJ; 2636 mp->mnt_flag &= ~MNT_SOFTDEP; 2637 MNT_IUNLOCK(mp); 2638 /* 2639 * Only validate the journal contents if the 2640 * filesystem is clean, otherwise we write the logs 2641 * but they'll never be used. If the filesystem was 2642 * still dirty when we mounted it the journal is 2643 * invalid and a new journal can only be valid if it 2644 * starts from a clean mount. 2645 */ 2646 if (fs->fs_clean) { 2647 DIP_SET(ip, i_modrev, fs->fs_mtime); 2648 ip->i_flags |= IN_MODIFIED; 2649 ffs_update(vp, 1); 2650 } 2651 } 2652 vput(vp); 2653 return (error); 2654} 2655 2656static void 2657journal_unmount(mp) 2658 struct mount *mp; 2659{ 2660 struct ufsmount *ump; 2661 2662 ump = VFSTOUFS(mp); 2663 if (ump->softdep_jblocks) 2664 jblocks_destroy(ump->softdep_jblocks); 2665 ump->softdep_jblocks = NULL; 2666} 2667 2668/* 2669 * Called when a journal record is ready to be written. Space is allocated 2670 * and the journal entry is created when the journal is flushed to stable 2671 * store. 2672 */ 2673static void 2674add_to_journal(wk) 2675 struct worklist *wk; 2676{ 2677 struct ufsmount *ump; 2678 2679 mtx_assert(&lk, MA_OWNED); 2680 ump = VFSTOUFS(wk->wk_mp); 2681 if (wk->wk_state & ONWORKLIST) 2682 panic("add_to_journal: %s(0x%X) already on list", 2683 TYPENAME(wk->wk_type), wk->wk_state); 2684 wk->wk_state |= ONWORKLIST | DEPCOMPLETE; 2685 if (LIST_EMPTY(&ump->softdep_journal_pending)) { 2686 ump->softdep_jblocks->jb_age = ticks; 2687 LIST_INSERT_HEAD(&ump->softdep_journal_pending, wk, wk_list); 2688 } else 2689 LIST_INSERT_AFTER(ump->softdep_journal_tail, wk, wk_list); 2690 ump->softdep_journal_tail = wk; 2691 ump->softdep_on_journal += 1; 2692} 2693 2694/* 2695 * Remove an arbitrary item for the journal worklist maintain the tail 2696 * pointer. This happens when a new operation obviates the need to 2697 * journal an old operation. 2698 */ 2699static void 2700remove_from_journal(wk) 2701 struct worklist *wk; 2702{ 2703 struct ufsmount *ump; 2704 2705 mtx_assert(&lk, MA_OWNED); 2706 ump = VFSTOUFS(wk->wk_mp); 2707#ifdef SUJ_DEBUG 2708 { 2709 struct worklist *wkn; 2710 2711 LIST_FOREACH(wkn, &ump->softdep_journal_pending, wk_list) 2712 if (wkn == wk) 2713 break; 2714 if (wkn == NULL) 2715 panic("remove_from_journal: %p is not in journal", wk); 2716 } 2717#endif 2718 /* 2719 * We emulate a TAILQ to save space in most structures which do not 2720 * require TAILQ semantics. Here we must update the tail position 2721 * when removing the tail which is not the final entry. This works 2722 * only if the worklist linkage are at the beginning of the structure. 2723 */ 2724 if (ump->softdep_journal_tail == wk) 2725 ump->softdep_journal_tail = 2726 (struct worklist *)wk->wk_list.le_prev; 2727 2728 WORKLIST_REMOVE(wk); 2729 ump->softdep_on_journal -= 1; 2730} 2731 2732/* 2733 * Check for journal space as well as dependency limits so the prelink 2734 * code can throttle both journaled and non-journaled filesystems. 2735 * Threshold is 0 for low and 1 for min. 2736 */ 2737static int 2738journal_space(ump, thresh) 2739 struct ufsmount *ump; 2740 int thresh; 2741{ 2742 struct jblocks *jblocks; 2743 int avail; 2744 2745 jblocks = ump->softdep_jblocks; 2746 if (jblocks == NULL) 2747 return (1); 2748 /* 2749 * We use a tighter restriction here to prevent request_cleanup() 2750 * running in threads from running into locks we currently hold. 2751 */ 2752 if (dep_current[D_INODEDEP] > (max_softdeps / 10) * 9) 2753 return (0); 2754 if (thresh) 2755 thresh = jblocks->jb_min; 2756 else 2757 thresh = jblocks->jb_low; 2758 avail = (ump->softdep_on_journal * JREC_SIZE) / DEV_BSIZE; 2759 avail = jblocks->jb_free - avail; 2760 2761 return (avail > thresh); 2762} 2763 2764static void 2765journal_suspend(ump) 2766 struct ufsmount *ump; 2767{ 2768 struct jblocks *jblocks; 2769 struct mount *mp; 2770 2771 mp = UFSTOVFS(ump); 2772 jblocks = ump->softdep_jblocks; 2773 MNT_ILOCK(mp); 2774 if ((mp->mnt_kern_flag & MNTK_SUSPEND) == 0) { 2775 stat_journal_min++; 2776 mp->mnt_kern_flag |= MNTK_SUSPEND; 2777 mp->mnt_susp_owner = FIRST_THREAD_IN_PROC(softdepproc); 2778 } 2779 jblocks->jb_suspended = 1; 2780 MNT_IUNLOCK(mp); 2781} 2782 2783static int 2784journal_unsuspend(struct ufsmount *ump) 2785{ 2786 struct jblocks *jblocks; 2787 struct mount *mp; 2788 2789 mp = UFSTOVFS(ump); 2790 jblocks = ump->softdep_jblocks; 2791 2792 if (jblocks != NULL && jblocks->jb_suspended && 2793 journal_space(ump, jblocks->jb_min)) { 2794 jblocks->jb_suspended = 0; 2795 FREE_LOCK(&lk); 2796 mp->mnt_susp_owner = curthread; 2797 vfs_write_resume(mp); 2798 ACQUIRE_LOCK(&lk); 2799 return (1); 2800 } 2801 return (0); 2802} 2803 2804/* 2805 * Called before any allocation function to be certain that there is 2806 * sufficient space in the journal prior to creating any new records. 2807 * Since in the case of block allocation we may have multiple locked 2808 * buffers at the time of the actual allocation we can not block 2809 * when the journal records are created. Doing so would create a deadlock 2810 * if any of these buffers needed to be flushed to reclaim space. Instead 2811 * we require a sufficiently large amount of available space such that 2812 * each thread in the system could have passed this allocation check and 2813 * still have sufficient free space. With 20% of a minimum journal size 2814 * of 1MB we have 6553 records available. 2815 */ 2816int 2817softdep_prealloc(vp, waitok) 2818 struct vnode *vp; 2819 int waitok; 2820{ 2821 struct ufsmount *ump; 2822 2823 /* 2824 * Nothing to do if we are not running journaled soft updates. 2825 * If we currently hold the snapshot lock, we must avoid handling 2826 * other resources that could cause deadlock. 2827 */ 2828 if (DOINGSUJ(vp) == 0 || IS_SNAPSHOT(VTOI(vp))) 2829 return (0); 2830 ump = VFSTOUFS(vp->v_mount); 2831 ACQUIRE_LOCK(&lk); 2832 if (journal_space(ump, 0)) { 2833 FREE_LOCK(&lk); 2834 return (0); 2835 } 2836 stat_journal_low++; 2837 FREE_LOCK(&lk); 2838 if (waitok == MNT_NOWAIT) 2839 return (ENOSPC); 2840 /* 2841 * Attempt to sync this vnode once to flush any journal 2842 * work attached to it. 2843 */ 2844 if ((curthread->td_pflags & TDP_COWINPROGRESS) == 0) 2845 ffs_syncvnode(vp, waitok, 0); 2846 ACQUIRE_LOCK(&lk); 2847 process_removes(vp); 2848 process_truncates(vp); 2849 if (journal_space(ump, 0) == 0) { 2850 softdep_speedup(); 2851 if (journal_space(ump, 1) == 0) 2852 journal_suspend(ump); 2853 } 2854 FREE_LOCK(&lk); 2855 2856 return (0); 2857} 2858 2859/* 2860 * Before adjusting a link count on a vnode verify that we have sufficient 2861 * journal space. If not, process operations that depend on the currently 2862 * locked pair of vnodes to try to flush space as the syncer, buf daemon, 2863 * and softdep flush threads can not acquire these locks to reclaim space. 2864 */ 2865static void 2866softdep_prelink(dvp, vp) 2867 struct vnode *dvp; 2868 struct vnode *vp; 2869{ 2870 struct ufsmount *ump; 2871 2872 ump = VFSTOUFS(dvp->v_mount); 2873 mtx_assert(&lk, MA_OWNED); 2874 /* 2875 * Nothing to do if we have sufficient journal space. 2876 * If we currently hold the snapshot lock, we must avoid 2877 * handling other resources that could cause deadlock. 2878 */ 2879 if (journal_space(ump, 0) || (vp && IS_SNAPSHOT(VTOI(vp)))) 2880 return; 2881 stat_journal_low++; 2882 FREE_LOCK(&lk); 2883 if (vp) 2884 ffs_syncvnode(vp, MNT_NOWAIT, 0); 2885 ffs_syncvnode(dvp, MNT_WAIT, 0); 2886 ACQUIRE_LOCK(&lk); 2887 /* Process vp before dvp as it may create .. removes. */ 2888 if (vp) { 2889 process_removes(vp); 2890 process_truncates(vp); 2891 } 2892 process_removes(dvp); 2893 process_truncates(dvp); 2894 softdep_speedup(); 2895 process_worklist_item(UFSTOVFS(ump), 2, LK_NOWAIT); 2896 if (journal_space(ump, 0) == 0) { 2897 softdep_speedup(); 2898 if (journal_space(ump, 1) == 0) 2899 journal_suspend(ump); 2900 } 2901} 2902 2903static void 2904jseg_write(ump, jseg, data) 2905 struct ufsmount *ump; 2906 struct jseg *jseg; 2907 uint8_t *data; 2908{ 2909 struct jsegrec *rec; 2910 2911 rec = (struct jsegrec *)data; 2912 rec->jsr_seq = jseg->js_seq; 2913 rec->jsr_oldest = jseg->js_oldseq; 2914 rec->jsr_cnt = jseg->js_cnt; 2915 rec->jsr_blocks = jseg->js_size / ump->um_devvp->v_bufobj.bo_bsize; 2916 rec->jsr_crc = 0; 2917 rec->jsr_time = ump->um_fs->fs_mtime; 2918} 2919 2920static inline void 2921inoref_write(inoref, jseg, rec) 2922 struct inoref *inoref; 2923 struct jseg *jseg; 2924 struct jrefrec *rec; 2925{ 2926 2927 inoref->if_jsegdep->jd_seg = jseg; 2928 rec->jr_ino = inoref->if_ino; 2929 rec->jr_parent = inoref->if_parent; 2930 rec->jr_nlink = inoref->if_nlink; 2931 rec->jr_mode = inoref->if_mode; 2932 rec->jr_diroff = inoref->if_diroff; 2933} 2934 2935static void 2936jaddref_write(jaddref, jseg, data) 2937 struct jaddref *jaddref; 2938 struct jseg *jseg; 2939 uint8_t *data; 2940{ 2941 struct jrefrec *rec; 2942 2943 rec = (struct jrefrec *)data; 2944 rec->jr_op = JOP_ADDREF; 2945 inoref_write(&jaddref->ja_ref, jseg, rec); 2946} 2947 2948static void 2949jremref_write(jremref, jseg, data) 2950 struct jremref *jremref; 2951 struct jseg *jseg; 2952 uint8_t *data; 2953{ 2954 struct jrefrec *rec; 2955 2956 rec = (struct jrefrec *)data; 2957 rec->jr_op = JOP_REMREF; 2958 inoref_write(&jremref->jr_ref, jseg, rec); 2959} 2960 2961static void 2962jmvref_write(jmvref, jseg, data) 2963 struct jmvref *jmvref; 2964 struct jseg *jseg; 2965 uint8_t *data; 2966{ 2967 struct jmvrec *rec; 2968 2969 rec = (struct jmvrec *)data; 2970 rec->jm_op = JOP_MVREF; 2971 rec->jm_ino = jmvref->jm_ino; 2972 rec->jm_parent = jmvref->jm_parent; 2973 rec->jm_oldoff = jmvref->jm_oldoff; 2974 rec->jm_newoff = jmvref->jm_newoff; 2975} 2976 2977static void 2978jnewblk_write(jnewblk, jseg, data) 2979 struct jnewblk *jnewblk; 2980 struct jseg *jseg; 2981 uint8_t *data; 2982{ 2983 struct jblkrec *rec; 2984 2985 jnewblk->jn_jsegdep->jd_seg = jseg; 2986 rec = (struct jblkrec *)data; 2987 rec->jb_op = JOP_NEWBLK; 2988 rec->jb_ino = jnewblk->jn_ino; 2989 rec->jb_blkno = jnewblk->jn_blkno; 2990 rec->jb_lbn = jnewblk->jn_lbn; 2991 rec->jb_frags = jnewblk->jn_frags; 2992 rec->jb_oldfrags = jnewblk->jn_oldfrags; 2993} 2994 2995static void 2996jfreeblk_write(jfreeblk, jseg, data) 2997 struct jfreeblk *jfreeblk; 2998 struct jseg *jseg; 2999 uint8_t *data; 3000{ 3001 struct jblkrec *rec; 3002 3003 jfreeblk->jf_dep.jb_jsegdep->jd_seg = jseg; 3004 rec = (struct jblkrec *)data; 3005 rec->jb_op = JOP_FREEBLK; 3006 rec->jb_ino = jfreeblk->jf_ino; 3007 rec->jb_blkno = jfreeblk->jf_blkno; 3008 rec->jb_lbn = jfreeblk->jf_lbn; 3009 rec->jb_frags = jfreeblk->jf_frags; 3010 rec->jb_oldfrags = 0; 3011} 3012 3013static void 3014jfreefrag_write(jfreefrag, jseg, data) 3015 struct jfreefrag *jfreefrag; 3016 struct jseg *jseg; 3017 uint8_t *data; 3018{ 3019 struct jblkrec *rec; 3020 3021 jfreefrag->fr_jsegdep->jd_seg = jseg; 3022 rec = (struct jblkrec *)data; 3023 rec->jb_op = JOP_FREEBLK; 3024 rec->jb_ino = jfreefrag->fr_ino; 3025 rec->jb_blkno = jfreefrag->fr_blkno; 3026 rec->jb_lbn = jfreefrag->fr_lbn; 3027 rec->jb_frags = jfreefrag->fr_frags; 3028 rec->jb_oldfrags = 0; 3029} 3030 3031static void 3032jtrunc_write(jtrunc, jseg, data) 3033 struct jtrunc *jtrunc; 3034 struct jseg *jseg; 3035 uint8_t *data; 3036{ 3037 struct jtrncrec *rec; 3038 3039 jtrunc->jt_dep.jb_jsegdep->jd_seg = jseg; 3040 rec = (struct jtrncrec *)data; 3041 rec->jt_op = JOP_TRUNC; 3042 rec->jt_ino = jtrunc->jt_ino; 3043 rec->jt_size = jtrunc->jt_size; 3044 rec->jt_extsize = jtrunc->jt_extsize; 3045} 3046 3047static void 3048jfsync_write(jfsync, jseg, data) 3049 struct jfsync *jfsync; 3050 struct jseg *jseg; 3051 uint8_t *data; 3052{ 3053 struct jtrncrec *rec; 3054 3055 rec = (struct jtrncrec *)data; 3056 rec->jt_op = JOP_SYNC; 3057 rec->jt_ino = jfsync->jfs_ino; 3058 rec->jt_size = jfsync->jfs_size; 3059 rec->jt_extsize = jfsync->jfs_extsize; 3060} 3061 3062static void 3063softdep_flushjournal(mp) 3064 struct mount *mp; 3065{ 3066 struct jblocks *jblocks; 3067 struct ufsmount *ump; 3068 3069 if (MOUNTEDSUJ(mp) == 0) 3070 return; 3071 ump = VFSTOUFS(mp); 3072 jblocks = ump->softdep_jblocks; 3073 ACQUIRE_LOCK(&lk); 3074 while (ump->softdep_on_journal) { 3075 jblocks->jb_needseg = 1; 3076 softdep_process_journal(mp, NULL, MNT_WAIT); 3077 } 3078 FREE_LOCK(&lk); 3079} 3080 3081/* 3082 * Flush some journal records to disk. 3083 */ 3084static void 3085softdep_process_journal(mp, needwk, flags) 3086 struct mount *mp; 3087 struct worklist *needwk; 3088 int flags; 3089{ 3090 struct jblocks *jblocks; 3091 struct ufsmount *ump; 3092 struct worklist *wk; 3093 struct jseg *jseg; 3094 struct buf *bp; 3095 uint8_t *data; 3096 struct fs *fs; 3097 int segwritten; 3098 int jrecmin; /* Minimum records per block. */ 3099 int jrecmax; /* Maximum records per block. */ 3100 int size; 3101 int cnt; 3102 int off; 3103 int devbsize; 3104 3105 if (MOUNTEDSUJ(mp) == 0) 3106 return; 3107 ump = VFSTOUFS(mp); 3108 fs = ump->um_fs; 3109 jblocks = ump->softdep_jblocks; 3110 devbsize = ump->um_devvp->v_bufobj.bo_bsize; 3111 /* 3112 * We write anywhere between a disk block and fs block. The upper 3113 * bound is picked to prevent buffer cache fragmentation and limit 3114 * processing time per I/O. 3115 */ 3116 jrecmin = (devbsize / JREC_SIZE) - 1; /* -1 for seg header */ 3117 jrecmax = (fs->fs_bsize / devbsize) * jrecmin; 3118 segwritten = 0; 3119 for (;;) { 3120 cnt = ump->softdep_on_journal; 3121 /* 3122 * Criteria for writing a segment: 3123 * 1) We have a full block. 3124 * 2) We're called from jwait() and haven't found the 3125 * journal item yet. 3126 * 3) Always write if needseg is set. 3127 * 4) If we are called from process_worklist and have 3128 * not yet written anything we write a partial block 3129 * to enforce a 1 second maximum latency on journal 3130 * entries. 3131 */ 3132 if (cnt < (jrecmax - 1) && needwk == NULL && 3133 jblocks->jb_needseg == 0 && (segwritten || cnt == 0)) 3134 break; 3135 cnt++; 3136 /* 3137 * Verify some free journal space. softdep_prealloc() should 3138 * guarantee that we don't run out so this is indicative of 3139 * a problem with the flow control. Try to recover 3140 * gracefully in any event. 3141 */ 3142 while (jblocks->jb_free == 0) { 3143 if (flags != MNT_WAIT) 3144 break; 3145 printf("softdep: Out of journal space!\n"); 3146 softdep_speedup(); 3147 msleep(jblocks, &lk, PRIBIO, "jblocks", hz); 3148 } 3149 FREE_LOCK(&lk); 3150 jseg = malloc(sizeof(*jseg), M_JSEG, M_SOFTDEP_FLAGS); 3151 workitem_alloc(&jseg->js_list, D_JSEG, mp); 3152 LIST_INIT(&jseg->js_entries); 3153 LIST_INIT(&jseg->js_indirs); 3154 jseg->js_state = ATTACHED; 3155 jseg->js_jblocks = jblocks; 3156 bp = geteblk(fs->fs_bsize, 0); 3157 ACQUIRE_LOCK(&lk); 3158 /* 3159 * If there was a race while we were allocating the block 3160 * and jseg the entry we care about was likely written. 3161 * We bail out in both the WAIT and NOWAIT case and assume 3162 * the caller will loop if the entry it cares about is 3163 * not written. 3164 */ 3165 cnt = ump->softdep_on_journal; 3166 if (cnt + jblocks->jb_needseg == 0 || jblocks->jb_free == 0) { 3167 bp->b_flags |= B_INVAL | B_NOCACHE; 3168 WORKITEM_FREE(jseg, D_JSEG); 3169 FREE_LOCK(&lk); 3170 brelse(bp); 3171 ACQUIRE_LOCK(&lk); 3172 break; 3173 } 3174 /* 3175 * Calculate the disk block size required for the available 3176 * records rounded to the min size. 3177 */ 3178 if (cnt == 0) 3179 size = devbsize; 3180 else if (cnt < jrecmax) 3181 size = howmany(cnt, jrecmin) * devbsize; 3182 else 3183 size = fs->fs_bsize; 3184 /* 3185 * Allocate a disk block for this journal data and account 3186 * for truncation of the requested size if enough contiguous 3187 * space was not available. 3188 */ 3189 bp->b_blkno = jblocks_alloc(jblocks, size, &size); 3190 bp->b_lblkno = bp->b_blkno; 3191 bp->b_offset = bp->b_blkno * DEV_BSIZE; 3192 bp->b_bcount = size; 3193 bp->b_bufobj = &ump->um_devvp->v_bufobj; 3194 bp->b_flags &= ~B_INVAL; 3195 bp->b_flags |= B_VALIDSUSPWRT | B_NOCOPY; 3196 /* 3197 * Initialize our jseg with cnt records. Assign the next 3198 * sequence number to it and link it in-order. 3199 */ 3200 cnt = MIN(cnt, (size / devbsize) * jrecmin); 3201 jseg->js_buf = bp; 3202 jseg->js_cnt = cnt; 3203 jseg->js_refs = cnt + 1; /* Self ref. */ 3204 jseg->js_size = size; 3205 jseg->js_seq = jblocks->jb_nextseq++; 3206 if (jblocks->jb_oldestseg == NULL) 3207 jblocks->jb_oldestseg = jseg; 3208 jseg->js_oldseq = jblocks->jb_oldestseg->js_seq; 3209 TAILQ_INSERT_TAIL(&jblocks->jb_segs, jseg, js_next); 3210 if (jblocks->jb_writeseg == NULL) 3211 jblocks->jb_writeseg = jseg; 3212 /* 3213 * Start filling in records from the pending list. 3214 */ 3215 data = bp->b_data; 3216 off = 0; 3217 while ((wk = LIST_FIRST(&ump->softdep_journal_pending)) 3218 != NULL) { 3219 if (cnt == 0) 3220 break; 3221 /* Place a segment header on every device block. */ 3222 if ((off % devbsize) == 0) { 3223 jseg_write(ump, jseg, data); 3224 off += JREC_SIZE; 3225 data = bp->b_data + off; 3226 } 3227 if (wk == needwk) 3228 needwk = NULL; 3229 remove_from_journal(wk); 3230 wk->wk_state |= INPROGRESS; 3231 WORKLIST_INSERT(&jseg->js_entries, wk); 3232 switch (wk->wk_type) { 3233 case D_JADDREF: 3234 jaddref_write(WK_JADDREF(wk), jseg, data); 3235 break; 3236 case D_JREMREF: 3237 jremref_write(WK_JREMREF(wk), jseg, data); 3238 break; 3239 case D_JMVREF: 3240 jmvref_write(WK_JMVREF(wk), jseg, data); 3241 break; 3242 case D_JNEWBLK: 3243 jnewblk_write(WK_JNEWBLK(wk), jseg, data); 3244 break; 3245 case D_JFREEBLK: 3246 jfreeblk_write(WK_JFREEBLK(wk), jseg, data); 3247 break; 3248 case D_JFREEFRAG: 3249 jfreefrag_write(WK_JFREEFRAG(wk), jseg, data); 3250 break; 3251 case D_JTRUNC: 3252 jtrunc_write(WK_JTRUNC(wk), jseg, data); 3253 break; 3254 case D_JFSYNC: 3255 jfsync_write(WK_JFSYNC(wk), jseg, data); 3256 break; 3257 default: 3258 panic("process_journal: Unknown type %s", 3259 TYPENAME(wk->wk_type)); 3260 /* NOTREACHED */ 3261 } 3262 off += JREC_SIZE; 3263 data = bp->b_data + off; 3264 cnt--; 3265 } 3266 /* 3267 * Write this one buffer and continue. 3268 */ 3269 segwritten = 1; 3270 jblocks->jb_needseg = 0; 3271 WORKLIST_INSERT(&bp->b_dep, &jseg->js_list); 3272 FREE_LOCK(&lk); 3273 BO_LOCK(bp->b_bufobj); 3274 bgetvp(ump->um_devvp, bp); 3275 BO_UNLOCK(bp->b_bufobj); 3276 /* 3277 * We only do the blocking wait once we find the journal 3278 * entry we're looking for. 3279 */ 3280 if (needwk == NULL && flags == MNT_WAIT) 3281 bwrite(bp); 3282 else 3283 bawrite(bp); 3284 ACQUIRE_LOCK(&lk); 3285 } 3286 /* 3287 * If we've suspended the filesystem because we ran out of journal 3288 * space either try to sync it here to make some progress or 3289 * unsuspend it if we already have. 3290 */ 3291 if (flags == 0 && jblocks->jb_suspended) { 3292 if (journal_unsuspend(ump)) 3293 return; 3294 FREE_LOCK(&lk); 3295 VFS_SYNC(mp, MNT_NOWAIT); 3296 ffs_sbupdate(ump, MNT_WAIT, 0); 3297 ACQUIRE_LOCK(&lk); 3298 } 3299} 3300 3301/* 3302 * Complete a jseg, allowing all dependencies awaiting journal writes 3303 * to proceed. Each journal dependency also attaches a jsegdep to dependent 3304 * structures so that the journal segment can be freed to reclaim space. 3305 */ 3306static void 3307complete_jseg(jseg) 3308 struct jseg *jseg; 3309{ 3310 struct worklist *wk; 3311 struct jmvref *jmvref; 3312 int waiting; 3313#ifdef INVARIANTS 3314 int i = 0; 3315#endif 3316 3317 while ((wk = LIST_FIRST(&jseg->js_entries)) != NULL) { 3318 WORKLIST_REMOVE(wk); 3319 waiting = wk->wk_state & IOWAITING; 3320 wk->wk_state &= ~(INPROGRESS | IOWAITING); 3321 wk->wk_state |= COMPLETE; 3322 KASSERT(i++ < jseg->js_cnt, 3323 ("handle_written_jseg: overflow %d >= %d", 3324 i - 1, jseg->js_cnt)); 3325 switch (wk->wk_type) { 3326 case D_JADDREF: 3327 handle_written_jaddref(WK_JADDREF(wk)); 3328 break; 3329 case D_JREMREF: 3330 handle_written_jremref(WK_JREMREF(wk)); 3331 break; 3332 case D_JMVREF: 3333 rele_jseg(jseg); /* No jsegdep. */ 3334 jmvref = WK_JMVREF(wk); 3335 LIST_REMOVE(jmvref, jm_deps); 3336 if ((jmvref->jm_pagedep->pd_state & ONWORKLIST) == 0) 3337 free_pagedep(jmvref->jm_pagedep); 3338 WORKITEM_FREE(jmvref, D_JMVREF); 3339 break; 3340 case D_JNEWBLK: 3341 handle_written_jnewblk(WK_JNEWBLK(wk)); 3342 break; 3343 case D_JFREEBLK: 3344 handle_written_jblkdep(&WK_JFREEBLK(wk)->jf_dep); 3345 break; 3346 case D_JTRUNC: 3347 handle_written_jblkdep(&WK_JTRUNC(wk)->jt_dep); 3348 break; 3349 case D_JFSYNC: 3350 rele_jseg(jseg); /* No jsegdep. */ 3351 WORKITEM_FREE(wk, D_JFSYNC); 3352 break; 3353 case D_JFREEFRAG: 3354 handle_written_jfreefrag(WK_JFREEFRAG(wk)); 3355 break; 3356 default: 3357 panic("handle_written_jseg: Unknown type %s", 3358 TYPENAME(wk->wk_type)); 3359 /* NOTREACHED */ 3360 } 3361 if (waiting) 3362 wakeup(wk); 3363 } 3364 /* Release the self reference so the structure may be freed. */ 3365 rele_jseg(jseg); 3366} 3367 3368/* 3369 * Mark a jseg as DEPCOMPLETE and throw away the buffer. Handle jseg 3370 * completions in order only. 3371 */ 3372static void 3373handle_written_jseg(jseg, bp) 3374 struct jseg *jseg; 3375 struct buf *bp; 3376{ 3377 struct jblocks *jblocks; 3378 struct jseg *jsegn; 3379 3380 if (jseg->js_refs == 0) 3381 panic("handle_written_jseg: No self-reference on %p", jseg); 3382 jseg->js_state |= DEPCOMPLETE; 3383 /* 3384 * We'll never need this buffer again, set flags so it will be 3385 * discarded. 3386 */ 3387 bp->b_flags |= B_INVAL | B_NOCACHE; 3388 jblocks = jseg->js_jblocks; 3389 /* 3390 * Don't allow out of order completions. If this isn't the first 3391 * block wait for it to write before we're done. 3392 */ 3393 if (jseg != jblocks->jb_writeseg) 3394 return; 3395 /* Iterate through available jsegs processing their entries. */ 3396 do { 3397 jblocks->jb_oldestwrseq = jseg->js_oldseq; 3398 jsegn = TAILQ_NEXT(jseg, js_next); 3399 complete_jseg(jseg); 3400 jseg = jsegn; 3401 } while (jseg && jseg->js_state & DEPCOMPLETE); 3402 jblocks->jb_writeseg = jseg; 3403 /* 3404 * Attempt to free jsegs now that oldestwrseq may have advanced. 3405 */ 3406 free_jsegs(jblocks); 3407} 3408 3409static inline struct jsegdep * 3410inoref_jseg(inoref) 3411 struct inoref *inoref; 3412{ 3413 struct jsegdep *jsegdep; 3414 3415 jsegdep = inoref->if_jsegdep; 3416 inoref->if_jsegdep = NULL; 3417 3418 return (jsegdep); 3419} 3420 3421/* 3422 * Called once a jremref has made it to stable store. The jremref is marked 3423 * complete and we attempt to free it. Any pagedeps writes sleeping waiting 3424 * for the jremref to complete will be awoken by free_jremref. 3425 */ 3426static void 3427handle_written_jremref(jremref) 3428 struct jremref *jremref; 3429{ 3430 struct inodedep *inodedep; 3431 struct jsegdep *jsegdep; 3432 struct dirrem *dirrem; 3433 3434 /* Grab the jsegdep. */ 3435 jsegdep = inoref_jseg(&jremref->jr_ref); 3436 /* 3437 * Remove us from the inoref list. 3438 */ 3439 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 3440 0, &inodedep) == 0) 3441 panic("handle_written_jremref: Lost inodedep"); 3442 TAILQ_REMOVE(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps); 3443 /* 3444 * Complete the dirrem. 3445 */ 3446 dirrem = jremref->jr_dirrem; 3447 jremref->jr_dirrem = NULL; 3448 LIST_REMOVE(jremref, jr_deps); 3449 jsegdep->jd_state |= jremref->jr_state & MKDIR_PARENT; 3450 jwork_insert(&dirrem->dm_jwork, jsegdep); 3451 if (LIST_EMPTY(&dirrem->dm_jremrefhd) && 3452 (dirrem->dm_state & COMPLETE) != 0) 3453 add_to_worklist(&dirrem->dm_list, 0); 3454 free_jremref(jremref); 3455} 3456 3457/* 3458 * Called once a jaddref has made it to stable store. The dependency is 3459 * marked complete and any dependent structures are added to the inode 3460 * bufwait list to be completed as soon as it is written. If a bitmap write 3461 * depends on this entry we move the inode into the inodedephd of the 3462 * bmsafemap dependency and attempt to remove the jaddref from the bmsafemap. 3463 */ 3464static void 3465handle_written_jaddref(jaddref) 3466 struct jaddref *jaddref; 3467{ 3468 struct jsegdep *jsegdep; 3469 struct inodedep *inodedep; 3470 struct diradd *diradd; 3471 struct mkdir *mkdir; 3472 3473 /* Grab the jsegdep. */ 3474 jsegdep = inoref_jseg(&jaddref->ja_ref); 3475 mkdir = NULL; 3476 diradd = NULL; 3477 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino, 3478 0, &inodedep) == 0) 3479 panic("handle_written_jaddref: Lost inodedep."); 3480 if (jaddref->ja_diradd == NULL) 3481 panic("handle_written_jaddref: No dependency"); 3482 if (jaddref->ja_diradd->da_list.wk_type == D_DIRADD) { 3483 diradd = jaddref->ja_diradd; 3484 WORKLIST_INSERT(&inodedep->id_bufwait, &diradd->da_list); 3485 } else if (jaddref->ja_state & MKDIR_PARENT) { 3486 mkdir = jaddref->ja_mkdir; 3487 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir->md_list); 3488 } else if (jaddref->ja_state & MKDIR_BODY) 3489 mkdir = jaddref->ja_mkdir; 3490 else 3491 panic("handle_written_jaddref: Unknown dependency %p", 3492 jaddref->ja_diradd); 3493 jaddref->ja_diradd = NULL; /* also clears ja_mkdir */ 3494 /* 3495 * Remove us from the inode list. 3496 */ 3497 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, if_deps); 3498 /* 3499 * The mkdir may be waiting on the jaddref to clear before freeing. 3500 */ 3501 if (mkdir) { 3502 KASSERT(mkdir->md_list.wk_type == D_MKDIR, 3503 ("handle_written_jaddref: Incorrect type for mkdir %s", 3504 TYPENAME(mkdir->md_list.wk_type))); 3505 mkdir->md_jaddref = NULL; 3506 diradd = mkdir->md_diradd; 3507 mkdir->md_state |= DEPCOMPLETE; 3508 complete_mkdir(mkdir); 3509 } 3510 jwork_insert(&diradd->da_jwork, jsegdep); 3511 if (jaddref->ja_state & NEWBLOCK) { 3512 inodedep->id_state |= ONDEPLIST; 3513 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_inodedephd, 3514 inodedep, id_deps); 3515 } 3516 free_jaddref(jaddref); 3517} 3518 3519/* 3520 * Called once a jnewblk journal is written. The allocdirect or allocindir 3521 * is placed in the bmsafemap to await notification of a written bitmap. If 3522 * the operation was canceled we add the segdep to the appropriate 3523 * dependency to free the journal space once the canceling operation 3524 * completes. 3525 */ 3526static void 3527handle_written_jnewblk(jnewblk) 3528 struct jnewblk *jnewblk; 3529{ 3530 struct bmsafemap *bmsafemap; 3531 struct freefrag *freefrag; 3532 struct freework *freework; 3533 struct jsegdep *jsegdep; 3534 struct newblk *newblk; 3535 3536 /* Grab the jsegdep. */ 3537 jsegdep = jnewblk->jn_jsegdep; 3538 jnewblk->jn_jsegdep = NULL; 3539 if (jnewblk->jn_dep == NULL) 3540 panic("handle_written_jnewblk: No dependency for the segdep."); 3541 switch (jnewblk->jn_dep->wk_type) { 3542 case D_NEWBLK: 3543 case D_ALLOCDIRECT: 3544 case D_ALLOCINDIR: 3545 /* 3546 * Add the written block to the bmsafemap so it can 3547 * be notified when the bitmap is on disk. 3548 */ 3549 newblk = WK_NEWBLK(jnewblk->jn_dep); 3550 newblk->nb_jnewblk = NULL; 3551 if ((newblk->nb_state & GOINGAWAY) == 0) { 3552 bmsafemap = newblk->nb_bmsafemap; 3553 newblk->nb_state |= ONDEPLIST; 3554 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, 3555 nb_deps); 3556 } 3557 jwork_insert(&newblk->nb_jwork, jsegdep); 3558 break; 3559 case D_FREEFRAG: 3560 /* 3561 * A newblock being removed by a freefrag when replaced by 3562 * frag extension. 3563 */ 3564 freefrag = WK_FREEFRAG(jnewblk->jn_dep); 3565 freefrag->ff_jdep = NULL; 3566 WORKLIST_INSERT(&freefrag->ff_jwork, &jsegdep->jd_list); 3567 break; 3568 case D_FREEWORK: 3569 /* 3570 * A direct block was removed by truncate. 3571 */ 3572 freework = WK_FREEWORK(jnewblk->jn_dep); 3573 freework->fw_jnewblk = NULL; 3574 WORKLIST_INSERT(&freework->fw_freeblks->fb_jwork, 3575 &jsegdep->jd_list); 3576 break; 3577 default: 3578 panic("handle_written_jnewblk: Unknown type %d.", 3579 jnewblk->jn_dep->wk_type); 3580 } 3581 jnewblk->jn_dep = NULL; 3582 free_jnewblk(jnewblk); 3583} 3584 3585/* 3586 * Cancel a jfreefrag that won't be needed, probably due to colliding with 3587 * an in-flight allocation that has not yet been committed. Divorce us 3588 * from the freefrag and mark it DEPCOMPLETE so that it may be added 3589 * to the worklist. 3590 */ 3591static void 3592cancel_jfreefrag(jfreefrag) 3593 struct jfreefrag *jfreefrag; 3594{ 3595 struct freefrag *freefrag; 3596 3597 if (jfreefrag->fr_jsegdep) { 3598 free_jsegdep(jfreefrag->fr_jsegdep); 3599 jfreefrag->fr_jsegdep = NULL; 3600 } 3601 freefrag = jfreefrag->fr_freefrag; 3602 jfreefrag->fr_freefrag = NULL; 3603 free_jfreefrag(jfreefrag); 3604 freefrag->ff_state |= DEPCOMPLETE; 3605} 3606 3607/* 3608 * Free a jfreefrag when the parent freefrag is rendered obsolete. 3609 */ 3610static void 3611free_jfreefrag(jfreefrag) 3612 struct jfreefrag *jfreefrag; 3613{ 3614 3615 if (jfreefrag->fr_state & INPROGRESS) 3616 WORKLIST_REMOVE(&jfreefrag->fr_list); 3617 else if (jfreefrag->fr_state & ONWORKLIST) 3618 remove_from_journal(&jfreefrag->fr_list); 3619 if (jfreefrag->fr_freefrag != NULL) 3620 panic("free_jfreefrag: Still attached to a freefrag."); 3621 WORKITEM_FREE(jfreefrag, D_JFREEFRAG); 3622} 3623 3624/* 3625 * Called when the journal write for a jfreefrag completes. The parent 3626 * freefrag is added to the worklist if this completes its dependencies. 3627 */ 3628static void 3629handle_written_jfreefrag(jfreefrag) 3630 struct jfreefrag *jfreefrag; 3631{ 3632 struct jsegdep *jsegdep; 3633 struct freefrag *freefrag; 3634 3635 /* Grab the jsegdep. */ 3636 jsegdep = jfreefrag->fr_jsegdep; 3637 jfreefrag->fr_jsegdep = NULL; 3638 freefrag = jfreefrag->fr_freefrag; 3639 if (freefrag == NULL) 3640 panic("handle_written_jfreefrag: No freefrag."); 3641 freefrag->ff_state |= DEPCOMPLETE; 3642 freefrag->ff_jdep = NULL; 3643 jwork_insert(&freefrag->ff_jwork, jsegdep); 3644 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE) 3645 add_to_worklist(&freefrag->ff_list, 0); 3646 jfreefrag->fr_freefrag = NULL; 3647 free_jfreefrag(jfreefrag); 3648} 3649 3650/* 3651 * Called when the journal write for a jfreeblk completes. The jfreeblk 3652 * is removed from the freeblks list of pending journal writes and the 3653 * jsegdep is moved to the freeblks jwork to be completed when all blocks 3654 * have been reclaimed. 3655 */ 3656static void 3657handle_written_jblkdep(jblkdep) 3658 struct jblkdep *jblkdep; 3659{ 3660 struct freeblks *freeblks; 3661 struct jsegdep *jsegdep; 3662 3663 /* Grab the jsegdep. */ 3664 jsegdep = jblkdep->jb_jsegdep; 3665 jblkdep->jb_jsegdep = NULL; 3666 freeblks = jblkdep->jb_freeblks; 3667 LIST_REMOVE(jblkdep, jb_deps); 3668 WORKLIST_INSERT(&freeblks->fb_jwork, &jsegdep->jd_list); 3669 /* 3670 * If the freeblks is all journaled, we can add it to the worklist. 3671 */ 3672 if (LIST_EMPTY(&freeblks->fb_jblkdephd) && 3673 (freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 3674 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 3675 3676 free_jblkdep(jblkdep); 3677} 3678 3679static struct jsegdep * 3680newjsegdep(struct worklist *wk) 3681{ 3682 struct jsegdep *jsegdep; 3683 3684 jsegdep = malloc(sizeof(*jsegdep), M_JSEGDEP, M_SOFTDEP_FLAGS); 3685 workitem_alloc(&jsegdep->jd_list, D_JSEGDEP, wk->wk_mp); 3686 jsegdep->jd_seg = NULL; 3687 3688 return (jsegdep); 3689} 3690 3691static struct jmvref * 3692newjmvref(dp, ino, oldoff, newoff) 3693 struct inode *dp; 3694 ino_t ino; 3695 off_t oldoff; 3696 off_t newoff; 3697{ 3698 struct jmvref *jmvref; 3699 3700 jmvref = malloc(sizeof(*jmvref), M_JMVREF, M_SOFTDEP_FLAGS); 3701 workitem_alloc(&jmvref->jm_list, D_JMVREF, UFSTOVFS(dp->i_ump)); 3702 jmvref->jm_list.wk_state = ATTACHED | DEPCOMPLETE; 3703 jmvref->jm_parent = dp->i_number; 3704 jmvref->jm_ino = ino; 3705 jmvref->jm_oldoff = oldoff; 3706 jmvref->jm_newoff = newoff; 3707 3708 return (jmvref); 3709} 3710 3711/* 3712 * Allocate a new jremref that tracks the removal of ip from dp with the 3713 * directory entry offset of diroff. Mark the entry as ATTACHED and 3714 * DEPCOMPLETE as we have all the information required for the journal write 3715 * and the directory has already been removed from the buffer. The caller 3716 * is responsible for linking the jremref into the pagedep and adding it 3717 * to the journal to write. The MKDIR_PARENT flag is set if we're doing 3718 * a DOTDOT addition so handle_workitem_remove() can properly assign 3719 * the jsegdep when we're done. 3720 */ 3721static struct jremref * 3722newjremref(struct dirrem *dirrem, struct inode *dp, struct inode *ip, 3723 off_t diroff, nlink_t nlink) 3724{ 3725 struct jremref *jremref; 3726 3727 jremref = malloc(sizeof(*jremref), M_JREMREF, M_SOFTDEP_FLAGS); 3728 workitem_alloc(&jremref->jr_list, D_JREMREF, UFSTOVFS(dp->i_ump)); 3729 jremref->jr_state = ATTACHED; 3730 newinoref(&jremref->jr_ref, ip->i_number, dp->i_number, diroff, 3731 nlink, ip->i_mode); 3732 jremref->jr_dirrem = dirrem; 3733 3734 return (jremref); 3735} 3736 3737static inline void 3738newinoref(struct inoref *inoref, ino_t ino, ino_t parent, off_t diroff, 3739 nlink_t nlink, uint16_t mode) 3740{ 3741 3742 inoref->if_jsegdep = newjsegdep(&inoref->if_list); 3743 inoref->if_diroff = diroff; 3744 inoref->if_ino = ino; 3745 inoref->if_parent = parent; 3746 inoref->if_nlink = nlink; 3747 inoref->if_mode = mode; 3748} 3749 3750/* 3751 * Allocate a new jaddref to track the addition of ino to dp at diroff. The 3752 * directory offset may not be known until later. The caller is responsible 3753 * adding the entry to the journal when this information is available. nlink 3754 * should be the link count prior to the addition and mode is only required 3755 * to have the correct FMT. 3756 */ 3757static struct jaddref * 3758newjaddref(struct inode *dp, ino_t ino, off_t diroff, int16_t nlink, 3759 uint16_t mode) 3760{ 3761 struct jaddref *jaddref; 3762 3763 jaddref = malloc(sizeof(*jaddref), M_JADDREF, M_SOFTDEP_FLAGS); 3764 workitem_alloc(&jaddref->ja_list, D_JADDREF, UFSTOVFS(dp->i_ump)); 3765 jaddref->ja_state = ATTACHED; 3766 jaddref->ja_mkdir = NULL; 3767 newinoref(&jaddref->ja_ref, ino, dp->i_number, diroff, nlink, mode); 3768 3769 return (jaddref); 3770} 3771 3772/* 3773 * Create a new free dependency for a freework. The caller is responsible 3774 * for adjusting the reference count when it has the lock held. The freedep 3775 * will track an outstanding bitmap write that will ultimately clear the 3776 * freework to continue. 3777 */ 3778static struct freedep * 3779newfreedep(struct freework *freework) 3780{ 3781 struct freedep *freedep; 3782 3783 freedep = malloc(sizeof(*freedep), M_FREEDEP, M_SOFTDEP_FLAGS); 3784 workitem_alloc(&freedep->fd_list, D_FREEDEP, freework->fw_list.wk_mp); 3785 freedep->fd_freework = freework; 3786 3787 return (freedep); 3788} 3789 3790/* 3791 * Free a freedep structure once the buffer it is linked to is written. If 3792 * this is the last reference to the freework schedule it for completion. 3793 */ 3794static void 3795free_freedep(freedep) 3796 struct freedep *freedep; 3797{ 3798 struct freework *freework; 3799 3800 freework = freedep->fd_freework; 3801 freework->fw_freeblks->fb_cgwait--; 3802 if (--freework->fw_ref == 0) 3803 freework_enqueue(freework); 3804 WORKITEM_FREE(freedep, D_FREEDEP); 3805} 3806 3807/* 3808 * Allocate a new freework structure that may be a level in an indirect 3809 * when parent is not NULL or a top level block when it is. The top level 3810 * freework structures are allocated without lk held and before the freeblks 3811 * is visible outside of softdep_setup_freeblocks(). 3812 */ 3813static struct freework * 3814newfreework(ump, freeblks, parent, lbn, nb, frags, off, journal) 3815 struct ufsmount *ump; 3816 struct freeblks *freeblks; 3817 struct freework *parent; 3818 ufs_lbn_t lbn; 3819 ufs2_daddr_t nb; 3820 int frags; 3821 int off; 3822 int journal; 3823{ 3824 struct freework *freework; 3825 3826 freework = malloc(sizeof(*freework), M_FREEWORK, M_SOFTDEP_FLAGS); 3827 workitem_alloc(&freework->fw_list, D_FREEWORK, freeblks->fb_list.wk_mp); 3828 freework->fw_state = ATTACHED; 3829 freework->fw_jnewblk = NULL; 3830 freework->fw_freeblks = freeblks; 3831 freework->fw_parent = parent; 3832 freework->fw_lbn = lbn; 3833 freework->fw_blkno = nb; 3834 freework->fw_frags = frags; 3835 freework->fw_indir = NULL; 3836 freework->fw_ref = (MOUNTEDSUJ(UFSTOVFS(ump)) == 0 || lbn >= -NXADDR) 3837 ? 0 : NINDIR(ump->um_fs) + 1; 3838 freework->fw_start = freework->fw_off = off; 3839 if (journal) 3840 newjfreeblk(freeblks, lbn, nb, frags); 3841 if (parent == NULL) { 3842 ACQUIRE_LOCK(&lk); 3843 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list); 3844 freeblks->fb_ref++; 3845 FREE_LOCK(&lk); 3846 } 3847 3848 return (freework); 3849} 3850 3851/* 3852 * Eliminate a jfreeblk for a block that does not need journaling. 3853 */ 3854static void 3855cancel_jfreeblk(freeblks, blkno) 3856 struct freeblks *freeblks; 3857 ufs2_daddr_t blkno; 3858{ 3859 struct jfreeblk *jfreeblk; 3860 struct jblkdep *jblkdep; 3861 3862 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) { 3863 if (jblkdep->jb_list.wk_type != D_JFREEBLK) 3864 continue; 3865 jfreeblk = WK_JFREEBLK(&jblkdep->jb_list); 3866 if (jfreeblk->jf_blkno == blkno) 3867 break; 3868 } 3869 if (jblkdep == NULL) 3870 return; 3871 free_jsegdep(jblkdep->jb_jsegdep); 3872 LIST_REMOVE(jblkdep, jb_deps); 3873 WORKITEM_FREE(jfreeblk, D_JFREEBLK); 3874} 3875 3876/* 3877 * Allocate a new jfreeblk to journal top level block pointer when truncating 3878 * a file. The caller must add this to the worklist when lk is held. 3879 */ 3880static struct jfreeblk * 3881newjfreeblk(freeblks, lbn, blkno, frags) 3882 struct freeblks *freeblks; 3883 ufs_lbn_t lbn; 3884 ufs2_daddr_t blkno; 3885 int frags; 3886{ 3887 struct jfreeblk *jfreeblk; 3888 3889 jfreeblk = malloc(sizeof(*jfreeblk), M_JFREEBLK, M_SOFTDEP_FLAGS); 3890 workitem_alloc(&jfreeblk->jf_dep.jb_list, D_JFREEBLK, 3891 freeblks->fb_list.wk_mp); 3892 jfreeblk->jf_dep.jb_jsegdep = newjsegdep(&jfreeblk->jf_dep.jb_list); 3893 jfreeblk->jf_dep.jb_freeblks = freeblks; 3894 jfreeblk->jf_ino = freeblks->fb_inum; 3895 jfreeblk->jf_lbn = lbn; 3896 jfreeblk->jf_blkno = blkno; 3897 jfreeblk->jf_frags = frags; 3898 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jfreeblk->jf_dep, jb_deps); 3899 3900 return (jfreeblk); 3901} 3902 3903/* 3904 * Allocate a new jtrunc to track a partial truncation. 3905 */ 3906static struct jtrunc * 3907newjtrunc(freeblks, size, extsize) 3908 struct freeblks *freeblks; 3909 off_t size; 3910 int extsize; 3911{ 3912 struct jtrunc *jtrunc; 3913 3914 jtrunc = malloc(sizeof(*jtrunc), M_JTRUNC, M_SOFTDEP_FLAGS); 3915 workitem_alloc(&jtrunc->jt_dep.jb_list, D_JTRUNC, 3916 freeblks->fb_list.wk_mp); 3917 jtrunc->jt_dep.jb_jsegdep = newjsegdep(&jtrunc->jt_dep.jb_list); 3918 jtrunc->jt_dep.jb_freeblks = freeblks; 3919 jtrunc->jt_ino = freeblks->fb_inum; 3920 jtrunc->jt_size = size; 3921 jtrunc->jt_extsize = extsize; 3922 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jtrunc->jt_dep, jb_deps); 3923 3924 return (jtrunc); 3925} 3926 3927/* 3928 * If we're canceling a new bitmap we have to search for another ref 3929 * to move into the bmsafemap dep. This might be better expressed 3930 * with another structure. 3931 */ 3932static void 3933move_newblock_dep(jaddref, inodedep) 3934 struct jaddref *jaddref; 3935 struct inodedep *inodedep; 3936{ 3937 struct inoref *inoref; 3938 struct jaddref *jaddrefn; 3939 3940 jaddrefn = NULL; 3941 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref; 3942 inoref = TAILQ_NEXT(inoref, if_deps)) { 3943 if ((jaddref->ja_state & NEWBLOCK) && 3944 inoref->if_list.wk_type == D_JADDREF) { 3945 jaddrefn = (struct jaddref *)inoref; 3946 break; 3947 } 3948 } 3949 if (jaddrefn == NULL) 3950 return; 3951 jaddrefn->ja_state &= ~(ATTACHED | UNDONE); 3952 jaddrefn->ja_state |= jaddref->ja_state & 3953 (ATTACHED | UNDONE | NEWBLOCK); 3954 jaddref->ja_state &= ~(ATTACHED | UNDONE | NEWBLOCK); 3955 jaddref->ja_state |= ATTACHED; 3956 LIST_REMOVE(jaddref, ja_bmdeps); 3957 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_jaddrefhd, jaddrefn, 3958 ja_bmdeps); 3959} 3960 3961/* 3962 * Cancel a jaddref either before it has been written or while it is being 3963 * written. This happens when a link is removed before the add reaches 3964 * the disk. The jaddref dependency is kept linked into the bmsafemap 3965 * and inode to prevent the link count or bitmap from reaching the disk 3966 * until handle_workitem_remove() re-adjusts the counts and bitmaps as 3967 * required. 3968 * 3969 * Returns 1 if the canceled addref requires journaling of the remove and 3970 * 0 otherwise. 3971 */ 3972static int 3973cancel_jaddref(jaddref, inodedep, wkhd) 3974 struct jaddref *jaddref; 3975 struct inodedep *inodedep; 3976 struct workhead *wkhd; 3977{ 3978 struct inoref *inoref; 3979 struct jsegdep *jsegdep; 3980 int needsj; 3981 3982 KASSERT((jaddref->ja_state & COMPLETE) == 0, 3983 ("cancel_jaddref: Canceling complete jaddref")); 3984 if (jaddref->ja_state & (INPROGRESS | COMPLETE)) 3985 needsj = 1; 3986 else 3987 needsj = 0; 3988 if (inodedep == NULL) 3989 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino, 3990 0, &inodedep) == 0) 3991 panic("cancel_jaddref: Lost inodedep"); 3992 /* 3993 * We must adjust the nlink of any reference operation that follows 3994 * us so that it is consistent with the in-memory reference. This 3995 * ensures that inode nlink rollbacks always have the correct link. 3996 */ 3997 if (needsj == 0) { 3998 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref; 3999 inoref = TAILQ_NEXT(inoref, if_deps)) { 4000 if (inoref->if_state & GOINGAWAY) 4001 break; 4002 inoref->if_nlink--; 4003 } 4004 } 4005 jsegdep = inoref_jseg(&jaddref->ja_ref); 4006 if (jaddref->ja_state & NEWBLOCK) 4007 move_newblock_dep(jaddref, inodedep); 4008 wake_worklist(&jaddref->ja_list); 4009 jaddref->ja_mkdir = NULL; 4010 if (jaddref->ja_state & INPROGRESS) { 4011 jaddref->ja_state &= ~INPROGRESS; 4012 WORKLIST_REMOVE(&jaddref->ja_list); 4013 jwork_insert(wkhd, jsegdep); 4014 } else { 4015 free_jsegdep(jsegdep); 4016 if (jaddref->ja_state & DEPCOMPLETE) 4017 remove_from_journal(&jaddref->ja_list); 4018 } 4019 jaddref->ja_state |= (GOINGAWAY | DEPCOMPLETE); 4020 /* 4021 * Leave NEWBLOCK jaddrefs on the inodedep so handle_workitem_remove 4022 * can arrange for them to be freed with the bitmap. Otherwise we 4023 * no longer need this addref attached to the inoreflst and it 4024 * will incorrectly adjust nlink if we leave it. 4025 */ 4026 if ((jaddref->ja_state & NEWBLOCK) == 0) { 4027 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, 4028 if_deps); 4029 jaddref->ja_state |= COMPLETE; 4030 free_jaddref(jaddref); 4031 return (needsj); 4032 } 4033 /* 4034 * Leave the head of the list for jsegdeps for fast merging. 4035 */ 4036 if (LIST_FIRST(wkhd) != NULL) { 4037 jaddref->ja_state |= ONWORKLIST; 4038 LIST_INSERT_AFTER(LIST_FIRST(wkhd), &jaddref->ja_list, wk_list); 4039 } else 4040 WORKLIST_INSERT(wkhd, &jaddref->ja_list); 4041 4042 return (needsj); 4043} 4044 4045/* 4046 * Attempt to free a jaddref structure when some work completes. This 4047 * should only succeed once the entry is written and all dependencies have 4048 * been notified. 4049 */ 4050static void 4051free_jaddref(jaddref) 4052 struct jaddref *jaddref; 4053{ 4054 4055 if ((jaddref->ja_state & ALLCOMPLETE) != ALLCOMPLETE) 4056 return; 4057 if (jaddref->ja_ref.if_jsegdep) 4058 panic("free_jaddref: segdep attached to jaddref %p(0x%X)\n", 4059 jaddref, jaddref->ja_state); 4060 if (jaddref->ja_state & NEWBLOCK) 4061 LIST_REMOVE(jaddref, ja_bmdeps); 4062 if (jaddref->ja_state & (INPROGRESS | ONWORKLIST)) 4063 panic("free_jaddref: Bad state %p(0x%X)", 4064 jaddref, jaddref->ja_state); 4065 if (jaddref->ja_mkdir != NULL) 4066 panic("free_jaddref: Work pending, 0x%X\n", jaddref->ja_state); 4067 WORKITEM_FREE(jaddref, D_JADDREF); 4068} 4069 4070/* 4071 * Free a jremref structure once it has been written or discarded. 4072 */ 4073static void 4074free_jremref(jremref) 4075 struct jremref *jremref; 4076{ 4077 4078 if (jremref->jr_ref.if_jsegdep) 4079 free_jsegdep(jremref->jr_ref.if_jsegdep); 4080 if (jremref->jr_state & INPROGRESS) 4081 panic("free_jremref: IO still pending"); 4082 WORKITEM_FREE(jremref, D_JREMREF); 4083} 4084 4085/* 4086 * Free a jnewblk structure. 4087 */ 4088static void 4089free_jnewblk(jnewblk) 4090 struct jnewblk *jnewblk; 4091{ 4092 4093 if ((jnewblk->jn_state & ALLCOMPLETE) != ALLCOMPLETE) 4094 return; 4095 LIST_REMOVE(jnewblk, jn_deps); 4096 if (jnewblk->jn_dep != NULL) 4097 panic("free_jnewblk: Dependency still attached."); 4098 WORKITEM_FREE(jnewblk, D_JNEWBLK); 4099} 4100 4101/* 4102 * Cancel a jnewblk which has been been made redundant by frag extension. 4103 */ 4104static void 4105cancel_jnewblk(jnewblk, wkhd) 4106 struct jnewblk *jnewblk; 4107 struct workhead *wkhd; 4108{ 4109 struct jsegdep *jsegdep; 4110 4111 jsegdep = jnewblk->jn_jsegdep; 4112 if (jnewblk->jn_jsegdep == NULL || jnewblk->jn_dep == NULL) 4113 panic("cancel_jnewblk: Invalid state"); 4114 jnewblk->jn_jsegdep = NULL; 4115 jnewblk->jn_dep = NULL; 4116 jnewblk->jn_state |= GOINGAWAY; 4117 if (jnewblk->jn_state & INPROGRESS) { 4118 jnewblk->jn_state &= ~INPROGRESS; 4119 WORKLIST_REMOVE(&jnewblk->jn_list); 4120 jwork_insert(wkhd, jsegdep); 4121 } else { 4122 free_jsegdep(jsegdep); 4123 remove_from_journal(&jnewblk->jn_list); 4124 } 4125 wake_worklist(&jnewblk->jn_list); 4126 WORKLIST_INSERT(wkhd, &jnewblk->jn_list); 4127} 4128 4129static void 4130free_jblkdep(jblkdep) 4131 struct jblkdep *jblkdep; 4132{ 4133 4134 if (jblkdep->jb_list.wk_type == D_JFREEBLK) 4135 WORKITEM_FREE(jblkdep, D_JFREEBLK); 4136 else if (jblkdep->jb_list.wk_type == D_JTRUNC) 4137 WORKITEM_FREE(jblkdep, D_JTRUNC); 4138 else 4139 panic("free_jblkdep: Unexpected type %s", 4140 TYPENAME(jblkdep->jb_list.wk_type)); 4141} 4142 4143/* 4144 * Free a single jseg once it is no longer referenced in memory or on 4145 * disk. Reclaim journal blocks and dependencies waiting for the segment 4146 * to disappear. 4147 */ 4148static void 4149free_jseg(jseg, jblocks) 4150 struct jseg *jseg; 4151 struct jblocks *jblocks; 4152{ 4153 struct freework *freework; 4154 4155 /* 4156 * Free freework structures that were lingering to indicate freed 4157 * indirect blocks that forced journal write ordering on reallocate. 4158 */ 4159 while ((freework = LIST_FIRST(&jseg->js_indirs)) != NULL) 4160 indirblk_remove(freework); 4161 if (jblocks->jb_oldestseg == jseg) 4162 jblocks->jb_oldestseg = TAILQ_NEXT(jseg, js_next); 4163 TAILQ_REMOVE(&jblocks->jb_segs, jseg, js_next); 4164 jblocks_free(jblocks, jseg->js_list.wk_mp, jseg->js_size); 4165 KASSERT(LIST_EMPTY(&jseg->js_entries), 4166 ("free_jseg: Freed jseg has valid entries.")); 4167 WORKITEM_FREE(jseg, D_JSEG); 4168} 4169 4170/* 4171 * Free all jsegs that meet the criteria for being reclaimed and update 4172 * oldestseg. 4173 */ 4174static void 4175free_jsegs(jblocks) 4176 struct jblocks *jblocks; 4177{ 4178 struct jseg *jseg; 4179 4180 /* 4181 * Free only those jsegs which have none allocated before them to 4182 * preserve the journal space ordering. 4183 */ 4184 while ((jseg = TAILQ_FIRST(&jblocks->jb_segs)) != NULL) { 4185 /* 4186 * Only reclaim space when nothing depends on this journal 4187 * set and another set has written that it is no longer 4188 * valid. 4189 */ 4190 if (jseg->js_refs != 0) { 4191 jblocks->jb_oldestseg = jseg; 4192 return; 4193 } 4194 if (!LIST_EMPTY(&jseg->js_indirs) && 4195 jseg->js_seq >= jblocks->jb_oldestwrseq) 4196 break; 4197 free_jseg(jseg, jblocks); 4198 } 4199 /* 4200 * If we exited the loop above we still must discover the 4201 * oldest valid segment. 4202 */ 4203 if (jseg) 4204 for (jseg = jblocks->jb_oldestseg; jseg != NULL; 4205 jseg = TAILQ_NEXT(jseg, js_next)) 4206 if (jseg->js_refs != 0) 4207 break; 4208 jblocks->jb_oldestseg = jseg; 4209 /* 4210 * The journal has no valid records but some jsegs may still be 4211 * waiting on oldestwrseq to advance. We force a small record 4212 * out to permit these lingering records to be reclaimed. 4213 */ 4214 if (jblocks->jb_oldestseg == NULL && !TAILQ_EMPTY(&jblocks->jb_segs)) 4215 jblocks->jb_needseg = 1; 4216} 4217 4218/* 4219 * Release one reference to a jseg and free it if the count reaches 0. This 4220 * should eventually reclaim journal space as well. 4221 */ 4222static void 4223rele_jseg(jseg) 4224 struct jseg *jseg; 4225{ 4226 4227 KASSERT(jseg->js_refs > 0, 4228 ("free_jseg: Invalid refcnt %d", jseg->js_refs)); 4229 if (--jseg->js_refs != 0) 4230 return; 4231 free_jsegs(jseg->js_jblocks); 4232} 4233 4234/* 4235 * Release a jsegdep and decrement the jseg count. 4236 */ 4237static void 4238free_jsegdep(jsegdep) 4239 struct jsegdep *jsegdep; 4240{ 4241 4242 if (jsegdep->jd_seg) 4243 rele_jseg(jsegdep->jd_seg); 4244 WORKITEM_FREE(jsegdep, D_JSEGDEP); 4245} 4246 4247/* 4248 * Wait for a journal item to make it to disk. Initiate journal processing 4249 * if required. 4250 */ 4251static int 4252jwait(wk, waitfor) 4253 struct worklist *wk; 4254 int waitfor; 4255{ 4256 4257 /* 4258 * Blocking journal waits cause slow synchronous behavior. Record 4259 * stats on the frequency of these blocking operations. 4260 */ 4261 if (waitfor == MNT_WAIT) { 4262 stat_journal_wait++; 4263 switch (wk->wk_type) { 4264 case D_JREMREF: 4265 case D_JMVREF: 4266 stat_jwait_filepage++; 4267 break; 4268 case D_JTRUNC: 4269 case D_JFREEBLK: 4270 stat_jwait_freeblks++; 4271 break; 4272 case D_JNEWBLK: 4273 stat_jwait_newblk++; 4274 break; 4275 case D_JADDREF: 4276 stat_jwait_inode++; 4277 break; 4278 default: 4279 break; 4280 } 4281 } 4282 /* 4283 * If IO has not started we process the journal. We can't mark the 4284 * worklist item as IOWAITING because we drop the lock while 4285 * processing the journal and the worklist entry may be freed after 4286 * this point. The caller may call back in and re-issue the request. 4287 */ 4288 if ((wk->wk_state & INPROGRESS) == 0) { 4289 softdep_process_journal(wk->wk_mp, wk, waitfor); 4290 if (waitfor != MNT_WAIT) 4291 return (EBUSY); 4292 return (0); 4293 } 4294 if (waitfor != MNT_WAIT) 4295 return (EBUSY); 4296 wait_worklist(wk, "jwait"); 4297 return (0); 4298} 4299 4300/* 4301 * Lookup an inodedep based on an inode pointer and set the nlinkdelta as 4302 * appropriate. This is a convenience function to reduce duplicate code 4303 * for the setup and revert functions below. 4304 */ 4305static struct inodedep * 4306inodedep_lookup_ip(ip) 4307 struct inode *ip; 4308{ 4309 struct inodedep *inodedep; 4310 int dflags; 4311 4312 KASSERT(ip->i_nlink >= ip->i_effnlink, 4313 ("inodedep_lookup_ip: bad delta")); 4314 dflags = DEPALLOC; 4315 if (IS_SNAPSHOT(ip)) 4316 dflags |= NODELAY; 4317 (void) inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, dflags, 4318 &inodedep); 4319 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 4320 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked")); 4321 4322 return (inodedep); 4323} 4324 4325/* 4326 * Called prior to creating a new inode and linking it to a directory. The 4327 * jaddref structure must already be allocated by softdep_setup_inomapdep 4328 * and it is discovered here so we can initialize the mode and update 4329 * nlinkdelta. 4330 */ 4331void 4332softdep_setup_create(dp, ip) 4333 struct inode *dp; 4334 struct inode *ip; 4335{ 4336 struct inodedep *inodedep; 4337 struct jaddref *jaddref; 4338 struct vnode *dvp; 4339 4340 KASSERT(ip->i_nlink == 1, 4341 ("softdep_setup_create: Invalid link count.")); 4342 dvp = ITOV(dp); 4343 ACQUIRE_LOCK(&lk); 4344 inodedep = inodedep_lookup_ip(ip); 4345 if (DOINGSUJ(dvp)) { 4346 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4347 inoreflst); 4348 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 4349 ("softdep_setup_create: No addref structure present.")); 4350 } 4351 softdep_prelink(dvp, NULL); 4352 FREE_LOCK(&lk); 4353} 4354 4355/* 4356 * Create a jaddref structure to track the addition of a DOTDOT link when 4357 * we are reparenting an inode as part of a rename. This jaddref will be 4358 * found by softdep_setup_directory_change. Adjusts nlinkdelta for 4359 * non-journaling softdep. 4360 */ 4361void 4362softdep_setup_dotdot_link(dp, ip) 4363 struct inode *dp; 4364 struct inode *ip; 4365{ 4366 struct inodedep *inodedep; 4367 struct jaddref *jaddref; 4368 struct vnode *dvp; 4369 struct vnode *vp; 4370 4371 dvp = ITOV(dp); 4372 vp = ITOV(ip); 4373 jaddref = NULL; 4374 /* 4375 * We don't set MKDIR_PARENT as this is not tied to a mkdir and 4376 * is used as a normal link would be. 4377 */ 4378 if (DOINGSUJ(dvp)) 4379 jaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET, 4380 dp->i_effnlink - 1, dp->i_mode); 4381 ACQUIRE_LOCK(&lk); 4382 inodedep = inodedep_lookup_ip(dp); 4383 if (jaddref) 4384 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 4385 if_deps); 4386 softdep_prelink(dvp, ITOV(ip)); 4387 FREE_LOCK(&lk); 4388} 4389 4390/* 4391 * Create a jaddref structure to track a new link to an inode. The directory 4392 * offset is not known until softdep_setup_directory_add or 4393 * softdep_setup_directory_change. Adjusts nlinkdelta for non-journaling 4394 * softdep. 4395 */ 4396void 4397softdep_setup_link(dp, ip) 4398 struct inode *dp; 4399 struct inode *ip; 4400{ 4401 struct inodedep *inodedep; 4402 struct jaddref *jaddref; 4403 struct vnode *dvp; 4404 4405 dvp = ITOV(dp); 4406 jaddref = NULL; 4407 if (DOINGSUJ(dvp)) 4408 jaddref = newjaddref(dp, ip->i_number, 0, ip->i_effnlink - 1, 4409 ip->i_mode); 4410 ACQUIRE_LOCK(&lk); 4411 inodedep = inodedep_lookup_ip(ip); 4412 if (jaddref) 4413 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 4414 if_deps); 4415 softdep_prelink(dvp, ITOV(ip)); 4416 FREE_LOCK(&lk); 4417} 4418 4419/* 4420 * Called to create the jaddref structures to track . and .. references as 4421 * well as lookup and further initialize the incomplete jaddref created 4422 * by softdep_setup_inomapdep when the inode was allocated. Adjusts 4423 * nlinkdelta for non-journaling softdep. 4424 */ 4425void 4426softdep_setup_mkdir(dp, ip) 4427 struct inode *dp; 4428 struct inode *ip; 4429{ 4430 struct inodedep *inodedep; 4431 struct jaddref *dotdotaddref; 4432 struct jaddref *dotaddref; 4433 struct jaddref *jaddref; 4434 struct vnode *dvp; 4435 4436 dvp = ITOV(dp); 4437 dotaddref = dotdotaddref = NULL; 4438 if (DOINGSUJ(dvp)) { 4439 dotaddref = newjaddref(ip, ip->i_number, DOT_OFFSET, 1, 4440 ip->i_mode); 4441 dotaddref->ja_state |= MKDIR_BODY; 4442 dotdotaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET, 4443 dp->i_effnlink - 1, dp->i_mode); 4444 dotdotaddref->ja_state |= MKDIR_PARENT; 4445 } 4446 ACQUIRE_LOCK(&lk); 4447 inodedep = inodedep_lookup_ip(ip); 4448 if (DOINGSUJ(dvp)) { 4449 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4450 inoreflst); 4451 KASSERT(jaddref != NULL, 4452 ("softdep_setup_mkdir: No addref structure present.")); 4453 KASSERT(jaddref->ja_parent == dp->i_number, 4454 ("softdep_setup_mkdir: bad parent %ju", 4455 (uintmax_t)jaddref->ja_parent)); 4456 TAILQ_INSERT_BEFORE(&jaddref->ja_ref, &dotaddref->ja_ref, 4457 if_deps); 4458 } 4459 inodedep = inodedep_lookup_ip(dp); 4460 if (DOINGSUJ(dvp)) 4461 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, 4462 &dotdotaddref->ja_ref, if_deps); 4463 softdep_prelink(ITOV(dp), NULL); 4464 FREE_LOCK(&lk); 4465} 4466 4467/* 4468 * Called to track nlinkdelta of the inode and parent directories prior to 4469 * unlinking a directory. 4470 */ 4471void 4472softdep_setup_rmdir(dp, ip) 4473 struct inode *dp; 4474 struct inode *ip; 4475{ 4476 struct vnode *dvp; 4477 4478 dvp = ITOV(dp); 4479 ACQUIRE_LOCK(&lk); 4480 (void) inodedep_lookup_ip(ip); 4481 (void) inodedep_lookup_ip(dp); 4482 softdep_prelink(dvp, ITOV(ip)); 4483 FREE_LOCK(&lk); 4484} 4485 4486/* 4487 * Called to track nlinkdelta of the inode and parent directories prior to 4488 * unlink. 4489 */ 4490void 4491softdep_setup_unlink(dp, ip) 4492 struct inode *dp; 4493 struct inode *ip; 4494{ 4495 struct vnode *dvp; 4496 4497 dvp = ITOV(dp); 4498 ACQUIRE_LOCK(&lk); 4499 (void) inodedep_lookup_ip(ip); 4500 (void) inodedep_lookup_ip(dp); 4501 softdep_prelink(dvp, ITOV(ip)); 4502 FREE_LOCK(&lk); 4503} 4504 4505/* 4506 * Called to release the journal structures created by a failed non-directory 4507 * creation. Adjusts nlinkdelta for non-journaling softdep. 4508 */ 4509void 4510softdep_revert_create(dp, ip) 4511 struct inode *dp; 4512 struct inode *ip; 4513{ 4514 struct inodedep *inodedep; 4515 struct jaddref *jaddref; 4516 struct vnode *dvp; 4517 4518 dvp = ITOV(dp); 4519 ACQUIRE_LOCK(&lk); 4520 inodedep = inodedep_lookup_ip(ip); 4521 if (DOINGSUJ(dvp)) { 4522 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4523 inoreflst); 4524 KASSERT(jaddref->ja_parent == dp->i_number, 4525 ("softdep_revert_create: addref parent mismatch")); 4526 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4527 } 4528 FREE_LOCK(&lk); 4529} 4530 4531/* 4532 * Called to release the journal structures created by a failed dotdot link 4533 * creation. Adjusts nlinkdelta for non-journaling softdep. 4534 */ 4535void 4536softdep_revert_dotdot_link(dp, ip) 4537 struct inode *dp; 4538 struct inode *ip; 4539{ 4540 struct inodedep *inodedep; 4541 struct jaddref *jaddref; 4542 struct vnode *dvp; 4543 4544 dvp = ITOV(dp); 4545 ACQUIRE_LOCK(&lk); 4546 inodedep = inodedep_lookup_ip(dp); 4547 if (DOINGSUJ(dvp)) { 4548 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4549 inoreflst); 4550 KASSERT(jaddref->ja_parent == ip->i_number, 4551 ("softdep_revert_dotdot_link: addref parent mismatch")); 4552 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4553 } 4554 FREE_LOCK(&lk); 4555} 4556 4557/* 4558 * Called to release the journal structures created by a failed link 4559 * addition. Adjusts nlinkdelta for non-journaling softdep. 4560 */ 4561void 4562softdep_revert_link(dp, ip) 4563 struct inode *dp; 4564 struct inode *ip; 4565{ 4566 struct inodedep *inodedep; 4567 struct jaddref *jaddref; 4568 struct vnode *dvp; 4569 4570 dvp = ITOV(dp); 4571 ACQUIRE_LOCK(&lk); 4572 inodedep = inodedep_lookup_ip(ip); 4573 if (DOINGSUJ(dvp)) { 4574 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4575 inoreflst); 4576 KASSERT(jaddref->ja_parent == dp->i_number, 4577 ("softdep_revert_link: addref parent mismatch")); 4578 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4579 } 4580 FREE_LOCK(&lk); 4581} 4582 4583/* 4584 * Called to release the journal structures created by a failed mkdir 4585 * attempt. Adjusts nlinkdelta for non-journaling softdep. 4586 */ 4587void 4588softdep_revert_mkdir(dp, ip) 4589 struct inode *dp; 4590 struct inode *ip; 4591{ 4592 struct inodedep *inodedep; 4593 struct jaddref *jaddref; 4594 struct jaddref *dotaddref; 4595 struct vnode *dvp; 4596 4597 dvp = ITOV(dp); 4598 4599 ACQUIRE_LOCK(&lk); 4600 inodedep = inodedep_lookup_ip(dp); 4601 if (DOINGSUJ(dvp)) { 4602 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4603 inoreflst); 4604 KASSERT(jaddref->ja_parent == ip->i_number, 4605 ("softdep_revert_mkdir: dotdot addref parent mismatch")); 4606 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4607 } 4608 inodedep = inodedep_lookup_ip(ip); 4609 if (DOINGSUJ(dvp)) { 4610 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4611 inoreflst); 4612 KASSERT(jaddref->ja_parent == dp->i_number, 4613 ("softdep_revert_mkdir: addref parent mismatch")); 4614 dotaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref, 4615 inoreflst, if_deps); 4616 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4617 KASSERT(dotaddref->ja_parent == ip->i_number, 4618 ("softdep_revert_mkdir: dot addref parent mismatch")); 4619 cancel_jaddref(dotaddref, inodedep, &inodedep->id_inowait); 4620 } 4621 FREE_LOCK(&lk); 4622} 4623 4624/* 4625 * Called to correct nlinkdelta after a failed rmdir. 4626 */ 4627void 4628softdep_revert_rmdir(dp, ip) 4629 struct inode *dp; 4630 struct inode *ip; 4631{ 4632 4633 ACQUIRE_LOCK(&lk); 4634 (void) inodedep_lookup_ip(ip); 4635 (void) inodedep_lookup_ip(dp); 4636 FREE_LOCK(&lk); 4637} 4638 4639/* 4640 * Protecting the freemaps (or bitmaps). 4641 * 4642 * To eliminate the need to execute fsck before mounting a filesystem 4643 * after a power failure, one must (conservatively) guarantee that the 4644 * on-disk copy of the bitmaps never indicate that a live inode or block is 4645 * free. So, when a block or inode is allocated, the bitmap should be 4646 * updated (on disk) before any new pointers. When a block or inode is 4647 * freed, the bitmap should not be updated until all pointers have been 4648 * reset. The latter dependency is handled by the delayed de-allocation 4649 * approach described below for block and inode de-allocation. The former 4650 * dependency is handled by calling the following procedure when a block or 4651 * inode is allocated. When an inode is allocated an "inodedep" is created 4652 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 4653 * Each "inodedep" is also inserted into the hash indexing structure so 4654 * that any additional link additions can be made dependent on the inode 4655 * allocation. 4656 * 4657 * The ufs filesystem maintains a number of free block counts (e.g., per 4658 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 4659 * in addition to the bitmaps. These counts are used to improve efficiency 4660 * during allocation and therefore must be consistent with the bitmaps. 4661 * There is no convenient way to guarantee post-crash consistency of these 4662 * counts with simple update ordering, for two main reasons: (1) The counts 4663 * and bitmaps for a single cylinder group block are not in the same disk 4664 * sector. If a disk write is interrupted (e.g., by power failure), one may 4665 * be written and the other not. (2) Some of the counts are located in the 4666 * superblock rather than the cylinder group block. So, we focus our soft 4667 * updates implementation on protecting the bitmaps. When mounting a 4668 * filesystem, we recompute the auxiliary counts from the bitmaps. 4669 */ 4670 4671/* 4672 * Called just after updating the cylinder group block to allocate an inode. 4673 */ 4674void 4675softdep_setup_inomapdep(bp, ip, newinum, mode) 4676 struct buf *bp; /* buffer for cylgroup block with inode map */ 4677 struct inode *ip; /* inode related to allocation */ 4678 ino_t newinum; /* new inode number being allocated */ 4679 int mode; 4680{ 4681 struct inodedep *inodedep; 4682 struct bmsafemap *bmsafemap; 4683 struct jaddref *jaddref; 4684 struct mount *mp; 4685 struct fs *fs; 4686 4687 mp = UFSTOVFS(ip->i_ump); 4688 fs = ip->i_ump->um_fs; 4689 jaddref = NULL; 4690 4691 /* 4692 * Allocate the journal reference add structure so that the bitmap 4693 * can be dependent on it. 4694 */ 4695 if (MOUNTEDSUJ(mp)) { 4696 jaddref = newjaddref(ip, newinum, 0, 0, mode); 4697 jaddref->ja_state |= NEWBLOCK; 4698 } 4699 4700 /* 4701 * Create a dependency for the newly allocated inode. 4702 * Panic if it already exists as something is seriously wrong. 4703 * Otherwise add it to the dependency list for the buffer holding 4704 * the cylinder group map from which it was allocated. 4705 * 4706 * We have to preallocate a bmsafemap entry in case it is needed 4707 * in bmsafemap_lookup since once we allocate the inodedep, we 4708 * have to finish initializing it before we can FREE_LOCK(). 4709 * By preallocating, we avoid FREE_LOCK() while doing a malloc 4710 * in bmsafemap_lookup. We cannot call bmsafemap_lookup before 4711 * creating the inodedep as it can be freed during the time 4712 * that we FREE_LOCK() while allocating the inodedep. We must 4713 * call workitem_alloc() before entering the locked section as 4714 * it also acquires the lock and we must avoid trying doing so 4715 * recursively. 4716 */ 4717 bmsafemap = malloc(sizeof(struct bmsafemap), 4718 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 4719 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); 4720 ACQUIRE_LOCK(&lk); 4721 if ((inodedep_lookup(mp, newinum, DEPALLOC | NODELAY, &inodedep))) 4722 panic("softdep_setup_inomapdep: dependency %p for new" 4723 "inode already exists", inodedep); 4724 bmsafemap = bmsafemap_lookup(mp, bp, ino_to_cg(fs, newinum), bmsafemap); 4725 if (jaddref) { 4726 LIST_INSERT_HEAD(&bmsafemap->sm_jaddrefhd, jaddref, ja_bmdeps); 4727 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 4728 if_deps); 4729 } else { 4730 inodedep->id_state |= ONDEPLIST; 4731 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 4732 } 4733 inodedep->id_bmsafemap = bmsafemap; 4734 inodedep->id_state &= ~DEPCOMPLETE; 4735 FREE_LOCK(&lk); 4736} 4737 4738/* 4739 * Called just after updating the cylinder group block to 4740 * allocate block or fragment. 4741 */ 4742void 4743softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags) 4744 struct buf *bp; /* buffer for cylgroup block with block map */ 4745 struct mount *mp; /* filesystem doing allocation */ 4746 ufs2_daddr_t newblkno; /* number of newly allocated block */ 4747 int frags; /* Number of fragments. */ 4748 int oldfrags; /* Previous number of fragments for extend. */ 4749{ 4750 struct newblk *newblk; 4751 struct bmsafemap *bmsafemap; 4752 struct jnewblk *jnewblk; 4753 struct fs *fs; 4754 4755 fs = VFSTOUFS(mp)->um_fs; 4756 jnewblk = NULL; 4757 /* 4758 * Create a dependency for the newly allocated block. 4759 * Add it to the dependency list for the buffer holding 4760 * the cylinder group map from which it was allocated. 4761 */ 4762 if (MOUNTEDSUJ(mp)) { 4763 jnewblk = malloc(sizeof(*jnewblk), M_JNEWBLK, M_SOFTDEP_FLAGS); 4764 workitem_alloc(&jnewblk->jn_list, D_JNEWBLK, mp); 4765 jnewblk->jn_jsegdep = newjsegdep(&jnewblk->jn_list); 4766 jnewblk->jn_state = ATTACHED; 4767 jnewblk->jn_blkno = newblkno; 4768 jnewblk->jn_frags = frags; 4769 jnewblk->jn_oldfrags = oldfrags; 4770#ifdef SUJ_DEBUG 4771 { 4772 struct cg *cgp; 4773 uint8_t *blksfree; 4774 long bno; 4775 int i; 4776 4777 cgp = (struct cg *)bp->b_data; 4778 blksfree = cg_blksfree(cgp); 4779 bno = dtogd(fs, jnewblk->jn_blkno); 4780 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; 4781 i++) { 4782 if (isset(blksfree, bno + i)) 4783 panic("softdep_setup_blkmapdep: " 4784 "free fragment %d from %d-%d " 4785 "state 0x%X dep %p", i, 4786 jnewblk->jn_oldfrags, 4787 jnewblk->jn_frags, 4788 jnewblk->jn_state, 4789 jnewblk->jn_dep); 4790 } 4791 } 4792#endif 4793 } 4794 ACQUIRE_LOCK(&lk); 4795 if (newblk_lookup(mp, newblkno, DEPALLOC, &newblk) != 0) 4796 panic("softdep_setup_blkmapdep: found block"); 4797 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp, 4798 dtog(fs, newblkno), NULL); 4799 if (jnewblk) { 4800 jnewblk->jn_dep = (struct worklist *)newblk; 4801 LIST_INSERT_HEAD(&bmsafemap->sm_jnewblkhd, jnewblk, jn_deps); 4802 } else { 4803 newblk->nb_state |= ONDEPLIST; 4804 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 4805 } 4806 newblk->nb_bmsafemap = bmsafemap; 4807 newblk->nb_jnewblk = jnewblk; 4808 FREE_LOCK(&lk); 4809} 4810 4811#define BMSAFEMAP_HASH(fs, cg) \ 4812 (&bmsafemap_hashtbl[((((register_t)(fs)) >> 13) + (cg)) & bmsafemap_hash]) 4813 4814static int 4815bmsafemap_find(bmsafemaphd, mp, cg, bmsafemapp) 4816 struct bmsafemap_hashhead *bmsafemaphd; 4817 struct mount *mp; 4818 int cg; 4819 struct bmsafemap **bmsafemapp; 4820{ 4821 struct bmsafemap *bmsafemap; 4822 4823 LIST_FOREACH(bmsafemap, bmsafemaphd, sm_hash) 4824 if (bmsafemap->sm_list.wk_mp == mp && bmsafemap->sm_cg == cg) 4825 break; 4826 if (bmsafemap) { 4827 *bmsafemapp = bmsafemap; 4828 return (1); 4829 } 4830 *bmsafemapp = NULL; 4831 4832 return (0); 4833} 4834 4835/* 4836 * Find the bmsafemap associated with a cylinder group buffer. 4837 * If none exists, create one. The buffer must be locked when 4838 * this routine is called and this routine must be called with 4839 * the softdep lock held. To avoid giving up the lock while 4840 * allocating a new bmsafemap, a preallocated bmsafemap may be 4841 * provided. If it is provided but not needed, it is freed. 4842 */ 4843static struct bmsafemap * 4844bmsafemap_lookup(mp, bp, cg, newbmsafemap) 4845 struct mount *mp; 4846 struct buf *bp; 4847 int cg; 4848 struct bmsafemap *newbmsafemap; 4849{ 4850 struct bmsafemap_hashhead *bmsafemaphd; 4851 struct bmsafemap *bmsafemap, *collision; 4852 struct worklist *wk; 4853 struct fs *fs; 4854 4855 mtx_assert(&lk, MA_OWNED); 4856 if (bp) 4857 LIST_FOREACH(wk, &bp->b_dep, wk_list) 4858 if (wk->wk_type == D_BMSAFEMAP) { 4859 if (newbmsafemap) 4860 WORKITEM_FREE(newbmsafemap,D_BMSAFEMAP); 4861 return (WK_BMSAFEMAP(wk)); 4862 } 4863 fs = VFSTOUFS(mp)->um_fs; 4864 bmsafemaphd = BMSAFEMAP_HASH(fs, cg); 4865 if (bmsafemap_find(bmsafemaphd, mp, cg, &bmsafemap) == 1) { 4866 if (newbmsafemap) 4867 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP); 4868 return (bmsafemap); 4869 } 4870 if (newbmsafemap) { 4871 bmsafemap = newbmsafemap; 4872 } else { 4873 FREE_LOCK(&lk); 4874 bmsafemap = malloc(sizeof(struct bmsafemap), 4875 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 4876 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); 4877 ACQUIRE_LOCK(&lk); 4878 } 4879 bmsafemap->sm_buf = bp; 4880 LIST_INIT(&bmsafemap->sm_inodedephd); 4881 LIST_INIT(&bmsafemap->sm_inodedepwr); 4882 LIST_INIT(&bmsafemap->sm_newblkhd); 4883 LIST_INIT(&bmsafemap->sm_newblkwr); 4884 LIST_INIT(&bmsafemap->sm_jaddrefhd); 4885 LIST_INIT(&bmsafemap->sm_jnewblkhd); 4886 LIST_INIT(&bmsafemap->sm_freehd); 4887 LIST_INIT(&bmsafemap->sm_freewr); 4888 if (bmsafemap_find(bmsafemaphd, mp, cg, &collision) == 1) { 4889 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 4890 return (collision); 4891 } 4892 bmsafemap->sm_cg = cg; 4893 LIST_INSERT_HEAD(bmsafemaphd, bmsafemap, sm_hash); 4894 LIST_INSERT_HEAD(&VFSTOUFS(mp)->softdep_dirtycg, bmsafemap, sm_next); 4895 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 4896 return (bmsafemap); 4897} 4898 4899/* 4900 * Direct block allocation dependencies. 4901 * 4902 * When a new block is allocated, the corresponding disk locations must be 4903 * initialized (with zeros or new data) before the on-disk inode points to 4904 * them. Also, the freemap from which the block was allocated must be 4905 * updated (on disk) before the inode's pointer. These two dependencies are 4906 * independent of each other and are needed for all file blocks and indirect 4907 * blocks that are pointed to directly by the inode. Just before the 4908 * "in-core" version of the inode is updated with a newly allocated block 4909 * number, a procedure (below) is called to setup allocation dependency 4910 * structures. These structures are removed when the corresponding 4911 * dependencies are satisfied or when the block allocation becomes obsolete 4912 * (i.e., the file is deleted, the block is de-allocated, or the block is a 4913 * fragment that gets upgraded). All of these cases are handled in 4914 * procedures described later. 4915 * 4916 * When a file extension causes a fragment to be upgraded, either to a larger 4917 * fragment or to a full block, the on-disk location may change (if the 4918 * previous fragment could not simply be extended). In this case, the old 4919 * fragment must be de-allocated, but not until after the inode's pointer has 4920 * been updated. In most cases, this is handled by later procedures, which 4921 * will construct a "freefrag" structure to be added to the workitem queue 4922 * when the inode update is complete (or obsolete). The main exception to 4923 * this is when an allocation occurs while a pending allocation dependency 4924 * (for the same block pointer) remains. This case is handled in the main 4925 * allocation dependency setup procedure by immediately freeing the 4926 * unreferenced fragments. 4927 */ 4928void 4929softdep_setup_allocdirect(ip, off, newblkno, oldblkno, newsize, oldsize, bp) 4930 struct inode *ip; /* inode to which block is being added */ 4931 ufs_lbn_t off; /* block pointer within inode */ 4932 ufs2_daddr_t newblkno; /* disk block number being added */ 4933 ufs2_daddr_t oldblkno; /* previous block number, 0 unless frag */ 4934 long newsize; /* size of new block */ 4935 long oldsize; /* size of new block */ 4936 struct buf *bp; /* bp for allocated block */ 4937{ 4938 struct allocdirect *adp, *oldadp; 4939 struct allocdirectlst *adphead; 4940 struct freefrag *freefrag; 4941 struct inodedep *inodedep; 4942 struct pagedep *pagedep; 4943 struct jnewblk *jnewblk; 4944 struct newblk *newblk; 4945 struct mount *mp; 4946 ufs_lbn_t lbn; 4947 4948 lbn = bp->b_lblkno; 4949 mp = UFSTOVFS(ip->i_ump); 4950 if (oldblkno && oldblkno != newblkno) 4951 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn); 4952 else 4953 freefrag = NULL; 4954 4955 ACQUIRE_LOCK(&lk); 4956 if (off >= NDADDR) { 4957 if (lbn > 0) 4958 panic("softdep_setup_allocdirect: bad lbn %jd, off %jd", 4959 lbn, off); 4960 /* allocating an indirect block */ 4961 if (oldblkno != 0) 4962 panic("softdep_setup_allocdirect: non-zero indir"); 4963 } else { 4964 if (off != lbn) 4965 panic("softdep_setup_allocdirect: lbn %jd != off %jd", 4966 lbn, off); 4967 /* 4968 * Allocating a direct block. 4969 * 4970 * If we are allocating a directory block, then we must 4971 * allocate an associated pagedep to track additions and 4972 * deletions. 4973 */ 4974 if ((ip->i_mode & IFMT) == IFDIR) 4975 pagedep_lookup(mp, bp, ip->i_number, off, DEPALLOC, 4976 &pagedep); 4977 } 4978 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0) 4979 panic("softdep_setup_allocdirect: lost block"); 4980 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 4981 ("softdep_setup_allocdirect: newblk already initialized")); 4982 /* 4983 * Convert the newblk to an allocdirect. 4984 */ 4985 newblk->nb_list.wk_type = D_ALLOCDIRECT; 4986 adp = (struct allocdirect *)newblk; 4987 newblk->nb_freefrag = freefrag; 4988 adp->ad_offset = off; 4989 adp->ad_oldblkno = oldblkno; 4990 adp->ad_newsize = newsize; 4991 adp->ad_oldsize = oldsize; 4992 4993 /* 4994 * Finish initializing the journal. 4995 */ 4996 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 4997 jnewblk->jn_ino = ip->i_number; 4998 jnewblk->jn_lbn = lbn; 4999 add_to_journal(&jnewblk->jn_list); 5000 } 5001 if (freefrag && freefrag->ff_jdep != NULL && 5002 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5003 add_to_journal(freefrag->ff_jdep); 5004 inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep); 5005 adp->ad_inodedep = inodedep; 5006 5007 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list); 5008 /* 5009 * The list of allocdirects must be kept in sorted and ascending 5010 * order so that the rollback routines can quickly determine the 5011 * first uncommitted block (the size of the file stored on disk 5012 * ends at the end of the lowest committed fragment, or if there 5013 * are no fragments, at the end of the highest committed block). 5014 * Since files generally grow, the typical case is that the new 5015 * block is to be added at the end of the list. We speed this 5016 * special case by checking against the last allocdirect in the 5017 * list before laboriously traversing the list looking for the 5018 * insertion point. 5019 */ 5020 adphead = &inodedep->id_newinoupdt; 5021 oldadp = TAILQ_LAST(adphead, allocdirectlst); 5022 if (oldadp == NULL || oldadp->ad_offset <= off) { 5023 /* insert at end of list */ 5024 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 5025 if (oldadp != NULL && oldadp->ad_offset == off) 5026 allocdirect_merge(adphead, adp, oldadp); 5027 FREE_LOCK(&lk); 5028 return; 5029 } 5030 TAILQ_FOREACH(oldadp, adphead, ad_next) { 5031 if (oldadp->ad_offset >= off) 5032 break; 5033 } 5034 if (oldadp == NULL) 5035 panic("softdep_setup_allocdirect: lost entry"); 5036 /* insert in middle of list */ 5037 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 5038 if (oldadp->ad_offset == off) 5039 allocdirect_merge(adphead, adp, oldadp); 5040 5041 FREE_LOCK(&lk); 5042} 5043 5044/* 5045 * Merge a newer and older journal record to be stored either in a 5046 * newblock or freefrag. This handles aggregating journal records for 5047 * fragment allocation into a second record as well as replacing a 5048 * journal free with an aborted journal allocation. A segment for the 5049 * oldest record will be placed on wkhd if it has been written. If not 5050 * the segment for the newer record will suffice. 5051 */ 5052static struct worklist * 5053jnewblk_merge(new, old, wkhd) 5054 struct worklist *new; 5055 struct worklist *old; 5056 struct workhead *wkhd; 5057{ 5058 struct jnewblk *njnewblk; 5059 struct jnewblk *jnewblk; 5060 5061 /* Handle NULLs to simplify callers. */ 5062 if (new == NULL) 5063 return (old); 5064 if (old == NULL) 5065 return (new); 5066 /* Replace a jfreefrag with a jnewblk. */ 5067 if (new->wk_type == D_JFREEFRAG) { 5068 cancel_jfreefrag(WK_JFREEFRAG(new)); 5069 return (old); 5070 } 5071 /* 5072 * Handle merging of two jnewblk records that describe 5073 * different sets of fragments in the same block. 5074 */ 5075 jnewblk = WK_JNEWBLK(old); 5076 njnewblk = WK_JNEWBLK(new); 5077 if (jnewblk->jn_blkno != njnewblk->jn_blkno) 5078 panic("jnewblk_merge: Merging disparate blocks."); 5079 /* 5080 * The record may be rolled back in the cg. 5081 */ 5082 if (jnewblk->jn_state & UNDONE) { 5083 jnewblk->jn_state &= ~UNDONE; 5084 njnewblk->jn_state |= UNDONE; 5085 njnewblk->jn_state &= ~ATTACHED; 5086 } 5087 /* 5088 * We modify the newer addref and free the older so that if neither 5089 * has been written the most up-to-date copy will be on disk. If 5090 * both have been written but rolled back we only temporarily need 5091 * one of them to fix the bits when the cg write completes. 5092 */ 5093 jnewblk->jn_state |= ATTACHED | COMPLETE; 5094 njnewblk->jn_oldfrags = jnewblk->jn_oldfrags; 5095 cancel_jnewblk(jnewblk, wkhd); 5096 WORKLIST_REMOVE(&jnewblk->jn_list); 5097 free_jnewblk(jnewblk); 5098 return (new); 5099} 5100 5101/* 5102 * Replace an old allocdirect dependency with a newer one. 5103 * This routine must be called with splbio interrupts blocked. 5104 */ 5105static void 5106allocdirect_merge(adphead, newadp, oldadp) 5107 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 5108 struct allocdirect *newadp; /* allocdirect being added */ 5109 struct allocdirect *oldadp; /* existing allocdirect being checked */ 5110{ 5111 struct worklist *wk; 5112 struct freefrag *freefrag; 5113 5114 freefrag = NULL; 5115 mtx_assert(&lk, MA_OWNED); 5116 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 5117 newadp->ad_oldsize != oldadp->ad_newsize || 5118 newadp->ad_offset >= NDADDR) 5119 panic("%s %jd != new %jd || old size %ld != new %ld", 5120 "allocdirect_merge: old blkno", 5121 (intmax_t)newadp->ad_oldblkno, 5122 (intmax_t)oldadp->ad_newblkno, 5123 newadp->ad_oldsize, oldadp->ad_newsize); 5124 newadp->ad_oldblkno = oldadp->ad_oldblkno; 5125 newadp->ad_oldsize = oldadp->ad_oldsize; 5126 /* 5127 * If the old dependency had a fragment to free or had never 5128 * previously had a block allocated, then the new dependency 5129 * can immediately post its freefrag and adopt the old freefrag. 5130 * This action is done by swapping the freefrag dependencies. 5131 * The new dependency gains the old one's freefrag, and the 5132 * old one gets the new one and then immediately puts it on 5133 * the worklist when it is freed by free_newblk. It is 5134 * not possible to do this swap when the old dependency had a 5135 * non-zero size but no previous fragment to free. This condition 5136 * arises when the new block is an extension of the old block. 5137 * Here, the first part of the fragment allocated to the new 5138 * dependency is part of the block currently claimed on disk by 5139 * the old dependency, so cannot legitimately be freed until the 5140 * conditions for the new dependency are fulfilled. 5141 */ 5142 freefrag = newadp->ad_freefrag; 5143 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 5144 newadp->ad_freefrag = oldadp->ad_freefrag; 5145 oldadp->ad_freefrag = freefrag; 5146 } 5147 /* 5148 * If we are tracking a new directory-block allocation, 5149 * move it from the old allocdirect to the new allocdirect. 5150 */ 5151 if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) { 5152 WORKLIST_REMOVE(wk); 5153 if (!LIST_EMPTY(&oldadp->ad_newdirblk)) 5154 panic("allocdirect_merge: extra newdirblk"); 5155 WORKLIST_INSERT(&newadp->ad_newdirblk, wk); 5156 } 5157 TAILQ_REMOVE(adphead, oldadp, ad_next); 5158 /* 5159 * We need to move any journal dependencies over to the freefrag 5160 * that releases this block if it exists. Otherwise we are 5161 * extending an existing block and we'll wait until that is 5162 * complete to release the journal space and extend the 5163 * new journal to cover this old space as well. 5164 */ 5165 if (freefrag == NULL) { 5166 if (oldadp->ad_newblkno != newadp->ad_newblkno) 5167 panic("allocdirect_merge: %jd != %jd", 5168 oldadp->ad_newblkno, newadp->ad_newblkno); 5169 newadp->ad_block.nb_jnewblk = (struct jnewblk *) 5170 jnewblk_merge(&newadp->ad_block.nb_jnewblk->jn_list, 5171 &oldadp->ad_block.nb_jnewblk->jn_list, 5172 &newadp->ad_block.nb_jwork); 5173 oldadp->ad_block.nb_jnewblk = NULL; 5174 cancel_newblk(&oldadp->ad_block, NULL, 5175 &newadp->ad_block.nb_jwork); 5176 } else { 5177 wk = (struct worklist *) cancel_newblk(&oldadp->ad_block, 5178 &freefrag->ff_list, &freefrag->ff_jwork); 5179 freefrag->ff_jdep = jnewblk_merge(freefrag->ff_jdep, wk, 5180 &freefrag->ff_jwork); 5181 } 5182 free_newblk(&oldadp->ad_block); 5183} 5184 5185/* 5186 * Allocate a jfreefrag structure to journal a single block free. 5187 */ 5188static struct jfreefrag * 5189newjfreefrag(freefrag, ip, blkno, size, lbn) 5190 struct freefrag *freefrag; 5191 struct inode *ip; 5192 ufs2_daddr_t blkno; 5193 long size; 5194 ufs_lbn_t lbn; 5195{ 5196 struct jfreefrag *jfreefrag; 5197 struct fs *fs; 5198 5199 fs = ip->i_fs; 5200 jfreefrag = malloc(sizeof(struct jfreefrag), M_JFREEFRAG, 5201 M_SOFTDEP_FLAGS); 5202 workitem_alloc(&jfreefrag->fr_list, D_JFREEFRAG, UFSTOVFS(ip->i_ump)); 5203 jfreefrag->fr_jsegdep = newjsegdep(&jfreefrag->fr_list); 5204 jfreefrag->fr_state = ATTACHED | DEPCOMPLETE; 5205 jfreefrag->fr_ino = ip->i_number; 5206 jfreefrag->fr_lbn = lbn; 5207 jfreefrag->fr_blkno = blkno; 5208 jfreefrag->fr_frags = numfrags(fs, size); 5209 jfreefrag->fr_freefrag = freefrag; 5210 5211 return (jfreefrag); 5212} 5213 5214/* 5215 * Allocate a new freefrag structure. 5216 */ 5217static struct freefrag * 5218newfreefrag(ip, blkno, size, lbn) 5219 struct inode *ip; 5220 ufs2_daddr_t blkno; 5221 long size; 5222 ufs_lbn_t lbn; 5223{ 5224 struct freefrag *freefrag; 5225 struct fs *fs; 5226 5227 fs = ip->i_fs; 5228 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 5229 panic("newfreefrag: frag size"); 5230 freefrag = malloc(sizeof(struct freefrag), 5231 M_FREEFRAG, M_SOFTDEP_FLAGS); 5232 workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ip->i_ump)); 5233 freefrag->ff_state = ATTACHED; 5234 LIST_INIT(&freefrag->ff_jwork); 5235 freefrag->ff_inum = ip->i_number; 5236 freefrag->ff_vtype = ITOV(ip)->v_type; 5237 freefrag->ff_blkno = blkno; 5238 freefrag->ff_fragsize = size; 5239 5240 if (MOUNTEDSUJ(UFSTOVFS(ip->i_ump))) { 5241 freefrag->ff_jdep = (struct worklist *) 5242 newjfreefrag(freefrag, ip, blkno, size, lbn); 5243 } else { 5244 freefrag->ff_state |= DEPCOMPLETE; 5245 freefrag->ff_jdep = NULL; 5246 } 5247 5248 return (freefrag); 5249} 5250 5251/* 5252 * This workitem de-allocates fragments that were replaced during 5253 * file block allocation. 5254 */ 5255static void 5256handle_workitem_freefrag(freefrag) 5257 struct freefrag *freefrag; 5258{ 5259 struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp); 5260 struct workhead wkhd; 5261 5262 /* 5263 * It would be illegal to add new completion items to the 5264 * freefrag after it was schedule to be done so it must be 5265 * safe to modify the list head here. 5266 */ 5267 LIST_INIT(&wkhd); 5268 ACQUIRE_LOCK(&lk); 5269 LIST_SWAP(&freefrag->ff_jwork, &wkhd, worklist, wk_list); 5270 /* 5271 * If the journal has not been written we must cancel it here. 5272 */ 5273 if (freefrag->ff_jdep) { 5274 if (freefrag->ff_jdep->wk_type != D_JNEWBLK) 5275 panic("handle_workitem_freefrag: Unexpected type %d\n", 5276 freefrag->ff_jdep->wk_type); 5277 cancel_jnewblk(WK_JNEWBLK(freefrag->ff_jdep), &wkhd); 5278 } 5279 FREE_LOCK(&lk); 5280 ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno, 5281 freefrag->ff_fragsize, freefrag->ff_inum, freefrag->ff_vtype, &wkhd); 5282 ACQUIRE_LOCK(&lk); 5283 WORKITEM_FREE(freefrag, D_FREEFRAG); 5284 FREE_LOCK(&lk); 5285} 5286 5287/* 5288 * Set up a dependency structure for an external attributes data block. 5289 * This routine follows much of the structure of softdep_setup_allocdirect. 5290 * See the description of softdep_setup_allocdirect above for details. 5291 */ 5292void 5293softdep_setup_allocext(ip, off, newblkno, oldblkno, newsize, oldsize, bp) 5294 struct inode *ip; 5295 ufs_lbn_t off; 5296 ufs2_daddr_t newblkno; 5297 ufs2_daddr_t oldblkno; 5298 long newsize; 5299 long oldsize; 5300 struct buf *bp; 5301{ 5302 struct allocdirect *adp, *oldadp; 5303 struct allocdirectlst *adphead; 5304 struct freefrag *freefrag; 5305 struct inodedep *inodedep; 5306 struct jnewblk *jnewblk; 5307 struct newblk *newblk; 5308 struct mount *mp; 5309 ufs_lbn_t lbn; 5310 5311 if (off >= NXADDR) 5312 panic("softdep_setup_allocext: lbn %lld > NXADDR", 5313 (long long)off); 5314 5315 lbn = bp->b_lblkno; 5316 mp = UFSTOVFS(ip->i_ump); 5317 if (oldblkno && oldblkno != newblkno) 5318 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn); 5319 else 5320 freefrag = NULL; 5321 5322 ACQUIRE_LOCK(&lk); 5323 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0) 5324 panic("softdep_setup_allocext: lost block"); 5325 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5326 ("softdep_setup_allocext: newblk already initialized")); 5327 /* 5328 * Convert the newblk to an allocdirect. 5329 */ 5330 newblk->nb_list.wk_type = D_ALLOCDIRECT; 5331 adp = (struct allocdirect *)newblk; 5332 newblk->nb_freefrag = freefrag; 5333 adp->ad_offset = off; 5334 adp->ad_oldblkno = oldblkno; 5335 adp->ad_newsize = newsize; 5336 adp->ad_oldsize = oldsize; 5337 adp->ad_state |= EXTDATA; 5338 5339 /* 5340 * Finish initializing the journal. 5341 */ 5342 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5343 jnewblk->jn_ino = ip->i_number; 5344 jnewblk->jn_lbn = lbn; 5345 add_to_journal(&jnewblk->jn_list); 5346 } 5347 if (freefrag && freefrag->ff_jdep != NULL && 5348 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5349 add_to_journal(freefrag->ff_jdep); 5350 inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep); 5351 adp->ad_inodedep = inodedep; 5352 5353 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list); 5354 /* 5355 * The list of allocdirects must be kept in sorted and ascending 5356 * order so that the rollback routines can quickly determine the 5357 * first uncommitted block (the size of the file stored on disk 5358 * ends at the end of the lowest committed fragment, or if there 5359 * are no fragments, at the end of the highest committed block). 5360 * Since files generally grow, the typical case is that the new 5361 * block is to be added at the end of the list. We speed this 5362 * special case by checking against the last allocdirect in the 5363 * list before laboriously traversing the list looking for the 5364 * insertion point. 5365 */ 5366 adphead = &inodedep->id_newextupdt; 5367 oldadp = TAILQ_LAST(adphead, allocdirectlst); 5368 if (oldadp == NULL || oldadp->ad_offset <= off) { 5369 /* insert at end of list */ 5370 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 5371 if (oldadp != NULL && oldadp->ad_offset == off) 5372 allocdirect_merge(adphead, adp, oldadp); 5373 FREE_LOCK(&lk); 5374 return; 5375 } 5376 TAILQ_FOREACH(oldadp, adphead, ad_next) { 5377 if (oldadp->ad_offset >= off) 5378 break; 5379 } 5380 if (oldadp == NULL) 5381 panic("softdep_setup_allocext: lost entry"); 5382 /* insert in middle of list */ 5383 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 5384 if (oldadp->ad_offset == off) 5385 allocdirect_merge(adphead, adp, oldadp); 5386 FREE_LOCK(&lk); 5387} 5388 5389/* 5390 * Indirect block allocation dependencies. 5391 * 5392 * The same dependencies that exist for a direct block also exist when 5393 * a new block is allocated and pointed to by an entry in a block of 5394 * indirect pointers. The undo/redo states described above are also 5395 * used here. Because an indirect block contains many pointers that 5396 * may have dependencies, a second copy of the entire in-memory indirect 5397 * block is kept. The buffer cache copy is always completely up-to-date. 5398 * The second copy, which is used only as a source for disk writes, 5399 * contains only the safe pointers (i.e., those that have no remaining 5400 * update dependencies). The second copy is freed when all pointers 5401 * are safe. The cache is not allowed to replace indirect blocks with 5402 * pending update dependencies. If a buffer containing an indirect 5403 * block with dependencies is written, these routines will mark it 5404 * dirty again. It can only be successfully written once all the 5405 * dependencies are removed. The ffs_fsync routine in conjunction with 5406 * softdep_sync_metadata work together to get all the dependencies 5407 * removed so that a file can be successfully written to disk. Three 5408 * procedures are used when setting up indirect block pointer 5409 * dependencies. The division is necessary because of the organization 5410 * of the "balloc" routine and because of the distinction between file 5411 * pages and file metadata blocks. 5412 */ 5413 5414/* 5415 * Allocate a new allocindir structure. 5416 */ 5417static struct allocindir * 5418newallocindir(ip, ptrno, newblkno, oldblkno, lbn) 5419 struct inode *ip; /* inode for file being extended */ 5420 int ptrno; /* offset of pointer in indirect block */ 5421 ufs2_daddr_t newblkno; /* disk block number being added */ 5422 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 5423 ufs_lbn_t lbn; 5424{ 5425 struct newblk *newblk; 5426 struct allocindir *aip; 5427 struct freefrag *freefrag; 5428 struct jnewblk *jnewblk; 5429 5430 if (oldblkno) 5431 freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize, lbn); 5432 else 5433 freefrag = NULL; 5434 ACQUIRE_LOCK(&lk); 5435 if (newblk_lookup(UFSTOVFS(ip->i_ump), newblkno, 0, &newblk) == 0) 5436 panic("new_allocindir: lost block"); 5437 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5438 ("newallocindir: newblk already initialized")); 5439 newblk->nb_list.wk_type = D_ALLOCINDIR; 5440 newblk->nb_freefrag = freefrag; 5441 aip = (struct allocindir *)newblk; 5442 aip->ai_offset = ptrno; 5443 aip->ai_oldblkno = oldblkno; 5444 aip->ai_lbn = lbn; 5445 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5446 jnewblk->jn_ino = ip->i_number; 5447 jnewblk->jn_lbn = lbn; 5448 add_to_journal(&jnewblk->jn_list); 5449 } 5450 if (freefrag && freefrag->ff_jdep != NULL && 5451 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5452 add_to_journal(freefrag->ff_jdep); 5453 return (aip); 5454} 5455 5456/* 5457 * Called just before setting an indirect block pointer 5458 * to a newly allocated file page. 5459 */ 5460void 5461softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 5462 struct inode *ip; /* inode for file being extended */ 5463 ufs_lbn_t lbn; /* allocated block number within file */ 5464 struct buf *bp; /* buffer with indirect blk referencing page */ 5465 int ptrno; /* offset of pointer in indirect block */ 5466 ufs2_daddr_t newblkno; /* disk block number being added */ 5467 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 5468 struct buf *nbp; /* buffer holding allocated page */ 5469{ 5470 struct inodedep *inodedep; 5471 struct freefrag *freefrag; 5472 struct allocindir *aip; 5473 struct pagedep *pagedep; 5474 struct mount *mp; 5475 int dflags; 5476 5477 if (lbn != nbp->b_lblkno) 5478 panic("softdep_setup_allocindir_page: lbn %jd != lblkno %jd", 5479 lbn, bp->b_lblkno); 5480 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page"); 5481 mp = UFSTOVFS(ip->i_ump); 5482 aip = newallocindir(ip, ptrno, newblkno, oldblkno, lbn); 5483 dflags = DEPALLOC; 5484 if (IS_SNAPSHOT(ip)) 5485 dflags |= NODELAY; 5486 (void) inodedep_lookup(mp, ip->i_number, dflags, &inodedep); 5487 /* 5488 * If we are allocating a directory page, then we must 5489 * allocate an associated pagedep to track additions and 5490 * deletions. 5491 */ 5492 if ((ip->i_mode & IFMT) == IFDIR) 5493 pagedep_lookup(mp, nbp, ip->i_number, lbn, DEPALLOC, &pagedep); 5494 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list); 5495 freefrag = setup_allocindir_phase2(bp, ip, inodedep, aip, lbn); 5496 FREE_LOCK(&lk); 5497 if (freefrag) 5498 handle_workitem_freefrag(freefrag); 5499} 5500 5501/* 5502 * Called just before setting an indirect block pointer to a 5503 * newly allocated indirect block. 5504 */ 5505void 5506softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 5507 struct buf *nbp; /* newly allocated indirect block */ 5508 struct inode *ip; /* inode for file being extended */ 5509 struct buf *bp; /* indirect block referencing allocated block */ 5510 int ptrno; /* offset of pointer in indirect block */ 5511 ufs2_daddr_t newblkno; /* disk block number being added */ 5512{ 5513 struct inodedep *inodedep; 5514 struct allocindir *aip; 5515 ufs_lbn_t lbn; 5516 int dflags; 5517 5518 lbn = nbp->b_lblkno; 5519 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta"); 5520 aip = newallocindir(ip, ptrno, newblkno, 0, lbn); 5521 dflags = DEPALLOC; 5522 if (IS_SNAPSHOT(ip)) 5523 dflags |= NODELAY; 5524 inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, dflags, &inodedep); 5525 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list); 5526 if (setup_allocindir_phase2(bp, ip, inodedep, aip, lbn)) 5527 panic("softdep_setup_allocindir_meta: Block already existed"); 5528 FREE_LOCK(&lk); 5529} 5530 5531static void 5532indirdep_complete(indirdep) 5533 struct indirdep *indirdep; 5534{ 5535 struct allocindir *aip; 5536 5537 LIST_REMOVE(indirdep, ir_next); 5538 indirdep->ir_state |= DEPCOMPLETE; 5539 5540 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) { 5541 LIST_REMOVE(aip, ai_next); 5542 free_newblk(&aip->ai_block); 5543 } 5544 /* 5545 * If this indirdep is not attached to a buf it was simply waiting 5546 * on completion to clear completehd. free_indirdep() asserts 5547 * that nothing is dangling. 5548 */ 5549 if ((indirdep->ir_state & ONWORKLIST) == 0) 5550 free_indirdep(indirdep); 5551} 5552 5553static struct indirdep * 5554indirdep_lookup(mp, ip, bp) 5555 struct mount *mp; 5556 struct inode *ip; 5557 struct buf *bp; 5558{ 5559 struct indirdep *indirdep, *newindirdep; 5560 struct newblk *newblk; 5561 struct worklist *wk; 5562 struct fs *fs; 5563 ufs2_daddr_t blkno; 5564 5565 mtx_assert(&lk, MA_OWNED); 5566 indirdep = NULL; 5567 newindirdep = NULL; 5568 fs = ip->i_fs; 5569 for (;;) { 5570 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5571 if (wk->wk_type != D_INDIRDEP) 5572 continue; 5573 indirdep = WK_INDIRDEP(wk); 5574 break; 5575 } 5576 /* Found on the buffer worklist, no new structure to free. */ 5577 if (indirdep != NULL && newindirdep == NULL) 5578 return (indirdep); 5579 if (indirdep != NULL && newindirdep != NULL) 5580 panic("indirdep_lookup: simultaneous create"); 5581 /* None found on the buffer and a new structure is ready. */ 5582 if (indirdep == NULL && newindirdep != NULL) 5583 break; 5584 /* None found and no new structure available. */ 5585 FREE_LOCK(&lk); 5586 newindirdep = malloc(sizeof(struct indirdep), 5587 M_INDIRDEP, M_SOFTDEP_FLAGS); 5588 workitem_alloc(&newindirdep->ir_list, D_INDIRDEP, mp); 5589 newindirdep->ir_state = ATTACHED; 5590 if (ip->i_ump->um_fstype == UFS1) 5591 newindirdep->ir_state |= UFS1FMT; 5592 TAILQ_INIT(&newindirdep->ir_trunc); 5593 newindirdep->ir_saveddata = NULL; 5594 LIST_INIT(&newindirdep->ir_deplisthd); 5595 LIST_INIT(&newindirdep->ir_donehd); 5596 LIST_INIT(&newindirdep->ir_writehd); 5597 LIST_INIT(&newindirdep->ir_completehd); 5598 if (bp->b_blkno == bp->b_lblkno) { 5599 ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp, 5600 NULL, NULL); 5601 bp->b_blkno = blkno; 5602 } 5603 newindirdep->ir_freeblks = NULL; 5604 newindirdep->ir_savebp = 5605 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0); 5606 newindirdep->ir_bp = bp; 5607 BUF_KERNPROC(newindirdep->ir_savebp); 5608 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 5609 ACQUIRE_LOCK(&lk); 5610 } 5611 indirdep = newindirdep; 5612 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 5613 /* 5614 * If the block is not yet allocated we don't set DEPCOMPLETE so 5615 * that we don't free dependencies until the pointers are valid. 5616 * This could search b_dep for D_ALLOCDIRECT/D_ALLOCINDIR rather 5617 * than using the hash. 5618 */ 5619 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk)) 5620 LIST_INSERT_HEAD(&newblk->nb_indirdeps, indirdep, ir_next); 5621 else 5622 indirdep->ir_state |= DEPCOMPLETE; 5623 return (indirdep); 5624} 5625 5626/* 5627 * Called to finish the allocation of the "aip" allocated 5628 * by one of the two routines above. 5629 */ 5630static struct freefrag * 5631setup_allocindir_phase2(bp, ip, inodedep, aip, lbn) 5632 struct buf *bp; /* in-memory copy of the indirect block */ 5633 struct inode *ip; /* inode for file being extended */ 5634 struct inodedep *inodedep; /* Inodedep for ip */ 5635 struct allocindir *aip; /* allocindir allocated by the above routines */ 5636 ufs_lbn_t lbn; /* Logical block number for this block. */ 5637{ 5638 struct fs *fs; 5639 struct indirdep *indirdep; 5640 struct allocindir *oldaip; 5641 struct freefrag *freefrag; 5642 struct mount *mp; 5643 5644 mtx_assert(&lk, MA_OWNED); 5645 mp = UFSTOVFS(ip->i_ump); 5646 fs = ip->i_fs; 5647 if (bp->b_lblkno >= 0) 5648 panic("setup_allocindir_phase2: not indir blk"); 5649 KASSERT(aip->ai_offset >= 0 && aip->ai_offset < NINDIR(fs), 5650 ("setup_allocindir_phase2: Bad offset %d", aip->ai_offset)); 5651 indirdep = indirdep_lookup(mp, ip, bp); 5652 KASSERT(indirdep->ir_savebp != NULL, 5653 ("setup_allocindir_phase2 NULL ir_savebp")); 5654 aip->ai_indirdep = indirdep; 5655 /* 5656 * Check for an unwritten dependency for this indirect offset. If 5657 * there is, merge the old dependency into the new one. This happens 5658 * as a result of reallocblk only. 5659 */ 5660 freefrag = NULL; 5661 if (aip->ai_oldblkno != 0) { 5662 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) { 5663 if (oldaip->ai_offset == aip->ai_offset) { 5664 freefrag = allocindir_merge(aip, oldaip); 5665 goto done; 5666 } 5667 } 5668 LIST_FOREACH(oldaip, &indirdep->ir_donehd, ai_next) { 5669 if (oldaip->ai_offset == aip->ai_offset) { 5670 freefrag = allocindir_merge(aip, oldaip); 5671 goto done; 5672 } 5673 } 5674 } 5675done: 5676 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 5677 return (freefrag); 5678} 5679 5680/* 5681 * Merge two allocindirs which refer to the same block. Move newblock 5682 * dependencies and setup the freefrags appropriately. 5683 */ 5684static struct freefrag * 5685allocindir_merge(aip, oldaip) 5686 struct allocindir *aip; 5687 struct allocindir *oldaip; 5688{ 5689 struct freefrag *freefrag; 5690 struct worklist *wk; 5691 5692 if (oldaip->ai_newblkno != aip->ai_oldblkno) 5693 panic("allocindir_merge: blkno"); 5694 aip->ai_oldblkno = oldaip->ai_oldblkno; 5695 freefrag = aip->ai_freefrag; 5696 aip->ai_freefrag = oldaip->ai_freefrag; 5697 oldaip->ai_freefrag = NULL; 5698 KASSERT(freefrag != NULL, ("setup_allocindir_phase2: No freefrag")); 5699 /* 5700 * If we are tracking a new directory-block allocation, 5701 * move it from the old allocindir to the new allocindir. 5702 */ 5703 if ((wk = LIST_FIRST(&oldaip->ai_newdirblk)) != NULL) { 5704 WORKLIST_REMOVE(wk); 5705 if (!LIST_EMPTY(&oldaip->ai_newdirblk)) 5706 panic("allocindir_merge: extra newdirblk"); 5707 WORKLIST_INSERT(&aip->ai_newdirblk, wk); 5708 } 5709 /* 5710 * We can skip journaling for this freefrag and just complete 5711 * any pending journal work for the allocindir that is being 5712 * removed after the freefrag completes. 5713 */ 5714 if (freefrag->ff_jdep) 5715 cancel_jfreefrag(WK_JFREEFRAG(freefrag->ff_jdep)); 5716 LIST_REMOVE(oldaip, ai_next); 5717 freefrag->ff_jdep = (struct worklist *)cancel_newblk(&oldaip->ai_block, 5718 &freefrag->ff_list, &freefrag->ff_jwork); 5719 free_newblk(&oldaip->ai_block); 5720 5721 return (freefrag); 5722} 5723 5724static inline void 5725setup_freedirect(freeblks, ip, i, needj) 5726 struct freeblks *freeblks; 5727 struct inode *ip; 5728 int i; 5729 int needj; 5730{ 5731 ufs2_daddr_t blkno; 5732 int frags; 5733 5734 blkno = DIP(ip, i_db[i]); 5735 if (blkno == 0) 5736 return; 5737 DIP_SET(ip, i_db[i], 0); 5738 frags = sblksize(ip->i_fs, ip->i_size, i); 5739 frags = numfrags(ip->i_fs, frags); 5740 newfreework(ip->i_ump, freeblks, NULL, i, blkno, frags, 0, needj); 5741} 5742 5743static inline void 5744setup_freeext(freeblks, ip, i, needj) 5745 struct freeblks *freeblks; 5746 struct inode *ip; 5747 int i; 5748 int needj; 5749{ 5750 ufs2_daddr_t blkno; 5751 int frags; 5752 5753 blkno = ip->i_din2->di_extb[i]; 5754 if (blkno == 0) 5755 return; 5756 ip->i_din2->di_extb[i] = 0; 5757 frags = sblksize(ip->i_fs, ip->i_din2->di_extsize, i); 5758 frags = numfrags(ip->i_fs, frags); 5759 newfreework(ip->i_ump, freeblks, NULL, -1 - i, blkno, frags, 0, needj); 5760} 5761 5762static inline void 5763setup_freeindir(freeblks, ip, i, lbn, needj) 5764 struct freeblks *freeblks; 5765 struct inode *ip; 5766 int i; 5767 ufs_lbn_t lbn; 5768 int needj; 5769{ 5770 ufs2_daddr_t blkno; 5771 5772 blkno = DIP(ip, i_ib[i]); 5773 if (blkno == 0) 5774 return; 5775 DIP_SET(ip, i_ib[i], 0); 5776 newfreework(ip->i_ump, freeblks, NULL, lbn, blkno, ip->i_fs->fs_frag, 5777 0, needj); 5778} 5779 5780static inline struct freeblks * 5781newfreeblks(mp, ip) 5782 struct mount *mp; 5783 struct inode *ip; 5784{ 5785 struct freeblks *freeblks; 5786 5787 freeblks = malloc(sizeof(struct freeblks), 5788 M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO); 5789 workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp); 5790 LIST_INIT(&freeblks->fb_jblkdephd); 5791 LIST_INIT(&freeblks->fb_jwork); 5792 freeblks->fb_ref = 0; 5793 freeblks->fb_cgwait = 0; 5794 freeblks->fb_state = ATTACHED; 5795 freeblks->fb_uid = ip->i_uid; 5796 freeblks->fb_inum = ip->i_number; 5797 freeblks->fb_vtype = ITOV(ip)->v_type; 5798 freeblks->fb_modrev = DIP(ip, i_modrev); 5799 freeblks->fb_devvp = ip->i_devvp; 5800 freeblks->fb_chkcnt = 0; 5801 freeblks->fb_len = 0; 5802 5803 return (freeblks); 5804} 5805 5806static void 5807trunc_indirdep(indirdep, freeblks, bp, off) 5808 struct indirdep *indirdep; 5809 struct freeblks *freeblks; 5810 struct buf *bp; 5811 int off; 5812{ 5813 struct allocindir *aip, *aipn; 5814 5815 /* 5816 * The first set of allocindirs won't be in savedbp. 5817 */ 5818 LIST_FOREACH_SAFE(aip, &indirdep->ir_deplisthd, ai_next, aipn) 5819 if (aip->ai_offset > off) 5820 cancel_allocindir(aip, bp, freeblks, 1); 5821 LIST_FOREACH_SAFE(aip, &indirdep->ir_donehd, ai_next, aipn) 5822 if (aip->ai_offset > off) 5823 cancel_allocindir(aip, bp, freeblks, 1); 5824 /* 5825 * These will exist in savedbp. 5826 */ 5827 LIST_FOREACH_SAFE(aip, &indirdep->ir_writehd, ai_next, aipn) 5828 if (aip->ai_offset > off) 5829 cancel_allocindir(aip, NULL, freeblks, 0); 5830 LIST_FOREACH_SAFE(aip, &indirdep->ir_completehd, ai_next, aipn) 5831 if (aip->ai_offset > off) 5832 cancel_allocindir(aip, NULL, freeblks, 0); 5833} 5834 5835/* 5836 * Follow the chain of indirects down to lastlbn creating a freework 5837 * structure for each. This will be used to start indir_trunc() at 5838 * the right offset and create the journal records for the parrtial 5839 * truncation. A second step will handle the truncated dependencies. 5840 */ 5841static int 5842setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno) 5843 struct freeblks *freeblks; 5844 struct inode *ip; 5845 ufs_lbn_t lbn; 5846 ufs_lbn_t lastlbn; 5847 ufs2_daddr_t blkno; 5848{ 5849 struct indirdep *indirdep; 5850 struct indirdep *indirn; 5851 struct freework *freework; 5852 struct newblk *newblk; 5853 struct mount *mp; 5854 struct buf *bp; 5855 uint8_t *start; 5856 uint8_t *end; 5857 ufs_lbn_t lbnadd; 5858 int level; 5859 int error; 5860 int off; 5861 5862 5863 freework = NULL; 5864 if (blkno == 0) 5865 return (0); 5866 mp = freeblks->fb_list.wk_mp; 5867 bp = getblk(ITOV(ip), lbn, mp->mnt_stat.f_iosize, 0, 0, 0); 5868 if ((bp->b_flags & B_CACHE) == 0) { 5869 bp->b_blkno = blkptrtodb(VFSTOUFS(mp), blkno); 5870 bp->b_iocmd = BIO_READ; 5871 bp->b_flags &= ~B_INVAL; 5872 bp->b_ioflags &= ~BIO_ERROR; 5873 vfs_busy_pages(bp, 0); 5874 bp->b_iooffset = dbtob(bp->b_blkno); 5875 bstrategy(bp); 5876 curthread->td_ru.ru_inblock++; 5877 error = bufwait(bp); 5878 if (error) { 5879 brelse(bp); 5880 return (error); 5881 } 5882 } 5883 level = lbn_level(lbn); 5884 lbnadd = lbn_offset(ip->i_fs, level); 5885 /* 5886 * Compute the offset of the last block we want to keep. Store 5887 * in the freework the first block we want to completely free. 5888 */ 5889 off = (lastlbn - -(lbn + level)) / lbnadd; 5890 if (off + 1 == NINDIR(ip->i_fs)) 5891 goto nowork; 5892 freework = newfreework(ip->i_ump, freeblks, NULL, lbn, blkno, 0, off+1, 5893 0); 5894 /* 5895 * Link the freework into the indirdep. This will prevent any new 5896 * allocations from proceeding until we are finished with the 5897 * truncate and the block is written. 5898 */ 5899 ACQUIRE_LOCK(&lk); 5900 indirdep = indirdep_lookup(mp, ip, bp); 5901 if (indirdep->ir_freeblks) 5902 panic("setup_trunc_indir: indirdep already truncated."); 5903 TAILQ_INSERT_TAIL(&indirdep->ir_trunc, freework, fw_next); 5904 freework->fw_indir = indirdep; 5905 /* 5906 * Cancel any allocindirs that will not make it to disk. 5907 * We have to do this for all copies of the indirdep that 5908 * live on this newblk. 5909 */ 5910 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 5911 newblk_lookup(mp, dbtofsb(ip->i_fs, bp->b_blkno), 0, &newblk); 5912 LIST_FOREACH(indirn, &newblk->nb_indirdeps, ir_next) 5913 trunc_indirdep(indirn, freeblks, bp, off); 5914 } else 5915 trunc_indirdep(indirdep, freeblks, bp, off); 5916 FREE_LOCK(&lk); 5917 /* 5918 * Creation is protected by the buf lock. The saveddata is only 5919 * needed if a full truncation follows a partial truncation but it 5920 * is difficult to allocate in that case so we fetch it anyway. 5921 */ 5922 if (indirdep->ir_saveddata == NULL) 5923 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP, 5924 M_SOFTDEP_FLAGS); 5925nowork: 5926 /* Fetch the blkno of the child and the zero start offset. */ 5927 if (ip->i_ump->um_fstype == UFS1) { 5928 blkno = ((ufs1_daddr_t *)bp->b_data)[off]; 5929 start = (uint8_t *)&((ufs1_daddr_t *)bp->b_data)[off+1]; 5930 } else { 5931 blkno = ((ufs2_daddr_t *)bp->b_data)[off]; 5932 start = (uint8_t *)&((ufs2_daddr_t *)bp->b_data)[off+1]; 5933 } 5934 if (freework) { 5935 /* Zero the truncated pointers. */ 5936 end = bp->b_data + bp->b_bcount; 5937 bzero(start, end - start); 5938 bdwrite(bp); 5939 } else 5940 bqrelse(bp); 5941 if (level == 0) 5942 return (0); 5943 lbn++; /* adjust level */ 5944 lbn -= (off * lbnadd); 5945 return setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno); 5946} 5947 5948/* 5949 * Complete the partial truncation of an indirect block setup by 5950 * setup_trunc_indir(). This zeros the truncated pointers in the saved 5951 * copy and writes them to disk before the freeblks is allowed to complete. 5952 */ 5953static void 5954complete_trunc_indir(freework) 5955 struct freework *freework; 5956{ 5957 struct freework *fwn; 5958 struct indirdep *indirdep; 5959 struct buf *bp; 5960 uintptr_t start; 5961 int count; 5962 5963 indirdep = freework->fw_indir; 5964 for (;;) { 5965 bp = indirdep->ir_bp; 5966 /* See if the block was discarded. */ 5967 if (bp == NULL) 5968 break; 5969 /* Inline part of getdirtybuf(). We dont want bremfree. */ 5970 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0) 5971 break; 5972 if (BUF_LOCK(bp, 5973 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, &lk) == 0) 5974 BUF_UNLOCK(bp); 5975 ACQUIRE_LOCK(&lk); 5976 } 5977 mtx_assert(&lk, MA_OWNED); 5978 freework->fw_state |= DEPCOMPLETE; 5979 TAILQ_REMOVE(&indirdep->ir_trunc, freework, fw_next); 5980 /* 5981 * Zero the pointers in the saved copy. 5982 */ 5983 if (indirdep->ir_state & UFS1FMT) 5984 start = sizeof(ufs1_daddr_t); 5985 else 5986 start = sizeof(ufs2_daddr_t); 5987 start *= freework->fw_start; 5988 count = indirdep->ir_savebp->b_bcount - start; 5989 start += (uintptr_t)indirdep->ir_savebp->b_data; 5990 bzero((char *)start, count); 5991 /* 5992 * We need to start the next truncation in the list if it has not 5993 * been started yet. 5994 */ 5995 fwn = TAILQ_FIRST(&indirdep->ir_trunc); 5996 if (fwn != NULL) { 5997 if (fwn->fw_freeblks == indirdep->ir_freeblks) 5998 TAILQ_REMOVE(&indirdep->ir_trunc, fwn, fw_next); 5999 if ((fwn->fw_state & ONWORKLIST) == 0) 6000 freework_enqueue(fwn); 6001 } 6002 /* 6003 * If bp is NULL the block was fully truncated, restore 6004 * the saved block list otherwise free it if it is no 6005 * longer needed. 6006 */ 6007 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 6008 if (bp == NULL) 6009 bcopy(indirdep->ir_saveddata, 6010 indirdep->ir_savebp->b_data, 6011 indirdep->ir_savebp->b_bcount); 6012 free(indirdep->ir_saveddata, M_INDIRDEP); 6013 indirdep->ir_saveddata = NULL; 6014 } 6015 /* 6016 * When bp is NULL there is a full truncation pending. We 6017 * must wait for this full truncation to be journaled before 6018 * we can release this freework because the disk pointers will 6019 * never be written as zero. 6020 */ 6021 if (bp == NULL) { 6022 if (LIST_EMPTY(&indirdep->ir_freeblks->fb_jblkdephd)) 6023 handle_written_freework(freework); 6024 else 6025 WORKLIST_INSERT(&indirdep->ir_freeblks->fb_freeworkhd, 6026 &freework->fw_list); 6027 } else { 6028 /* Complete when the real copy is written. */ 6029 WORKLIST_INSERT(&bp->b_dep, &freework->fw_list); 6030 BUF_UNLOCK(bp); 6031 } 6032} 6033 6034/* 6035 * Calculate the number of blocks we are going to release where datablocks 6036 * is the current total and length is the new file size. 6037 */ 6038ufs2_daddr_t 6039blkcount(fs, datablocks, length) 6040 struct fs *fs; 6041 ufs2_daddr_t datablocks; 6042 off_t length; 6043{ 6044 off_t totblks, numblks; 6045 6046 totblks = 0; 6047 numblks = howmany(length, fs->fs_bsize); 6048 if (numblks <= NDADDR) { 6049 totblks = howmany(length, fs->fs_fsize); 6050 goto out; 6051 } 6052 totblks = blkstofrags(fs, numblks); 6053 numblks -= NDADDR; 6054 /* 6055 * Count all single, then double, then triple indirects required. 6056 * Subtracting one indirects worth of blocks for each pass 6057 * acknowledges one of each pointed to by the inode. 6058 */ 6059 for (;;) { 6060 totblks += blkstofrags(fs, howmany(numblks, NINDIR(fs))); 6061 numblks -= NINDIR(fs); 6062 if (numblks <= 0) 6063 break; 6064 numblks = howmany(numblks, NINDIR(fs)); 6065 } 6066out: 6067 totblks = fsbtodb(fs, totblks); 6068 /* 6069 * Handle sparse files. We can't reclaim more blocks than the inode 6070 * references. We will correct it later in handle_complete_freeblks() 6071 * when we know the real count. 6072 */ 6073 if (totblks > datablocks) 6074 return (0); 6075 return (datablocks - totblks); 6076} 6077 6078/* 6079 * Handle freeblocks for journaled softupdate filesystems. 6080 * 6081 * Contrary to normal softupdates, we must preserve the block pointers in 6082 * indirects until their subordinates are free. This is to avoid journaling 6083 * every block that is freed which may consume more space than the journal 6084 * itself. The recovery program will see the free block journals at the 6085 * base of the truncated area and traverse them to reclaim space. The 6086 * pointers in the inode may be cleared immediately after the journal 6087 * records are written because each direct and indirect pointer in the 6088 * inode is recorded in a journal. This permits full truncation to proceed 6089 * asynchronously. The write order is journal -> inode -> cgs -> indirects. 6090 * 6091 * The algorithm is as follows: 6092 * 1) Traverse the in-memory state and create journal entries to release 6093 * the relevant blocks and full indirect trees. 6094 * 2) Traverse the indirect block chain adding partial truncation freework 6095 * records to indirects in the path to lastlbn. The freework will 6096 * prevent new allocation dependencies from being satisfied in this 6097 * indirect until the truncation completes. 6098 * 3) Read and lock the inode block, performing an update with the new size 6099 * and pointers. This prevents truncated data from becoming valid on 6100 * disk through step 4. 6101 * 4) Reap unsatisfied dependencies that are beyond the truncated area, 6102 * eliminate journal work for those records that do not require it. 6103 * 5) Schedule the journal records to be written followed by the inode block. 6104 * 6) Allocate any necessary frags for the end of file. 6105 * 7) Zero any partially truncated blocks. 6106 * 6107 * From this truncation proceeds asynchronously using the freework and 6108 * indir_trunc machinery. The file will not be extended again into a 6109 * partially truncated indirect block until all work is completed but 6110 * the normal dependency mechanism ensures that it is rolled back/forward 6111 * as appropriate. Further truncation may occur without delay and is 6112 * serialized in indir_trunc(). 6113 */ 6114void 6115softdep_journal_freeblocks(ip, cred, length, flags) 6116 struct inode *ip; /* The inode whose length is to be reduced */ 6117 struct ucred *cred; 6118 off_t length; /* The new length for the file */ 6119 int flags; /* IO_EXT and/or IO_NORMAL */ 6120{ 6121 struct freeblks *freeblks, *fbn; 6122 struct inodedep *inodedep; 6123 struct jblkdep *jblkdep; 6124 struct allocdirect *adp, *adpn; 6125 struct fs *fs; 6126 struct buf *bp; 6127 struct vnode *vp; 6128 struct mount *mp; 6129 ufs2_daddr_t extblocks, datablocks; 6130 ufs_lbn_t tmpval, lbn, lastlbn; 6131 int frags, lastoff, iboff, allocblock, needj, dflags, error, i; 6132 6133 fs = ip->i_fs; 6134 mp = UFSTOVFS(ip->i_ump); 6135 vp = ITOV(ip); 6136 needj = 1; 6137 iboff = -1; 6138 allocblock = 0; 6139 extblocks = 0; 6140 datablocks = 0; 6141 frags = 0; 6142 freeblks = newfreeblks(mp, ip); 6143 ACQUIRE_LOCK(&lk); 6144 /* 6145 * If we're truncating a removed file that will never be written 6146 * we don't need to journal the block frees. The canceled journals 6147 * for the allocations will suffice. 6148 */ 6149 dflags = DEPALLOC; 6150 if (IS_SNAPSHOT(ip)) 6151 dflags |= NODELAY; 6152 inodedep_lookup(mp, ip->i_number, dflags, &inodedep); 6153 if ((inodedep->id_state & (UNLINKED | DEPCOMPLETE)) == UNLINKED && 6154 length == 0) 6155 needj = 0; 6156 FREE_LOCK(&lk); 6157 /* 6158 * Calculate the lbn that we are truncating to. This results in -1 6159 * if we're truncating the 0 bytes. So it is the last lbn we want 6160 * to keep, not the first lbn we want to truncate. 6161 */ 6162 lastlbn = lblkno(fs, length + fs->fs_bsize - 1) - 1; 6163 lastoff = blkoff(fs, length); 6164 /* 6165 * Compute frags we are keeping in lastlbn. 0 means all. 6166 */ 6167 if (lastlbn >= 0 && lastlbn < NDADDR) { 6168 frags = fragroundup(fs, lastoff); 6169 /* adp offset of last valid allocdirect. */ 6170 iboff = lastlbn; 6171 } else if (lastlbn > 0) 6172 iboff = NDADDR; 6173 if (fs->fs_magic == FS_UFS2_MAGIC) 6174 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 6175 /* 6176 * Handle normal data blocks and indirects. This section saves 6177 * values used after the inode update to complete frag and indirect 6178 * truncation. 6179 */ 6180 if ((flags & IO_NORMAL) != 0) { 6181 /* 6182 * Handle truncation of whole direct and indirect blocks. 6183 */ 6184 for (i = iboff + 1; i < NDADDR; i++) 6185 setup_freedirect(freeblks, ip, i, needj); 6186 for (i = 0, tmpval = NINDIR(fs), lbn = NDADDR; i < NIADDR; 6187 i++, lbn += tmpval, tmpval *= NINDIR(fs)) { 6188 /* Release a whole indirect tree. */ 6189 if (lbn > lastlbn) { 6190 setup_freeindir(freeblks, ip, i, -lbn -i, 6191 needj); 6192 continue; 6193 } 6194 iboff = i + NDADDR; 6195 /* 6196 * Traverse partially truncated indirect tree. 6197 */ 6198 if (lbn <= lastlbn && lbn + tmpval - 1 > lastlbn) 6199 setup_trunc_indir(freeblks, ip, -lbn - i, 6200 lastlbn, DIP(ip, i_ib[i])); 6201 } 6202 /* 6203 * Handle partial truncation to a frag boundary. 6204 */ 6205 if (frags) { 6206 ufs2_daddr_t blkno; 6207 long oldfrags; 6208 6209 oldfrags = blksize(fs, ip, lastlbn); 6210 blkno = DIP(ip, i_db[lastlbn]); 6211 if (blkno && oldfrags != frags) { 6212 oldfrags -= frags; 6213 oldfrags = numfrags(ip->i_fs, oldfrags); 6214 blkno += numfrags(ip->i_fs, frags); 6215 newfreework(ip->i_ump, freeblks, NULL, lastlbn, 6216 blkno, oldfrags, 0, needj); 6217 } else if (blkno == 0) 6218 allocblock = 1; 6219 } 6220 /* 6221 * Add a journal record for partial truncate if we are 6222 * handling indirect blocks. Non-indirects need no extra 6223 * journaling. 6224 */ 6225 if (length != 0 && lastlbn >= NDADDR) { 6226 ip->i_flag |= IN_TRUNCATED; 6227 newjtrunc(freeblks, length, 0); 6228 } 6229 ip->i_size = length; 6230 DIP_SET(ip, i_size, ip->i_size); 6231 datablocks = DIP(ip, i_blocks) - extblocks; 6232 if (length != 0) 6233 datablocks = blkcount(ip->i_fs, datablocks, length); 6234 freeblks->fb_len = length; 6235 } 6236 if ((flags & IO_EXT) != 0) { 6237 for (i = 0; i < NXADDR; i++) 6238 setup_freeext(freeblks, ip, i, needj); 6239 ip->i_din2->di_extsize = 0; 6240 datablocks += extblocks; 6241 } 6242#ifdef QUOTA 6243 /* Reference the quotas in case the block count is wrong in the end. */ 6244 quotaref(vp, freeblks->fb_quota); 6245 (void) chkdq(ip, -datablocks, NOCRED, 0); 6246#endif 6247 freeblks->fb_chkcnt = -datablocks; 6248 UFS_LOCK(ip->i_ump); 6249 fs->fs_pendingblocks += datablocks; 6250 UFS_UNLOCK(ip->i_ump); 6251 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks); 6252 /* 6253 * Handle truncation of incomplete alloc direct dependencies. We 6254 * hold the inode block locked to prevent incomplete dependencies 6255 * from reaching the disk while we are eliminating those that 6256 * have been truncated. This is a partially inlined ffs_update(). 6257 */ 6258 ufs_itimes(vp); 6259 ip->i_flag &= ~(IN_LAZYACCESS | IN_LAZYMOD | IN_MODIFIED); 6260 error = bread(ip->i_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 6261 (int)fs->fs_bsize, cred, &bp); 6262 if (error) { 6263 brelse(bp); 6264 softdep_error("softdep_journal_freeblocks", error); 6265 return; 6266 } 6267 if (bp->b_bufsize == fs->fs_bsize) 6268 bp->b_flags |= B_CLUSTEROK; 6269 softdep_update_inodeblock(ip, bp, 0); 6270 if (ip->i_ump->um_fstype == UFS1) 6271 *((struct ufs1_dinode *)bp->b_data + 6272 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1; 6273 else 6274 *((struct ufs2_dinode *)bp->b_data + 6275 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2; 6276 ACQUIRE_LOCK(&lk); 6277 (void) inodedep_lookup(mp, ip->i_number, dflags, &inodedep); 6278 if ((inodedep->id_state & IOSTARTED) != 0) 6279 panic("softdep_setup_freeblocks: inode busy"); 6280 /* 6281 * Add the freeblks structure to the list of operations that 6282 * must await the zero'ed inode being written to disk. If we 6283 * still have a bitmap dependency (needj), then the inode 6284 * has never been written to disk, so we can process the 6285 * freeblks below once we have deleted the dependencies. 6286 */ 6287 if (needj) 6288 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list); 6289 else 6290 freeblks->fb_state |= COMPLETE; 6291 if ((flags & IO_NORMAL) != 0) { 6292 TAILQ_FOREACH_SAFE(adp, &inodedep->id_inoupdt, ad_next, adpn) { 6293 if (adp->ad_offset > iboff) 6294 cancel_allocdirect(&inodedep->id_inoupdt, adp, 6295 freeblks); 6296 /* 6297 * Truncate the allocdirect. We could eliminate 6298 * or modify journal records as well. 6299 */ 6300 else if (adp->ad_offset == iboff && frags) 6301 adp->ad_newsize = frags; 6302 } 6303 } 6304 if ((flags & IO_EXT) != 0) 6305 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0) 6306 cancel_allocdirect(&inodedep->id_extupdt, adp, 6307 freeblks); 6308 /* 6309 * Add journal work. 6310 */ 6311 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) 6312 add_to_journal(&jblkdep->jb_list); 6313 FREE_LOCK(&lk); 6314 bdwrite(bp); 6315 /* 6316 * Truncate dependency structures beyond length. 6317 */ 6318 trunc_dependencies(ip, freeblks, lastlbn, frags, flags); 6319 /* 6320 * This is only set when we need to allocate a fragment because 6321 * none existed at the end of a frag-sized file. It handles only 6322 * allocating a new, zero filled block. 6323 */ 6324 if (allocblock) { 6325 ip->i_size = length - lastoff; 6326 DIP_SET(ip, i_size, ip->i_size); 6327 error = UFS_BALLOC(vp, length - 1, 1, cred, BA_CLRBUF, &bp); 6328 if (error != 0) { 6329 softdep_error("softdep_journal_freeblks", error); 6330 return; 6331 } 6332 ip->i_size = length; 6333 DIP_SET(ip, i_size, length); 6334 ip->i_flag |= IN_CHANGE | IN_UPDATE; 6335 allocbuf(bp, frags); 6336 ffs_update(vp, 0); 6337 bawrite(bp); 6338 } else if (lastoff != 0 && vp->v_type != VDIR) { 6339 int size; 6340 6341 /* 6342 * Zero the end of a truncated frag or block. 6343 */ 6344 size = sblksize(fs, length, lastlbn); 6345 error = bread(vp, lastlbn, size, cred, &bp); 6346 if (error) { 6347 softdep_error("softdep_journal_freeblks", error); 6348 return; 6349 } 6350 bzero((char *)bp->b_data + lastoff, size - lastoff); 6351 bawrite(bp); 6352 6353 } 6354 ACQUIRE_LOCK(&lk); 6355 inodedep_lookup(mp, ip->i_number, dflags, &inodedep); 6356 TAILQ_INSERT_TAIL(&inodedep->id_freeblklst, freeblks, fb_next); 6357 freeblks->fb_state |= DEPCOMPLETE | ONDEPLIST; 6358 /* 6359 * We zero earlier truncations so they don't erroneously 6360 * update i_blocks. 6361 */ 6362 if (freeblks->fb_len == 0 && (flags & IO_NORMAL) != 0) 6363 TAILQ_FOREACH(fbn, &inodedep->id_freeblklst, fb_next) 6364 fbn->fb_len = 0; 6365 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE && 6366 LIST_EMPTY(&freeblks->fb_jblkdephd)) 6367 freeblks->fb_state |= INPROGRESS; 6368 else 6369 freeblks = NULL; 6370 FREE_LOCK(&lk); 6371 if (freeblks) 6372 handle_workitem_freeblocks(freeblks, 0); 6373 trunc_pages(ip, length, extblocks, flags); 6374 6375} 6376 6377/* 6378 * Flush a JOP_SYNC to the journal. 6379 */ 6380void 6381softdep_journal_fsync(ip) 6382 struct inode *ip; 6383{ 6384 struct jfsync *jfsync; 6385 6386 if ((ip->i_flag & IN_TRUNCATED) == 0) 6387 return; 6388 ip->i_flag &= ~IN_TRUNCATED; 6389 jfsync = malloc(sizeof(*jfsync), M_JFSYNC, M_SOFTDEP_FLAGS | M_ZERO); 6390 workitem_alloc(&jfsync->jfs_list, D_JFSYNC, UFSTOVFS(ip->i_ump)); 6391 jfsync->jfs_size = ip->i_size; 6392 jfsync->jfs_ino = ip->i_number; 6393 ACQUIRE_LOCK(&lk); 6394 add_to_journal(&jfsync->jfs_list); 6395 jwait(&jfsync->jfs_list, MNT_WAIT); 6396 FREE_LOCK(&lk); 6397} 6398 6399/* 6400 * Block de-allocation dependencies. 6401 * 6402 * When blocks are de-allocated, the on-disk pointers must be nullified before 6403 * the blocks are made available for use by other files. (The true 6404 * requirement is that old pointers must be nullified before new on-disk 6405 * pointers are set. We chose this slightly more stringent requirement to 6406 * reduce complexity.) Our implementation handles this dependency by updating 6407 * the inode (or indirect block) appropriately but delaying the actual block 6408 * de-allocation (i.e., freemap and free space count manipulation) until 6409 * after the updated versions reach stable storage. After the disk is 6410 * updated, the blocks can be safely de-allocated whenever it is convenient. 6411 * This implementation handles only the common case of reducing a file's 6412 * length to zero. Other cases are handled by the conventional synchronous 6413 * write approach. 6414 * 6415 * The ffs implementation with which we worked double-checks 6416 * the state of the block pointers and file size as it reduces 6417 * a file's length. Some of this code is replicated here in our 6418 * soft updates implementation. The freeblks->fb_chkcnt field is 6419 * used to transfer a part of this information to the procedure 6420 * that eventually de-allocates the blocks. 6421 * 6422 * This routine should be called from the routine that shortens 6423 * a file's length, before the inode's size or block pointers 6424 * are modified. It will save the block pointer information for 6425 * later release and zero the inode so that the calling routine 6426 * can release it. 6427 */ 6428void 6429softdep_setup_freeblocks(ip, length, flags) 6430 struct inode *ip; /* The inode whose length is to be reduced */ 6431 off_t length; /* The new length for the file */ 6432 int flags; /* IO_EXT and/or IO_NORMAL */ 6433{ 6434 struct ufs1_dinode *dp1; 6435 struct ufs2_dinode *dp2; 6436 struct freeblks *freeblks; 6437 struct inodedep *inodedep; 6438 struct allocdirect *adp; 6439 struct buf *bp; 6440 struct fs *fs; 6441 ufs2_daddr_t extblocks, datablocks; 6442 struct mount *mp; 6443 int i, delay, error, dflags; 6444 ufs_lbn_t tmpval; 6445 ufs_lbn_t lbn; 6446 6447 fs = ip->i_fs; 6448 mp = UFSTOVFS(ip->i_ump); 6449 if (length != 0) 6450 panic("softdep_setup_freeblocks: non-zero length"); 6451 freeblks = newfreeblks(mp, ip); 6452 extblocks = 0; 6453 datablocks = 0; 6454 if (fs->fs_magic == FS_UFS2_MAGIC) 6455 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 6456 if ((flags & IO_NORMAL) != 0) { 6457 for (i = 0; i < NDADDR; i++) 6458 setup_freedirect(freeblks, ip, i, 0); 6459 for (i = 0, tmpval = NINDIR(fs), lbn = NDADDR; i < NIADDR; 6460 i++, lbn += tmpval, tmpval *= NINDIR(fs)) 6461 setup_freeindir(freeblks, ip, i, -lbn -i, 0); 6462 ip->i_size = 0; 6463 DIP_SET(ip, i_size, 0); 6464 datablocks = DIP(ip, i_blocks) - extblocks; 6465 } 6466 if ((flags & IO_EXT) != 0) { 6467 for (i = 0; i < NXADDR; i++) 6468 setup_freeext(freeblks, ip, i, 0); 6469 ip->i_din2->di_extsize = 0; 6470 datablocks += extblocks; 6471 } 6472#ifdef QUOTA 6473 /* Reference the quotas in case the block count is wrong in the end. */ 6474 quotaref(ITOV(ip), freeblks->fb_quota); 6475 (void) chkdq(ip, -datablocks, NOCRED, 0); 6476#endif 6477 freeblks->fb_chkcnt = -datablocks; 6478 UFS_LOCK(ip->i_ump); 6479 fs->fs_pendingblocks += datablocks; 6480 UFS_UNLOCK(ip->i_ump); 6481 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks); 6482 /* 6483 * Push the zero'ed inode to to its disk buffer so that we are free 6484 * to delete its dependencies below. Once the dependencies are gone 6485 * the buffer can be safely released. 6486 */ 6487 if ((error = bread(ip->i_devvp, 6488 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 6489 (int)fs->fs_bsize, NOCRED, &bp)) != 0) { 6490 brelse(bp); 6491 softdep_error("softdep_setup_freeblocks", error); 6492 } 6493 if (ip->i_ump->um_fstype == UFS1) { 6494 dp1 = ((struct ufs1_dinode *)bp->b_data + 6495 ino_to_fsbo(fs, ip->i_number)); 6496 ip->i_din1->di_freelink = dp1->di_freelink; 6497 *dp1 = *ip->i_din1; 6498 } else { 6499 dp2 = ((struct ufs2_dinode *)bp->b_data + 6500 ino_to_fsbo(fs, ip->i_number)); 6501 ip->i_din2->di_freelink = dp2->di_freelink; 6502 *dp2 = *ip->i_din2; 6503 } 6504 /* 6505 * Find and eliminate any inode dependencies. 6506 */ 6507 ACQUIRE_LOCK(&lk); 6508 dflags = DEPALLOC; 6509 if (IS_SNAPSHOT(ip)) 6510 dflags |= NODELAY; 6511 (void) inodedep_lookup(mp, ip->i_number, dflags, &inodedep); 6512 if ((inodedep->id_state & IOSTARTED) != 0) 6513 panic("softdep_setup_freeblocks: inode busy"); 6514 /* 6515 * Add the freeblks structure to the list of operations that 6516 * must await the zero'ed inode being written to disk. If we 6517 * still have a bitmap dependency (delay == 0), then the inode 6518 * has never been written to disk, so we can process the 6519 * freeblks below once we have deleted the dependencies. 6520 */ 6521 delay = (inodedep->id_state & DEPCOMPLETE); 6522 if (delay) 6523 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list); 6524 else 6525 freeblks->fb_state |= COMPLETE; 6526 /* 6527 * Because the file length has been truncated to zero, any 6528 * pending block allocation dependency structures associated 6529 * with this inode are obsolete and can simply be de-allocated. 6530 * We must first merge the two dependency lists to get rid of 6531 * any duplicate freefrag structures, then purge the merged list. 6532 * If we still have a bitmap dependency, then the inode has never 6533 * been written to disk, so we can free any fragments without delay. 6534 */ 6535 if (flags & IO_NORMAL) { 6536 merge_inode_lists(&inodedep->id_newinoupdt, 6537 &inodedep->id_inoupdt); 6538 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 6539 cancel_allocdirect(&inodedep->id_inoupdt, adp, 6540 freeblks); 6541 } 6542 if (flags & IO_EXT) { 6543 merge_inode_lists(&inodedep->id_newextupdt, 6544 &inodedep->id_extupdt); 6545 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0) 6546 cancel_allocdirect(&inodedep->id_extupdt, adp, 6547 freeblks); 6548 } 6549 FREE_LOCK(&lk); 6550 bdwrite(bp); 6551 trunc_dependencies(ip, freeblks, -1, 0, flags); 6552 ACQUIRE_LOCK(&lk); 6553 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 6554 (void) free_inodedep(inodedep); 6555 freeblks->fb_state |= DEPCOMPLETE; 6556 /* 6557 * If the inode with zeroed block pointers is now on disk 6558 * we can start freeing blocks. 6559 */ 6560 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 6561 freeblks->fb_state |= INPROGRESS; 6562 else 6563 freeblks = NULL; 6564 FREE_LOCK(&lk); 6565 if (freeblks) 6566 handle_workitem_freeblocks(freeblks, 0); 6567 trunc_pages(ip, length, extblocks, flags); 6568} 6569 6570/* 6571 * Eliminate pages from the page cache that back parts of this inode and 6572 * adjust the vnode pager's idea of our size. This prevents stale data 6573 * from hanging around in the page cache. 6574 */ 6575static void 6576trunc_pages(ip, length, extblocks, flags) 6577 struct inode *ip; 6578 off_t length; 6579 ufs2_daddr_t extblocks; 6580 int flags; 6581{ 6582 struct vnode *vp; 6583 struct fs *fs; 6584 ufs_lbn_t lbn; 6585 off_t end, extend; 6586 6587 vp = ITOV(ip); 6588 fs = ip->i_fs; 6589 extend = OFF_TO_IDX(lblktosize(fs, -extblocks)); 6590 if ((flags & IO_EXT) != 0) 6591 vn_pages_remove(vp, extend, 0); 6592 if ((flags & IO_NORMAL) == 0) 6593 return; 6594 BO_LOCK(&vp->v_bufobj); 6595 drain_output(vp); 6596 BO_UNLOCK(&vp->v_bufobj); 6597 /* 6598 * The vnode pager eliminates file pages we eliminate indirects 6599 * below. 6600 */ 6601 vnode_pager_setsize(vp, length); 6602 /* 6603 * Calculate the end based on the last indirect we want to keep. If 6604 * the block extends into indirects we can just use the negative of 6605 * its lbn. Doubles and triples exist at lower numbers so we must 6606 * be careful not to remove those, if they exist. double and triple 6607 * indirect lbns do not overlap with others so it is not important 6608 * to verify how many levels are required. 6609 */ 6610 lbn = lblkno(fs, length); 6611 if (lbn >= NDADDR) { 6612 /* Calculate the virtual lbn of the triple indirect. */ 6613 lbn = -lbn - (NIADDR - 1); 6614 end = OFF_TO_IDX(lblktosize(fs, lbn)); 6615 } else 6616 end = extend; 6617 vn_pages_remove(vp, OFF_TO_IDX(OFF_MAX), end); 6618} 6619 6620/* 6621 * See if the buf bp is in the range eliminated by truncation. 6622 */ 6623static int 6624trunc_check_buf(bp, blkoffp, lastlbn, lastoff, flags) 6625 struct buf *bp; 6626 int *blkoffp; 6627 ufs_lbn_t lastlbn; 6628 int lastoff; 6629 int flags; 6630{ 6631 ufs_lbn_t lbn; 6632 6633 *blkoffp = 0; 6634 /* Only match ext/normal blocks as appropriate. */ 6635 if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) || 6636 ((flags & IO_NORMAL) == 0 && (bp->b_xflags & BX_ALTDATA) == 0)) 6637 return (0); 6638 /* ALTDATA is always a full truncation. */ 6639 if ((bp->b_xflags & BX_ALTDATA) != 0) 6640 return (1); 6641 /* -1 is full truncation. */ 6642 if (lastlbn == -1) 6643 return (1); 6644 /* 6645 * If this is a partial truncate we only want those 6646 * blocks and indirect blocks that cover the range 6647 * we're after. 6648 */ 6649 lbn = bp->b_lblkno; 6650 if (lbn < 0) 6651 lbn = -(lbn + lbn_level(lbn)); 6652 if (lbn < lastlbn) 6653 return (0); 6654 /* Here we only truncate lblkno if it's partial. */ 6655 if (lbn == lastlbn) { 6656 if (lastoff == 0) 6657 return (0); 6658 *blkoffp = lastoff; 6659 } 6660 return (1); 6661} 6662 6663/* 6664 * Eliminate any dependencies that exist in memory beyond lblkno:off 6665 */ 6666static void 6667trunc_dependencies(ip, freeblks, lastlbn, lastoff, flags) 6668 struct inode *ip; 6669 struct freeblks *freeblks; 6670 ufs_lbn_t lastlbn; 6671 int lastoff; 6672 int flags; 6673{ 6674 struct bufobj *bo; 6675 struct vnode *vp; 6676 struct buf *bp; 6677 struct fs *fs; 6678 int blkoff; 6679 6680 /* 6681 * We must wait for any I/O in progress to finish so that 6682 * all potential buffers on the dirty list will be visible. 6683 * Once they are all there, walk the list and get rid of 6684 * any dependencies. 6685 */ 6686 fs = ip->i_fs; 6687 vp = ITOV(ip); 6688 bo = &vp->v_bufobj; 6689 BO_LOCK(bo); 6690 drain_output(vp); 6691 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) 6692 bp->b_vflags &= ~BV_SCANNED; 6693restart: 6694 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) { 6695 if (bp->b_vflags & BV_SCANNED) 6696 continue; 6697 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) { 6698 bp->b_vflags |= BV_SCANNED; 6699 continue; 6700 } 6701 if ((bp = getdirtybuf(bp, BO_MTX(bo), MNT_WAIT)) == NULL) 6702 goto restart; 6703 BO_UNLOCK(bo); 6704 if (deallocate_dependencies(bp, freeblks, blkoff)) 6705 bqrelse(bp); 6706 else 6707 brelse(bp); 6708 BO_LOCK(bo); 6709 goto restart; 6710 } 6711 /* 6712 * Now do the work of vtruncbuf while also matching indirect blocks. 6713 */ 6714 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) 6715 bp->b_vflags &= ~BV_SCANNED; 6716cleanrestart: 6717 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) { 6718 if (bp->b_vflags & BV_SCANNED) 6719 continue; 6720 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) { 6721 bp->b_vflags |= BV_SCANNED; 6722 continue; 6723 } 6724 if (BUF_LOCK(bp, 6725 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 6726 BO_MTX(bo)) == ENOLCK) { 6727 BO_LOCK(bo); 6728 goto cleanrestart; 6729 } 6730 bp->b_vflags |= BV_SCANNED; 6731 BO_LOCK(bo); 6732 bremfree(bp); 6733 BO_UNLOCK(bo); 6734 if (blkoff != 0) { 6735 allocbuf(bp, blkoff); 6736 bqrelse(bp); 6737 } else { 6738 bp->b_flags |= B_INVAL | B_NOCACHE | B_RELBUF; 6739 brelse(bp); 6740 } 6741 BO_LOCK(bo); 6742 goto cleanrestart; 6743 } 6744 drain_output(vp); 6745 BO_UNLOCK(bo); 6746} 6747 6748static int 6749cancel_pagedep(pagedep, freeblks, blkoff) 6750 struct pagedep *pagedep; 6751 struct freeblks *freeblks; 6752 int blkoff; 6753{ 6754 struct jremref *jremref; 6755 struct jmvref *jmvref; 6756 struct dirrem *dirrem, *tmp; 6757 int i; 6758 6759 /* 6760 * Copy any directory remove dependencies to the list 6761 * to be processed after the freeblks proceeds. If 6762 * directory entry never made it to disk they 6763 * can be dumped directly onto the work list. 6764 */ 6765 LIST_FOREACH_SAFE(dirrem, &pagedep->pd_dirremhd, dm_next, tmp) { 6766 /* Skip this directory removal if it is intended to remain. */ 6767 if (dirrem->dm_offset < blkoff) 6768 continue; 6769 /* 6770 * If there are any dirrems we wait for the journal write 6771 * to complete and then restart the buf scan as the lock 6772 * has been dropped. 6773 */ 6774 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) { 6775 jwait(&jremref->jr_list, MNT_WAIT); 6776 return (ERESTART); 6777 } 6778 LIST_REMOVE(dirrem, dm_next); 6779 dirrem->dm_dirinum = pagedep->pd_ino; 6780 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &dirrem->dm_list); 6781 } 6782 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) { 6783 jwait(&jmvref->jm_list, MNT_WAIT); 6784 return (ERESTART); 6785 } 6786 /* 6787 * When we're partially truncating a pagedep we just want to flush 6788 * journal entries and return. There can not be any adds in the 6789 * truncated portion of the directory and newblk must remain if 6790 * part of the block remains. 6791 */ 6792 if (blkoff != 0) { 6793 struct diradd *dap; 6794 6795 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 6796 if (dap->da_offset > blkoff) 6797 panic("cancel_pagedep: diradd %p off %d > %d", 6798 dap, dap->da_offset, blkoff); 6799 for (i = 0; i < DAHASHSZ; i++) 6800 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) 6801 if (dap->da_offset > blkoff) 6802 panic("cancel_pagedep: diradd %p off %d > %d", 6803 dap, dap->da_offset, blkoff); 6804 return (0); 6805 } 6806 /* 6807 * There should be no directory add dependencies present 6808 * as the directory could not be truncated until all 6809 * children were removed. 6810 */ 6811 KASSERT(LIST_FIRST(&pagedep->pd_pendinghd) == NULL, 6812 ("deallocate_dependencies: pendinghd != NULL")); 6813 for (i = 0; i < DAHASHSZ; i++) 6814 KASSERT(LIST_FIRST(&pagedep->pd_diraddhd[i]) == NULL, 6815 ("deallocate_dependencies: diraddhd != NULL")); 6816 if ((pagedep->pd_state & NEWBLOCK) != 0) 6817 free_newdirblk(pagedep->pd_newdirblk); 6818 if (free_pagedep(pagedep) == 0) 6819 panic("Failed to free pagedep %p", pagedep); 6820 return (0); 6821} 6822 6823/* 6824 * Reclaim any dependency structures from a buffer that is about to 6825 * be reallocated to a new vnode. The buffer must be locked, thus, 6826 * no I/O completion operations can occur while we are manipulating 6827 * its associated dependencies. The mutex is held so that other I/O's 6828 * associated with related dependencies do not occur. 6829 */ 6830static int 6831deallocate_dependencies(bp, freeblks, off) 6832 struct buf *bp; 6833 struct freeblks *freeblks; 6834 int off; 6835{ 6836 struct indirdep *indirdep; 6837 struct pagedep *pagedep; 6838 struct allocdirect *adp; 6839 struct worklist *wk, *wkn; 6840 6841 ACQUIRE_LOCK(&lk); 6842 LIST_FOREACH_SAFE(wk, &bp->b_dep, wk_list, wkn) { 6843 switch (wk->wk_type) { 6844 case D_INDIRDEP: 6845 indirdep = WK_INDIRDEP(wk); 6846 if (bp->b_lblkno >= 0 || 6847 bp->b_blkno != indirdep->ir_savebp->b_lblkno) 6848 panic("deallocate_dependencies: not indir"); 6849 cancel_indirdep(indirdep, bp, freeblks); 6850 continue; 6851 6852 case D_PAGEDEP: 6853 pagedep = WK_PAGEDEP(wk); 6854 if (cancel_pagedep(pagedep, freeblks, off)) { 6855 FREE_LOCK(&lk); 6856 return (ERESTART); 6857 } 6858 continue; 6859 6860 case D_ALLOCINDIR: 6861 /* 6862 * Simply remove the allocindir, we'll find it via 6863 * the indirdep where we can clear pointers if 6864 * needed. 6865 */ 6866 WORKLIST_REMOVE(wk); 6867 continue; 6868 6869 case D_FREEWORK: 6870 /* 6871 * A truncation is waiting for the zero'd pointers 6872 * to be written. It can be freed when the freeblks 6873 * is journaled. 6874 */ 6875 WORKLIST_REMOVE(wk); 6876 wk->wk_state |= ONDEPLIST; 6877 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk); 6878 break; 6879 6880 case D_ALLOCDIRECT: 6881 adp = WK_ALLOCDIRECT(wk); 6882 if (off != 0) 6883 continue; 6884 /* FALLTHROUGH */ 6885 default: 6886 panic("deallocate_dependencies: Unexpected type %s", 6887 TYPENAME(wk->wk_type)); 6888 /* NOTREACHED */ 6889 } 6890 } 6891 FREE_LOCK(&lk); 6892 /* 6893 * Don't throw away this buf, we were partially truncating and 6894 * some deps may always remain. 6895 */ 6896 if (off) { 6897 allocbuf(bp, off); 6898 bp->b_vflags |= BV_SCANNED; 6899 return (EBUSY); 6900 } 6901 bp->b_flags |= B_INVAL | B_NOCACHE; 6902 6903 return (0); 6904} 6905 6906/* 6907 * An allocdirect is being canceled due to a truncate. We must make sure 6908 * the journal entry is released in concert with the blkfree that releases 6909 * the storage. Completed journal entries must not be released until the 6910 * space is no longer pointed to by the inode or in the bitmap. 6911 */ 6912static void 6913cancel_allocdirect(adphead, adp, freeblks) 6914 struct allocdirectlst *adphead; 6915 struct allocdirect *adp; 6916 struct freeblks *freeblks; 6917{ 6918 struct freework *freework; 6919 struct newblk *newblk; 6920 struct worklist *wk; 6921 6922 TAILQ_REMOVE(adphead, adp, ad_next); 6923 newblk = (struct newblk *)adp; 6924 freework = NULL; 6925 /* 6926 * Find the correct freework structure. 6927 */ 6928 LIST_FOREACH(wk, &freeblks->fb_freeworkhd, wk_list) { 6929 if (wk->wk_type != D_FREEWORK) 6930 continue; 6931 freework = WK_FREEWORK(wk); 6932 if (freework->fw_blkno == newblk->nb_newblkno) 6933 break; 6934 } 6935 if (freework == NULL) 6936 panic("cancel_allocdirect: Freework not found"); 6937 /* 6938 * If a newblk exists at all we still have the journal entry that 6939 * initiated the allocation so we do not need to journal the free. 6940 */ 6941 cancel_jfreeblk(freeblks, freework->fw_blkno); 6942 /* 6943 * If the journal hasn't been written the jnewblk must be passed 6944 * to the call to ffs_blkfree that reclaims the space. We accomplish 6945 * this by linking the journal dependency into the freework to be 6946 * freed when freework_freeblock() is called. If the journal has 6947 * been written we can simply reclaim the journal space when the 6948 * freeblks work is complete. 6949 */ 6950 freework->fw_jnewblk = cancel_newblk(newblk, &freework->fw_list, 6951 &freeblks->fb_jwork); 6952 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list); 6953} 6954 6955 6956/* 6957 * Cancel a new block allocation. May be an indirect or direct block. We 6958 * remove it from various lists and return any journal record that needs to 6959 * be resolved by the caller. 6960 * 6961 * A special consideration is made for indirects which were never pointed 6962 * at on disk and will never be found once this block is released. 6963 */ 6964static struct jnewblk * 6965cancel_newblk(newblk, wk, wkhd) 6966 struct newblk *newblk; 6967 struct worklist *wk; 6968 struct workhead *wkhd; 6969{ 6970 struct jnewblk *jnewblk; 6971 6972 newblk->nb_state |= GOINGAWAY; 6973 /* 6974 * Previously we traversed the completedhd on each indirdep 6975 * attached to this newblk to cancel them and gather journal 6976 * work. Since we need only the oldest journal segment and 6977 * the lowest point on the tree will always have the oldest 6978 * journal segment we are free to release the segments 6979 * of any subordinates and may leave the indirdep list to 6980 * indirdep_complete() when this newblk is freed. 6981 */ 6982 if (newblk->nb_state & ONDEPLIST) { 6983 newblk->nb_state &= ~ONDEPLIST; 6984 LIST_REMOVE(newblk, nb_deps); 6985 } 6986 if (newblk->nb_state & ONWORKLIST) 6987 WORKLIST_REMOVE(&newblk->nb_list); 6988 /* 6989 * If the journal entry hasn't been written we save a pointer to 6990 * the dependency that frees it until it is written or the 6991 * superseding operation completes. 6992 */ 6993 jnewblk = newblk->nb_jnewblk; 6994 if (jnewblk != NULL && wk != NULL) { 6995 newblk->nb_jnewblk = NULL; 6996 jnewblk->jn_dep = wk; 6997 } 6998 if (!LIST_EMPTY(&newblk->nb_jwork)) 6999 jwork_move(wkhd, &newblk->nb_jwork); 7000 /* 7001 * When truncating we must free the newdirblk early to remove 7002 * the pagedep from the hash before returning. 7003 */ 7004 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL) 7005 free_newdirblk(WK_NEWDIRBLK(wk)); 7006 if (!LIST_EMPTY(&newblk->nb_newdirblk)) 7007 panic("cancel_newblk: extra newdirblk"); 7008 7009 return (jnewblk); 7010} 7011 7012/* 7013 * Schedule the freefrag associated with a newblk to be released once 7014 * the pointers are written and the previous block is no longer needed. 7015 */ 7016static void 7017newblk_freefrag(newblk) 7018 struct newblk *newblk; 7019{ 7020 struct freefrag *freefrag; 7021 7022 if (newblk->nb_freefrag == NULL) 7023 return; 7024 freefrag = newblk->nb_freefrag; 7025 newblk->nb_freefrag = NULL; 7026 freefrag->ff_state |= COMPLETE; 7027 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE) 7028 add_to_worklist(&freefrag->ff_list, 0); 7029} 7030 7031/* 7032 * Free a newblk. Generate a new freefrag work request if appropriate. 7033 * This must be called after the inode pointer and any direct block pointers 7034 * are valid or fully removed via truncate or frag extension. 7035 */ 7036static void 7037free_newblk(newblk) 7038 struct newblk *newblk; 7039{ 7040 struct indirdep *indirdep; 7041 struct worklist *wk; 7042 7043 KASSERT(newblk->nb_jnewblk == NULL, 7044 ("free_newblk; jnewblk %p still attached", newblk->nb_jnewblk)); 7045 mtx_assert(&lk, MA_OWNED); 7046 newblk_freefrag(newblk); 7047 if (newblk->nb_state & ONDEPLIST) 7048 LIST_REMOVE(newblk, nb_deps); 7049 if (newblk->nb_state & ONWORKLIST) 7050 WORKLIST_REMOVE(&newblk->nb_list); 7051 LIST_REMOVE(newblk, nb_hash); 7052 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL) 7053 free_newdirblk(WK_NEWDIRBLK(wk)); 7054 if (!LIST_EMPTY(&newblk->nb_newdirblk)) 7055 panic("free_newblk: extra newdirblk"); 7056 while ((indirdep = LIST_FIRST(&newblk->nb_indirdeps)) != NULL) 7057 indirdep_complete(indirdep); 7058 handle_jwork(&newblk->nb_jwork); 7059 newblk->nb_list.wk_type = D_NEWBLK; 7060 WORKITEM_FREE(newblk, D_NEWBLK); 7061} 7062 7063/* 7064 * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep. 7065 * This routine must be called with splbio interrupts blocked. 7066 */ 7067static void 7068free_newdirblk(newdirblk) 7069 struct newdirblk *newdirblk; 7070{ 7071 struct pagedep *pagedep; 7072 struct diradd *dap; 7073 struct worklist *wk; 7074 7075 mtx_assert(&lk, MA_OWNED); 7076 WORKLIST_REMOVE(&newdirblk->db_list); 7077 /* 7078 * If the pagedep is still linked onto the directory buffer 7079 * dependency chain, then some of the entries on the 7080 * pd_pendinghd list may not be committed to disk yet. In 7081 * this case, we will simply clear the NEWBLOCK flag and 7082 * let the pd_pendinghd list be processed when the pagedep 7083 * is next written. If the pagedep is no longer on the buffer 7084 * dependency chain, then all the entries on the pd_pending 7085 * list are committed to disk and we can free them here. 7086 */ 7087 pagedep = newdirblk->db_pagedep; 7088 pagedep->pd_state &= ~NEWBLOCK; 7089 if ((pagedep->pd_state & ONWORKLIST) == 0) { 7090 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 7091 free_diradd(dap, NULL); 7092 /* 7093 * If no dependencies remain, the pagedep will be freed. 7094 */ 7095 free_pagedep(pagedep); 7096 } 7097 /* Should only ever be one item in the list. */ 7098 while ((wk = LIST_FIRST(&newdirblk->db_mkdir)) != NULL) { 7099 WORKLIST_REMOVE(wk); 7100 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 7101 } 7102 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 7103} 7104 7105/* 7106 * Prepare an inode to be freed. The actual free operation is not 7107 * done until the zero'ed inode has been written to disk. 7108 */ 7109void 7110softdep_freefile(pvp, ino, mode) 7111 struct vnode *pvp; 7112 ino_t ino; 7113 int mode; 7114{ 7115 struct inode *ip = VTOI(pvp); 7116 struct inodedep *inodedep; 7117 struct freefile *freefile; 7118 struct freeblks *freeblks; 7119 7120 /* 7121 * This sets up the inode de-allocation dependency. 7122 */ 7123 freefile = malloc(sizeof(struct freefile), 7124 M_FREEFILE, M_SOFTDEP_FLAGS); 7125 workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount); 7126 freefile->fx_mode = mode; 7127 freefile->fx_oldinum = ino; 7128 freefile->fx_devvp = ip->i_devvp; 7129 LIST_INIT(&freefile->fx_jwork); 7130 UFS_LOCK(ip->i_ump); 7131 ip->i_fs->fs_pendinginodes += 1; 7132 UFS_UNLOCK(ip->i_ump); 7133 7134 /* 7135 * If the inodedep does not exist, then the zero'ed inode has 7136 * been written to disk. If the allocated inode has never been 7137 * written to disk, then the on-disk inode is zero'ed. In either 7138 * case we can free the file immediately. If the journal was 7139 * canceled before being written the inode will never make it to 7140 * disk and we must send the canceled journal entrys to 7141 * ffs_freefile() to be cleared in conjunction with the bitmap. 7142 * Any blocks waiting on the inode to write can be safely freed 7143 * here as it will never been written. 7144 */ 7145 ACQUIRE_LOCK(&lk); 7146 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep); 7147 if (inodedep) { 7148 /* 7149 * Clear out freeblks that no longer need to reference 7150 * this inode. 7151 */ 7152 while ((freeblks = 7153 TAILQ_FIRST(&inodedep->id_freeblklst)) != NULL) { 7154 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, 7155 fb_next); 7156 freeblks->fb_state &= ~ONDEPLIST; 7157 } 7158 /* 7159 * Remove this inode from the unlinked list. 7160 */ 7161 if (inodedep->id_state & UNLINKED) { 7162 /* 7163 * Save the journal work to be freed with the bitmap 7164 * before we clear UNLINKED. Otherwise it can be lost 7165 * if the inode block is written. 7166 */ 7167 handle_bufwait(inodedep, &freefile->fx_jwork); 7168 clear_unlinked_inodedep(inodedep); 7169 /* Re-acquire inodedep as we've dropped lk. */ 7170 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep); 7171 } 7172 } 7173 if (inodedep == NULL || check_inode_unwritten(inodedep)) { 7174 FREE_LOCK(&lk); 7175 handle_workitem_freefile(freefile); 7176 return; 7177 } 7178 if ((inodedep->id_state & DEPCOMPLETE) == 0) 7179 inodedep->id_state |= GOINGAWAY; 7180 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 7181 FREE_LOCK(&lk); 7182 if (ip->i_number == ino) 7183 ip->i_flag |= IN_MODIFIED; 7184} 7185 7186/* 7187 * Check to see if an inode has never been written to disk. If 7188 * so free the inodedep and return success, otherwise return failure. 7189 * This routine must be called with splbio interrupts blocked. 7190 * 7191 * If we still have a bitmap dependency, then the inode has never 7192 * been written to disk. Drop the dependency as it is no longer 7193 * necessary since the inode is being deallocated. We set the 7194 * ALLCOMPLETE flags since the bitmap now properly shows that the 7195 * inode is not allocated. Even if the inode is actively being 7196 * written, it has been rolled back to its zero'ed state, so we 7197 * are ensured that a zero inode is what is on the disk. For short 7198 * lived files, this change will usually result in removing all the 7199 * dependencies from the inode so that it can be freed immediately. 7200 */ 7201static int 7202check_inode_unwritten(inodedep) 7203 struct inodedep *inodedep; 7204{ 7205 7206 mtx_assert(&lk, MA_OWNED); 7207 7208 if ((inodedep->id_state & (DEPCOMPLETE | UNLINKED)) != 0 || 7209 !LIST_EMPTY(&inodedep->id_dirremhd) || 7210 !LIST_EMPTY(&inodedep->id_pendinghd) || 7211 !LIST_EMPTY(&inodedep->id_bufwait) || 7212 !LIST_EMPTY(&inodedep->id_inowait) || 7213 !TAILQ_EMPTY(&inodedep->id_inoreflst) || 7214 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 7215 !TAILQ_EMPTY(&inodedep->id_newinoupdt) || 7216 !TAILQ_EMPTY(&inodedep->id_extupdt) || 7217 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 7218 !TAILQ_EMPTY(&inodedep->id_freeblklst) || 7219 inodedep->id_mkdiradd != NULL || 7220 inodedep->id_nlinkdelta != 0) 7221 return (0); 7222 /* 7223 * Another process might be in initiate_write_inodeblock_ufs[12] 7224 * trying to allocate memory without holding "Softdep Lock". 7225 */ 7226 if ((inodedep->id_state & IOSTARTED) != 0 && 7227 inodedep->id_savedino1 == NULL) 7228 return (0); 7229 7230 if (inodedep->id_state & ONDEPLIST) 7231 LIST_REMOVE(inodedep, id_deps); 7232 inodedep->id_state &= ~ONDEPLIST; 7233 inodedep->id_state |= ALLCOMPLETE; 7234 inodedep->id_bmsafemap = NULL; 7235 if (inodedep->id_state & ONWORKLIST) 7236 WORKLIST_REMOVE(&inodedep->id_list); 7237 if (inodedep->id_savedino1 != NULL) { 7238 free(inodedep->id_savedino1, M_SAVEDINO); 7239 inodedep->id_savedino1 = NULL; 7240 } 7241 if (free_inodedep(inodedep) == 0) 7242 panic("check_inode_unwritten: busy inode"); 7243 return (1); 7244} 7245 7246/* 7247 * Try to free an inodedep structure. Return 1 if it could be freed. 7248 */ 7249static int 7250free_inodedep(inodedep) 7251 struct inodedep *inodedep; 7252{ 7253 7254 mtx_assert(&lk, MA_OWNED); 7255 if ((inodedep->id_state & (ONWORKLIST | UNLINKED)) != 0 || 7256 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 7257 !LIST_EMPTY(&inodedep->id_dirremhd) || 7258 !LIST_EMPTY(&inodedep->id_pendinghd) || 7259 !LIST_EMPTY(&inodedep->id_bufwait) || 7260 !LIST_EMPTY(&inodedep->id_inowait) || 7261 !TAILQ_EMPTY(&inodedep->id_inoreflst) || 7262 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 7263 !TAILQ_EMPTY(&inodedep->id_newinoupdt) || 7264 !TAILQ_EMPTY(&inodedep->id_extupdt) || 7265 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 7266 !TAILQ_EMPTY(&inodedep->id_freeblklst) || 7267 inodedep->id_mkdiradd != NULL || 7268 inodedep->id_nlinkdelta != 0 || 7269 inodedep->id_savedino1 != NULL) 7270 return (0); 7271 if (inodedep->id_state & ONDEPLIST) 7272 LIST_REMOVE(inodedep, id_deps); 7273 LIST_REMOVE(inodedep, id_hash); 7274 WORKITEM_FREE(inodedep, D_INODEDEP); 7275 return (1); 7276} 7277 7278/* 7279 * Free the block referenced by a freework structure. The parent freeblks 7280 * structure is released and completed when the final cg bitmap reaches 7281 * the disk. This routine may be freeing a jnewblk which never made it to 7282 * disk in which case we do not have to wait as the operation is undone 7283 * in memory immediately. 7284 */ 7285static void 7286freework_freeblock(freework) 7287 struct freework *freework; 7288{ 7289 struct freeblks *freeblks; 7290 struct jnewblk *jnewblk; 7291 struct ufsmount *ump; 7292 struct workhead wkhd; 7293 struct fs *fs; 7294 int bsize; 7295 int needj; 7296 7297 mtx_assert(&lk, MA_OWNED); 7298 /* 7299 * Handle partial truncate separately. 7300 */ 7301 if (freework->fw_indir) { 7302 complete_trunc_indir(freework); 7303 return; 7304 } 7305 freeblks = freework->fw_freeblks; 7306 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7307 fs = ump->um_fs; 7308 needj = MOUNTEDSUJ(freeblks->fb_list.wk_mp) != 0; 7309 bsize = lfragtosize(fs, freework->fw_frags); 7310 LIST_INIT(&wkhd); 7311 /* 7312 * DEPCOMPLETE is cleared in indirblk_insert() if the block lives 7313 * on the indirblk hashtable and prevents premature freeing. 7314 */ 7315 freework->fw_state |= DEPCOMPLETE; 7316 /* 7317 * SUJ needs to wait for the segment referencing freed indirect 7318 * blocks to expire so that we know the checker will not confuse 7319 * a re-allocated indirect block with its old contents. 7320 */ 7321 if (needj && freework->fw_lbn <= -NDADDR) 7322 indirblk_insert(freework); 7323 /* 7324 * If we are canceling an existing jnewblk pass it to the free 7325 * routine, otherwise pass the freeblk which will ultimately 7326 * release the freeblks. If we're not journaling, we can just 7327 * free the freeblks immediately. 7328 */ 7329 jnewblk = freework->fw_jnewblk; 7330 if (jnewblk != NULL) { 7331 cancel_jnewblk(jnewblk, &wkhd); 7332 needj = 0; 7333 } else if (needj) { 7334 freework->fw_state |= DELAYEDFREE; 7335 freeblks->fb_cgwait++; 7336 WORKLIST_INSERT(&wkhd, &freework->fw_list); 7337 } 7338 FREE_LOCK(&lk); 7339 freeblks_free(ump, freeblks, btodb(bsize)); 7340 ffs_blkfree(ump, fs, freeblks->fb_devvp, freework->fw_blkno, bsize, 7341 freeblks->fb_inum, freeblks->fb_vtype, &wkhd); 7342 ACQUIRE_LOCK(&lk); 7343 /* 7344 * The jnewblk will be discarded and the bits in the map never 7345 * made it to disk. We can immediately free the freeblk. 7346 */ 7347 if (needj == 0) 7348 handle_written_freework(freework); 7349} 7350 7351/* 7352 * We enqueue freework items that need processing back on the freeblks and 7353 * add the freeblks to the worklist. This makes it easier to find all work 7354 * required to flush a truncation in process_truncates(). 7355 */ 7356static void 7357freework_enqueue(freework) 7358 struct freework *freework; 7359{ 7360 struct freeblks *freeblks; 7361 7362 freeblks = freework->fw_freeblks; 7363 if ((freework->fw_state & INPROGRESS) == 0) 7364 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list); 7365 if ((freeblks->fb_state & 7366 (ONWORKLIST | INPROGRESS | ALLCOMPLETE)) == ALLCOMPLETE && 7367 LIST_EMPTY(&freeblks->fb_jblkdephd)) 7368 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 7369} 7370 7371/* 7372 * Start, continue, or finish the process of freeing an indirect block tree. 7373 * The free operation may be paused at any point with fw_off containing the 7374 * offset to restart from. This enables us to implement some flow control 7375 * for large truncates which may fan out and generate a huge number of 7376 * dependencies. 7377 */ 7378static void 7379handle_workitem_indirblk(freework) 7380 struct freework *freework; 7381{ 7382 struct freeblks *freeblks; 7383 struct ufsmount *ump; 7384 struct fs *fs; 7385 7386 freeblks = freework->fw_freeblks; 7387 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7388 fs = ump->um_fs; 7389 if (freework->fw_state & DEPCOMPLETE) { 7390 handle_written_freework(freework); 7391 return; 7392 } 7393 if (freework->fw_off == NINDIR(fs)) { 7394 freework_freeblock(freework); 7395 return; 7396 } 7397 freework->fw_state |= INPROGRESS; 7398 FREE_LOCK(&lk); 7399 indir_trunc(freework, fsbtodb(fs, freework->fw_blkno), 7400 freework->fw_lbn); 7401 ACQUIRE_LOCK(&lk); 7402} 7403 7404/* 7405 * Called when a freework structure attached to a cg buf is written. The 7406 * ref on either the parent or the freeblks structure is released and 7407 * the freeblks is added back to the worklist if there is more work to do. 7408 */ 7409static void 7410handle_written_freework(freework) 7411 struct freework *freework; 7412{ 7413 struct freeblks *freeblks; 7414 struct freework *parent; 7415 7416 freeblks = freework->fw_freeblks; 7417 parent = freework->fw_parent; 7418 if (freework->fw_state & DELAYEDFREE) 7419 freeblks->fb_cgwait--; 7420 freework->fw_state |= COMPLETE; 7421 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE) 7422 WORKITEM_FREE(freework, D_FREEWORK); 7423 if (parent) { 7424 if (--parent->fw_ref == 0) 7425 freework_enqueue(parent); 7426 return; 7427 } 7428 if (--freeblks->fb_ref != 0) 7429 return; 7430 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST | INPROGRESS)) == 7431 ALLCOMPLETE && LIST_EMPTY(&freeblks->fb_jblkdephd)) 7432 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 7433} 7434 7435/* 7436 * This workitem routine performs the block de-allocation. 7437 * The workitem is added to the pending list after the updated 7438 * inode block has been written to disk. As mentioned above, 7439 * checks regarding the number of blocks de-allocated (compared 7440 * to the number of blocks allocated for the file) are also 7441 * performed in this function. 7442 */ 7443static int 7444handle_workitem_freeblocks(freeblks, flags) 7445 struct freeblks *freeblks; 7446 int flags; 7447{ 7448 struct freework *freework; 7449 struct newblk *newblk; 7450 struct allocindir *aip; 7451 struct ufsmount *ump; 7452 struct worklist *wk; 7453 7454 KASSERT(LIST_EMPTY(&freeblks->fb_jblkdephd), 7455 ("handle_workitem_freeblocks: Journal entries not written.")); 7456 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7457 ACQUIRE_LOCK(&lk); 7458 while ((wk = LIST_FIRST(&freeblks->fb_freeworkhd)) != NULL) { 7459 WORKLIST_REMOVE(wk); 7460 switch (wk->wk_type) { 7461 case D_DIRREM: 7462 wk->wk_state |= COMPLETE; 7463 add_to_worklist(wk, 0); 7464 continue; 7465 7466 case D_ALLOCDIRECT: 7467 free_newblk(WK_NEWBLK(wk)); 7468 continue; 7469 7470 case D_ALLOCINDIR: 7471 aip = WK_ALLOCINDIR(wk); 7472 freework = NULL; 7473 if (aip->ai_state & DELAYEDFREE) { 7474 FREE_LOCK(&lk); 7475 freework = newfreework(ump, freeblks, NULL, 7476 aip->ai_lbn, aip->ai_newblkno, 7477 ump->um_fs->fs_frag, 0, 0); 7478 ACQUIRE_LOCK(&lk); 7479 } 7480 newblk = WK_NEWBLK(wk); 7481 if (newblk->nb_jnewblk) { 7482 freework->fw_jnewblk = newblk->nb_jnewblk; 7483 newblk->nb_jnewblk->jn_dep = &freework->fw_list; 7484 newblk->nb_jnewblk = NULL; 7485 } 7486 free_newblk(newblk); 7487 continue; 7488 7489 case D_FREEWORK: 7490 freework = WK_FREEWORK(wk); 7491 if (freework->fw_lbn <= -NDADDR) 7492 handle_workitem_indirblk(freework); 7493 else 7494 freework_freeblock(freework); 7495 continue; 7496 default: 7497 panic("handle_workitem_freeblocks: Unknown type %s", 7498 TYPENAME(wk->wk_type)); 7499 } 7500 } 7501 if (freeblks->fb_ref != 0) { 7502 freeblks->fb_state &= ~INPROGRESS; 7503 wake_worklist(&freeblks->fb_list); 7504 freeblks = NULL; 7505 } 7506 FREE_LOCK(&lk); 7507 if (freeblks) 7508 return handle_complete_freeblocks(freeblks, flags); 7509 return (0); 7510} 7511 7512/* 7513 * Handle completion of block free via truncate. This allows fs_pending 7514 * to track the actual free block count more closely than if we only updated 7515 * it at the end. We must be careful to handle cases where the block count 7516 * on free was incorrect. 7517 */ 7518static void 7519freeblks_free(ump, freeblks, blocks) 7520 struct ufsmount *ump; 7521 struct freeblks *freeblks; 7522 int blocks; 7523{ 7524 struct fs *fs; 7525 ufs2_daddr_t remain; 7526 7527 UFS_LOCK(ump); 7528 remain = -freeblks->fb_chkcnt; 7529 freeblks->fb_chkcnt += blocks; 7530 if (remain > 0) { 7531 if (remain < blocks) 7532 blocks = remain; 7533 fs = ump->um_fs; 7534 fs->fs_pendingblocks -= blocks; 7535 } 7536 UFS_UNLOCK(ump); 7537} 7538 7539/* 7540 * Once all of the freework workitems are complete we can retire the 7541 * freeblocks dependency and any journal work awaiting completion. This 7542 * can not be called until all other dependencies are stable on disk. 7543 */ 7544static int 7545handle_complete_freeblocks(freeblks, flags) 7546 struct freeblks *freeblks; 7547 int flags; 7548{ 7549 struct inodedep *inodedep; 7550 struct inode *ip; 7551 struct vnode *vp; 7552 struct fs *fs; 7553 struct ufsmount *ump; 7554 ufs2_daddr_t spare; 7555 7556 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7557 fs = ump->um_fs; 7558 flags = LK_EXCLUSIVE | flags; 7559 spare = freeblks->fb_chkcnt; 7560 7561 /* 7562 * If we did not release the expected number of blocks we may have 7563 * to adjust the inode block count here. Only do so if it wasn't 7564 * a truncation to zero and the modrev still matches. 7565 */ 7566 if (spare && freeblks->fb_len != 0) { 7567 if (ffs_vgetf(freeblks->fb_list.wk_mp, freeblks->fb_inum, 7568 flags, &vp, FFSV_FORCEINSMQ) != 0) 7569 return (EBUSY); 7570 ip = VTOI(vp); 7571 if (DIP(ip, i_modrev) == freeblks->fb_modrev) { 7572 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - spare); 7573 ip->i_flag |= IN_CHANGE; 7574 /* 7575 * We must wait so this happens before the 7576 * journal is reclaimed. 7577 */ 7578 ffs_update(vp, 1); 7579 } 7580 vput(vp); 7581 } 7582 if (spare < 0) { 7583 UFS_LOCK(ump); 7584 fs->fs_pendingblocks += spare; 7585 UFS_UNLOCK(ump); 7586 } 7587#ifdef QUOTA 7588 /* Handle spare. */ 7589 if (spare) 7590 quotaadj(freeblks->fb_quota, ump, -spare); 7591 quotarele(freeblks->fb_quota); 7592#endif 7593 ACQUIRE_LOCK(&lk); 7594 if (freeblks->fb_state & ONDEPLIST) { 7595 inodedep_lookup(freeblks->fb_list.wk_mp, freeblks->fb_inum, 7596 0, &inodedep); 7597 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, fb_next); 7598 freeblks->fb_state &= ~ONDEPLIST; 7599 if (TAILQ_EMPTY(&inodedep->id_freeblklst)) 7600 free_inodedep(inodedep); 7601 } 7602 /* 7603 * All of the freeblock deps must be complete prior to this call 7604 * so it's now safe to complete earlier outstanding journal entries. 7605 */ 7606 handle_jwork(&freeblks->fb_jwork); 7607 WORKITEM_FREE(freeblks, D_FREEBLKS); 7608 FREE_LOCK(&lk); 7609 return (0); 7610} 7611 7612/* 7613 * Release blocks associated with the freeblks and stored in the indirect 7614 * block dbn. If level is greater than SINGLE, the block is an indirect block 7615 * and recursive calls to indirtrunc must be used to cleanse other indirect 7616 * blocks. 7617 * 7618 * This handles partial and complete truncation of blocks. Partial is noted 7619 * with goingaway == 0. In this case the freework is completed after the 7620 * zero'd indirects are written to disk. For full truncation the freework 7621 * is completed after the block is freed. 7622 */ 7623static void 7624indir_trunc(freework, dbn, lbn) 7625 struct freework *freework; 7626 ufs2_daddr_t dbn; 7627 ufs_lbn_t lbn; 7628{ 7629 struct freework *nfreework; 7630 struct workhead wkhd; 7631 struct freeblks *freeblks; 7632 struct buf *bp; 7633 struct fs *fs; 7634 struct indirdep *indirdep; 7635 struct ufsmount *ump; 7636 ufs1_daddr_t *bap1 = 0; 7637 ufs2_daddr_t nb, nnb, *bap2 = 0; 7638 ufs_lbn_t lbnadd, nlbn; 7639 int i, nblocks, ufs1fmt; 7640 int freedblocks; 7641 int goingaway; 7642 int freedeps; 7643 int needj; 7644 int level; 7645 int cnt; 7646 7647 freeblks = freework->fw_freeblks; 7648 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7649 fs = ump->um_fs; 7650 /* 7651 * Get buffer of block pointers to be freed. There are three cases: 7652 * 7653 * 1) Partial truncate caches the indirdep pointer in the freework 7654 * which provides us a back copy to the save bp which holds the 7655 * pointers we want to clear. When this completes the zero 7656 * pointers are written to the real copy. 7657 * 2) The indirect is being completely truncated, cancel_indirdep() 7658 * eliminated the real copy and placed the indirdep on the saved 7659 * copy. The indirdep and buf are discarded when this completes. 7660 * 3) The indirect was not in memory, we read a copy off of the disk 7661 * using the devvp and drop and invalidate the buffer when we're 7662 * done. 7663 */ 7664 goingaway = 1; 7665 indirdep = NULL; 7666 if (freework->fw_indir != NULL) { 7667 goingaway = 0; 7668 indirdep = freework->fw_indir; 7669 bp = indirdep->ir_savebp; 7670 if (bp == NULL || bp->b_blkno != dbn) 7671 panic("indir_trunc: Bad saved buf %p blkno %jd", 7672 bp, (intmax_t)dbn); 7673 } else if ((bp = incore(&freeblks->fb_devvp->v_bufobj, dbn)) != NULL) { 7674 /* 7675 * The lock prevents the buf dep list from changing and 7676 * indirects on devvp should only ever have one dependency. 7677 */ 7678 indirdep = WK_INDIRDEP(LIST_FIRST(&bp->b_dep)); 7679 if (indirdep == NULL || (indirdep->ir_state & GOINGAWAY) == 0) 7680 panic("indir_trunc: Bad indirdep %p from buf %p", 7681 indirdep, bp); 7682 } else if (bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 7683 NOCRED, &bp) != 0) { 7684 brelse(bp); 7685 return; 7686 } 7687 ACQUIRE_LOCK(&lk); 7688 /* Protects against a race with complete_trunc_indir(). */ 7689 freework->fw_state &= ~INPROGRESS; 7690 /* 7691 * If we have an indirdep we need to enforce the truncation order 7692 * and discard it when it is complete. 7693 */ 7694 if (indirdep) { 7695 if (freework != TAILQ_FIRST(&indirdep->ir_trunc) && 7696 !TAILQ_EMPTY(&indirdep->ir_trunc)) { 7697 /* 7698 * Add the complete truncate to the list on the 7699 * indirdep to enforce in-order processing. 7700 */ 7701 if (freework->fw_indir == NULL) 7702 TAILQ_INSERT_TAIL(&indirdep->ir_trunc, 7703 freework, fw_next); 7704 FREE_LOCK(&lk); 7705 return; 7706 } 7707 /* 7708 * If we're goingaway, free the indirdep. Otherwise it will 7709 * linger until the write completes. 7710 */ 7711 if (goingaway) { 7712 free_indirdep(indirdep); 7713 ump->um_numindirdeps -= 1; 7714 } 7715 } 7716 FREE_LOCK(&lk); 7717 /* Initialize pointers depending on block size. */ 7718 if (ump->um_fstype == UFS1) { 7719 bap1 = (ufs1_daddr_t *)bp->b_data; 7720 nb = bap1[freework->fw_off]; 7721 ufs1fmt = 1; 7722 } else { 7723 bap2 = (ufs2_daddr_t *)bp->b_data; 7724 nb = bap2[freework->fw_off]; 7725 ufs1fmt = 0; 7726 } 7727 level = lbn_level(lbn); 7728 needj = MOUNTEDSUJ(UFSTOVFS(ump)) != 0; 7729 lbnadd = lbn_offset(fs, level); 7730 nblocks = btodb(fs->fs_bsize); 7731 nfreework = freework; 7732 freedeps = 0; 7733 cnt = 0; 7734 /* 7735 * Reclaim blocks. Traverses into nested indirect levels and 7736 * arranges for the current level to be freed when subordinates 7737 * are free when journaling. 7738 */ 7739 for (i = freework->fw_off; i < NINDIR(fs); i++, nb = nnb) { 7740 if (i != NINDIR(fs) - 1) { 7741 if (ufs1fmt) 7742 nnb = bap1[i+1]; 7743 else 7744 nnb = bap2[i+1]; 7745 } else 7746 nnb = 0; 7747 if (nb == 0) 7748 continue; 7749 cnt++; 7750 if (level != 0) { 7751 nlbn = (lbn + 1) - (i * lbnadd); 7752 if (needj != 0) { 7753 nfreework = newfreework(ump, freeblks, freework, 7754 nlbn, nb, fs->fs_frag, 0, 0); 7755 freedeps++; 7756 } 7757 indir_trunc(nfreework, fsbtodb(fs, nb), nlbn); 7758 } else { 7759 struct freedep *freedep; 7760 7761 /* 7762 * Attempt to aggregate freedep dependencies for 7763 * all blocks being released to the same CG. 7764 */ 7765 LIST_INIT(&wkhd); 7766 if (needj != 0 && 7767 (nnb == 0 || (dtog(fs, nb) != dtog(fs, nnb)))) { 7768 freedep = newfreedep(freework); 7769 WORKLIST_INSERT_UNLOCKED(&wkhd, 7770 &freedep->fd_list); 7771 freedeps++; 7772 } 7773 ffs_blkfree(ump, fs, freeblks->fb_devvp, nb, 7774 fs->fs_bsize, freeblks->fb_inum, 7775 freeblks->fb_vtype, &wkhd); 7776 } 7777 } 7778 if (goingaway) { 7779 bp->b_flags |= B_INVAL | B_NOCACHE; 7780 brelse(bp); 7781 } 7782 freedblocks = 0; 7783 if (level == 0) 7784 freedblocks = (nblocks * cnt); 7785 if (needj == 0) 7786 freedblocks += nblocks; 7787 freeblks_free(ump, freeblks, freedblocks); 7788 /* 7789 * If we are journaling set up the ref counts and offset so this 7790 * indirect can be completed when its children are free. 7791 */ 7792 if (needj) { 7793 ACQUIRE_LOCK(&lk); 7794 freework->fw_off = i; 7795 freework->fw_ref += freedeps; 7796 freework->fw_ref -= NINDIR(fs) + 1; 7797 if (level == 0) 7798 freeblks->fb_cgwait += freedeps; 7799 if (freework->fw_ref == 0) 7800 freework_freeblock(freework); 7801 FREE_LOCK(&lk); 7802 return; 7803 } 7804 /* 7805 * If we're not journaling we can free the indirect now. 7806 */ 7807 dbn = dbtofsb(fs, dbn); 7808 ffs_blkfree(ump, fs, freeblks->fb_devvp, dbn, fs->fs_bsize, 7809 freeblks->fb_inum, freeblks->fb_vtype, NULL); 7810 /* Non SUJ softdep does single-threaded truncations. */ 7811 if (freework->fw_blkno == dbn) { 7812 freework->fw_state |= ALLCOMPLETE; 7813 ACQUIRE_LOCK(&lk); 7814 handle_written_freework(freework); 7815 FREE_LOCK(&lk); 7816 } 7817 return; 7818} 7819 7820/* 7821 * Cancel an allocindir when it is removed via truncation. When bp is not 7822 * NULL the indirect never appeared on disk and is scheduled to be freed 7823 * independently of the indir so we can more easily track journal work. 7824 */ 7825static void 7826cancel_allocindir(aip, bp, freeblks, trunc) 7827 struct allocindir *aip; 7828 struct buf *bp; 7829 struct freeblks *freeblks; 7830 int trunc; 7831{ 7832 struct indirdep *indirdep; 7833 struct freefrag *freefrag; 7834 struct newblk *newblk; 7835 7836 newblk = (struct newblk *)aip; 7837 LIST_REMOVE(aip, ai_next); 7838 /* 7839 * We must eliminate the pointer in bp if it must be freed on its 7840 * own due to partial truncate or pending journal work. 7841 */ 7842 if (bp && (trunc || newblk->nb_jnewblk)) { 7843 /* 7844 * Clear the pointer and mark the aip to be freed 7845 * directly if it never existed on disk. 7846 */ 7847 aip->ai_state |= DELAYEDFREE; 7848 indirdep = aip->ai_indirdep; 7849 if (indirdep->ir_state & UFS1FMT) 7850 ((ufs1_daddr_t *)bp->b_data)[aip->ai_offset] = 0; 7851 else 7852 ((ufs2_daddr_t *)bp->b_data)[aip->ai_offset] = 0; 7853 } 7854 /* 7855 * When truncating the previous pointer will be freed via 7856 * savedbp. Eliminate the freefrag which would dup free. 7857 */ 7858 if (trunc && (freefrag = newblk->nb_freefrag) != NULL) { 7859 newblk->nb_freefrag = NULL; 7860 if (freefrag->ff_jdep) 7861 cancel_jfreefrag( 7862 WK_JFREEFRAG(freefrag->ff_jdep)); 7863 jwork_move(&freeblks->fb_jwork, &freefrag->ff_jwork); 7864 WORKITEM_FREE(freefrag, D_FREEFRAG); 7865 } 7866 /* 7867 * If the journal hasn't been written the jnewblk must be passed 7868 * to the call to ffs_blkfree that reclaims the space. We accomplish 7869 * this by leaving the journal dependency on the newblk to be freed 7870 * when a freework is created in handle_workitem_freeblocks(). 7871 */ 7872 cancel_newblk(newblk, NULL, &freeblks->fb_jwork); 7873 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list); 7874} 7875 7876/* 7877 * Create the mkdir dependencies for . and .. in a new directory. Link them 7878 * in to a newdirblk so any subsequent additions are tracked properly. The 7879 * caller is responsible for adding the mkdir1 dependency to the journal 7880 * and updating id_mkdiradd. This function returns with lk held. 7881 */ 7882static struct mkdir * 7883setup_newdir(dap, newinum, dinum, newdirbp, mkdirp) 7884 struct diradd *dap; 7885 ino_t newinum; 7886 ino_t dinum; 7887 struct buf *newdirbp; 7888 struct mkdir **mkdirp; 7889{ 7890 struct newblk *newblk; 7891 struct pagedep *pagedep; 7892 struct inodedep *inodedep; 7893 struct newdirblk *newdirblk = 0; 7894 struct mkdir *mkdir1, *mkdir2; 7895 struct worklist *wk; 7896 struct jaddref *jaddref; 7897 struct mount *mp; 7898 7899 mp = dap->da_list.wk_mp; 7900 newdirblk = malloc(sizeof(struct newdirblk), M_NEWDIRBLK, 7901 M_SOFTDEP_FLAGS); 7902 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 7903 LIST_INIT(&newdirblk->db_mkdir); 7904 mkdir1 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS); 7905 workitem_alloc(&mkdir1->md_list, D_MKDIR, mp); 7906 mkdir1->md_state = ATTACHED | MKDIR_BODY; 7907 mkdir1->md_diradd = dap; 7908 mkdir1->md_jaddref = NULL; 7909 mkdir2 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS); 7910 workitem_alloc(&mkdir2->md_list, D_MKDIR, mp); 7911 mkdir2->md_state = ATTACHED | MKDIR_PARENT; 7912 mkdir2->md_diradd = dap; 7913 mkdir2->md_jaddref = NULL; 7914 if (MOUNTEDSUJ(mp) == 0) { 7915 mkdir1->md_state |= DEPCOMPLETE; 7916 mkdir2->md_state |= DEPCOMPLETE; 7917 } 7918 /* 7919 * Dependency on "." and ".." being written to disk. 7920 */ 7921 mkdir1->md_buf = newdirbp; 7922 ACQUIRE_LOCK(&lk); 7923 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 7924 /* 7925 * We must link the pagedep, allocdirect, and newdirblk for 7926 * the initial file page so the pointer to the new directory 7927 * is not written until the directory contents are live and 7928 * any subsequent additions are not marked live until the 7929 * block is reachable via the inode. 7930 */ 7931 if (pagedep_lookup(mp, newdirbp, newinum, 0, 0, &pagedep) == 0) 7932 panic("setup_newdir: lost pagedep"); 7933 LIST_FOREACH(wk, &newdirbp->b_dep, wk_list) 7934 if (wk->wk_type == D_ALLOCDIRECT) 7935 break; 7936 if (wk == NULL) 7937 panic("setup_newdir: lost allocdirect"); 7938 if (pagedep->pd_state & NEWBLOCK) 7939 panic("setup_newdir: NEWBLOCK already set"); 7940 newblk = WK_NEWBLK(wk); 7941 pagedep->pd_state |= NEWBLOCK; 7942 pagedep->pd_newdirblk = newdirblk; 7943 newdirblk->db_pagedep = pagedep; 7944 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list); 7945 WORKLIST_INSERT(&newdirblk->db_mkdir, &mkdir1->md_list); 7946 /* 7947 * Look up the inodedep for the parent directory so that we 7948 * can link mkdir2 into the pending dotdot jaddref or 7949 * the inode write if there is none. If the inode is 7950 * ALLCOMPLETE and no jaddref is present all dependencies have 7951 * been satisfied and mkdir2 can be freed. 7952 */ 7953 inodedep_lookup(mp, dinum, 0, &inodedep); 7954 if (MOUNTEDSUJ(mp)) { 7955 if (inodedep == NULL) 7956 panic("setup_newdir: Lost parent."); 7957 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 7958 inoreflst); 7959 KASSERT(jaddref != NULL && jaddref->ja_parent == newinum && 7960 (jaddref->ja_state & MKDIR_PARENT), 7961 ("setup_newdir: bad dotdot jaddref %p", jaddref)); 7962 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 7963 mkdir2->md_jaddref = jaddref; 7964 jaddref->ja_mkdir = mkdir2; 7965 } else if (inodedep == NULL || 7966 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 7967 dap->da_state &= ~MKDIR_PARENT; 7968 WORKITEM_FREE(mkdir2, D_MKDIR); 7969 } else { 7970 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 7971 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir2->md_list); 7972 } 7973 *mkdirp = mkdir2; 7974 7975 return (mkdir1); 7976} 7977 7978/* 7979 * Directory entry addition dependencies. 7980 * 7981 * When adding a new directory entry, the inode (with its incremented link 7982 * count) must be written to disk before the directory entry's pointer to it. 7983 * Also, if the inode is newly allocated, the corresponding freemap must be 7984 * updated (on disk) before the directory entry's pointer. These requirements 7985 * are met via undo/redo on the directory entry's pointer, which consists 7986 * simply of the inode number. 7987 * 7988 * As directory entries are added and deleted, the free space within a 7989 * directory block can become fragmented. The ufs filesystem will compact 7990 * a fragmented directory block to make space for a new entry. When this 7991 * occurs, the offsets of previously added entries change. Any "diradd" 7992 * dependency structures corresponding to these entries must be updated with 7993 * the new offsets. 7994 */ 7995 7996/* 7997 * This routine is called after the in-memory inode's link 7998 * count has been incremented, but before the directory entry's 7999 * pointer to the inode has been set. 8000 */ 8001int 8002softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 8003 struct buf *bp; /* buffer containing directory block */ 8004 struct inode *dp; /* inode for directory */ 8005 off_t diroffset; /* offset of new entry in directory */ 8006 ino_t newinum; /* inode referenced by new directory entry */ 8007 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 8008 int isnewblk; /* entry is in a newly allocated block */ 8009{ 8010 int offset; /* offset of new entry within directory block */ 8011 ufs_lbn_t lbn; /* block in directory containing new entry */ 8012 struct fs *fs; 8013 struct diradd *dap; 8014 struct newblk *newblk; 8015 struct pagedep *pagedep; 8016 struct inodedep *inodedep; 8017 struct newdirblk *newdirblk = 0; 8018 struct mkdir *mkdir1, *mkdir2; 8019 struct jaddref *jaddref; 8020 struct mount *mp; 8021 int isindir; 8022 8023 /* 8024 * Whiteouts have no dependencies. 8025 */ 8026 if (newinum == WINO) { 8027 if (newdirbp != NULL) 8028 bdwrite(newdirbp); 8029 return (0); 8030 } 8031 jaddref = NULL; 8032 mkdir1 = mkdir2 = NULL; 8033 mp = UFSTOVFS(dp->i_ump); 8034 fs = dp->i_fs; 8035 lbn = lblkno(fs, diroffset); 8036 offset = blkoff(fs, diroffset); 8037 dap = malloc(sizeof(struct diradd), M_DIRADD, 8038 M_SOFTDEP_FLAGS|M_ZERO); 8039 workitem_alloc(&dap->da_list, D_DIRADD, mp); 8040 dap->da_offset = offset; 8041 dap->da_newinum = newinum; 8042 dap->da_state = ATTACHED; 8043 LIST_INIT(&dap->da_jwork); 8044 isindir = bp->b_lblkno >= NDADDR; 8045 if (isnewblk && 8046 (isindir ? blkoff(fs, diroffset) : fragoff(fs, diroffset)) == 0) { 8047 newdirblk = malloc(sizeof(struct newdirblk), 8048 M_NEWDIRBLK, M_SOFTDEP_FLAGS); 8049 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 8050 LIST_INIT(&newdirblk->db_mkdir); 8051 } 8052 /* 8053 * If we're creating a new directory setup the dependencies and set 8054 * the dap state to wait for them. Otherwise it's COMPLETE and 8055 * we can move on. 8056 */ 8057 if (newdirbp == NULL) { 8058 dap->da_state |= DEPCOMPLETE; 8059 ACQUIRE_LOCK(&lk); 8060 } else { 8061 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 8062 mkdir1 = setup_newdir(dap, newinum, dp->i_number, newdirbp, 8063 &mkdir2); 8064 } 8065 /* 8066 * Link into parent directory pagedep to await its being written. 8067 */ 8068 pagedep_lookup(mp, bp, dp->i_number, lbn, DEPALLOC, &pagedep); 8069#ifdef DEBUG 8070 if (diradd_lookup(pagedep, offset) != NULL) 8071 panic("softdep_setup_directory_add: %p already at off %d\n", 8072 diradd_lookup(pagedep, offset), offset); 8073#endif 8074 dap->da_pagedep = pagedep; 8075 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 8076 da_pdlist); 8077 inodedep_lookup(mp, newinum, DEPALLOC | NODELAY, &inodedep); 8078 /* 8079 * If we're journaling, link the diradd into the jaddref so it 8080 * may be completed after the journal entry is written. Otherwise, 8081 * link the diradd into its inodedep. If the inode is not yet 8082 * written place it on the bufwait list, otherwise do the post-inode 8083 * write processing to put it on the id_pendinghd list. 8084 */ 8085 if (MOUNTEDSUJ(mp)) { 8086 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 8087 inoreflst); 8088 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 8089 ("softdep_setup_directory_add: bad jaddref %p", jaddref)); 8090 jaddref->ja_diroff = diroffset; 8091 jaddref->ja_diradd = dap; 8092 add_to_journal(&jaddref->ja_list); 8093 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 8094 diradd_inode_written(dap, inodedep); 8095 else 8096 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 8097 /* 8098 * Add the journal entries for . and .. links now that the primary 8099 * link is written. 8100 */ 8101 if (mkdir1 != NULL && MOUNTEDSUJ(mp)) { 8102 jaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref, 8103 inoreflst, if_deps); 8104 KASSERT(jaddref != NULL && 8105 jaddref->ja_ino == jaddref->ja_parent && 8106 (jaddref->ja_state & MKDIR_BODY), 8107 ("softdep_setup_directory_add: bad dot jaddref %p", 8108 jaddref)); 8109 mkdir1->md_jaddref = jaddref; 8110 jaddref->ja_mkdir = mkdir1; 8111 /* 8112 * It is important that the dotdot journal entry 8113 * is added prior to the dot entry since dot writes 8114 * both the dot and dotdot links. These both must 8115 * be added after the primary link for the journal 8116 * to remain consistent. 8117 */ 8118 add_to_journal(&mkdir2->md_jaddref->ja_list); 8119 add_to_journal(&jaddref->ja_list); 8120 } 8121 /* 8122 * If we are adding a new directory remember this diradd so that if 8123 * we rename it we can keep the dot and dotdot dependencies. If 8124 * we are adding a new name for an inode that has a mkdiradd we 8125 * must be in rename and we have to move the dot and dotdot 8126 * dependencies to this new name. The old name is being orphaned 8127 * soon. 8128 */ 8129 if (mkdir1 != NULL) { 8130 if (inodedep->id_mkdiradd != NULL) 8131 panic("softdep_setup_directory_add: Existing mkdir"); 8132 inodedep->id_mkdiradd = dap; 8133 } else if (inodedep->id_mkdiradd) 8134 merge_diradd(inodedep, dap); 8135 if (newdirblk) { 8136 /* 8137 * There is nothing to do if we are already tracking 8138 * this block. 8139 */ 8140 if ((pagedep->pd_state & NEWBLOCK) != 0) { 8141 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 8142 FREE_LOCK(&lk); 8143 return (0); 8144 } 8145 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk) 8146 == 0) 8147 panic("softdep_setup_directory_add: lost entry"); 8148 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list); 8149 pagedep->pd_state |= NEWBLOCK; 8150 pagedep->pd_newdirblk = newdirblk; 8151 newdirblk->db_pagedep = pagedep; 8152 FREE_LOCK(&lk); 8153 /* 8154 * If we extended into an indirect signal direnter to sync. 8155 */ 8156 if (isindir) 8157 return (1); 8158 return (0); 8159 } 8160 FREE_LOCK(&lk); 8161 return (0); 8162} 8163 8164/* 8165 * This procedure is called to change the offset of a directory 8166 * entry when compacting a directory block which must be owned 8167 * exclusively by the caller. Note that the actual entry movement 8168 * must be done in this procedure to ensure that no I/O completions 8169 * occur while the move is in progress. 8170 */ 8171void 8172softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize) 8173 struct buf *bp; /* Buffer holding directory block. */ 8174 struct inode *dp; /* inode for directory */ 8175 caddr_t base; /* address of dp->i_offset */ 8176 caddr_t oldloc; /* address of old directory location */ 8177 caddr_t newloc; /* address of new directory location */ 8178 int entrysize; /* size of directory entry */ 8179{ 8180 int offset, oldoffset, newoffset; 8181 struct pagedep *pagedep; 8182 struct jmvref *jmvref; 8183 struct diradd *dap; 8184 struct direct *de; 8185 struct mount *mp; 8186 ufs_lbn_t lbn; 8187 int flags; 8188 8189 mp = UFSTOVFS(dp->i_ump); 8190 de = (struct direct *)oldloc; 8191 jmvref = NULL; 8192 flags = 0; 8193 /* 8194 * Moves are always journaled as it would be too complex to 8195 * determine if any affected adds or removes are present in the 8196 * journal. 8197 */ 8198 if (MOUNTEDSUJ(mp)) { 8199 flags = DEPALLOC; 8200 jmvref = newjmvref(dp, de->d_ino, 8201 dp->i_offset + (oldloc - base), 8202 dp->i_offset + (newloc - base)); 8203 } 8204 lbn = lblkno(dp->i_fs, dp->i_offset); 8205 offset = blkoff(dp->i_fs, dp->i_offset); 8206 oldoffset = offset + (oldloc - base); 8207 newoffset = offset + (newloc - base); 8208 ACQUIRE_LOCK(&lk); 8209 if (pagedep_lookup(mp, bp, dp->i_number, lbn, flags, &pagedep) == 0) 8210 goto done; 8211 dap = diradd_lookup(pagedep, oldoffset); 8212 if (dap) { 8213 dap->da_offset = newoffset; 8214 newoffset = DIRADDHASH(newoffset); 8215 oldoffset = DIRADDHASH(oldoffset); 8216 if ((dap->da_state & ALLCOMPLETE) != ALLCOMPLETE && 8217 newoffset != oldoffset) { 8218 LIST_REMOVE(dap, da_pdlist); 8219 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[newoffset], 8220 dap, da_pdlist); 8221 } 8222 } 8223done: 8224 if (jmvref) { 8225 jmvref->jm_pagedep = pagedep; 8226 LIST_INSERT_HEAD(&pagedep->pd_jmvrefhd, jmvref, jm_deps); 8227 add_to_journal(&jmvref->jm_list); 8228 } 8229 bcopy(oldloc, newloc, entrysize); 8230 FREE_LOCK(&lk); 8231} 8232 8233/* 8234 * Move the mkdir dependencies and journal work from one diradd to another 8235 * when renaming a directory. The new name must depend on the mkdir deps 8236 * completing as the old name did. Directories can only have one valid link 8237 * at a time so one must be canonical. 8238 */ 8239static void 8240merge_diradd(inodedep, newdap) 8241 struct inodedep *inodedep; 8242 struct diradd *newdap; 8243{ 8244 struct diradd *olddap; 8245 struct mkdir *mkdir, *nextmd; 8246 short state; 8247 8248 olddap = inodedep->id_mkdiradd; 8249 inodedep->id_mkdiradd = newdap; 8250 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8251 newdap->da_state &= ~DEPCOMPLETE; 8252 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 8253 nextmd = LIST_NEXT(mkdir, md_mkdirs); 8254 if (mkdir->md_diradd != olddap) 8255 continue; 8256 mkdir->md_diradd = newdap; 8257 state = mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY); 8258 newdap->da_state |= state; 8259 olddap->da_state &= ~state; 8260 if ((olddap->da_state & 8261 (MKDIR_PARENT | MKDIR_BODY)) == 0) 8262 break; 8263 } 8264 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 8265 panic("merge_diradd: unfound ref"); 8266 } 8267 /* 8268 * Any mkdir related journal items are not safe to be freed until 8269 * the new name is stable. 8270 */ 8271 jwork_move(&newdap->da_jwork, &olddap->da_jwork); 8272 olddap->da_state |= DEPCOMPLETE; 8273 complete_diradd(olddap); 8274} 8275 8276/* 8277 * Move the diradd to the pending list when all diradd dependencies are 8278 * complete. 8279 */ 8280static void 8281complete_diradd(dap) 8282 struct diradd *dap; 8283{ 8284 struct pagedep *pagedep; 8285 8286 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 8287 if (dap->da_state & DIRCHG) 8288 pagedep = dap->da_previous->dm_pagedep; 8289 else 8290 pagedep = dap->da_pagedep; 8291 LIST_REMOVE(dap, da_pdlist); 8292 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 8293 } 8294} 8295 8296/* 8297 * Cancel a diradd when a dirrem overlaps with it. We must cancel the journal 8298 * add entries and conditonally journal the remove. 8299 */ 8300static void 8301cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref) 8302 struct diradd *dap; 8303 struct dirrem *dirrem; 8304 struct jremref *jremref; 8305 struct jremref *dotremref; 8306 struct jremref *dotdotremref; 8307{ 8308 struct inodedep *inodedep; 8309 struct jaddref *jaddref; 8310 struct inoref *inoref; 8311 struct mkdir *mkdir; 8312 8313 /* 8314 * If no remove references were allocated we're on a non-journaled 8315 * filesystem and can skip the cancel step. 8316 */ 8317 if (jremref == NULL) { 8318 free_diradd(dap, NULL); 8319 return; 8320 } 8321 /* 8322 * Cancel the primary name an free it if it does not require 8323 * journaling. 8324 */ 8325 if (inodedep_lookup(dap->da_list.wk_mp, dap->da_newinum, 8326 0, &inodedep) != 0) { 8327 /* Abort the addref that reference this diradd. */ 8328 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 8329 if (inoref->if_list.wk_type != D_JADDREF) 8330 continue; 8331 jaddref = (struct jaddref *)inoref; 8332 if (jaddref->ja_diradd != dap) 8333 continue; 8334 if (cancel_jaddref(jaddref, inodedep, 8335 &dirrem->dm_jwork) == 0) { 8336 free_jremref(jremref); 8337 jremref = NULL; 8338 } 8339 break; 8340 } 8341 } 8342 /* 8343 * Cancel subordinate names and free them if they do not require 8344 * journaling. 8345 */ 8346 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8347 LIST_FOREACH(mkdir, &mkdirlisthd, md_mkdirs) { 8348 if (mkdir->md_diradd != dap) 8349 continue; 8350 if ((jaddref = mkdir->md_jaddref) == NULL) 8351 continue; 8352 mkdir->md_jaddref = NULL; 8353 if (mkdir->md_state & MKDIR_PARENT) { 8354 if (cancel_jaddref(jaddref, NULL, 8355 &dirrem->dm_jwork) == 0) { 8356 free_jremref(dotdotremref); 8357 dotdotremref = NULL; 8358 } 8359 } else { 8360 if (cancel_jaddref(jaddref, inodedep, 8361 &dirrem->dm_jwork) == 0) { 8362 free_jremref(dotremref); 8363 dotremref = NULL; 8364 } 8365 } 8366 } 8367 } 8368 8369 if (jremref) 8370 journal_jremref(dirrem, jremref, inodedep); 8371 if (dotremref) 8372 journal_jremref(dirrem, dotremref, inodedep); 8373 if (dotdotremref) 8374 journal_jremref(dirrem, dotdotremref, NULL); 8375 jwork_move(&dirrem->dm_jwork, &dap->da_jwork); 8376 free_diradd(dap, &dirrem->dm_jwork); 8377} 8378 8379/* 8380 * Free a diradd dependency structure. This routine must be called 8381 * with splbio interrupts blocked. 8382 */ 8383static void 8384free_diradd(dap, wkhd) 8385 struct diradd *dap; 8386 struct workhead *wkhd; 8387{ 8388 struct dirrem *dirrem; 8389 struct pagedep *pagedep; 8390 struct inodedep *inodedep; 8391 struct mkdir *mkdir, *nextmd; 8392 8393 mtx_assert(&lk, MA_OWNED); 8394 LIST_REMOVE(dap, da_pdlist); 8395 if (dap->da_state & ONWORKLIST) 8396 WORKLIST_REMOVE(&dap->da_list); 8397 if ((dap->da_state & DIRCHG) == 0) { 8398 pagedep = dap->da_pagedep; 8399 } else { 8400 dirrem = dap->da_previous; 8401 pagedep = dirrem->dm_pagedep; 8402 dirrem->dm_dirinum = pagedep->pd_ino; 8403 dirrem->dm_state |= COMPLETE; 8404 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 8405 add_to_worklist(&dirrem->dm_list, 0); 8406 } 8407 if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum, 8408 0, &inodedep) != 0) 8409 if (inodedep->id_mkdiradd == dap) 8410 inodedep->id_mkdiradd = NULL; 8411 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8412 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 8413 nextmd = LIST_NEXT(mkdir, md_mkdirs); 8414 if (mkdir->md_diradd != dap) 8415 continue; 8416 dap->da_state &= 8417 ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)); 8418 LIST_REMOVE(mkdir, md_mkdirs); 8419 if (mkdir->md_state & ONWORKLIST) 8420 WORKLIST_REMOVE(&mkdir->md_list); 8421 if (mkdir->md_jaddref != NULL) 8422 panic("free_diradd: Unexpected jaddref"); 8423 WORKITEM_FREE(mkdir, D_MKDIR); 8424 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 8425 break; 8426 } 8427 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 8428 panic("free_diradd: unfound ref"); 8429 } 8430 if (inodedep) 8431 free_inodedep(inodedep); 8432 /* 8433 * Free any journal segments waiting for the directory write. 8434 */ 8435 handle_jwork(&dap->da_jwork); 8436 WORKITEM_FREE(dap, D_DIRADD); 8437} 8438 8439/* 8440 * Directory entry removal dependencies. 8441 * 8442 * When removing a directory entry, the entry's inode pointer must be 8443 * zero'ed on disk before the corresponding inode's link count is decremented 8444 * (possibly freeing the inode for re-use). This dependency is handled by 8445 * updating the directory entry but delaying the inode count reduction until 8446 * after the directory block has been written to disk. After this point, the 8447 * inode count can be decremented whenever it is convenient. 8448 */ 8449 8450/* 8451 * This routine should be called immediately after removing 8452 * a directory entry. The inode's link count should not be 8453 * decremented by the calling procedure -- the soft updates 8454 * code will do this task when it is safe. 8455 */ 8456void 8457softdep_setup_remove(bp, dp, ip, isrmdir) 8458 struct buf *bp; /* buffer containing directory block */ 8459 struct inode *dp; /* inode for the directory being modified */ 8460 struct inode *ip; /* inode for directory entry being removed */ 8461 int isrmdir; /* indicates if doing RMDIR */ 8462{ 8463 struct dirrem *dirrem, *prevdirrem; 8464 struct inodedep *inodedep; 8465 int direct; 8466 8467 /* 8468 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. We want 8469 * newdirrem() to setup the full directory remove which requires 8470 * isrmdir > 1. 8471 */ 8472 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 8473 /* 8474 * Add the dirrem to the inodedep's pending remove list for quick 8475 * discovery later. 8476 */ 8477 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, 8478 &inodedep) == 0) 8479 panic("softdep_setup_remove: Lost inodedep."); 8480 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked")); 8481 dirrem->dm_state |= ONDEPLIST; 8482 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 8483 8484 /* 8485 * If the COMPLETE flag is clear, then there were no active 8486 * entries and we want to roll back to a zeroed entry until 8487 * the new inode is committed to disk. If the COMPLETE flag is 8488 * set then we have deleted an entry that never made it to 8489 * disk. If the entry we deleted resulted from a name change, 8490 * then the old name still resides on disk. We cannot delete 8491 * its inode (returned to us in prevdirrem) until the zeroed 8492 * directory entry gets to disk. The new inode has never been 8493 * referenced on the disk, so can be deleted immediately. 8494 */ 8495 if ((dirrem->dm_state & COMPLETE) == 0) { 8496 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 8497 dm_next); 8498 FREE_LOCK(&lk); 8499 } else { 8500 if (prevdirrem != NULL) 8501 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 8502 prevdirrem, dm_next); 8503 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 8504 direct = LIST_EMPTY(&dirrem->dm_jremrefhd); 8505 FREE_LOCK(&lk); 8506 if (direct) 8507 handle_workitem_remove(dirrem, 0); 8508 } 8509} 8510 8511/* 8512 * Check for an entry matching 'offset' on both the pd_dirraddhd list and the 8513 * pd_pendinghd list of a pagedep. 8514 */ 8515static struct diradd * 8516diradd_lookup(pagedep, offset) 8517 struct pagedep *pagedep; 8518 int offset; 8519{ 8520 struct diradd *dap; 8521 8522 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 8523 if (dap->da_offset == offset) 8524 return (dap); 8525 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 8526 if (dap->da_offset == offset) 8527 return (dap); 8528 return (NULL); 8529} 8530 8531/* 8532 * Search for a .. diradd dependency in a directory that is being removed. 8533 * If the directory was renamed to a new parent we have a diradd rather 8534 * than a mkdir for the .. entry. We need to cancel it now before 8535 * it is found in truncate(). 8536 */ 8537static struct jremref * 8538cancel_diradd_dotdot(ip, dirrem, jremref) 8539 struct inode *ip; 8540 struct dirrem *dirrem; 8541 struct jremref *jremref; 8542{ 8543 struct pagedep *pagedep; 8544 struct diradd *dap; 8545 struct worklist *wk; 8546 8547 if (pagedep_lookup(UFSTOVFS(ip->i_ump), NULL, ip->i_number, 0, 0, 8548 &pagedep) == 0) 8549 return (jremref); 8550 dap = diradd_lookup(pagedep, DOTDOT_OFFSET); 8551 if (dap == NULL) 8552 return (jremref); 8553 cancel_diradd(dap, dirrem, jremref, NULL, NULL); 8554 /* 8555 * Mark any journal work as belonging to the parent so it is freed 8556 * with the .. reference. 8557 */ 8558 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list) 8559 wk->wk_state |= MKDIR_PARENT; 8560 return (NULL); 8561} 8562 8563/* 8564 * Cancel the MKDIR_PARENT mkdir component of a diradd when we're going to 8565 * replace it with a dirrem/diradd pair as a result of re-parenting a 8566 * directory. This ensures that we don't simultaneously have a mkdir and 8567 * a diradd for the same .. entry. 8568 */ 8569static struct jremref * 8570cancel_mkdir_dotdot(ip, dirrem, jremref) 8571 struct inode *ip; 8572 struct dirrem *dirrem; 8573 struct jremref *jremref; 8574{ 8575 struct inodedep *inodedep; 8576 struct jaddref *jaddref; 8577 struct mkdir *mkdir; 8578 struct diradd *dap; 8579 8580 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, 8581 &inodedep) == 0) 8582 panic("cancel_mkdir_dotdot: Lost inodedep"); 8583 dap = inodedep->id_mkdiradd; 8584 if (dap == NULL || (dap->da_state & MKDIR_PARENT) == 0) 8585 return (jremref); 8586 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; 8587 mkdir = LIST_NEXT(mkdir, md_mkdirs)) 8588 if (mkdir->md_diradd == dap && mkdir->md_state & MKDIR_PARENT) 8589 break; 8590 if (mkdir == NULL) 8591 panic("cancel_mkdir_dotdot: Unable to find mkdir\n"); 8592 if ((jaddref = mkdir->md_jaddref) != NULL) { 8593 mkdir->md_jaddref = NULL; 8594 jaddref->ja_state &= ~MKDIR_PARENT; 8595 if (inodedep_lookup(UFSTOVFS(ip->i_ump), jaddref->ja_ino, 0, 8596 &inodedep) == 0) 8597 panic("cancel_mkdir_dotdot: Lost parent inodedep"); 8598 if (cancel_jaddref(jaddref, inodedep, &dirrem->dm_jwork)) { 8599 journal_jremref(dirrem, jremref, inodedep); 8600 jremref = NULL; 8601 } 8602 } 8603 if (mkdir->md_state & ONWORKLIST) 8604 WORKLIST_REMOVE(&mkdir->md_list); 8605 mkdir->md_state |= ALLCOMPLETE; 8606 complete_mkdir(mkdir); 8607 return (jremref); 8608} 8609 8610static void 8611journal_jremref(dirrem, jremref, inodedep) 8612 struct dirrem *dirrem; 8613 struct jremref *jremref; 8614 struct inodedep *inodedep; 8615{ 8616 8617 if (inodedep == NULL) 8618 if (inodedep_lookup(jremref->jr_list.wk_mp, 8619 jremref->jr_ref.if_ino, 0, &inodedep) == 0) 8620 panic("journal_jremref: Lost inodedep"); 8621 LIST_INSERT_HEAD(&dirrem->dm_jremrefhd, jremref, jr_deps); 8622 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps); 8623 add_to_journal(&jremref->jr_list); 8624} 8625 8626static void 8627dirrem_journal(dirrem, jremref, dotremref, dotdotremref) 8628 struct dirrem *dirrem; 8629 struct jremref *jremref; 8630 struct jremref *dotremref; 8631 struct jremref *dotdotremref; 8632{ 8633 struct inodedep *inodedep; 8634 8635 8636 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 0, 8637 &inodedep) == 0) 8638 panic("dirrem_journal: Lost inodedep"); 8639 journal_jremref(dirrem, jremref, inodedep); 8640 if (dotremref) 8641 journal_jremref(dirrem, dotremref, inodedep); 8642 if (dotdotremref) 8643 journal_jremref(dirrem, dotdotremref, NULL); 8644} 8645 8646/* 8647 * Allocate a new dirrem if appropriate and return it along with 8648 * its associated pagedep. Called without a lock, returns with lock. 8649 */ 8650static struct dirrem * 8651newdirrem(bp, dp, ip, isrmdir, prevdirremp) 8652 struct buf *bp; /* buffer containing directory block */ 8653 struct inode *dp; /* inode for the directory being modified */ 8654 struct inode *ip; /* inode for directory entry being removed */ 8655 int isrmdir; /* indicates if doing RMDIR */ 8656 struct dirrem **prevdirremp; /* previously referenced inode, if any */ 8657{ 8658 int offset; 8659 ufs_lbn_t lbn; 8660 struct diradd *dap; 8661 struct dirrem *dirrem; 8662 struct pagedep *pagedep; 8663 struct jremref *jremref; 8664 struct jremref *dotremref; 8665 struct jremref *dotdotremref; 8666 struct vnode *dvp; 8667 8668 /* 8669 * Whiteouts have no deletion dependencies. 8670 */ 8671 if (ip == NULL) 8672 panic("newdirrem: whiteout"); 8673 dvp = ITOV(dp); 8674 /* 8675 * If we are over our limit, try to improve the situation. 8676 * Limiting the number of dirrem structures will also limit 8677 * the number of freefile and freeblks structures. 8678 */ 8679 ACQUIRE_LOCK(&lk); 8680 if (!IS_SNAPSHOT(ip) && dep_current[D_DIRREM] > max_softdeps / 2) 8681 (void) request_cleanup(ITOV(dp)->v_mount, FLUSH_BLOCKS); 8682 FREE_LOCK(&lk); 8683 dirrem = malloc(sizeof(struct dirrem), 8684 M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO); 8685 workitem_alloc(&dirrem->dm_list, D_DIRREM, dvp->v_mount); 8686 LIST_INIT(&dirrem->dm_jremrefhd); 8687 LIST_INIT(&dirrem->dm_jwork); 8688 dirrem->dm_state = isrmdir ? RMDIR : 0; 8689 dirrem->dm_oldinum = ip->i_number; 8690 *prevdirremp = NULL; 8691 /* 8692 * Allocate remove reference structures to track journal write 8693 * dependencies. We will always have one for the link and 8694 * when doing directories we will always have one more for dot. 8695 * When renaming a directory we skip the dotdot link change so 8696 * this is not needed. 8697 */ 8698 jremref = dotremref = dotdotremref = NULL; 8699 if (DOINGSUJ(dvp)) { 8700 if (isrmdir) { 8701 jremref = newjremref(dirrem, dp, ip, dp->i_offset, 8702 ip->i_effnlink + 2); 8703 dotremref = newjremref(dirrem, ip, ip, DOT_OFFSET, 8704 ip->i_effnlink + 1); 8705 dotdotremref = newjremref(dirrem, ip, dp, DOTDOT_OFFSET, 8706 dp->i_effnlink + 1); 8707 dotdotremref->jr_state |= MKDIR_PARENT; 8708 } else 8709 jremref = newjremref(dirrem, dp, ip, dp->i_offset, 8710 ip->i_effnlink + 1); 8711 } 8712 ACQUIRE_LOCK(&lk); 8713 lbn = lblkno(dp->i_fs, dp->i_offset); 8714 offset = blkoff(dp->i_fs, dp->i_offset); 8715 pagedep_lookup(UFSTOVFS(dp->i_ump), bp, dp->i_number, lbn, DEPALLOC, 8716 &pagedep); 8717 dirrem->dm_pagedep = pagedep; 8718 dirrem->dm_offset = offset; 8719 /* 8720 * If we're renaming a .. link to a new directory, cancel any 8721 * existing MKDIR_PARENT mkdir. If it has already been canceled 8722 * the jremref is preserved for any potential diradd in this 8723 * location. This can not coincide with a rmdir. 8724 */ 8725 if (dp->i_offset == DOTDOT_OFFSET) { 8726 if (isrmdir) 8727 panic("newdirrem: .. directory change during remove?"); 8728 jremref = cancel_mkdir_dotdot(dp, dirrem, jremref); 8729 } 8730 /* 8731 * If we're removing a directory search for the .. dependency now and 8732 * cancel it. Any pending journal work will be added to the dirrem 8733 * to be completed when the workitem remove completes. 8734 */ 8735 if (isrmdir) 8736 dotdotremref = cancel_diradd_dotdot(ip, dirrem, dotdotremref); 8737 /* 8738 * Check for a diradd dependency for the same directory entry. 8739 * If present, then both dependencies become obsolete and can 8740 * be de-allocated. 8741 */ 8742 dap = diradd_lookup(pagedep, offset); 8743 if (dap == NULL) { 8744 /* 8745 * Link the jremref structures into the dirrem so they are 8746 * written prior to the pagedep. 8747 */ 8748 if (jremref) 8749 dirrem_journal(dirrem, jremref, dotremref, 8750 dotdotremref); 8751 return (dirrem); 8752 } 8753 /* 8754 * Must be ATTACHED at this point. 8755 */ 8756 if ((dap->da_state & ATTACHED) == 0) 8757 panic("newdirrem: not ATTACHED"); 8758 if (dap->da_newinum != ip->i_number) 8759 panic("newdirrem: inum %ju should be %ju", 8760 (uintmax_t)ip->i_number, (uintmax_t)dap->da_newinum); 8761 /* 8762 * If we are deleting a changed name that never made it to disk, 8763 * then return the dirrem describing the previous inode (which 8764 * represents the inode currently referenced from this entry on disk). 8765 */ 8766 if ((dap->da_state & DIRCHG) != 0) { 8767 *prevdirremp = dap->da_previous; 8768 dap->da_state &= ~DIRCHG; 8769 dap->da_pagedep = pagedep; 8770 } 8771 /* 8772 * We are deleting an entry that never made it to disk. 8773 * Mark it COMPLETE so we can delete its inode immediately. 8774 */ 8775 dirrem->dm_state |= COMPLETE; 8776 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref); 8777#ifdef SUJ_DEBUG 8778 if (isrmdir == 0) { 8779 struct worklist *wk; 8780 8781 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list) 8782 if (wk->wk_state & (MKDIR_BODY | MKDIR_PARENT)) 8783 panic("bad wk %p (0x%X)\n", wk, wk->wk_state); 8784 } 8785#endif 8786 8787 return (dirrem); 8788} 8789 8790/* 8791 * Directory entry change dependencies. 8792 * 8793 * Changing an existing directory entry requires that an add operation 8794 * be completed first followed by a deletion. The semantics for the addition 8795 * are identical to the description of adding a new entry above except 8796 * that the rollback is to the old inode number rather than zero. Once 8797 * the addition dependency is completed, the removal is done as described 8798 * in the removal routine above. 8799 */ 8800 8801/* 8802 * This routine should be called immediately after changing 8803 * a directory entry. The inode's link count should not be 8804 * decremented by the calling procedure -- the soft updates 8805 * code will perform this task when it is safe. 8806 */ 8807void 8808softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 8809 struct buf *bp; /* buffer containing directory block */ 8810 struct inode *dp; /* inode for the directory being modified */ 8811 struct inode *ip; /* inode for directory entry being removed */ 8812 ino_t newinum; /* new inode number for changed entry */ 8813 int isrmdir; /* indicates if doing RMDIR */ 8814{ 8815 int offset; 8816 struct diradd *dap = NULL; 8817 struct dirrem *dirrem, *prevdirrem; 8818 struct pagedep *pagedep; 8819 struct inodedep *inodedep; 8820 struct jaddref *jaddref; 8821 struct mount *mp; 8822 8823 offset = blkoff(dp->i_fs, dp->i_offset); 8824 mp = UFSTOVFS(dp->i_ump); 8825 8826 /* 8827 * Whiteouts do not need diradd dependencies. 8828 */ 8829 if (newinum != WINO) { 8830 dap = malloc(sizeof(struct diradd), 8831 M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO); 8832 workitem_alloc(&dap->da_list, D_DIRADD, mp); 8833 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 8834 dap->da_offset = offset; 8835 dap->da_newinum = newinum; 8836 LIST_INIT(&dap->da_jwork); 8837 } 8838 8839 /* 8840 * Allocate a new dirrem and ACQUIRE_LOCK. 8841 */ 8842 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 8843 pagedep = dirrem->dm_pagedep; 8844 /* 8845 * The possible values for isrmdir: 8846 * 0 - non-directory file rename 8847 * 1 - directory rename within same directory 8848 * inum - directory rename to new directory of given inode number 8849 * When renaming to a new directory, we are both deleting and 8850 * creating a new directory entry, so the link count on the new 8851 * directory should not change. Thus we do not need the followup 8852 * dirrem which is usually done in handle_workitem_remove. We set 8853 * the DIRCHG flag to tell handle_workitem_remove to skip the 8854 * followup dirrem. 8855 */ 8856 if (isrmdir > 1) 8857 dirrem->dm_state |= DIRCHG; 8858 8859 /* 8860 * Whiteouts have no additional dependencies, 8861 * so just put the dirrem on the correct list. 8862 */ 8863 if (newinum == WINO) { 8864 if ((dirrem->dm_state & COMPLETE) == 0) { 8865 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 8866 dm_next); 8867 } else { 8868 dirrem->dm_dirinum = pagedep->pd_ino; 8869 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 8870 add_to_worklist(&dirrem->dm_list, 0); 8871 } 8872 FREE_LOCK(&lk); 8873 return; 8874 } 8875 /* 8876 * Add the dirrem to the inodedep's pending remove list for quick 8877 * discovery later. A valid nlinkdelta ensures that this lookup 8878 * will not fail. 8879 */ 8880 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 8881 panic("softdep_setup_directory_change: Lost inodedep."); 8882 dirrem->dm_state |= ONDEPLIST; 8883 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 8884 8885 /* 8886 * If the COMPLETE flag is clear, then there were no active 8887 * entries and we want to roll back to the previous inode until 8888 * the new inode is committed to disk. If the COMPLETE flag is 8889 * set, then we have deleted an entry that never made it to disk. 8890 * If the entry we deleted resulted from a name change, then the old 8891 * inode reference still resides on disk. Any rollback that we do 8892 * needs to be to that old inode (returned to us in prevdirrem). If 8893 * the entry we deleted resulted from a create, then there is 8894 * no entry on the disk, so we want to roll back to zero rather 8895 * than the uncommitted inode. In either of the COMPLETE cases we 8896 * want to immediately free the unwritten and unreferenced inode. 8897 */ 8898 if ((dirrem->dm_state & COMPLETE) == 0) { 8899 dap->da_previous = dirrem; 8900 } else { 8901 if (prevdirrem != NULL) { 8902 dap->da_previous = prevdirrem; 8903 } else { 8904 dap->da_state &= ~DIRCHG; 8905 dap->da_pagedep = pagedep; 8906 } 8907 dirrem->dm_dirinum = pagedep->pd_ino; 8908 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 8909 add_to_worklist(&dirrem->dm_list, 0); 8910 } 8911 /* 8912 * Lookup the jaddref for this journal entry. We must finish 8913 * initializing it and make the diradd write dependent on it. 8914 * If we're not journaling, put it on the id_bufwait list if the 8915 * inode is not yet written. If it is written, do the post-inode 8916 * write processing to put it on the id_pendinghd list. 8917 */ 8918 inodedep_lookup(mp, newinum, DEPALLOC | NODELAY, &inodedep); 8919 if (MOUNTEDSUJ(mp)) { 8920 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 8921 inoreflst); 8922 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 8923 ("softdep_setup_directory_change: bad jaddref %p", 8924 jaddref)); 8925 jaddref->ja_diroff = dp->i_offset; 8926 jaddref->ja_diradd = dap; 8927 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 8928 dap, da_pdlist); 8929 add_to_journal(&jaddref->ja_list); 8930 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 8931 dap->da_state |= COMPLETE; 8932 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 8933 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 8934 } else { 8935 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 8936 dap, da_pdlist); 8937 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 8938 } 8939 /* 8940 * If we're making a new name for a directory that has not been 8941 * committed when need to move the dot and dotdot references to 8942 * this new name. 8943 */ 8944 if (inodedep->id_mkdiradd && dp->i_offset != DOTDOT_OFFSET) 8945 merge_diradd(inodedep, dap); 8946 FREE_LOCK(&lk); 8947} 8948 8949/* 8950 * Called whenever the link count on an inode is changed. 8951 * It creates an inode dependency so that the new reference(s) 8952 * to the inode cannot be committed to disk until the updated 8953 * inode has been written. 8954 */ 8955void 8956softdep_change_linkcnt(ip) 8957 struct inode *ip; /* the inode with the increased link count */ 8958{ 8959 struct inodedep *inodedep; 8960 int dflags; 8961 8962 ACQUIRE_LOCK(&lk); 8963 dflags = DEPALLOC; 8964 if (IS_SNAPSHOT(ip)) 8965 dflags |= NODELAY; 8966 inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, dflags, &inodedep); 8967 if (ip->i_nlink < ip->i_effnlink) 8968 panic("softdep_change_linkcnt: bad delta"); 8969 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 8970 FREE_LOCK(&lk); 8971} 8972 8973/* 8974 * Attach a sbdep dependency to the superblock buf so that we can keep 8975 * track of the head of the linked list of referenced but unlinked inodes. 8976 */ 8977void 8978softdep_setup_sbupdate(ump, fs, bp) 8979 struct ufsmount *ump; 8980 struct fs *fs; 8981 struct buf *bp; 8982{ 8983 struct sbdep *sbdep; 8984 struct worklist *wk; 8985 8986 if (MOUNTEDSUJ(UFSTOVFS(ump)) == 0) 8987 return; 8988 LIST_FOREACH(wk, &bp->b_dep, wk_list) 8989 if (wk->wk_type == D_SBDEP) 8990 break; 8991 if (wk != NULL) 8992 return; 8993 sbdep = malloc(sizeof(struct sbdep), M_SBDEP, M_SOFTDEP_FLAGS); 8994 workitem_alloc(&sbdep->sb_list, D_SBDEP, UFSTOVFS(ump)); 8995 sbdep->sb_fs = fs; 8996 sbdep->sb_ump = ump; 8997 ACQUIRE_LOCK(&lk); 8998 WORKLIST_INSERT(&bp->b_dep, &sbdep->sb_list); 8999 FREE_LOCK(&lk); 9000} 9001 9002/* 9003 * Return the first unlinked inodedep which is ready to be the head of the 9004 * list. The inodedep and all those after it must have valid next pointers. 9005 */ 9006static struct inodedep * 9007first_unlinked_inodedep(ump) 9008 struct ufsmount *ump; 9009{ 9010 struct inodedep *inodedep; 9011 struct inodedep *idp; 9012 9013 mtx_assert(&lk, MA_OWNED); 9014 for (inodedep = TAILQ_LAST(&ump->softdep_unlinked, inodedeplst); 9015 inodedep; inodedep = idp) { 9016 if ((inodedep->id_state & UNLINKNEXT) == 0) 9017 return (NULL); 9018 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9019 if (idp == NULL || (idp->id_state & UNLINKNEXT) == 0) 9020 break; 9021 if ((inodedep->id_state & UNLINKPREV) == 0) 9022 break; 9023 } 9024 return (inodedep); 9025} 9026 9027/* 9028 * Set the sujfree unlinked head pointer prior to writing a superblock. 9029 */ 9030static void 9031initiate_write_sbdep(sbdep) 9032 struct sbdep *sbdep; 9033{ 9034 struct inodedep *inodedep; 9035 struct fs *bpfs; 9036 struct fs *fs; 9037 9038 bpfs = sbdep->sb_fs; 9039 fs = sbdep->sb_ump->um_fs; 9040 inodedep = first_unlinked_inodedep(sbdep->sb_ump); 9041 if (inodedep) { 9042 fs->fs_sujfree = inodedep->id_ino; 9043 inodedep->id_state |= UNLINKPREV; 9044 } else 9045 fs->fs_sujfree = 0; 9046 bpfs->fs_sujfree = fs->fs_sujfree; 9047} 9048 9049/* 9050 * After a superblock is written determine whether it must be written again 9051 * due to a changing unlinked list head. 9052 */ 9053static int 9054handle_written_sbdep(sbdep, bp) 9055 struct sbdep *sbdep; 9056 struct buf *bp; 9057{ 9058 struct inodedep *inodedep; 9059 struct mount *mp; 9060 struct fs *fs; 9061 9062 mtx_assert(&lk, MA_OWNED); 9063 fs = sbdep->sb_fs; 9064 mp = UFSTOVFS(sbdep->sb_ump); 9065 /* 9066 * If the superblock doesn't match the in-memory list start over. 9067 */ 9068 inodedep = first_unlinked_inodedep(sbdep->sb_ump); 9069 if ((inodedep && fs->fs_sujfree != inodedep->id_ino) || 9070 (inodedep == NULL && fs->fs_sujfree != 0)) { 9071 bdirty(bp); 9072 return (1); 9073 } 9074 WORKITEM_FREE(sbdep, D_SBDEP); 9075 if (fs->fs_sujfree == 0) 9076 return (0); 9077 /* 9078 * Now that we have a record of this inode in stable store allow it 9079 * to be written to free up pending work. Inodes may see a lot of 9080 * write activity after they are unlinked which we must not hold up. 9081 */ 9082 for (; inodedep != NULL; inodedep = TAILQ_NEXT(inodedep, id_unlinked)) { 9083 if ((inodedep->id_state & UNLINKLINKS) != UNLINKLINKS) 9084 panic("handle_written_sbdep: Bad inodedep %p (0x%X)", 9085 inodedep, inodedep->id_state); 9086 if (inodedep->id_state & UNLINKONLIST) 9087 break; 9088 inodedep->id_state |= DEPCOMPLETE | UNLINKONLIST; 9089 } 9090 9091 return (0); 9092} 9093 9094/* 9095 * Mark an inodedep as unlinked and insert it into the in-memory unlinked list. 9096 */ 9097static void 9098unlinked_inodedep(mp, inodedep) 9099 struct mount *mp; 9100 struct inodedep *inodedep; 9101{ 9102 struct ufsmount *ump; 9103 9104 mtx_assert(&lk, MA_OWNED); 9105 if (MOUNTEDSUJ(mp) == 0) 9106 return; 9107 ump = VFSTOUFS(mp); 9108 ump->um_fs->fs_fmod = 1; 9109 if (inodedep->id_state & UNLINKED) 9110 panic("unlinked_inodedep: %p already unlinked\n", inodedep); 9111 inodedep->id_state |= UNLINKED; 9112 TAILQ_INSERT_HEAD(&ump->softdep_unlinked, inodedep, id_unlinked); 9113} 9114 9115/* 9116 * Remove an inodedep from the unlinked inodedep list. This may require 9117 * disk writes if the inode has made it that far. 9118 */ 9119static void 9120clear_unlinked_inodedep(inodedep) 9121 struct inodedep *inodedep; 9122{ 9123 struct ufsmount *ump; 9124 struct inodedep *idp; 9125 struct inodedep *idn; 9126 struct fs *fs; 9127 struct buf *bp; 9128 ino_t ino; 9129 ino_t nino; 9130 ino_t pino; 9131 int error; 9132 9133 ump = VFSTOUFS(inodedep->id_list.wk_mp); 9134 fs = ump->um_fs; 9135 ino = inodedep->id_ino; 9136 error = 0; 9137 for (;;) { 9138 mtx_assert(&lk, MA_OWNED); 9139 KASSERT((inodedep->id_state & UNLINKED) != 0, 9140 ("clear_unlinked_inodedep: inodedep %p not unlinked", 9141 inodedep)); 9142 /* 9143 * If nothing has yet been written simply remove us from 9144 * the in memory list and return. This is the most common 9145 * case where handle_workitem_remove() loses the final 9146 * reference. 9147 */ 9148 if ((inodedep->id_state & UNLINKLINKS) == 0) 9149 break; 9150 /* 9151 * If we have a NEXT pointer and no PREV pointer we can simply 9152 * clear NEXT's PREV and remove ourselves from the list. Be 9153 * careful not to clear PREV if the superblock points at 9154 * next as well. 9155 */ 9156 idn = TAILQ_NEXT(inodedep, id_unlinked); 9157 if ((inodedep->id_state & UNLINKLINKS) == UNLINKNEXT) { 9158 if (idn && fs->fs_sujfree != idn->id_ino) 9159 idn->id_state &= ~UNLINKPREV; 9160 break; 9161 } 9162 /* 9163 * Here we have an inodedep which is actually linked into 9164 * the list. We must remove it by forcing a write to the 9165 * link before us, whether it be the superblock or an inode. 9166 * Unfortunately the list may change while we're waiting 9167 * on the buf lock for either resource so we must loop until 9168 * we lock the right one. If both the superblock and an 9169 * inode point to this inode we must clear the inode first 9170 * followed by the superblock. 9171 */ 9172 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9173 pino = 0; 9174 if (idp && (idp->id_state & UNLINKNEXT)) 9175 pino = idp->id_ino; 9176 FREE_LOCK(&lk); 9177 if (pino == 0) 9178 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc), 9179 (int)fs->fs_sbsize, 0, 0, 0); 9180 else 9181 error = bread(ump->um_devvp, 9182 fsbtodb(fs, ino_to_fsba(fs, pino)), 9183 (int)fs->fs_bsize, NOCRED, &bp); 9184 ACQUIRE_LOCK(&lk); 9185 if (error) 9186 break; 9187 /* If the list has changed restart the loop. */ 9188 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9189 nino = 0; 9190 if (idp && (idp->id_state & UNLINKNEXT)) 9191 nino = idp->id_ino; 9192 if (nino != pino || 9193 (inodedep->id_state & UNLINKPREV) != UNLINKPREV) { 9194 FREE_LOCK(&lk); 9195 brelse(bp); 9196 ACQUIRE_LOCK(&lk); 9197 continue; 9198 } 9199 nino = 0; 9200 idn = TAILQ_NEXT(inodedep, id_unlinked); 9201 if (idn) 9202 nino = idn->id_ino; 9203 /* 9204 * Remove us from the in memory list. After this we cannot 9205 * access the inodedep. 9206 */ 9207 KASSERT((inodedep->id_state & UNLINKED) != 0, 9208 ("clear_unlinked_inodedep: inodedep %p not unlinked", 9209 inodedep)); 9210 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST); 9211 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked); 9212 FREE_LOCK(&lk); 9213 /* 9214 * The predecessor's next pointer is manually updated here 9215 * so that the NEXT flag is never cleared for an element 9216 * that is in the list. 9217 */ 9218 if (pino == 0) { 9219 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize); 9220 ffs_oldfscompat_write((struct fs *)bp->b_data, ump); 9221 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data, 9222 bp); 9223 } else if (fs->fs_magic == FS_UFS1_MAGIC) 9224 ((struct ufs1_dinode *)bp->b_data + 9225 ino_to_fsbo(fs, pino))->di_freelink = nino; 9226 else 9227 ((struct ufs2_dinode *)bp->b_data + 9228 ino_to_fsbo(fs, pino))->di_freelink = nino; 9229 /* 9230 * If the bwrite fails we have no recourse to recover. The 9231 * filesystem is corrupted already. 9232 */ 9233 bwrite(bp); 9234 ACQUIRE_LOCK(&lk); 9235 /* 9236 * If the superblock pointer still needs to be cleared force 9237 * a write here. 9238 */ 9239 if (fs->fs_sujfree == ino) { 9240 FREE_LOCK(&lk); 9241 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc), 9242 (int)fs->fs_sbsize, 0, 0, 0); 9243 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize); 9244 ffs_oldfscompat_write((struct fs *)bp->b_data, ump); 9245 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data, 9246 bp); 9247 bwrite(bp); 9248 ACQUIRE_LOCK(&lk); 9249 } 9250 9251 if (fs->fs_sujfree != ino) 9252 return; 9253 panic("clear_unlinked_inodedep: Failed to clear free head"); 9254 } 9255 if (inodedep->id_ino == fs->fs_sujfree) 9256 panic("clear_unlinked_inodedep: Freeing head of free list"); 9257 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST); 9258 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked); 9259 return; 9260} 9261 9262/* 9263 * This workitem decrements the inode's link count. 9264 * If the link count reaches zero, the file is removed. 9265 */ 9266static int 9267handle_workitem_remove(dirrem, flags) 9268 struct dirrem *dirrem; 9269 int flags; 9270{ 9271 struct inodedep *inodedep; 9272 struct workhead dotdotwk; 9273 struct worklist *wk; 9274 struct ufsmount *ump; 9275 struct mount *mp; 9276 struct vnode *vp; 9277 struct inode *ip; 9278 ino_t oldinum; 9279 9280 if (dirrem->dm_state & ONWORKLIST) 9281 panic("handle_workitem_remove: dirrem %p still on worklist", 9282 dirrem); 9283 oldinum = dirrem->dm_oldinum; 9284 mp = dirrem->dm_list.wk_mp; 9285 ump = VFSTOUFS(mp); 9286 flags |= LK_EXCLUSIVE; 9287 if (ffs_vgetf(mp, oldinum, flags, &vp, FFSV_FORCEINSMQ) != 0) 9288 return (EBUSY); 9289 ip = VTOI(vp); 9290 ACQUIRE_LOCK(&lk); 9291 if ((inodedep_lookup(mp, oldinum, 0, &inodedep)) == 0) 9292 panic("handle_workitem_remove: lost inodedep"); 9293 if (dirrem->dm_state & ONDEPLIST) 9294 LIST_REMOVE(dirrem, dm_inonext); 9295 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd), 9296 ("handle_workitem_remove: Journal entries not written.")); 9297 9298 /* 9299 * Move all dependencies waiting on the remove to complete 9300 * from the dirrem to the inode inowait list to be completed 9301 * after the inode has been updated and written to disk. Any 9302 * marked MKDIR_PARENT are saved to be completed when the .. ref 9303 * is removed. 9304 */ 9305 LIST_INIT(&dotdotwk); 9306 while ((wk = LIST_FIRST(&dirrem->dm_jwork)) != NULL) { 9307 WORKLIST_REMOVE(wk); 9308 if (wk->wk_state & MKDIR_PARENT) { 9309 wk->wk_state &= ~MKDIR_PARENT; 9310 WORKLIST_INSERT(&dotdotwk, wk); 9311 continue; 9312 } 9313 WORKLIST_INSERT(&inodedep->id_inowait, wk); 9314 } 9315 LIST_SWAP(&dirrem->dm_jwork, &dotdotwk, worklist, wk_list); 9316 /* 9317 * Normal file deletion. 9318 */ 9319 if ((dirrem->dm_state & RMDIR) == 0) { 9320 ip->i_nlink--; 9321 DIP_SET(ip, i_nlink, ip->i_nlink); 9322 ip->i_flag |= IN_CHANGE; 9323 if (ip->i_nlink < ip->i_effnlink) 9324 panic("handle_workitem_remove: bad file delta"); 9325 if (ip->i_nlink == 0) 9326 unlinked_inodedep(mp, inodedep); 9327 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9328 KASSERT(LIST_EMPTY(&dirrem->dm_jwork), 9329 ("handle_workitem_remove: worklist not empty. %s", 9330 TYPENAME(LIST_FIRST(&dirrem->dm_jwork)->wk_type))); 9331 WORKITEM_FREE(dirrem, D_DIRREM); 9332 FREE_LOCK(&lk); 9333 goto out; 9334 } 9335 /* 9336 * Directory deletion. Decrement reference count for both the 9337 * just deleted parent directory entry and the reference for ".". 9338 * Arrange to have the reference count on the parent decremented 9339 * to account for the loss of "..". 9340 */ 9341 ip->i_nlink -= 2; 9342 DIP_SET(ip, i_nlink, ip->i_nlink); 9343 ip->i_flag |= IN_CHANGE; 9344 if (ip->i_nlink < ip->i_effnlink) 9345 panic("handle_workitem_remove: bad dir delta"); 9346 if (ip->i_nlink == 0) 9347 unlinked_inodedep(mp, inodedep); 9348 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9349 /* 9350 * Rename a directory to a new parent. Since, we are both deleting 9351 * and creating a new directory entry, the link count on the new 9352 * directory should not change. Thus we skip the followup dirrem. 9353 */ 9354 if (dirrem->dm_state & DIRCHG) { 9355 KASSERT(LIST_EMPTY(&dirrem->dm_jwork), 9356 ("handle_workitem_remove: DIRCHG and worklist not empty.")); 9357 WORKITEM_FREE(dirrem, D_DIRREM); 9358 FREE_LOCK(&lk); 9359 goto out; 9360 } 9361 dirrem->dm_state = ONDEPLIST; 9362 dirrem->dm_oldinum = dirrem->dm_dirinum; 9363 /* 9364 * Place the dirrem on the parent's diremhd list. 9365 */ 9366 if (inodedep_lookup(mp, dirrem->dm_oldinum, 0, &inodedep) == 0) 9367 panic("handle_workitem_remove: lost dir inodedep"); 9368 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 9369 /* 9370 * If the allocated inode has never been written to disk, then 9371 * the on-disk inode is zero'ed and we can remove the file 9372 * immediately. When journaling if the inode has been marked 9373 * unlinked and not DEPCOMPLETE we know it can never be written. 9374 */ 9375 inodedep_lookup(mp, oldinum, 0, &inodedep); 9376 if (inodedep == NULL || 9377 (inodedep->id_state & (DEPCOMPLETE | UNLINKED)) == UNLINKED || 9378 check_inode_unwritten(inodedep)) { 9379 FREE_LOCK(&lk); 9380 vput(vp); 9381 return handle_workitem_remove(dirrem, flags); 9382 } 9383 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 9384 FREE_LOCK(&lk); 9385 ip->i_flag |= IN_CHANGE; 9386out: 9387 ffs_update(vp, 0); 9388 vput(vp); 9389 return (0); 9390} 9391 9392/* 9393 * Inode de-allocation dependencies. 9394 * 9395 * When an inode's link count is reduced to zero, it can be de-allocated. We 9396 * found it convenient to postpone de-allocation until after the inode is 9397 * written to disk with its new link count (zero). At this point, all of the 9398 * on-disk inode's block pointers are nullified and, with careful dependency 9399 * list ordering, all dependencies related to the inode will be satisfied and 9400 * the corresponding dependency structures de-allocated. So, if/when the 9401 * inode is reused, there will be no mixing of old dependencies with new 9402 * ones. This artificial dependency is set up by the block de-allocation 9403 * procedure above (softdep_setup_freeblocks) and completed by the 9404 * following procedure. 9405 */ 9406static void 9407handle_workitem_freefile(freefile) 9408 struct freefile *freefile; 9409{ 9410 struct workhead wkhd; 9411 struct fs *fs; 9412 struct inodedep *idp; 9413 struct ufsmount *ump; 9414 int error; 9415 9416 ump = VFSTOUFS(freefile->fx_list.wk_mp); 9417 fs = ump->um_fs; 9418#ifdef DEBUG 9419 ACQUIRE_LOCK(&lk); 9420 error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp); 9421 FREE_LOCK(&lk); 9422 if (error) 9423 panic("handle_workitem_freefile: inodedep %p survived", idp); 9424#endif 9425 UFS_LOCK(ump); 9426 fs->fs_pendinginodes -= 1; 9427 UFS_UNLOCK(ump); 9428 LIST_INIT(&wkhd); 9429 LIST_SWAP(&freefile->fx_jwork, &wkhd, worklist, wk_list); 9430 if ((error = ffs_freefile(ump, fs, freefile->fx_devvp, 9431 freefile->fx_oldinum, freefile->fx_mode, &wkhd)) != 0) 9432 softdep_error("handle_workitem_freefile", error); 9433 ACQUIRE_LOCK(&lk); 9434 WORKITEM_FREE(freefile, D_FREEFILE); 9435 FREE_LOCK(&lk); 9436} 9437 9438 9439/* 9440 * Helper function which unlinks marker element from work list and returns 9441 * the next element on the list. 9442 */ 9443static __inline struct worklist * 9444markernext(struct worklist *marker) 9445{ 9446 struct worklist *next; 9447 9448 next = LIST_NEXT(marker, wk_list); 9449 LIST_REMOVE(marker, wk_list); 9450 return next; 9451} 9452 9453/* 9454 * Disk writes. 9455 * 9456 * The dependency structures constructed above are most actively used when file 9457 * system blocks are written to disk. No constraints are placed on when a 9458 * block can be written, but unsatisfied update dependencies are made safe by 9459 * modifying (or replacing) the source memory for the duration of the disk 9460 * write. When the disk write completes, the memory block is again brought 9461 * up-to-date. 9462 * 9463 * In-core inode structure reclamation. 9464 * 9465 * Because there are a finite number of "in-core" inode structures, they are 9466 * reused regularly. By transferring all inode-related dependencies to the 9467 * in-memory inode block and indexing them separately (via "inodedep"s), we 9468 * can allow "in-core" inode structures to be reused at any time and avoid 9469 * any increase in contention. 9470 * 9471 * Called just before entering the device driver to initiate a new disk I/O. 9472 * The buffer must be locked, thus, no I/O completion operations can occur 9473 * while we are manipulating its associated dependencies. 9474 */ 9475static void 9476softdep_disk_io_initiation(bp) 9477 struct buf *bp; /* structure describing disk write to occur */ 9478{ 9479 struct worklist *wk; 9480 struct worklist marker; 9481 struct inodedep *inodedep; 9482 struct freeblks *freeblks; 9483 struct jblkdep *jblkdep; 9484 struct newblk *newblk; 9485 9486 /* 9487 * We only care about write operations. There should never 9488 * be dependencies for reads. 9489 */ 9490 if (bp->b_iocmd != BIO_WRITE) 9491 panic("softdep_disk_io_initiation: not write"); 9492 9493 if (bp->b_vflags & BV_BKGRDINPROG) 9494 panic("softdep_disk_io_initiation: Writing buffer with " 9495 "background write in progress: %p", bp); 9496 9497 marker.wk_type = D_LAST + 1; /* Not a normal workitem */ 9498 PHOLD(curproc); /* Don't swap out kernel stack */ 9499 9500 ACQUIRE_LOCK(&lk); 9501 /* 9502 * Do any necessary pre-I/O processing. 9503 */ 9504 for (wk = LIST_FIRST(&bp->b_dep); wk != NULL; 9505 wk = markernext(&marker)) { 9506 LIST_INSERT_AFTER(wk, &marker, wk_list); 9507 switch (wk->wk_type) { 9508 9509 case D_PAGEDEP: 9510 initiate_write_filepage(WK_PAGEDEP(wk), bp); 9511 continue; 9512 9513 case D_INODEDEP: 9514 inodedep = WK_INODEDEP(wk); 9515 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) 9516 initiate_write_inodeblock_ufs1(inodedep, bp); 9517 else 9518 initiate_write_inodeblock_ufs2(inodedep, bp); 9519 continue; 9520 9521 case D_INDIRDEP: 9522 initiate_write_indirdep(WK_INDIRDEP(wk), bp); 9523 continue; 9524 9525 case D_BMSAFEMAP: 9526 initiate_write_bmsafemap(WK_BMSAFEMAP(wk), bp); 9527 continue; 9528 9529 case D_JSEG: 9530 WK_JSEG(wk)->js_buf = NULL; 9531 continue; 9532 9533 case D_FREEBLKS: 9534 freeblks = WK_FREEBLKS(wk); 9535 jblkdep = LIST_FIRST(&freeblks->fb_jblkdephd); 9536 /* 9537 * We have to wait for the freeblks to be journaled 9538 * before we can write an inodeblock with updated 9539 * pointers. Be careful to arrange the marker so 9540 * we revisit the freeblks if it's not removed by 9541 * the first jwait(). 9542 */ 9543 if (jblkdep != NULL) { 9544 LIST_REMOVE(&marker, wk_list); 9545 LIST_INSERT_BEFORE(wk, &marker, wk_list); 9546 jwait(&jblkdep->jb_list, MNT_WAIT); 9547 } 9548 continue; 9549 case D_ALLOCDIRECT: 9550 case D_ALLOCINDIR: 9551 /* 9552 * We have to wait for the jnewblk to be journaled 9553 * before we can write to a block if the contents 9554 * may be confused with an earlier file's indirect 9555 * at recovery time. Handle the marker as described 9556 * above. 9557 */ 9558 newblk = WK_NEWBLK(wk); 9559 if (newblk->nb_jnewblk != NULL && 9560 indirblk_lookup(newblk->nb_list.wk_mp, 9561 newblk->nb_newblkno)) { 9562 LIST_REMOVE(&marker, wk_list); 9563 LIST_INSERT_BEFORE(wk, &marker, wk_list); 9564 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 9565 } 9566 continue; 9567 9568 case D_SBDEP: 9569 initiate_write_sbdep(WK_SBDEP(wk)); 9570 continue; 9571 9572 case D_MKDIR: 9573 case D_FREEWORK: 9574 case D_FREEDEP: 9575 case D_JSEGDEP: 9576 continue; 9577 9578 default: 9579 panic("handle_disk_io_initiation: Unexpected type %s", 9580 TYPENAME(wk->wk_type)); 9581 /* NOTREACHED */ 9582 } 9583 } 9584 FREE_LOCK(&lk); 9585 PRELE(curproc); /* Allow swapout of kernel stack */ 9586} 9587 9588/* 9589 * Called from within the procedure above to deal with unsatisfied 9590 * allocation dependencies in a directory. The buffer must be locked, 9591 * thus, no I/O completion operations can occur while we are 9592 * manipulating its associated dependencies. 9593 */ 9594static void 9595initiate_write_filepage(pagedep, bp) 9596 struct pagedep *pagedep; 9597 struct buf *bp; 9598{ 9599 struct jremref *jremref; 9600 struct jmvref *jmvref; 9601 struct dirrem *dirrem; 9602 struct diradd *dap; 9603 struct direct *ep; 9604 int i; 9605 9606 if (pagedep->pd_state & IOSTARTED) { 9607 /* 9608 * This can only happen if there is a driver that does not 9609 * understand chaining. Here biodone will reissue the call 9610 * to strategy for the incomplete buffers. 9611 */ 9612 printf("initiate_write_filepage: already started\n"); 9613 return; 9614 } 9615 pagedep->pd_state |= IOSTARTED; 9616 /* 9617 * Wait for all journal remove dependencies to hit the disk. 9618 * We can not allow any potentially conflicting directory adds 9619 * to be visible before removes and rollback is too difficult. 9620 * lk may be dropped and re-acquired, however we hold the buf 9621 * locked so the dependency can not go away. 9622 */ 9623 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) 9624 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) 9625 jwait(&jremref->jr_list, MNT_WAIT); 9626 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) 9627 jwait(&jmvref->jm_list, MNT_WAIT); 9628 for (i = 0; i < DAHASHSZ; i++) { 9629 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 9630 ep = (struct direct *) 9631 ((char *)bp->b_data + dap->da_offset); 9632 if (ep->d_ino != dap->da_newinum) 9633 panic("%s: dir inum %ju != new %ju", 9634 "initiate_write_filepage", 9635 (uintmax_t)ep->d_ino, 9636 (uintmax_t)dap->da_newinum); 9637 if (dap->da_state & DIRCHG) 9638 ep->d_ino = dap->da_previous->dm_oldinum; 9639 else 9640 ep->d_ino = 0; 9641 dap->da_state &= ~ATTACHED; 9642 dap->da_state |= UNDONE; 9643 } 9644 } 9645} 9646 9647/* 9648 * Version of initiate_write_inodeblock that handles UFS1 dinodes. 9649 * Note that any bug fixes made to this routine must be done in the 9650 * version found below. 9651 * 9652 * Called from within the procedure above to deal with unsatisfied 9653 * allocation dependencies in an inodeblock. The buffer must be 9654 * locked, thus, no I/O completion operations can occur while we 9655 * are manipulating its associated dependencies. 9656 */ 9657static void 9658initiate_write_inodeblock_ufs1(inodedep, bp) 9659 struct inodedep *inodedep; 9660 struct buf *bp; /* The inode block */ 9661{ 9662 struct allocdirect *adp, *lastadp; 9663 struct ufs1_dinode *dp; 9664 struct ufs1_dinode *sip; 9665 struct inoref *inoref; 9666 struct fs *fs; 9667 ufs_lbn_t i; 9668#ifdef INVARIANTS 9669 ufs_lbn_t prevlbn = 0; 9670#endif 9671 int deplist; 9672 9673 if (inodedep->id_state & IOSTARTED) 9674 panic("initiate_write_inodeblock_ufs1: already started"); 9675 inodedep->id_state |= IOSTARTED; 9676 fs = inodedep->id_fs; 9677 dp = (struct ufs1_dinode *)bp->b_data + 9678 ino_to_fsbo(fs, inodedep->id_ino); 9679 9680 /* 9681 * If we're on the unlinked list but have not yet written our 9682 * next pointer initialize it here. 9683 */ 9684 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) { 9685 struct inodedep *inon; 9686 9687 inon = TAILQ_NEXT(inodedep, id_unlinked); 9688 dp->di_freelink = inon ? inon->id_ino : 0; 9689 } 9690 /* 9691 * If the bitmap is not yet written, then the allocated 9692 * inode cannot be written to disk. 9693 */ 9694 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 9695 if (inodedep->id_savedino1 != NULL) 9696 panic("initiate_write_inodeblock_ufs1: I/O underway"); 9697 FREE_LOCK(&lk); 9698 sip = malloc(sizeof(struct ufs1_dinode), 9699 M_SAVEDINO, M_SOFTDEP_FLAGS); 9700 ACQUIRE_LOCK(&lk); 9701 inodedep->id_savedino1 = sip; 9702 *inodedep->id_savedino1 = *dp; 9703 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 9704 dp->di_gen = inodedep->id_savedino1->di_gen; 9705 dp->di_freelink = inodedep->id_savedino1->di_freelink; 9706 return; 9707 } 9708 /* 9709 * If no dependencies, then there is nothing to roll back. 9710 */ 9711 inodedep->id_savedsize = dp->di_size; 9712 inodedep->id_savedextsize = 0; 9713 inodedep->id_savednlink = dp->di_nlink; 9714 if (TAILQ_EMPTY(&inodedep->id_inoupdt) && 9715 TAILQ_EMPTY(&inodedep->id_inoreflst)) 9716 return; 9717 /* 9718 * Revert the link count to that of the first unwritten journal entry. 9719 */ 9720 inoref = TAILQ_FIRST(&inodedep->id_inoreflst); 9721 if (inoref) 9722 dp->di_nlink = inoref->if_nlink; 9723 /* 9724 * Set the dependencies to busy. 9725 */ 9726 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 9727 adp = TAILQ_NEXT(adp, ad_next)) { 9728#ifdef INVARIANTS 9729 if (deplist != 0 && prevlbn >= adp->ad_offset) 9730 panic("softdep_write_inodeblock: lbn order"); 9731 prevlbn = adp->ad_offset; 9732 if (adp->ad_offset < NDADDR && 9733 dp->di_db[adp->ad_offset] != adp->ad_newblkno) 9734 panic("%s: direct pointer #%jd mismatch %d != %jd", 9735 "softdep_write_inodeblock", 9736 (intmax_t)adp->ad_offset, 9737 dp->di_db[adp->ad_offset], 9738 (intmax_t)adp->ad_newblkno); 9739 if (adp->ad_offset >= NDADDR && 9740 dp->di_ib[adp->ad_offset - NDADDR] != adp->ad_newblkno) 9741 panic("%s: indirect pointer #%jd mismatch %d != %jd", 9742 "softdep_write_inodeblock", 9743 (intmax_t)adp->ad_offset - NDADDR, 9744 dp->di_ib[adp->ad_offset - NDADDR], 9745 (intmax_t)adp->ad_newblkno); 9746 deplist |= 1 << adp->ad_offset; 9747 if ((adp->ad_state & ATTACHED) == 0) 9748 panic("softdep_write_inodeblock: Unknown state 0x%x", 9749 adp->ad_state); 9750#endif /* INVARIANTS */ 9751 adp->ad_state &= ~ATTACHED; 9752 adp->ad_state |= UNDONE; 9753 } 9754 /* 9755 * The on-disk inode cannot claim to be any larger than the last 9756 * fragment that has been written. Otherwise, the on-disk inode 9757 * might have fragments that were not the last block in the file 9758 * which would corrupt the filesystem. 9759 */ 9760 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 9761 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 9762 if (adp->ad_offset >= NDADDR) 9763 break; 9764 dp->di_db[adp->ad_offset] = adp->ad_oldblkno; 9765 /* keep going until hitting a rollback to a frag */ 9766 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 9767 continue; 9768 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 9769 for (i = adp->ad_offset + 1; i < NDADDR; i++) { 9770#ifdef INVARIANTS 9771 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 9772 panic("softdep_write_inodeblock: lost dep1"); 9773#endif /* INVARIANTS */ 9774 dp->di_db[i] = 0; 9775 } 9776 for (i = 0; i < NIADDR; i++) { 9777#ifdef INVARIANTS 9778 if (dp->di_ib[i] != 0 && 9779 (deplist & ((1 << NDADDR) << i)) == 0) 9780 panic("softdep_write_inodeblock: lost dep2"); 9781#endif /* INVARIANTS */ 9782 dp->di_ib[i] = 0; 9783 } 9784 return; 9785 } 9786 /* 9787 * If we have zero'ed out the last allocated block of the file, 9788 * roll back the size to the last currently allocated block. 9789 * We know that this last allocated block is a full-sized as 9790 * we already checked for fragments in the loop above. 9791 */ 9792 if (lastadp != NULL && 9793 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 9794 for (i = lastadp->ad_offset; i >= 0; i--) 9795 if (dp->di_db[i] != 0) 9796 break; 9797 dp->di_size = (i + 1) * fs->fs_bsize; 9798 } 9799 /* 9800 * The only dependencies are for indirect blocks. 9801 * 9802 * The file size for indirect block additions is not guaranteed. 9803 * Such a guarantee would be non-trivial to achieve. The conventional 9804 * synchronous write implementation also does not make this guarantee. 9805 * Fsck should catch and fix discrepancies. Arguably, the file size 9806 * can be over-estimated without destroying integrity when the file 9807 * moves into the indirect blocks (i.e., is large). If we want to 9808 * postpone fsck, we are stuck with this argument. 9809 */ 9810 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 9811 dp->di_ib[adp->ad_offset - NDADDR] = 0; 9812} 9813 9814/* 9815 * Version of initiate_write_inodeblock that handles UFS2 dinodes. 9816 * Note that any bug fixes made to this routine must be done in the 9817 * version found above. 9818 * 9819 * Called from within the procedure above to deal with unsatisfied 9820 * allocation dependencies in an inodeblock. The buffer must be 9821 * locked, thus, no I/O completion operations can occur while we 9822 * are manipulating its associated dependencies. 9823 */ 9824static void 9825initiate_write_inodeblock_ufs2(inodedep, bp) 9826 struct inodedep *inodedep; 9827 struct buf *bp; /* The inode block */ 9828{ 9829 struct allocdirect *adp, *lastadp; 9830 struct ufs2_dinode *dp; 9831 struct ufs2_dinode *sip; 9832 struct inoref *inoref; 9833 struct fs *fs; 9834 ufs_lbn_t i; 9835#ifdef INVARIANTS 9836 ufs_lbn_t prevlbn = 0; 9837#endif 9838 int deplist; 9839 9840 if (inodedep->id_state & IOSTARTED) 9841 panic("initiate_write_inodeblock_ufs2: already started"); 9842 inodedep->id_state |= IOSTARTED; 9843 fs = inodedep->id_fs; 9844 dp = (struct ufs2_dinode *)bp->b_data + 9845 ino_to_fsbo(fs, inodedep->id_ino); 9846 9847 /* 9848 * If we're on the unlinked list but have not yet written our 9849 * next pointer initialize it here. 9850 */ 9851 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) { 9852 struct inodedep *inon; 9853 9854 inon = TAILQ_NEXT(inodedep, id_unlinked); 9855 dp->di_freelink = inon ? inon->id_ino : 0; 9856 } 9857 /* 9858 * If the bitmap is not yet written, then the allocated 9859 * inode cannot be written to disk. 9860 */ 9861 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 9862 if (inodedep->id_savedino2 != NULL) 9863 panic("initiate_write_inodeblock_ufs2: I/O underway"); 9864 FREE_LOCK(&lk); 9865 sip = malloc(sizeof(struct ufs2_dinode), 9866 M_SAVEDINO, M_SOFTDEP_FLAGS); 9867 ACQUIRE_LOCK(&lk); 9868 inodedep->id_savedino2 = sip; 9869 *inodedep->id_savedino2 = *dp; 9870 bzero((caddr_t)dp, sizeof(struct ufs2_dinode)); 9871 dp->di_gen = inodedep->id_savedino2->di_gen; 9872 dp->di_freelink = inodedep->id_savedino2->di_freelink; 9873 return; 9874 } 9875 /* 9876 * If no dependencies, then there is nothing to roll back. 9877 */ 9878 inodedep->id_savedsize = dp->di_size; 9879 inodedep->id_savedextsize = dp->di_extsize; 9880 inodedep->id_savednlink = dp->di_nlink; 9881 if (TAILQ_EMPTY(&inodedep->id_inoupdt) && 9882 TAILQ_EMPTY(&inodedep->id_extupdt) && 9883 TAILQ_EMPTY(&inodedep->id_inoreflst)) 9884 return; 9885 /* 9886 * Revert the link count to that of the first unwritten journal entry. 9887 */ 9888 inoref = TAILQ_FIRST(&inodedep->id_inoreflst); 9889 if (inoref) 9890 dp->di_nlink = inoref->if_nlink; 9891 9892 /* 9893 * Set the ext data dependencies to busy. 9894 */ 9895 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 9896 adp = TAILQ_NEXT(adp, ad_next)) { 9897#ifdef INVARIANTS 9898 if (deplist != 0 && prevlbn >= adp->ad_offset) 9899 panic("softdep_write_inodeblock: lbn order"); 9900 prevlbn = adp->ad_offset; 9901 if (dp->di_extb[adp->ad_offset] != adp->ad_newblkno) 9902 panic("%s: direct pointer #%jd mismatch %jd != %jd", 9903 "softdep_write_inodeblock", 9904 (intmax_t)adp->ad_offset, 9905 (intmax_t)dp->di_extb[adp->ad_offset], 9906 (intmax_t)adp->ad_newblkno); 9907 deplist |= 1 << adp->ad_offset; 9908 if ((adp->ad_state & ATTACHED) == 0) 9909 panic("softdep_write_inodeblock: Unknown state 0x%x", 9910 adp->ad_state); 9911#endif /* INVARIANTS */ 9912 adp->ad_state &= ~ATTACHED; 9913 adp->ad_state |= UNDONE; 9914 } 9915 /* 9916 * The on-disk inode cannot claim to be any larger than the last 9917 * fragment that has been written. Otherwise, the on-disk inode 9918 * might have fragments that were not the last block in the ext 9919 * data which would corrupt the filesystem. 9920 */ 9921 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 9922 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 9923 dp->di_extb[adp->ad_offset] = adp->ad_oldblkno; 9924 /* keep going until hitting a rollback to a frag */ 9925 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 9926 continue; 9927 dp->di_extsize = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 9928 for (i = adp->ad_offset + 1; i < NXADDR; i++) { 9929#ifdef INVARIANTS 9930 if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0) 9931 panic("softdep_write_inodeblock: lost dep1"); 9932#endif /* INVARIANTS */ 9933 dp->di_extb[i] = 0; 9934 } 9935 lastadp = NULL; 9936 break; 9937 } 9938 /* 9939 * If we have zero'ed out the last allocated block of the ext 9940 * data, roll back the size to the last currently allocated block. 9941 * We know that this last allocated block is a full-sized as 9942 * we already checked for fragments in the loop above. 9943 */ 9944 if (lastadp != NULL && 9945 dp->di_extsize <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 9946 for (i = lastadp->ad_offset; i >= 0; i--) 9947 if (dp->di_extb[i] != 0) 9948 break; 9949 dp->di_extsize = (i + 1) * fs->fs_bsize; 9950 } 9951 /* 9952 * Set the file data dependencies to busy. 9953 */ 9954 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 9955 adp = TAILQ_NEXT(adp, ad_next)) { 9956#ifdef INVARIANTS 9957 if (deplist != 0 && prevlbn >= adp->ad_offset) 9958 panic("softdep_write_inodeblock: lbn order"); 9959 if ((adp->ad_state & ATTACHED) == 0) 9960 panic("inodedep %p and adp %p not attached", inodedep, adp); 9961 prevlbn = adp->ad_offset; 9962 if (adp->ad_offset < NDADDR && 9963 dp->di_db[adp->ad_offset] != adp->ad_newblkno) 9964 panic("%s: direct pointer #%jd mismatch %jd != %jd", 9965 "softdep_write_inodeblock", 9966 (intmax_t)adp->ad_offset, 9967 (intmax_t)dp->di_db[adp->ad_offset], 9968 (intmax_t)adp->ad_newblkno); 9969 if (adp->ad_offset >= NDADDR && 9970 dp->di_ib[adp->ad_offset - NDADDR] != adp->ad_newblkno) 9971 panic("%s indirect pointer #%jd mismatch %jd != %jd", 9972 "softdep_write_inodeblock:", 9973 (intmax_t)adp->ad_offset - NDADDR, 9974 (intmax_t)dp->di_ib[adp->ad_offset - NDADDR], 9975 (intmax_t)adp->ad_newblkno); 9976 deplist |= 1 << adp->ad_offset; 9977 if ((adp->ad_state & ATTACHED) == 0) 9978 panic("softdep_write_inodeblock: Unknown state 0x%x", 9979 adp->ad_state); 9980#endif /* INVARIANTS */ 9981 adp->ad_state &= ~ATTACHED; 9982 adp->ad_state |= UNDONE; 9983 } 9984 /* 9985 * The on-disk inode cannot claim to be any larger than the last 9986 * fragment that has been written. Otherwise, the on-disk inode 9987 * might have fragments that were not the last block in the file 9988 * which would corrupt the filesystem. 9989 */ 9990 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 9991 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 9992 if (adp->ad_offset >= NDADDR) 9993 break; 9994 dp->di_db[adp->ad_offset] = adp->ad_oldblkno; 9995 /* keep going until hitting a rollback to a frag */ 9996 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 9997 continue; 9998 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 9999 for (i = adp->ad_offset + 1; i < NDADDR; i++) { 10000#ifdef INVARIANTS 10001 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 10002 panic("softdep_write_inodeblock: lost dep2"); 10003#endif /* INVARIANTS */ 10004 dp->di_db[i] = 0; 10005 } 10006 for (i = 0; i < NIADDR; i++) { 10007#ifdef INVARIANTS 10008 if (dp->di_ib[i] != 0 && 10009 (deplist & ((1 << NDADDR) << i)) == 0) 10010 panic("softdep_write_inodeblock: lost dep3"); 10011#endif /* INVARIANTS */ 10012 dp->di_ib[i] = 0; 10013 } 10014 return; 10015 } 10016 /* 10017 * If we have zero'ed out the last allocated block of the file, 10018 * roll back the size to the last currently allocated block. 10019 * We know that this last allocated block is a full-sized as 10020 * we already checked for fragments in the loop above. 10021 */ 10022 if (lastadp != NULL && 10023 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10024 for (i = lastadp->ad_offset; i >= 0; i--) 10025 if (dp->di_db[i] != 0) 10026 break; 10027 dp->di_size = (i + 1) * fs->fs_bsize; 10028 } 10029 /* 10030 * The only dependencies are for indirect blocks. 10031 * 10032 * The file size for indirect block additions is not guaranteed. 10033 * Such a guarantee would be non-trivial to achieve. The conventional 10034 * synchronous write implementation also does not make this guarantee. 10035 * Fsck should catch and fix discrepancies. Arguably, the file size 10036 * can be over-estimated without destroying integrity when the file 10037 * moves into the indirect blocks (i.e., is large). If we want to 10038 * postpone fsck, we are stuck with this argument. 10039 */ 10040 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 10041 dp->di_ib[adp->ad_offset - NDADDR] = 0; 10042} 10043 10044/* 10045 * Cancel an indirdep as a result of truncation. Release all of the 10046 * children allocindirs and place their journal work on the appropriate 10047 * list. 10048 */ 10049static void 10050cancel_indirdep(indirdep, bp, freeblks) 10051 struct indirdep *indirdep; 10052 struct buf *bp; 10053 struct freeblks *freeblks; 10054{ 10055 struct allocindir *aip; 10056 10057 /* 10058 * None of the indirect pointers will ever be visible, 10059 * so they can simply be tossed. GOINGAWAY ensures 10060 * that allocated pointers will be saved in the buffer 10061 * cache until they are freed. Note that they will 10062 * only be able to be found by their physical address 10063 * since the inode mapping the logical address will 10064 * be gone. The save buffer used for the safe copy 10065 * was allocated in setup_allocindir_phase2 using 10066 * the physical address so it could be used for this 10067 * purpose. Hence we swap the safe copy with the real 10068 * copy, allowing the safe copy to be freed and holding 10069 * on to the real copy for later use in indir_trunc. 10070 */ 10071 if (indirdep->ir_state & GOINGAWAY) 10072 panic("cancel_indirdep: already gone"); 10073 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 10074 indirdep->ir_state |= DEPCOMPLETE; 10075 LIST_REMOVE(indirdep, ir_next); 10076 } 10077 indirdep->ir_state |= GOINGAWAY; 10078 VFSTOUFS(indirdep->ir_list.wk_mp)->um_numindirdeps += 1; 10079 /* 10080 * Pass in bp for blocks still have journal writes 10081 * pending so we can cancel them on their own. 10082 */ 10083 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 10084 cancel_allocindir(aip, bp, freeblks, 0); 10085 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) 10086 cancel_allocindir(aip, NULL, freeblks, 0); 10087 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != 0) 10088 cancel_allocindir(aip, NULL, freeblks, 0); 10089 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != 0) 10090 cancel_allocindir(aip, NULL, freeblks, 0); 10091 /* 10092 * If there are pending partial truncations we need to keep the 10093 * old block copy around until they complete. This is because 10094 * the current b_data is not a perfect superset of the available 10095 * blocks. 10096 */ 10097 if (TAILQ_EMPTY(&indirdep->ir_trunc)) 10098 bcopy(bp->b_data, indirdep->ir_savebp->b_data, bp->b_bcount); 10099 else 10100 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 10101 WORKLIST_REMOVE(&indirdep->ir_list); 10102 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, &indirdep->ir_list); 10103 indirdep->ir_bp = NULL; 10104 indirdep->ir_freeblks = freeblks; 10105} 10106 10107/* 10108 * Free an indirdep once it no longer has new pointers to track. 10109 */ 10110static void 10111free_indirdep(indirdep) 10112 struct indirdep *indirdep; 10113{ 10114 10115 KASSERT(TAILQ_EMPTY(&indirdep->ir_trunc), 10116 ("free_indirdep: Indir trunc list not empty.")); 10117 KASSERT(LIST_EMPTY(&indirdep->ir_completehd), 10118 ("free_indirdep: Complete head not empty.")); 10119 KASSERT(LIST_EMPTY(&indirdep->ir_writehd), 10120 ("free_indirdep: write head not empty.")); 10121 KASSERT(LIST_EMPTY(&indirdep->ir_donehd), 10122 ("free_indirdep: done head not empty.")); 10123 KASSERT(LIST_EMPTY(&indirdep->ir_deplisthd), 10124 ("free_indirdep: deplist head not empty.")); 10125 KASSERT((indirdep->ir_state & DEPCOMPLETE), 10126 ("free_indirdep: %p still on newblk list.", indirdep)); 10127 KASSERT(indirdep->ir_saveddata == NULL, 10128 ("free_indirdep: %p still has saved data.", indirdep)); 10129 if (indirdep->ir_state & ONWORKLIST) 10130 WORKLIST_REMOVE(&indirdep->ir_list); 10131 WORKITEM_FREE(indirdep, D_INDIRDEP); 10132} 10133 10134/* 10135 * Called before a write to an indirdep. This routine is responsible for 10136 * rolling back pointers to a safe state which includes only those 10137 * allocindirs which have been completed. 10138 */ 10139static void 10140initiate_write_indirdep(indirdep, bp) 10141 struct indirdep *indirdep; 10142 struct buf *bp; 10143{ 10144 10145 indirdep->ir_state |= IOSTARTED; 10146 if (indirdep->ir_state & GOINGAWAY) 10147 panic("disk_io_initiation: indirdep gone"); 10148 /* 10149 * If there are no remaining dependencies, this will be writing 10150 * the real pointers. 10151 */ 10152 if (LIST_EMPTY(&indirdep->ir_deplisthd) && 10153 TAILQ_EMPTY(&indirdep->ir_trunc)) 10154 return; 10155 /* 10156 * Replace up-to-date version with safe version. 10157 */ 10158 if (indirdep->ir_saveddata == NULL) { 10159 FREE_LOCK(&lk); 10160 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP, 10161 M_SOFTDEP_FLAGS); 10162 ACQUIRE_LOCK(&lk); 10163 } 10164 indirdep->ir_state &= ~ATTACHED; 10165 indirdep->ir_state |= UNDONE; 10166 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 10167 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 10168 bp->b_bcount); 10169} 10170 10171/* 10172 * Called when an inode has been cleared in a cg bitmap. This finally 10173 * eliminates any canceled jaddrefs 10174 */ 10175void 10176softdep_setup_inofree(mp, bp, ino, wkhd) 10177 struct mount *mp; 10178 struct buf *bp; 10179 ino_t ino; 10180 struct workhead *wkhd; 10181{ 10182 struct worklist *wk, *wkn; 10183 struct inodedep *inodedep; 10184 uint8_t *inosused; 10185 struct cg *cgp; 10186 struct fs *fs; 10187 10188 ACQUIRE_LOCK(&lk); 10189 fs = VFSTOUFS(mp)->um_fs; 10190 cgp = (struct cg *)bp->b_data; 10191 inosused = cg_inosused(cgp); 10192 if (isset(inosused, ino % fs->fs_ipg)) 10193 panic("softdep_setup_inofree: inode %ju not freed.", 10194 (uintmax_t)ino); 10195 if (inodedep_lookup(mp, ino, 0, &inodedep)) 10196 panic("softdep_setup_inofree: ino %ju has existing inodedep %p", 10197 (uintmax_t)ino, inodedep); 10198 if (wkhd) { 10199 LIST_FOREACH_SAFE(wk, wkhd, wk_list, wkn) { 10200 if (wk->wk_type != D_JADDREF) 10201 continue; 10202 WORKLIST_REMOVE(wk); 10203 /* 10204 * We can free immediately even if the jaddref 10205 * isn't attached in a background write as now 10206 * the bitmaps are reconciled. 10207 */ 10208 wk->wk_state |= COMPLETE | ATTACHED; 10209 free_jaddref(WK_JADDREF(wk)); 10210 } 10211 jwork_move(&bp->b_dep, wkhd); 10212 } 10213 FREE_LOCK(&lk); 10214} 10215 10216 10217/* 10218 * Called via ffs_blkfree() after a set of frags has been cleared from a cg 10219 * map. Any dependencies waiting for the write to clear are added to the 10220 * buf's list and any jnewblks that are being canceled are discarded 10221 * immediately. 10222 */ 10223void 10224softdep_setup_blkfree(mp, bp, blkno, frags, wkhd) 10225 struct mount *mp; 10226 struct buf *bp; 10227 ufs2_daddr_t blkno; 10228 int frags; 10229 struct workhead *wkhd; 10230{ 10231 struct bmsafemap *bmsafemap; 10232 struct jnewblk *jnewblk; 10233 struct worklist *wk; 10234 struct fs *fs; 10235#ifdef SUJ_DEBUG 10236 uint8_t *blksfree; 10237 struct cg *cgp; 10238 ufs2_daddr_t jstart; 10239 ufs2_daddr_t jend; 10240 ufs2_daddr_t end; 10241 long bno; 10242 int i; 10243#endif 10244 10245 ACQUIRE_LOCK(&lk); 10246 /* Lookup the bmsafemap so we track when it is dirty. */ 10247 fs = VFSTOUFS(mp)->um_fs; 10248 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL); 10249 /* 10250 * Detach any jnewblks which have been canceled. They must linger 10251 * until the bitmap is cleared again by ffs_blkfree() to prevent 10252 * an unjournaled allocation from hitting the disk. 10253 */ 10254 if (wkhd) { 10255 while ((wk = LIST_FIRST(wkhd)) != NULL) { 10256 WORKLIST_REMOVE(wk); 10257 if (wk->wk_type != D_JNEWBLK) { 10258 WORKLIST_INSERT(&bmsafemap->sm_freehd, wk); 10259 continue; 10260 } 10261 jnewblk = WK_JNEWBLK(wk); 10262 KASSERT(jnewblk->jn_state & GOINGAWAY, 10263 ("softdep_setup_blkfree: jnewblk not canceled.")); 10264#ifdef SUJ_DEBUG 10265 /* 10266 * Assert that this block is free in the bitmap 10267 * before we discard the jnewblk. 10268 */ 10269 cgp = (struct cg *)bp->b_data; 10270 blksfree = cg_blksfree(cgp); 10271 bno = dtogd(fs, jnewblk->jn_blkno); 10272 for (i = jnewblk->jn_oldfrags; 10273 i < jnewblk->jn_frags; i++) { 10274 if (isset(blksfree, bno + i)) 10275 continue; 10276 panic("softdep_setup_blkfree: not free"); 10277 } 10278#endif 10279 /* 10280 * Even if it's not attached we can free immediately 10281 * as the new bitmap is correct. 10282 */ 10283 wk->wk_state |= COMPLETE | ATTACHED; 10284 free_jnewblk(jnewblk); 10285 } 10286 } 10287 10288#ifdef SUJ_DEBUG 10289 /* 10290 * Assert that we are not freeing a block which has an outstanding 10291 * allocation dependency. 10292 */ 10293 fs = VFSTOUFS(mp)->um_fs; 10294 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL); 10295 end = blkno + frags; 10296 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) { 10297 /* 10298 * Don't match against blocks that will be freed when the 10299 * background write is done. 10300 */ 10301 if ((jnewblk->jn_state & (ATTACHED | COMPLETE | DEPCOMPLETE)) == 10302 (COMPLETE | DEPCOMPLETE)) 10303 continue; 10304 jstart = jnewblk->jn_blkno + jnewblk->jn_oldfrags; 10305 jend = jnewblk->jn_blkno + jnewblk->jn_frags; 10306 if ((blkno >= jstart && blkno < jend) || 10307 (end > jstart && end <= jend)) { 10308 printf("state 0x%X %jd - %d %d dep %p\n", 10309 jnewblk->jn_state, jnewblk->jn_blkno, 10310 jnewblk->jn_oldfrags, jnewblk->jn_frags, 10311 jnewblk->jn_dep); 10312 panic("softdep_setup_blkfree: " 10313 "%jd-%jd(%d) overlaps with %jd-%jd", 10314 blkno, end, frags, jstart, jend); 10315 } 10316 } 10317#endif 10318 FREE_LOCK(&lk); 10319} 10320 10321/* 10322 * Revert a block allocation when the journal record that describes it 10323 * is not yet written. 10324 */ 10325int 10326jnewblk_rollback(jnewblk, fs, cgp, blksfree) 10327 struct jnewblk *jnewblk; 10328 struct fs *fs; 10329 struct cg *cgp; 10330 uint8_t *blksfree; 10331{ 10332 ufs1_daddr_t fragno; 10333 long cgbno, bbase; 10334 int frags, blk; 10335 int i; 10336 10337 frags = 0; 10338 cgbno = dtogd(fs, jnewblk->jn_blkno); 10339 /* 10340 * We have to test which frags need to be rolled back. We may 10341 * be operating on a stale copy when doing background writes. 10342 */ 10343 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) 10344 if (isclr(blksfree, cgbno + i)) 10345 frags++; 10346 if (frags == 0) 10347 return (0); 10348 /* 10349 * This is mostly ffs_blkfree() sans some validation and 10350 * superblock updates. 10351 */ 10352 if (frags == fs->fs_frag) { 10353 fragno = fragstoblks(fs, cgbno); 10354 ffs_setblock(fs, blksfree, fragno); 10355 ffs_clusteracct(fs, cgp, fragno, 1); 10356 cgp->cg_cs.cs_nbfree++; 10357 } else { 10358 cgbno += jnewblk->jn_oldfrags; 10359 bbase = cgbno - fragnum(fs, cgbno); 10360 /* Decrement the old frags. */ 10361 blk = blkmap(fs, blksfree, bbase); 10362 ffs_fragacct(fs, blk, cgp->cg_frsum, -1); 10363 /* Deallocate the fragment */ 10364 for (i = 0; i < frags; i++) 10365 setbit(blksfree, cgbno + i); 10366 cgp->cg_cs.cs_nffree += frags; 10367 /* Add back in counts associated with the new frags */ 10368 blk = blkmap(fs, blksfree, bbase); 10369 ffs_fragacct(fs, blk, cgp->cg_frsum, 1); 10370 /* If a complete block has been reassembled, account for it. */ 10371 fragno = fragstoblks(fs, bbase); 10372 if (ffs_isblock(fs, blksfree, fragno)) { 10373 cgp->cg_cs.cs_nffree -= fs->fs_frag; 10374 ffs_clusteracct(fs, cgp, fragno, 1); 10375 cgp->cg_cs.cs_nbfree++; 10376 } 10377 } 10378 stat_jnewblk++; 10379 jnewblk->jn_state &= ~ATTACHED; 10380 jnewblk->jn_state |= UNDONE; 10381 10382 return (frags); 10383} 10384 10385static void 10386initiate_write_bmsafemap(bmsafemap, bp) 10387 struct bmsafemap *bmsafemap; 10388 struct buf *bp; /* The cg block. */ 10389{ 10390 struct jaddref *jaddref; 10391 struct jnewblk *jnewblk; 10392 uint8_t *inosused; 10393 uint8_t *blksfree; 10394 struct cg *cgp; 10395 struct fs *fs; 10396 ino_t ino; 10397 10398 if (bmsafemap->sm_state & IOSTARTED) 10399 panic("initiate_write_bmsafemap: Already started\n"); 10400 bmsafemap->sm_state |= IOSTARTED; 10401 /* 10402 * Clear any inode allocations which are pending journal writes. 10403 */ 10404 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd) != NULL) { 10405 cgp = (struct cg *)bp->b_data; 10406 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 10407 inosused = cg_inosused(cgp); 10408 LIST_FOREACH(jaddref, &bmsafemap->sm_jaddrefhd, ja_bmdeps) { 10409 ino = jaddref->ja_ino % fs->fs_ipg; 10410 /* 10411 * If this is a background copy the inode may not 10412 * be marked used yet. 10413 */ 10414 if (isset(inosused, ino)) { 10415 if ((jaddref->ja_mode & IFMT) == IFDIR) 10416 cgp->cg_cs.cs_ndir--; 10417 cgp->cg_cs.cs_nifree++; 10418 clrbit(inosused, ino); 10419 jaddref->ja_state &= ~ATTACHED; 10420 jaddref->ja_state |= UNDONE; 10421 stat_jaddref++; 10422 } else if ((bp->b_xflags & BX_BKGRDMARKER) == 0) 10423 panic("initiate_write_bmsafemap: inode %ju " 10424 "marked free", (uintmax_t)jaddref->ja_ino); 10425 } 10426 } 10427 /* 10428 * Clear any block allocations which are pending journal writes. 10429 */ 10430 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) { 10431 cgp = (struct cg *)bp->b_data; 10432 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 10433 blksfree = cg_blksfree(cgp); 10434 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) { 10435 if (jnewblk_rollback(jnewblk, fs, cgp, blksfree)) 10436 continue; 10437 if ((bp->b_xflags & BX_BKGRDMARKER) == 0) 10438 panic("initiate_write_bmsafemap: block %jd " 10439 "marked free", jnewblk->jn_blkno); 10440 } 10441 } 10442 /* 10443 * Move allocation lists to the written lists so they can be 10444 * cleared once the block write is complete. 10445 */ 10446 LIST_SWAP(&bmsafemap->sm_inodedephd, &bmsafemap->sm_inodedepwr, 10447 inodedep, id_deps); 10448 LIST_SWAP(&bmsafemap->sm_newblkhd, &bmsafemap->sm_newblkwr, 10449 newblk, nb_deps); 10450 LIST_SWAP(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, worklist, 10451 wk_list); 10452} 10453 10454/* 10455 * This routine is called during the completion interrupt 10456 * service routine for a disk write (from the procedure called 10457 * by the device driver to inform the filesystem caches of 10458 * a request completion). It should be called early in this 10459 * procedure, before the block is made available to other 10460 * processes or other routines are called. 10461 * 10462 */ 10463static void 10464softdep_disk_write_complete(bp) 10465 struct buf *bp; /* describes the completed disk write */ 10466{ 10467 struct worklist *wk; 10468 struct worklist *owk; 10469 struct workhead reattach; 10470 struct freeblks *freeblks; 10471 struct buf *sbp; 10472 10473 /* 10474 * If an error occurred while doing the write, then the data 10475 * has not hit the disk and the dependencies cannot be unrolled. 10476 */ 10477 if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0) 10478 return; 10479 LIST_INIT(&reattach); 10480 /* 10481 * This lock must not be released anywhere in this code segment. 10482 */ 10483 sbp = NULL; 10484 owk = NULL; 10485 ACQUIRE_LOCK(&lk); 10486 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 10487 WORKLIST_REMOVE(wk); 10488 dep_write[wk->wk_type]++; 10489 if (wk == owk) 10490 panic("duplicate worklist: %p\n", wk); 10491 owk = wk; 10492 switch (wk->wk_type) { 10493 10494 case D_PAGEDEP: 10495 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 10496 WORKLIST_INSERT(&reattach, wk); 10497 continue; 10498 10499 case D_INODEDEP: 10500 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 10501 WORKLIST_INSERT(&reattach, wk); 10502 continue; 10503 10504 case D_BMSAFEMAP: 10505 if (handle_written_bmsafemap(WK_BMSAFEMAP(wk), bp)) 10506 WORKLIST_INSERT(&reattach, wk); 10507 continue; 10508 10509 case D_MKDIR: 10510 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 10511 continue; 10512 10513 case D_ALLOCDIRECT: 10514 wk->wk_state |= COMPLETE; 10515 handle_allocdirect_partdone(WK_ALLOCDIRECT(wk), NULL); 10516 continue; 10517 10518 case D_ALLOCINDIR: 10519 wk->wk_state |= COMPLETE; 10520 handle_allocindir_partdone(WK_ALLOCINDIR(wk)); 10521 continue; 10522 10523 case D_INDIRDEP: 10524 if (handle_written_indirdep(WK_INDIRDEP(wk), bp, &sbp)) 10525 WORKLIST_INSERT(&reattach, wk); 10526 continue; 10527 10528 case D_FREEBLKS: 10529 wk->wk_state |= COMPLETE; 10530 freeblks = WK_FREEBLKS(wk); 10531 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE && 10532 LIST_EMPTY(&freeblks->fb_jblkdephd)) 10533 add_to_worklist(wk, WK_NODELAY); 10534 continue; 10535 10536 case D_FREEWORK: 10537 handle_written_freework(WK_FREEWORK(wk)); 10538 break; 10539 10540 case D_JSEGDEP: 10541 free_jsegdep(WK_JSEGDEP(wk)); 10542 continue; 10543 10544 case D_JSEG: 10545 handle_written_jseg(WK_JSEG(wk), bp); 10546 continue; 10547 10548 case D_SBDEP: 10549 if (handle_written_sbdep(WK_SBDEP(wk), bp)) 10550 WORKLIST_INSERT(&reattach, wk); 10551 continue; 10552 10553 case D_FREEDEP: 10554 free_freedep(WK_FREEDEP(wk)); 10555 continue; 10556 10557 default: 10558 panic("handle_disk_write_complete: Unknown type %s", 10559 TYPENAME(wk->wk_type)); 10560 /* NOTREACHED */ 10561 } 10562 } 10563 /* 10564 * Reattach any requests that must be redone. 10565 */ 10566 while ((wk = LIST_FIRST(&reattach)) != NULL) { 10567 WORKLIST_REMOVE(wk); 10568 WORKLIST_INSERT(&bp->b_dep, wk); 10569 } 10570 FREE_LOCK(&lk); 10571 if (sbp) 10572 brelse(sbp); 10573} 10574 10575/* 10576 * Called from within softdep_disk_write_complete above. Note that 10577 * this routine is always called from interrupt level with further 10578 * splbio interrupts blocked. 10579 */ 10580static void 10581handle_allocdirect_partdone(adp, wkhd) 10582 struct allocdirect *adp; /* the completed allocdirect */ 10583 struct workhead *wkhd; /* Work to do when inode is writtne. */ 10584{ 10585 struct allocdirectlst *listhead; 10586 struct allocdirect *listadp; 10587 struct inodedep *inodedep; 10588 long bsize; 10589 10590 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 10591 return; 10592 /* 10593 * The on-disk inode cannot claim to be any larger than the last 10594 * fragment that has been written. Otherwise, the on-disk inode 10595 * might have fragments that were not the last block in the file 10596 * which would corrupt the filesystem. Thus, we cannot free any 10597 * allocdirects after one whose ad_oldblkno claims a fragment as 10598 * these blocks must be rolled back to zero before writing the inode. 10599 * We check the currently active set of allocdirects in id_inoupdt 10600 * or id_extupdt as appropriate. 10601 */ 10602 inodedep = adp->ad_inodedep; 10603 bsize = inodedep->id_fs->fs_bsize; 10604 if (adp->ad_state & EXTDATA) 10605 listhead = &inodedep->id_extupdt; 10606 else 10607 listhead = &inodedep->id_inoupdt; 10608 TAILQ_FOREACH(listadp, listhead, ad_next) { 10609 /* found our block */ 10610 if (listadp == adp) 10611 break; 10612 /* continue if ad_oldlbn is not a fragment */ 10613 if (listadp->ad_oldsize == 0 || 10614 listadp->ad_oldsize == bsize) 10615 continue; 10616 /* hit a fragment */ 10617 return; 10618 } 10619 /* 10620 * If we have reached the end of the current list without 10621 * finding the just finished dependency, then it must be 10622 * on the future dependency list. Future dependencies cannot 10623 * be freed until they are moved to the current list. 10624 */ 10625 if (listadp == NULL) { 10626#ifdef DEBUG 10627 if (adp->ad_state & EXTDATA) 10628 listhead = &inodedep->id_newextupdt; 10629 else 10630 listhead = &inodedep->id_newinoupdt; 10631 TAILQ_FOREACH(listadp, listhead, ad_next) 10632 /* found our block */ 10633 if (listadp == adp) 10634 break; 10635 if (listadp == NULL) 10636 panic("handle_allocdirect_partdone: lost dep"); 10637#endif /* DEBUG */ 10638 return; 10639 } 10640 /* 10641 * If we have found the just finished dependency, then queue 10642 * it along with anything that follows it that is complete. 10643 * Since the pointer has not yet been written in the inode 10644 * as the dependency prevents it, place the allocdirect on the 10645 * bufwait list where it will be freed once the pointer is 10646 * valid. 10647 */ 10648 if (wkhd == NULL) 10649 wkhd = &inodedep->id_bufwait; 10650 for (; adp; adp = listadp) { 10651 listadp = TAILQ_NEXT(adp, ad_next); 10652 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 10653 return; 10654 TAILQ_REMOVE(listhead, adp, ad_next); 10655 WORKLIST_INSERT(wkhd, &adp->ad_block.nb_list); 10656 } 10657} 10658 10659/* 10660 * Called from within softdep_disk_write_complete above. This routine 10661 * completes successfully written allocindirs. 10662 */ 10663static void 10664handle_allocindir_partdone(aip) 10665 struct allocindir *aip; /* the completed allocindir */ 10666{ 10667 struct indirdep *indirdep; 10668 10669 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 10670 return; 10671 indirdep = aip->ai_indirdep; 10672 LIST_REMOVE(aip, ai_next); 10673 /* 10674 * Don't set a pointer while the buffer is undergoing IO or while 10675 * we have active truncations. 10676 */ 10677 if (indirdep->ir_state & UNDONE || !TAILQ_EMPTY(&indirdep->ir_trunc)) { 10678 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 10679 return; 10680 } 10681 if (indirdep->ir_state & UFS1FMT) 10682 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 10683 aip->ai_newblkno; 10684 else 10685 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 10686 aip->ai_newblkno; 10687 /* 10688 * Await the pointer write before freeing the allocindir. 10689 */ 10690 LIST_INSERT_HEAD(&indirdep->ir_writehd, aip, ai_next); 10691} 10692 10693/* 10694 * Release segments held on a jwork list. 10695 */ 10696static void 10697handle_jwork(wkhd) 10698 struct workhead *wkhd; 10699{ 10700 struct worklist *wk; 10701 10702 while ((wk = LIST_FIRST(wkhd)) != NULL) { 10703 WORKLIST_REMOVE(wk); 10704 switch (wk->wk_type) { 10705 case D_JSEGDEP: 10706 free_jsegdep(WK_JSEGDEP(wk)); 10707 continue; 10708 case D_FREEDEP: 10709 free_freedep(WK_FREEDEP(wk)); 10710 continue; 10711 case D_FREEFRAG: 10712 rele_jseg(WK_JSEG(WK_FREEFRAG(wk)->ff_jdep)); 10713 WORKITEM_FREE(wk, D_FREEFRAG); 10714 continue; 10715 case D_FREEWORK: 10716 handle_written_freework(WK_FREEWORK(wk)); 10717 continue; 10718 default: 10719 panic("handle_jwork: Unknown type %s\n", 10720 TYPENAME(wk->wk_type)); 10721 } 10722 } 10723} 10724 10725/* 10726 * Handle the bufwait list on an inode when it is safe to release items 10727 * held there. This normally happens after an inode block is written but 10728 * may be delayed and handled later if there are pending journal items that 10729 * are not yet safe to be released. 10730 */ 10731static struct freefile * 10732handle_bufwait(inodedep, refhd) 10733 struct inodedep *inodedep; 10734 struct workhead *refhd; 10735{ 10736 struct jaddref *jaddref; 10737 struct freefile *freefile; 10738 struct worklist *wk; 10739 10740 freefile = NULL; 10741 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 10742 WORKLIST_REMOVE(wk); 10743 switch (wk->wk_type) { 10744 case D_FREEFILE: 10745 /* 10746 * We defer adding freefile to the worklist 10747 * until all other additions have been made to 10748 * ensure that it will be done after all the 10749 * old blocks have been freed. 10750 */ 10751 if (freefile != NULL) 10752 panic("handle_bufwait: freefile"); 10753 freefile = WK_FREEFILE(wk); 10754 continue; 10755 10756 case D_MKDIR: 10757 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 10758 continue; 10759 10760 case D_DIRADD: 10761 diradd_inode_written(WK_DIRADD(wk), inodedep); 10762 continue; 10763 10764 case D_FREEFRAG: 10765 wk->wk_state |= COMPLETE; 10766 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE) 10767 add_to_worklist(wk, 0); 10768 continue; 10769 10770 case D_DIRREM: 10771 wk->wk_state |= COMPLETE; 10772 add_to_worklist(wk, 0); 10773 continue; 10774 10775 case D_ALLOCDIRECT: 10776 case D_ALLOCINDIR: 10777 free_newblk(WK_NEWBLK(wk)); 10778 continue; 10779 10780 case D_JNEWBLK: 10781 wk->wk_state |= COMPLETE; 10782 free_jnewblk(WK_JNEWBLK(wk)); 10783 continue; 10784 10785 /* 10786 * Save freed journal segments and add references on 10787 * the supplied list which will delay their release 10788 * until the cg bitmap is cleared on disk. 10789 */ 10790 case D_JSEGDEP: 10791 if (refhd == NULL) 10792 free_jsegdep(WK_JSEGDEP(wk)); 10793 else 10794 WORKLIST_INSERT(refhd, wk); 10795 continue; 10796 10797 case D_JADDREF: 10798 jaddref = WK_JADDREF(wk); 10799 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, 10800 if_deps); 10801 /* 10802 * Transfer any jaddrefs to the list to be freed with 10803 * the bitmap if we're handling a removed file. 10804 */ 10805 if (refhd == NULL) { 10806 wk->wk_state |= COMPLETE; 10807 free_jaddref(jaddref); 10808 } else 10809 WORKLIST_INSERT(refhd, wk); 10810 continue; 10811 10812 default: 10813 panic("handle_bufwait: Unknown type %p(%s)", 10814 wk, TYPENAME(wk->wk_type)); 10815 /* NOTREACHED */ 10816 } 10817 } 10818 return (freefile); 10819} 10820/* 10821 * Called from within softdep_disk_write_complete above to restore 10822 * in-memory inode block contents to their most up-to-date state. Note 10823 * that this routine is always called from interrupt level with further 10824 * splbio interrupts blocked. 10825 */ 10826static int 10827handle_written_inodeblock(inodedep, bp) 10828 struct inodedep *inodedep; 10829 struct buf *bp; /* buffer containing the inode block */ 10830{ 10831 struct freefile *freefile; 10832 struct allocdirect *adp, *nextadp; 10833 struct ufs1_dinode *dp1 = NULL; 10834 struct ufs2_dinode *dp2 = NULL; 10835 struct workhead wkhd; 10836 int hadchanges, fstype; 10837 ino_t freelink; 10838 10839 LIST_INIT(&wkhd); 10840 hadchanges = 0; 10841 freefile = NULL; 10842 if ((inodedep->id_state & IOSTARTED) == 0) 10843 panic("handle_written_inodeblock: not started"); 10844 inodedep->id_state &= ~IOSTARTED; 10845 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) { 10846 fstype = UFS1; 10847 dp1 = (struct ufs1_dinode *)bp->b_data + 10848 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 10849 freelink = dp1->di_freelink; 10850 } else { 10851 fstype = UFS2; 10852 dp2 = (struct ufs2_dinode *)bp->b_data + 10853 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 10854 freelink = dp2->di_freelink; 10855 } 10856 /* 10857 * Leave this inodeblock dirty until it's in the list. 10858 */ 10859 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) == UNLINKED) { 10860 struct inodedep *inon; 10861 10862 inon = TAILQ_NEXT(inodedep, id_unlinked); 10863 if ((inon == NULL && freelink == 0) || 10864 (inon && inon->id_ino == freelink)) { 10865 if (inon) 10866 inon->id_state |= UNLINKPREV; 10867 inodedep->id_state |= UNLINKNEXT; 10868 } 10869 hadchanges = 1; 10870 } 10871 /* 10872 * If we had to rollback the inode allocation because of 10873 * bitmaps being incomplete, then simply restore it. 10874 * Keep the block dirty so that it will not be reclaimed until 10875 * all associated dependencies have been cleared and the 10876 * corresponding updates written to disk. 10877 */ 10878 if (inodedep->id_savedino1 != NULL) { 10879 hadchanges = 1; 10880 if (fstype == UFS1) 10881 *dp1 = *inodedep->id_savedino1; 10882 else 10883 *dp2 = *inodedep->id_savedino2; 10884 free(inodedep->id_savedino1, M_SAVEDINO); 10885 inodedep->id_savedino1 = NULL; 10886 if ((bp->b_flags & B_DELWRI) == 0) 10887 stat_inode_bitmap++; 10888 bdirty(bp); 10889 /* 10890 * If the inode is clear here and GOINGAWAY it will never 10891 * be written. Process the bufwait and clear any pending 10892 * work which may include the freefile. 10893 */ 10894 if (inodedep->id_state & GOINGAWAY) 10895 goto bufwait; 10896 return (1); 10897 } 10898 inodedep->id_state |= COMPLETE; 10899 /* 10900 * Roll forward anything that had to be rolled back before 10901 * the inode could be updated. 10902 */ 10903 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 10904 nextadp = TAILQ_NEXT(adp, ad_next); 10905 if (adp->ad_state & ATTACHED) 10906 panic("handle_written_inodeblock: new entry"); 10907 if (fstype == UFS1) { 10908 if (adp->ad_offset < NDADDR) { 10909 if (dp1->di_db[adp->ad_offset]!=adp->ad_oldblkno) 10910 panic("%s %s #%jd mismatch %d != %jd", 10911 "handle_written_inodeblock:", 10912 "direct pointer", 10913 (intmax_t)adp->ad_offset, 10914 dp1->di_db[adp->ad_offset], 10915 (intmax_t)adp->ad_oldblkno); 10916 dp1->di_db[adp->ad_offset] = adp->ad_newblkno; 10917 } else { 10918 if (dp1->di_ib[adp->ad_offset - NDADDR] != 0) 10919 panic("%s: %s #%jd allocated as %d", 10920 "handle_written_inodeblock", 10921 "indirect pointer", 10922 (intmax_t)adp->ad_offset - NDADDR, 10923 dp1->di_ib[adp->ad_offset - NDADDR]); 10924 dp1->di_ib[adp->ad_offset - NDADDR] = 10925 adp->ad_newblkno; 10926 } 10927 } else { 10928 if (adp->ad_offset < NDADDR) { 10929 if (dp2->di_db[adp->ad_offset]!=adp->ad_oldblkno) 10930 panic("%s: %s #%jd %s %jd != %jd", 10931 "handle_written_inodeblock", 10932 "direct pointer", 10933 (intmax_t)adp->ad_offset, "mismatch", 10934 (intmax_t)dp2->di_db[adp->ad_offset], 10935 (intmax_t)adp->ad_oldblkno); 10936 dp2->di_db[adp->ad_offset] = adp->ad_newblkno; 10937 } else { 10938 if (dp2->di_ib[adp->ad_offset - NDADDR] != 0) 10939 panic("%s: %s #%jd allocated as %jd", 10940 "handle_written_inodeblock", 10941 "indirect pointer", 10942 (intmax_t)adp->ad_offset - NDADDR, 10943 (intmax_t) 10944 dp2->di_ib[adp->ad_offset - NDADDR]); 10945 dp2->di_ib[adp->ad_offset - NDADDR] = 10946 adp->ad_newblkno; 10947 } 10948 } 10949 adp->ad_state &= ~UNDONE; 10950 adp->ad_state |= ATTACHED; 10951 hadchanges = 1; 10952 } 10953 for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) { 10954 nextadp = TAILQ_NEXT(adp, ad_next); 10955 if (adp->ad_state & ATTACHED) 10956 panic("handle_written_inodeblock: new entry"); 10957 if (dp2->di_extb[adp->ad_offset] != adp->ad_oldblkno) 10958 panic("%s: direct pointers #%jd %s %jd != %jd", 10959 "handle_written_inodeblock", 10960 (intmax_t)adp->ad_offset, "mismatch", 10961 (intmax_t)dp2->di_extb[adp->ad_offset], 10962 (intmax_t)adp->ad_oldblkno); 10963 dp2->di_extb[adp->ad_offset] = adp->ad_newblkno; 10964 adp->ad_state &= ~UNDONE; 10965 adp->ad_state |= ATTACHED; 10966 hadchanges = 1; 10967 } 10968 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 10969 stat_direct_blk_ptrs++; 10970 /* 10971 * Reset the file size to its most up-to-date value. 10972 */ 10973 if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1) 10974 panic("handle_written_inodeblock: bad size"); 10975 if (inodedep->id_savednlink > LINK_MAX) 10976 panic("handle_written_inodeblock: Invalid link count " 10977 "%d for inodedep %p", inodedep->id_savednlink, inodedep); 10978 if (fstype == UFS1) { 10979 if (dp1->di_nlink != inodedep->id_savednlink) { 10980 dp1->di_nlink = inodedep->id_savednlink; 10981 hadchanges = 1; 10982 } 10983 if (dp1->di_size != inodedep->id_savedsize) { 10984 dp1->di_size = inodedep->id_savedsize; 10985 hadchanges = 1; 10986 } 10987 } else { 10988 if (dp2->di_nlink != inodedep->id_savednlink) { 10989 dp2->di_nlink = inodedep->id_savednlink; 10990 hadchanges = 1; 10991 } 10992 if (dp2->di_size != inodedep->id_savedsize) { 10993 dp2->di_size = inodedep->id_savedsize; 10994 hadchanges = 1; 10995 } 10996 if (dp2->di_extsize != inodedep->id_savedextsize) { 10997 dp2->di_extsize = inodedep->id_savedextsize; 10998 hadchanges = 1; 10999 } 11000 } 11001 inodedep->id_savedsize = -1; 11002 inodedep->id_savedextsize = -1; 11003 inodedep->id_savednlink = -1; 11004 /* 11005 * If there were any rollbacks in the inode block, then it must be 11006 * marked dirty so that its will eventually get written back in 11007 * its correct form. 11008 */ 11009 if (hadchanges) 11010 bdirty(bp); 11011bufwait: 11012 /* 11013 * Process any allocdirects that completed during the update. 11014 */ 11015 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 11016 handle_allocdirect_partdone(adp, &wkhd); 11017 if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 11018 handle_allocdirect_partdone(adp, &wkhd); 11019 /* 11020 * Process deallocations that were held pending until the 11021 * inode had been written to disk. Freeing of the inode 11022 * is delayed until after all blocks have been freed to 11023 * avoid creation of new <vfsid, inum, lbn> triples 11024 * before the old ones have been deleted. Completely 11025 * unlinked inodes are not processed until the unlinked 11026 * inode list is written or the last reference is removed. 11027 */ 11028 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) != UNLINKED) { 11029 freefile = handle_bufwait(inodedep, NULL); 11030 if (freefile && !LIST_EMPTY(&wkhd)) { 11031 WORKLIST_INSERT(&wkhd, &freefile->fx_list); 11032 freefile = NULL; 11033 } 11034 } 11035 /* 11036 * Move rolled forward dependency completions to the bufwait list 11037 * now that those that were already written have been processed. 11038 */ 11039 if (!LIST_EMPTY(&wkhd) && hadchanges == 0) 11040 panic("handle_written_inodeblock: bufwait but no changes"); 11041 jwork_move(&inodedep->id_bufwait, &wkhd); 11042 11043 if (freefile != NULL) { 11044 /* 11045 * If the inode is goingaway it was never written. Fake up 11046 * the state here so free_inodedep() can succeed. 11047 */ 11048 if (inodedep->id_state & GOINGAWAY) 11049 inodedep->id_state |= COMPLETE | DEPCOMPLETE; 11050 if (free_inodedep(inodedep) == 0) 11051 panic("handle_written_inodeblock: live inodedep %p", 11052 inodedep); 11053 add_to_worklist(&freefile->fx_list, 0); 11054 return (0); 11055 } 11056 11057 /* 11058 * If no outstanding dependencies, free it. 11059 */ 11060 if (free_inodedep(inodedep) || 11061 (TAILQ_FIRST(&inodedep->id_inoreflst) == 0 && 11062 TAILQ_FIRST(&inodedep->id_inoupdt) == 0 && 11063 TAILQ_FIRST(&inodedep->id_extupdt) == 0 && 11064 LIST_FIRST(&inodedep->id_bufwait) == 0)) 11065 return (0); 11066 return (hadchanges); 11067} 11068 11069static int 11070handle_written_indirdep(indirdep, bp, bpp) 11071 struct indirdep *indirdep; 11072 struct buf *bp; 11073 struct buf **bpp; 11074{ 11075 struct allocindir *aip; 11076 struct buf *sbp; 11077 int chgs; 11078 11079 if (indirdep->ir_state & GOINGAWAY) 11080 panic("handle_written_indirdep: indirdep gone"); 11081 if ((indirdep->ir_state & IOSTARTED) == 0) 11082 panic("handle_written_indirdep: IO not started"); 11083 chgs = 0; 11084 /* 11085 * If there were rollbacks revert them here. 11086 */ 11087 if (indirdep->ir_saveddata) { 11088 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 11089 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 11090 free(indirdep->ir_saveddata, M_INDIRDEP); 11091 indirdep->ir_saveddata = NULL; 11092 } 11093 chgs = 1; 11094 } 11095 indirdep->ir_state &= ~(UNDONE | IOSTARTED); 11096 indirdep->ir_state |= ATTACHED; 11097 /* 11098 * Move allocindirs with written pointers to the completehd if 11099 * the indirdep's pointer is not yet written. Otherwise 11100 * free them here. 11101 */ 11102 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != 0) { 11103 LIST_REMOVE(aip, ai_next); 11104 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 11105 LIST_INSERT_HEAD(&indirdep->ir_completehd, aip, 11106 ai_next); 11107 newblk_freefrag(&aip->ai_block); 11108 continue; 11109 } 11110 free_newblk(&aip->ai_block); 11111 } 11112 /* 11113 * Move allocindirs that have finished dependency processing from 11114 * the done list to the write list after updating the pointers. 11115 */ 11116 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 11117 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 11118 handle_allocindir_partdone(aip); 11119 if (aip == LIST_FIRST(&indirdep->ir_donehd)) 11120 panic("disk_write_complete: not gone"); 11121 chgs = 1; 11122 } 11123 } 11124 /* 11125 * Preserve the indirdep if there were any changes or if it is not 11126 * yet valid on disk. 11127 */ 11128 if (chgs) { 11129 stat_indir_blk_ptrs++; 11130 bdirty(bp); 11131 return (1); 11132 } 11133 /* 11134 * If there were no changes we can discard the savedbp and detach 11135 * ourselves from the buf. We are only carrying completed pointers 11136 * in this case. 11137 */ 11138 sbp = indirdep->ir_savebp; 11139 sbp->b_flags |= B_INVAL | B_NOCACHE; 11140 indirdep->ir_savebp = NULL; 11141 indirdep->ir_bp = NULL; 11142 if (*bpp != NULL) 11143 panic("handle_written_indirdep: bp already exists."); 11144 *bpp = sbp; 11145 /* 11146 * The indirdep may not be freed until its parent points at it. 11147 */ 11148 if (indirdep->ir_state & DEPCOMPLETE) 11149 free_indirdep(indirdep); 11150 11151 return (0); 11152} 11153 11154/* 11155 * Process a diradd entry after its dependent inode has been written. 11156 * This routine must be called with splbio interrupts blocked. 11157 */ 11158static void 11159diradd_inode_written(dap, inodedep) 11160 struct diradd *dap; 11161 struct inodedep *inodedep; 11162{ 11163 11164 dap->da_state |= COMPLETE; 11165 complete_diradd(dap); 11166 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 11167} 11168 11169/* 11170 * Returns true if the bmsafemap will have rollbacks when written. Must 11171 * only be called with lk and the buf lock on the cg held. 11172 */ 11173static int 11174bmsafemap_rollbacks(bmsafemap) 11175 struct bmsafemap *bmsafemap; 11176{ 11177 11178 return (!LIST_EMPTY(&bmsafemap->sm_jaddrefhd) | 11179 !LIST_EMPTY(&bmsafemap->sm_jnewblkhd)); 11180} 11181 11182/* 11183 * Re-apply an allocation when a cg write is complete. 11184 */ 11185static int 11186jnewblk_rollforward(jnewblk, fs, cgp, blksfree) 11187 struct jnewblk *jnewblk; 11188 struct fs *fs; 11189 struct cg *cgp; 11190 uint8_t *blksfree; 11191{ 11192 ufs1_daddr_t fragno; 11193 ufs2_daddr_t blkno; 11194 long cgbno, bbase; 11195 int frags, blk; 11196 int i; 11197 11198 frags = 0; 11199 cgbno = dtogd(fs, jnewblk->jn_blkno); 11200 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) { 11201 if (isclr(blksfree, cgbno + i)) 11202 panic("jnewblk_rollforward: re-allocated fragment"); 11203 frags++; 11204 } 11205 if (frags == fs->fs_frag) { 11206 blkno = fragstoblks(fs, cgbno); 11207 ffs_clrblock(fs, blksfree, (long)blkno); 11208 ffs_clusteracct(fs, cgp, blkno, -1); 11209 cgp->cg_cs.cs_nbfree--; 11210 } else { 11211 bbase = cgbno - fragnum(fs, cgbno); 11212 cgbno += jnewblk->jn_oldfrags; 11213 /* If a complete block had been reassembled, account for it. */ 11214 fragno = fragstoblks(fs, bbase); 11215 if (ffs_isblock(fs, blksfree, fragno)) { 11216 cgp->cg_cs.cs_nffree += fs->fs_frag; 11217 ffs_clusteracct(fs, cgp, fragno, -1); 11218 cgp->cg_cs.cs_nbfree--; 11219 } 11220 /* Decrement the old frags. */ 11221 blk = blkmap(fs, blksfree, bbase); 11222 ffs_fragacct(fs, blk, cgp->cg_frsum, -1); 11223 /* Allocate the fragment */ 11224 for (i = 0; i < frags; i++) 11225 clrbit(blksfree, cgbno + i); 11226 cgp->cg_cs.cs_nffree -= frags; 11227 /* Add back in counts associated with the new frags */ 11228 blk = blkmap(fs, blksfree, bbase); 11229 ffs_fragacct(fs, blk, cgp->cg_frsum, 1); 11230 } 11231 return (frags); 11232} 11233 11234/* 11235 * Complete a write to a bmsafemap structure. Roll forward any bitmap 11236 * changes if it's not a background write. Set all written dependencies 11237 * to DEPCOMPLETE and free the structure if possible. 11238 */ 11239static int 11240handle_written_bmsafemap(bmsafemap, bp) 11241 struct bmsafemap *bmsafemap; 11242 struct buf *bp; 11243{ 11244 struct newblk *newblk; 11245 struct inodedep *inodedep; 11246 struct jaddref *jaddref, *jatmp; 11247 struct jnewblk *jnewblk, *jntmp; 11248 struct ufsmount *ump; 11249 uint8_t *inosused; 11250 uint8_t *blksfree; 11251 struct cg *cgp; 11252 struct fs *fs; 11253 ino_t ino; 11254 int chgs; 11255 11256 if ((bmsafemap->sm_state & IOSTARTED) == 0) 11257 panic("initiate_write_bmsafemap: Not started\n"); 11258 ump = VFSTOUFS(bmsafemap->sm_list.wk_mp); 11259 chgs = 0; 11260 bmsafemap->sm_state &= ~IOSTARTED; 11261 /* 11262 * Release journal work that was waiting on the write. 11263 */ 11264 handle_jwork(&bmsafemap->sm_freewr); 11265 11266 /* 11267 * Restore unwritten inode allocation pending jaddref writes. 11268 */ 11269 if (!LIST_EMPTY(&bmsafemap->sm_jaddrefhd)) { 11270 cgp = (struct cg *)bp->b_data; 11271 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 11272 inosused = cg_inosused(cgp); 11273 LIST_FOREACH_SAFE(jaddref, &bmsafemap->sm_jaddrefhd, 11274 ja_bmdeps, jatmp) { 11275 if ((jaddref->ja_state & UNDONE) == 0) 11276 continue; 11277 ino = jaddref->ja_ino % fs->fs_ipg; 11278 if (isset(inosused, ino)) 11279 panic("handle_written_bmsafemap: " 11280 "re-allocated inode"); 11281 if ((bp->b_xflags & BX_BKGRDMARKER) == 0) { 11282 if ((jaddref->ja_mode & IFMT) == IFDIR) 11283 cgp->cg_cs.cs_ndir++; 11284 cgp->cg_cs.cs_nifree--; 11285 setbit(inosused, ino); 11286 chgs = 1; 11287 } 11288 jaddref->ja_state &= ~UNDONE; 11289 jaddref->ja_state |= ATTACHED; 11290 free_jaddref(jaddref); 11291 } 11292 } 11293 /* 11294 * Restore any block allocations which are pending journal writes. 11295 */ 11296 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) { 11297 cgp = (struct cg *)bp->b_data; 11298 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 11299 blksfree = cg_blksfree(cgp); 11300 LIST_FOREACH_SAFE(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps, 11301 jntmp) { 11302 if ((jnewblk->jn_state & UNDONE) == 0) 11303 continue; 11304 if ((bp->b_xflags & BX_BKGRDMARKER) == 0 && 11305 jnewblk_rollforward(jnewblk, fs, cgp, blksfree)) 11306 chgs = 1; 11307 jnewblk->jn_state &= ~(UNDONE | NEWBLOCK); 11308 jnewblk->jn_state |= ATTACHED; 11309 free_jnewblk(jnewblk); 11310 } 11311 } 11312 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkwr))) { 11313 newblk->nb_state |= DEPCOMPLETE; 11314 newblk->nb_state &= ~ONDEPLIST; 11315 newblk->nb_bmsafemap = NULL; 11316 LIST_REMOVE(newblk, nb_deps); 11317 if (newblk->nb_list.wk_type == D_ALLOCDIRECT) 11318 handle_allocdirect_partdone( 11319 WK_ALLOCDIRECT(&newblk->nb_list), NULL); 11320 else if (newblk->nb_list.wk_type == D_ALLOCINDIR) 11321 handle_allocindir_partdone( 11322 WK_ALLOCINDIR(&newblk->nb_list)); 11323 else if (newblk->nb_list.wk_type != D_NEWBLK) 11324 panic("handle_written_bmsafemap: Unexpected type: %s", 11325 TYPENAME(newblk->nb_list.wk_type)); 11326 } 11327 while ((inodedep = LIST_FIRST(&bmsafemap->sm_inodedepwr)) != NULL) { 11328 inodedep->id_state |= DEPCOMPLETE; 11329 inodedep->id_state &= ~ONDEPLIST; 11330 LIST_REMOVE(inodedep, id_deps); 11331 inodedep->id_bmsafemap = NULL; 11332 } 11333 LIST_REMOVE(bmsafemap, sm_next); 11334 if (chgs == 0 && LIST_EMPTY(&bmsafemap->sm_jaddrefhd) && 11335 LIST_EMPTY(&bmsafemap->sm_jnewblkhd) && 11336 LIST_EMPTY(&bmsafemap->sm_newblkhd) && 11337 LIST_EMPTY(&bmsafemap->sm_inodedephd) && 11338 LIST_EMPTY(&bmsafemap->sm_freehd)) { 11339 LIST_REMOVE(bmsafemap, sm_hash); 11340 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 11341 return (0); 11342 } 11343 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next); 11344 bdirty(bp); 11345 return (1); 11346} 11347 11348/* 11349 * Try to free a mkdir dependency. 11350 */ 11351static void 11352complete_mkdir(mkdir) 11353 struct mkdir *mkdir; 11354{ 11355 struct diradd *dap; 11356 11357 if ((mkdir->md_state & ALLCOMPLETE) != ALLCOMPLETE) 11358 return; 11359 LIST_REMOVE(mkdir, md_mkdirs); 11360 dap = mkdir->md_diradd; 11361 dap->da_state &= ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)); 11362 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) { 11363 dap->da_state |= DEPCOMPLETE; 11364 complete_diradd(dap); 11365 } 11366 WORKITEM_FREE(mkdir, D_MKDIR); 11367} 11368 11369/* 11370 * Handle the completion of a mkdir dependency. 11371 */ 11372static void 11373handle_written_mkdir(mkdir, type) 11374 struct mkdir *mkdir; 11375 int type; 11376{ 11377 11378 if ((mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)) != type) 11379 panic("handle_written_mkdir: bad type"); 11380 mkdir->md_state |= COMPLETE; 11381 complete_mkdir(mkdir); 11382} 11383 11384static int 11385free_pagedep(pagedep) 11386 struct pagedep *pagedep; 11387{ 11388 int i; 11389 11390 if (pagedep->pd_state & NEWBLOCK) 11391 return (0); 11392 if (!LIST_EMPTY(&pagedep->pd_dirremhd)) 11393 return (0); 11394 for (i = 0; i < DAHASHSZ; i++) 11395 if (!LIST_EMPTY(&pagedep->pd_diraddhd[i])) 11396 return (0); 11397 if (!LIST_EMPTY(&pagedep->pd_pendinghd)) 11398 return (0); 11399 if (!LIST_EMPTY(&pagedep->pd_jmvrefhd)) 11400 return (0); 11401 if (pagedep->pd_state & ONWORKLIST) 11402 WORKLIST_REMOVE(&pagedep->pd_list); 11403 LIST_REMOVE(pagedep, pd_hash); 11404 WORKITEM_FREE(pagedep, D_PAGEDEP); 11405 11406 return (1); 11407} 11408 11409/* 11410 * Called from within softdep_disk_write_complete above. 11411 * A write operation was just completed. Removed inodes can 11412 * now be freed and associated block pointers may be committed. 11413 * Note that this routine is always called from interrupt level 11414 * with further splbio interrupts blocked. 11415 */ 11416static int 11417handle_written_filepage(pagedep, bp) 11418 struct pagedep *pagedep; 11419 struct buf *bp; /* buffer containing the written page */ 11420{ 11421 struct dirrem *dirrem; 11422 struct diradd *dap, *nextdap; 11423 struct direct *ep; 11424 int i, chgs; 11425 11426 if ((pagedep->pd_state & IOSTARTED) == 0) 11427 panic("handle_written_filepage: not started"); 11428 pagedep->pd_state &= ~IOSTARTED; 11429 /* 11430 * Process any directory removals that have been committed. 11431 */ 11432 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 11433 LIST_REMOVE(dirrem, dm_next); 11434 dirrem->dm_state |= COMPLETE; 11435 dirrem->dm_dirinum = pagedep->pd_ino; 11436 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd), 11437 ("handle_written_filepage: Journal entries not written.")); 11438 add_to_worklist(&dirrem->dm_list, 0); 11439 } 11440 /* 11441 * Free any directory additions that have been committed. 11442 * If it is a newly allocated block, we have to wait until 11443 * the on-disk directory inode claims the new block. 11444 */ 11445 if ((pagedep->pd_state & NEWBLOCK) == 0) 11446 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 11447 free_diradd(dap, NULL); 11448 /* 11449 * Uncommitted directory entries must be restored. 11450 */ 11451 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 11452 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 11453 dap = nextdap) { 11454 nextdap = LIST_NEXT(dap, da_pdlist); 11455 if (dap->da_state & ATTACHED) 11456 panic("handle_written_filepage: attached"); 11457 ep = (struct direct *) 11458 ((char *)bp->b_data + dap->da_offset); 11459 ep->d_ino = dap->da_newinum; 11460 dap->da_state &= ~UNDONE; 11461 dap->da_state |= ATTACHED; 11462 chgs = 1; 11463 /* 11464 * If the inode referenced by the directory has 11465 * been written out, then the dependency can be 11466 * moved to the pending list. 11467 */ 11468 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 11469 LIST_REMOVE(dap, da_pdlist); 11470 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 11471 da_pdlist); 11472 } 11473 } 11474 } 11475 /* 11476 * If there were any rollbacks in the directory, then it must be 11477 * marked dirty so that its will eventually get written back in 11478 * its correct form. 11479 */ 11480 if (chgs) { 11481 if ((bp->b_flags & B_DELWRI) == 0) 11482 stat_dir_entry++; 11483 bdirty(bp); 11484 return (1); 11485 } 11486 /* 11487 * If we are not waiting for a new directory block to be 11488 * claimed by its inode, then the pagedep will be freed. 11489 * Otherwise it will remain to track any new entries on 11490 * the page in case they are fsync'ed. 11491 */ 11492 free_pagedep(pagedep); 11493 return (0); 11494} 11495 11496/* 11497 * Writing back in-core inode structures. 11498 * 11499 * The filesystem only accesses an inode's contents when it occupies an 11500 * "in-core" inode structure. These "in-core" structures are separate from 11501 * the page frames used to cache inode blocks. Only the latter are 11502 * transferred to/from the disk. So, when the updated contents of the 11503 * "in-core" inode structure are copied to the corresponding in-memory inode 11504 * block, the dependencies are also transferred. The following procedure is 11505 * called when copying a dirty "in-core" inode to a cached inode block. 11506 */ 11507 11508/* 11509 * Called when an inode is loaded from disk. If the effective link count 11510 * differed from the actual link count when it was last flushed, then we 11511 * need to ensure that the correct effective link count is put back. 11512 */ 11513void 11514softdep_load_inodeblock(ip) 11515 struct inode *ip; /* the "in_core" copy of the inode */ 11516{ 11517 struct inodedep *inodedep; 11518 11519 /* 11520 * Check for alternate nlink count. 11521 */ 11522 ip->i_effnlink = ip->i_nlink; 11523 ACQUIRE_LOCK(&lk); 11524 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, 11525 &inodedep) == 0) { 11526 FREE_LOCK(&lk); 11527 return; 11528 } 11529 ip->i_effnlink -= inodedep->id_nlinkdelta; 11530 FREE_LOCK(&lk); 11531} 11532 11533/* 11534 * This routine is called just before the "in-core" inode 11535 * information is to be copied to the in-memory inode block. 11536 * Recall that an inode block contains several inodes. If 11537 * the force flag is set, then the dependencies will be 11538 * cleared so that the update can always be made. Note that 11539 * the buffer is locked when this routine is called, so we 11540 * will never be in the middle of writing the inode block 11541 * to disk. 11542 */ 11543void 11544softdep_update_inodeblock(ip, bp, waitfor) 11545 struct inode *ip; /* the "in_core" copy of the inode */ 11546 struct buf *bp; /* the buffer containing the inode block */ 11547 int waitfor; /* nonzero => update must be allowed */ 11548{ 11549 struct inodedep *inodedep; 11550 struct inoref *inoref; 11551 struct worklist *wk; 11552 struct mount *mp; 11553 struct buf *ibp; 11554 struct fs *fs; 11555 int error; 11556 11557 mp = UFSTOVFS(ip->i_ump); 11558 fs = ip->i_fs; 11559 /* 11560 * Preserve the freelink that is on disk. clear_unlinked_inodedep() 11561 * does not have access to the in-core ip so must write directly into 11562 * the inode block buffer when setting freelink. 11563 */ 11564 if (fs->fs_magic == FS_UFS1_MAGIC) 11565 DIP_SET(ip, i_freelink, ((struct ufs1_dinode *)bp->b_data + 11566 ino_to_fsbo(fs, ip->i_number))->di_freelink); 11567 else 11568 DIP_SET(ip, i_freelink, ((struct ufs2_dinode *)bp->b_data + 11569 ino_to_fsbo(fs, ip->i_number))->di_freelink); 11570 /* 11571 * If the effective link count is not equal to the actual link 11572 * count, then we must track the difference in an inodedep while 11573 * the inode is (potentially) tossed out of the cache. Otherwise, 11574 * if there is no existing inodedep, then there are no dependencies 11575 * to track. 11576 */ 11577 ACQUIRE_LOCK(&lk); 11578again: 11579 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 11580 FREE_LOCK(&lk); 11581 if (ip->i_effnlink != ip->i_nlink) 11582 panic("softdep_update_inodeblock: bad link count"); 11583 return; 11584 } 11585 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) 11586 panic("softdep_update_inodeblock: bad delta"); 11587 /* 11588 * If we're flushing all dependencies we must also move any waiting 11589 * for journal writes onto the bufwait list prior to I/O. 11590 */ 11591 if (waitfor) { 11592 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 11593 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 11594 == DEPCOMPLETE) { 11595 jwait(&inoref->if_list, MNT_WAIT); 11596 goto again; 11597 } 11598 } 11599 } 11600 /* 11601 * Changes have been initiated. Anything depending on these 11602 * changes cannot occur until this inode has been written. 11603 */ 11604 inodedep->id_state &= ~COMPLETE; 11605 if ((inodedep->id_state & ONWORKLIST) == 0) 11606 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 11607 /* 11608 * Any new dependencies associated with the incore inode must 11609 * now be moved to the list associated with the buffer holding 11610 * the in-memory copy of the inode. Once merged process any 11611 * allocdirects that are completed by the merger. 11612 */ 11613 merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt); 11614 if (!TAILQ_EMPTY(&inodedep->id_inoupdt)) 11615 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt), 11616 NULL); 11617 merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt); 11618 if (!TAILQ_EMPTY(&inodedep->id_extupdt)) 11619 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt), 11620 NULL); 11621 /* 11622 * Now that the inode has been pushed into the buffer, the 11623 * operations dependent on the inode being written to disk 11624 * can be moved to the id_bufwait so that they will be 11625 * processed when the buffer I/O completes. 11626 */ 11627 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 11628 WORKLIST_REMOVE(wk); 11629 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 11630 } 11631 /* 11632 * Newly allocated inodes cannot be written until the bitmap 11633 * that allocates them have been written (indicated by 11634 * DEPCOMPLETE being set in id_state). If we are doing a 11635 * forced sync (e.g., an fsync on a file), we force the bitmap 11636 * to be written so that the update can be done. 11637 */ 11638 if (waitfor == 0) { 11639 FREE_LOCK(&lk); 11640 return; 11641 } 11642retry: 11643 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) != 0) { 11644 FREE_LOCK(&lk); 11645 return; 11646 } 11647 ibp = inodedep->id_bmsafemap->sm_buf; 11648 ibp = getdirtybuf(ibp, &lk, MNT_WAIT); 11649 if (ibp == NULL) { 11650 /* 11651 * If ibp came back as NULL, the dependency could have been 11652 * freed while we slept. Look it up again, and check to see 11653 * that it has completed. 11654 */ 11655 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 11656 goto retry; 11657 FREE_LOCK(&lk); 11658 return; 11659 } 11660 FREE_LOCK(&lk); 11661 if ((error = bwrite(ibp)) != 0) 11662 softdep_error("softdep_update_inodeblock: bwrite", error); 11663} 11664 11665/* 11666 * Merge the a new inode dependency list (such as id_newinoupdt) into an 11667 * old inode dependency list (such as id_inoupdt). This routine must be 11668 * called with splbio interrupts blocked. 11669 */ 11670static void 11671merge_inode_lists(newlisthead, oldlisthead) 11672 struct allocdirectlst *newlisthead; 11673 struct allocdirectlst *oldlisthead; 11674{ 11675 struct allocdirect *listadp, *newadp; 11676 11677 newadp = TAILQ_FIRST(newlisthead); 11678 for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) { 11679 if (listadp->ad_offset < newadp->ad_offset) { 11680 listadp = TAILQ_NEXT(listadp, ad_next); 11681 continue; 11682 } 11683 TAILQ_REMOVE(newlisthead, newadp, ad_next); 11684 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 11685 if (listadp->ad_offset == newadp->ad_offset) { 11686 allocdirect_merge(oldlisthead, newadp, 11687 listadp); 11688 listadp = newadp; 11689 } 11690 newadp = TAILQ_FIRST(newlisthead); 11691 } 11692 while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) { 11693 TAILQ_REMOVE(newlisthead, newadp, ad_next); 11694 TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next); 11695 } 11696} 11697 11698/* 11699 * If we are doing an fsync, then we must ensure that any directory 11700 * entries for the inode have been written after the inode gets to disk. 11701 */ 11702int 11703softdep_fsync(vp) 11704 struct vnode *vp; /* the "in_core" copy of the inode */ 11705{ 11706 struct inodedep *inodedep; 11707 struct pagedep *pagedep; 11708 struct inoref *inoref; 11709 struct worklist *wk; 11710 struct diradd *dap; 11711 struct mount *mp; 11712 struct vnode *pvp; 11713 struct inode *ip; 11714 struct buf *bp; 11715 struct fs *fs; 11716 struct thread *td = curthread; 11717 int error, flushparent, pagedep_new_block; 11718 ino_t parentino; 11719 ufs_lbn_t lbn; 11720 11721 ip = VTOI(vp); 11722 fs = ip->i_fs; 11723 mp = vp->v_mount; 11724 ACQUIRE_LOCK(&lk); 11725restart: 11726 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 11727 FREE_LOCK(&lk); 11728 return (0); 11729 } 11730 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 11731 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 11732 == DEPCOMPLETE) { 11733 jwait(&inoref->if_list, MNT_WAIT); 11734 goto restart; 11735 } 11736 } 11737 if (!LIST_EMPTY(&inodedep->id_inowait) || 11738 !TAILQ_EMPTY(&inodedep->id_extupdt) || 11739 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 11740 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 11741 !TAILQ_EMPTY(&inodedep->id_newinoupdt)) 11742 panic("softdep_fsync: pending ops %p", inodedep); 11743 for (error = 0, flushparent = 0; ; ) { 11744 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 11745 break; 11746 if (wk->wk_type != D_DIRADD) 11747 panic("softdep_fsync: Unexpected type %s", 11748 TYPENAME(wk->wk_type)); 11749 dap = WK_DIRADD(wk); 11750 /* 11751 * Flush our parent if this directory entry has a MKDIR_PARENT 11752 * dependency or is contained in a newly allocated block. 11753 */ 11754 if (dap->da_state & DIRCHG) 11755 pagedep = dap->da_previous->dm_pagedep; 11756 else 11757 pagedep = dap->da_pagedep; 11758 parentino = pagedep->pd_ino; 11759 lbn = pagedep->pd_lbn; 11760 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) 11761 panic("softdep_fsync: dirty"); 11762 if ((dap->da_state & MKDIR_PARENT) || 11763 (pagedep->pd_state & NEWBLOCK)) 11764 flushparent = 1; 11765 else 11766 flushparent = 0; 11767 /* 11768 * If we are being fsync'ed as part of vgone'ing this vnode, 11769 * then we will not be able to release and recover the 11770 * vnode below, so we just have to give up on writing its 11771 * directory entry out. It will eventually be written, just 11772 * not now, but then the user was not asking to have it 11773 * written, so we are not breaking any promises. 11774 */ 11775 if (vp->v_iflag & VI_DOOMED) 11776 break; 11777 /* 11778 * We prevent deadlock by always fetching inodes from the 11779 * root, moving down the directory tree. Thus, when fetching 11780 * our parent directory, we first try to get the lock. If 11781 * that fails, we must unlock ourselves before requesting 11782 * the lock on our parent. See the comment in ufs_lookup 11783 * for details on possible races. 11784 */ 11785 FREE_LOCK(&lk); 11786 if (ffs_vgetf(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp, 11787 FFSV_FORCEINSMQ)) { 11788 error = vfs_busy(mp, MBF_NOWAIT); 11789 if (error != 0) { 11790 vfs_ref(mp); 11791 VOP_UNLOCK(vp, 0); 11792 error = vfs_busy(mp, 0); 11793 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 11794 vfs_rel(mp); 11795 if (error != 0) 11796 return (ENOENT); 11797 if (vp->v_iflag & VI_DOOMED) { 11798 vfs_unbusy(mp); 11799 return (ENOENT); 11800 } 11801 } 11802 VOP_UNLOCK(vp, 0); 11803 error = ffs_vgetf(mp, parentino, LK_EXCLUSIVE, 11804 &pvp, FFSV_FORCEINSMQ); 11805 vfs_unbusy(mp); 11806 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 11807 if (vp->v_iflag & VI_DOOMED) { 11808 if (error == 0) 11809 vput(pvp); 11810 error = ENOENT; 11811 } 11812 if (error != 0) 11813 return (error); 11814 } 11815 /* 11816 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps 11817 * that are contained in direct blocks will be resolved by 11818 * doing a ffs_update. Pagedeps contained in indirect blocks 11819 * may require a complete sync'ing of the directory. So, we 11820 * try the cheap and fast ffs_update first, and if that fails, 11821 * then we do the slower ffs_syncvnode of the directory. 11822 */ 11823 if (flushparent) { 11824 int locked; 11825 11826 if ((error = ffs_update(pvp, 1)) != 0) { 11827 vput(pvp); 11828 return (error); 11829 } 11830 ACQUIRE_LOCK(&lk); 11831 locked = 1; 11832 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) { 11833 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) != NULL) { 11834 if (wk->wk_type != D_DIRADD) 11835 panic("softdep_fsync: Unexpected type %s", 11836 TYPENAME(wk->wk_type)); 11837 dap = WK_DIRADD(wk); 11838 if (dap->da_state & DIRCHG) 11839 pagedep = dap->da_previous->dm_pagedep; 11840 else 11841 pagedep = dap->da_pagedep; 11842 pagedep_new_block = pagedep->pd_state & NEWBLOCK; 11843 FREE_LOCK(&lk); 11844 locked = 0; 11845 if (pagedep_new_block && (error = 11846 ffs_syncvnode(pvp, MNT_WAIT, 0))) { 11847 vput(pvp); 11848 return (error); 11849 } 11850 } 11851 } 11852 if (locked) 11853 FREE_LOCK(&lk); 11854 } 11855 /* 11856 * Flush directory page containing the inode's name. 11857 */ 11858 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred, 11859 &bp); 11860 if (error == 0) 11861 error = bwrite(bp); 11862 else 11863 brelse(bp); 11864 vput(pvp); 11865 if (error != 0) 11866 return (error); 11867 ACQUIRE_LOCK(&lk); 11868 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 11869 break; 11870 } 11871 FREE_LOCK(&lk); 11872 return (0); 11873} 11874 11875/* 11876 * Flush all the dirty bitmaps associated with the block device 11877 * before flushing the rest of the dirty blocks so as to reduce 11878 * the number of dependencies that will have to be rolled back. 11879 * 11880 * XXX Unused? 11881 */ 11882void 11883softdep_fsync_mountdev(vp) 11884 struct vnode *vp; 11885{ 11886 struct buf *bp, *nbp; 11887 struct worklist *wk; 11888 struct bufobj *bo; 11889 11890 if (!vn_isdisk(vp, NULL)) 11891 panic("softdep_fsync_mountdev: vnode not a disk"); 11892 bo = &vp->v_bufobj; 11893restart: 11894 BO_LOCK(bo); 11895 ACQUIRE_LOCK(&lk); 11896 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 11897 /* 11898 * If it is already scheduled, skip to the next buffer. 11899 */ 11900 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 11901 continue; 11902 11903 if ((bp->b_flags & B_DELWRI) == 0) 11904 panic("softdep_fsync_mountdev: not dirty"); 11905 /* 11906 * We are only interested in bitmaps with outstanding 11907 * dependencies. 11908 */ 11909 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 11910 wk->wk_type != D_BMSAFEMAP || 11911 (bp->b_vflags & BV_BKGRDINPROG)) { 11912 BUF_UNLOCK(bp); 11913 continue; 11914 } 11915 FREE_LOCK(&lk); 11916 BO_UNLOCK(bo); 11917 bremfree(bp); 11918 (void) bawrite(bp); 11919 goto restart; 11920 } 11921 FREE_LOCK(&lk); 11922 drain_output(vp); 11923 BO_UNLOCK(bo); 11924} 11925 11926/* 11927 * Sync all cylinder groups that were dirty at the time this function is 11928 * called. Newly dirtied cgs will be inserted before the sintenel. This 11929 * is used to flush freedep activity that may be holding up writes to a 11930 * indirect block. 11931 */ 11932static int 11933sync_cgs(mp, waitfor) 11934 struct mount *mp; 11935 int waitfor; 11936{ 11937 struct bmsafemap *bmsafemap; 11938 struct bmsafemap *sintenel; 11939 struct ufsmount *ump; 11940 struct buf *bp; 11941 int error; 11942 11943 sintenel = malloc(sizeof(*sintenel), M_BMSAFEMAP, M_ZERO | M_WAITOK); 11944 sintenel->sm_cg = -1; 11945 ump = VFSTOUFS(mp); 11946 error = 0; 11947 ACQUIRE_LOCK(&lk); 11948 LIST_INSERT_HEAD(&ump->softdep_dirtycg, sintenel, sm_next); 11949 for (bmsafemap = LIST_NEXT(sintenel, sm_next); bmsafemap != NULL; 11950 bmsafemap = LIST_NEXT(sintenel, sm_next)) { 11951 /* Skip sintenels and cgs with no work to release. */ 11952 if (bmsafemap->sm_cg == -1 || 11953 (LIST_EMPTY(&bmsafemap->sm_freehd) && 11954 LIST_EMPTY(&bmsafemap->sm_freewr))) { 11955 LIST_REMOVE(sintenel, sm_next); 11956 LIST_INSERT_AFTER(bmsafemap, sintenel, sm_next); 11957 continue; 11958 } 11959 /* 11960 * If we don't get the lock and we're waiting try again, if 11961 * not move on to the next buf and try to sync it. 11962 */ 11963 bp = getdirtybuf(bmsafemap->sm_buf, &lk, waitfor); 11964 if (bp == NULL && waitfor == MNT_WAIT) 11965 continue; 11966 LIST_REMOVE(sintenel, sm_next); 11967 LIST_INSERT_AFTER(bmsafemap, sintenel, sm_next); 11968 if (bp == NULL) 11969 continue; 11970 FREE_LOCK(&lk); 11971 if (waitfor == MNT_NOWAIT) 11972 bawrite(bp); 11973 else 11974 error = bwrite(bp); 11975 ACQUIRE_LOCK(&lk); 11976 if (error) 11977 break; 11978 } 11979 LIST_REMOVE(sintenel, sm_next); 11980 FREE_LOCK(&lk); 11981 free(sintenel, M_BMSAFEMAP); 11982 return (error); 11983} 11984 11985/* 11986 * This routine is called when we are trying to synchronously flush a 11987 * file. This routine must eliminate any filesystem metadata dependencies 11988 * so that the syncing routine can succeed. 11989 */ 11990int 11991softdep_sync_metadata(struct vnode *vp) 11992{ 11993 int error; 11994 11995 /* 11996 * Ensure that any direct block dependencies have been cleared, 11997 * truncations are started, and inode references are journaled. 11998 */ 11999 ACQUIRE_LOCK(&lk); 12000 /* 12001 * Write all journal records to prevent rollbacks on devvp. 12002 */ 12003 if (vp->v_type == VCHR) 12004 softdep_flushjournal(vp->v_mount); 12005 error = flush_inodedep_deps(vp, vp->v_mount, VTOI(vp)->i_number); 12006 /* 12007 * Ensure that all truncates are written so we won't find deps on 12008 * indirect blocks. 12009 */ 12010 process_truncates(vp); 12011 FREE_LOCK(&lk); 12012 12013 return (error); 12014} 12015 12016/* 12017 * This routine is called when we are attempting to sync a buf with 12018 * dependencies. If waitfor is MNT_NOWAIT it attempts to schedule any 12019 * other IO it can but returns EBUSY if the buffer is not yet able to 12020 * be written. Dependencies which will not cause rollbacks will always 12021 * return 0. 12022 */ 12023int 12024softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor) 12025{ 12026 struct indirdep *indirdep; 12027 struct pagedep *pagedep; 12028 struct allocindir *aip; 12029 struct newblk *newblk; 12030 struct buf *nbp; 12031 struct worklist *wk; 12032 int i, error; 12033 12034 /* 12035 * For VCHR we just don't want to force flush any dependencies that 12036 * will cause rollbacks. 12037 */ 12038 if (vp->v_type == VCHR) { 12039 if (waitfor == MNT_NOWAIT && softdep_count_dependencies(bp, 0)) 12040 return (EBUSY); 12041 return (0); 12042 } 12043 ACQUIRE_LOCK(&lk); 12044 /* 12045 * As we hold the buffer locked, none of its dependencies 12046 * will disappear. 12047 */ 12048 error = 0; 12049top: 12050 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 12051 switch (wk->wk_type) { 12052 12053 case D_ALLOCDIRECT: 12054 case D_ALLOCINDIR: 12055 newblk = WK_NEWBLK(wk); 12056 if (newblk->nb_jnewblk != NULL) { 12057 if (waitfor == MNT_NOWAIT) { 12058 error = EBUSY; 12059 goto out_unlock; 12060 } 12061 jwait(&newblk->nb_jnewblk->jn_list, waitfor); 12062 goto top; 12063 } 12064 if (newblk->nb_state & DEPCOMPLETE || 12065 waitfor == MNT_NOWAIT) 12066 continue; 12067 nbp = newblk->nb_bmsafemap->sm_buf; 12068 nbp = getdirtybuf(nbp, &lk, waitfor); 12069 if (nbp == NULL) 12070 goto top; 12071 FREE_LOCK(&lk); 12072 if ((error = bwrite(nbp)) != 0) 12073 goto out; 12074 ACQUIRE_LOCK(&lk); 12075 continue; 12076 12077 case D_INDIRDEP: 12078 indirdep = WK_INDIRDEP(wk); 12079 if (waitfor == MNT_NOWAIT) { 12080 if (!TAILQ_EMPTY(&indirdep->ir_trunc) || 12081 !LIST_EMPTY(&indirdep->ir_deplisthd)) { 12082 error = EBUSY; 12083 goto out_unlock; 12084 } 12085 } 12086 if (!TAILQ_EMPTY(&indirdep->ir_trunc)) 12087 panic("softdep_sync_buf: truncation pending."); 12088 restart: 12089 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 12090 newblk = (struct newblk *)aip; 12091 if (newblk->nb_jnewblk != NULL) { 12092 jwait(&newblk->nb_jnewblk->jn_list, 12093 waitfor); 12094 goto restart; 12095 } 12096 if (newblk->nb_state & DEPCOMPLETE) 12097 continue; 12098 nbp = newblk->nb_bmsafemap->sm_buf; 12099 nbp = getdirtybuf(nbp, &lk, waitfor); 12100 if (nbp == NULL) 12101 goto restart; 12102 FREE_LOCK(&lk); 12103 if ((error = bwrite(nbp)) != 0) 12104 goto out; 12105 ACQUIRE_LOCK(&lk); 12106 goto restart; 12107 } 12108 continue; 12109 12110 case D_PAGEDEP: 12111 /* 12112 * Only flush directory entries in synchronous passes. 12113 */ 12114 if (waitfor != MNT_WAIT) { 12115 error = EBUSY; 12116 goto out_unlock; 12117 } 12118 /* 12119 * While syncing snapshots, we must allow recursive 12120 * lookups. 12121 */ 12122 BUF_AREC(bp); 12123 /* 12124 * We are trying to sync a directory that may 12125 * have dependencies on both its own metadata 12126 * and/or dependencies on the inodes of any 12127 * recently allocated files. We walk its diradd 12128 * lists pushing out the associated inode. 12129 */ 12130 pagedep = WK_PAGEDEP(wk); 12131 for (i = 0; i < DAHASHSZ; i++) { 12132 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 12133 continue; 12134 if ((error = flush_pagedep_deps(vp, wk->wk_mp, 12135 &pagedep->pd_diraddhd[i]))) { 12136 BUF_NOREC(bp); 12137 goto out_unlock; 12138 } 12139 } 12140 BUF_NOREC(bp); 12141 continue; 12142 12143 case D_FREEWORK: 12144 case D_FREEDEP: 12145 case D_JSEGDEP: 12146 case D_JNEWBLK: 12147 continue; 12148 12149 default: 12150 panic("softdep_sync_buf: Unknown type %s", 12151 TYPENAME(wk->wk_type)); 12152 /* NOTREACHED */ 12153 } 12154 } 12155out_unlock: 12156 FREE_LOCK(&lk); 12157out: 12158 return (error); 12159} 12160 12161/* 12162 * Flush the dependencies associated with an inodedep. 12163 * Called with splbio blocked. 12164 */ 12165static int 12166flush_inodedep_deps(vp, mp, ino) 12167 struct vnode *vp; 12168 struct mount *mp; 12169 ino_t ino; 12170{ 12171 struct inodedep *inodedep; 12172 struct inoref *inoref; 12173 int error, waitfor; 12174 12175 /* 12176 * This work is done in two passes. The first pass grabs most 12177 * of the buffers and begins asynchronously writing them. The 12178 * only way to wait for these asynchronous writes is to sleep 12179 * on the filesystem vnode which may stay busy for a long time 12180 * if the filesystem is active. So, instead, we make a second 12181 * pass over the dependencies blocking on each write. In the 12182 * usual case we will be blocking against a write that we 12183 * initiated, so when it is done the dependency will have been 12184 * resolved. Thus the second pass is expected to end quickly. 12185 * We give a brief window at the top of the loop to allow 12186 * any pending I/O to complete. 12187 */ 12188 for (error = 0, waitfor = MNT_NOWAIT; ; ) { 12189 if (error) 12190 return (error); 12191 FREE_LOCK(&lk); 12192 ACQUIRE_LOCK(&lk); 12193restart: 12194 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 12195 return (0); 12196 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12197 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12198 == DEPCOMPLETE) { 12199 jwait(&inoref->if_list, MNT_WAIT); 12200 goto restart; 12201 } 12202 } 12203 if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) || 12204 flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) || 12205 flush_deplist(&inodedep->id_extupdt, waitfor, &error) || 12206 flush_deplist(&inodedep->id_newextupdt, waitfor, &error)) 12207 continue; 12208 /* 12209 * If pass2, we are done, otherwise do pass 2. 12210 */ 12211 if (waitfor == MNT_WAIT) 12212 break; 12213 waitfor = MNT_WAIT; 12214 } 12215 /* 12216 * Try freeing inodedep in case all dependencies have been removed. 12217 */ 12218 if (inodedep_lookup(mp, ino, 0, &inodedep) != 0) 12219 (void) free_inodedep(inodedep); 12220 return (0); 12221} 12222 12223/* 12224 * Flush an inode dependency list. 12225 * Called with splbio blocked. 12226 */ 12227static int 12228flush_deplist(listhead, waitfor, errorp) 12229 struct allocdirectlst *listhead; 12230 int waitfor; 12231 int *errorp; 12232{ 12233 struct allocdirect *adp; 12234 struct newblk *newblk; 12235 struct buf *bp; 12236 12237 mtx_assert(&lk, MA_OWNED); 12238 TAILQ_FOREACH(adp, listhead, ad_next) { 12239 newblk = (struct newblk *)adp; 12240 if (newblk->nb_jnewblk != NULL) { 12241 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 12242 return (1); 12243 } 12244 if (newblk->nb_state & DEPCOMPLETE) 12245 continue; 12246 bp = newblk->nb_bmsafemap->sm_buf; 12247 bp = getdirtybuf(bp, &lk, waitfor); 12248 if (bp == NULL) { 12249 if (waitfor == MNT_NOWAIT) 12250 continue; 12251 return (1); 12252 } 12253 FREE_LOCK(&lk); 12254 if (waitfor == MNT_NOWAIT) 12255 bawrite(bp); 12256 else 12257 *errorp = bwrite(bp); 12258 ACQUIRE_LOCK(&lk); 12259 return (1); 12260 } 12261 return (0); 12262} 12263 12264/* 12265 * Flush dependencies associated with an allocdirect block. 12266 */ 12267static int 12268flush_newblk_dep(vp, mp, lbn) 12269 struct vnode *vp; 12270 struct mount *mp; 12271 ufs_lbn_t lbn; 12272{ 12273 struct newblk *newblk; 12274 struct bufobj *bo; 12275 struct inode *ip; 12276 struct buf *bp; 12277 ufs2_daddr_t blkno; 12278 int error; 12279 12280 error = 0; 12281 bo = &vp->v_bufobj; 12282 ip = VTOI(vp); 12283 blkno = DIP(ip, i_db[lbn]); 12284 if (blkno == 0) 12285 panic("flush_newblk_dep: Missing block"); 12286 ACQUIRE_LOCK(&lk); 12287 /* 12288 * Loop until all dependencies related to this block are satisfied. 12289 * We must be careful to restart after each sleep in case a write 12290 * completes some part of this process for us. 12291 */ 12292 for (;;) { 12293 if (newblk_lookup(mp, blkno, 0, &newblk) == 0) { 12294 FREE_LOCK(&lk); 12295 break; 12296 } 12297 if (newblk->nb_list.wk_type != D_ALLOCDIRECT) 12298 panic("flush_newblk_deps: Bad newblk %p", newblk); 12299 /* 12300 * Flush the journal. 12301 */ 12302 if (newblk->nb_jnewblk != NULL) { 12303 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 12304 continue; 12305 } 12306 /* 12307 * Write the bitmap dependency. 12308 */ 12309 if ((newblk->nb_state & DEPCOMPLETE) == 0) { 12310 bp = newblk->nb_bmsafemap->sm_buf; 12311 bp = getdirtybuf(bp, &lk, MNT_WAIT); 12312 if (bp == NULL) 12313 continue; 12314 FREE_LOCK(&lk); 12315 error = bwrite(bp); 12316 if (error) 12317 break; 12318 ACQUIRE_LOCK(&lk); 12319 continue; 12320 } 12321 /* 12322 * Write the buffer. 12323 */ 12324 FREE_LOCK(&lk); 12325 BO_LOCK(bo); 12326 bp = gbincore(bo, lbn); 12327 if (bp != NULL) { 12328 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | 12329 LK_INTERLOCK, BO_MTX(bo)); 12330 if (error == ENOLCK) { 12331 ACQUIRE_LOCK(&lk); 12332 continue; /* Slept, retry */ 12333 } 12334 if (error != 0) 12335 break; /* Failed */ 12336 if (bp->b_flags & B_DELWRI) { 12337 bremfree(bp); 12338 error = bwrite(bp); 12339 if (error) 12340 break; 12341 } else 12342 BUF_UNLOCK(bp); 12343 } else 12344 BO_UNLOCK(bo); 12345 /* 12346 * We have to wait for the direct pointers to 12347 * point at the newdirblk before the dependency 12348 * will go away. 12349 */ 12350 error = ffs_update(vp, 1); 12351 if (error) 12352 break; 12353 ACQUIRE_LOCK(&lk); 12354 } 12355 return (error); 12356} 12357 12358/* 12359 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 12360 * Called with splbio blocked. 12361 */ 12362static int 12363flush_pagedep_deps(pvp, mp, diraddhdp) 12364 struct vnode *pvp; 12365 struct mount *mp; 12366 struct diraddhd *diraddhdp; 12367{ 12368 struct inodedep *inodedep; 12369 struct inoref *inoref; 12370 struct ufsmount *ump; 12371 struct diradd *dap; 12372 struct vnode *vp; 12373 int error = 0; 12374 struct buf *bp; 12375 ino_t inum; 12376 12377 ump = VFSTOUFS(mp); 12378restart: 12379 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 12380 /* 12381 * Flush ourselves if this directory entry 12382 * has a MKDIR_PARENT dependency. 12383 */ 12384 if (dap->da_state & MKDIR_PARENT) { 12385 FREE_LOCK(&lk); 12386 if ((error = ffs_update(pvp, 1)) != 0) 12387 break; 12388 ACQUIRE_LOCK(&lk); 12389 /* 12390 * If that cleared dependencies, go on to next. 12391 */ 12392 if (dap != LIST_FIRST(diraddhdp)) 12393 continue; 12394 if (dap->da_state & MKDIR_PARENT) 12395 panic("flush_pagedep_deps: MKDIR_PARENT"); 12396 } 12397 /* 12398 * A newly allocated directory must have its "." and 12399 * ".." entries written out before its name can be 12400 * committed in its parent. 12401 */ 12402 inum = dap->da_newinum; 12403 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 12404 panic("flush_pagedep_deps: lost inode1"); 12405 /* 12406 * Wait for any pending journal adds to complete so we don't 12407 * cause rollbacks while syncing. 12408 */ 12409 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12410 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12411 == DEPCOMPLETE) { 12412 jwait(&inoref->if_list, MNT_WAIT); 12413 goto restart; 12414 } 12415 } 12416 if (dap->da_state & MKDIR_BODY) { 12417 FREE_LOCK(&lk); 12418 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp, 12419 FFSV_FORCEINSMQ))) 12420 break; 12421 error = flush_newblk_dep(vp, mp, 0); 12422 /* 12423 * If we still have the dependency we might need to 12424 * update the vnode to sync the new link count to 12425 * disk. 12426 */ 12427 if (error == 0 && dap == LIST_FIRST(diraddhdp)) 12428 error = ffs_update(vp, 1); 12429 vput(vp); 12430 if (error != 0) 12431 break; 12432 ACQUIRE_LOCK(&lk); 12433 /* 12434 * If that cleared dependencies, go on to next. 12435 */ 12436 if (dap != LIST_FIRST(diraddhdp)) 12437 continue; 12438 if (dap->da_state & MKDIR_BODY) { 12439 inodedep_lookup(UFSTOVFS(ump), inum, 0, 12440 &inodedep); 12441 panic("flush_pagedep_deps: MKDIR_BODY " 12442 "inodedep %p dap %p vp %p", 12443 inodedep, dap, vp); 12444 } 12445 } 12446 /* 12447 * Flush the inode on which the directory entry depends. 12448 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 12449 * the only remaining dependency is that the updated inode 12450 * count must get pushed to disk. The inode has already 12451 * been pushed into its inode buffer (via VOP_UPDATE) at 12452 * the time of the reference count change. So we need only 12453 * locate that buffer, ensure that there will be no rollback 12454 * caused by a bitmap dependency, then write the inode buffer. 12455 */ 12456retry: 12457 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 12458 panic("flush_pagedep_deps: lost inode"); 12459 /* 12460 * If the inode still has bitmap dependencies, 12461 * push them to disk. 12462 */ 12463 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) == 0) { 12464 bp = inodedep->id_bmsafemap->sm_buf; 12465 bp = getdirtybuf(bp, &lk, MNT_WAIT); 12466 if (bp == NULL) 12467 goto retry; 12468 FREE_LOCK(&lk); 12469 if ((error = bwrite(bp)) != 0) 12470 break; 12471 ACQUIRE_LOCK(&lk); 12472 if (dap != LIST_FIRST(diraddhdp)) 12473 continue; 12474 } 12475 /* 12476 * If the inode is still sitting in a buffer waiting 12477 * to be written or waiting for the link count to be 12478 * adjusted update it here to flush it to disk. 12479 */ 12480 if (dap == LIST_FIRST(diraddhdp)) { 12481 FREE_LOCK(&lk); 12482 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp, 12483 FFSV_FORCEINSMQ))) 12484 break; 12485 error = ffs_update(vp, 1); 12486 vput(vp); 12487 if (error) 12488 break; 12489 ACQUIRE_LOCK(&lk); 12490 } 12491 /* 12492 * If we have failed to get rid of all the dependencies 12493 * then something is seriously wrong. 12494 */ 12495 if (dap == LIST_FIRST(diraddhdp)) { 12496 inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep); 12497 panic("flush_pagedep_deps: failed to flush " 12498 "inodedep %p ino %ju dap %p", 12499 inodedep, (uintmax_t)inum, dap); 12500 } 12501 } 12502 if (error) 12503 ACQUIRE_LOCK(&lk); 12504 return (error); 12505} 12506 12507/* 12508 * A large burst of file addition or deletion activity can drive the 12509 * memory load excessively high. First attempt to slow things down 12510 * using the techniques below. If that fails, this routine requests 12511 * the offending operations to fall back to running synchronously 12512 * until the memory load returns to a reasonable level. 12513 */ 12514int 12515softdep_slowdown(vp) 12516 struct vnode *vp; 12517{ 12518 struct ufsmount *ump; 12519 int jlow; 12520 int max_softdeps_hard; 12521 12522 ACQUIRE_LOCK(&lk); 12523 jlow = 0; 12524 /* 12525 * Check for journal space if needed. 12526 */ 12527 if (DOINGSUJ(vp)) { 12528 ump = VFSTOUFS(vp->v_mount); 12529 if (journal_space(ump, 0) == 0) 12530 jlow = 1; 12531 } 12532 max_softdeps_hard = max_softdeps * 11 / 10; 12533 if (dep_current[D_DIRREM] < max_softdeps_hard / 2 && 12534 dep_current[D_INODEDEP] < max_softdeps_hard && 12535 VFSTOUFS(vp->v_mount)->um_numindirdeps < maxindirdeps && 12536 dep_current[D_FREEBLKS] < max_softdeps_hard && jlow == 0) { 12537 FREE_LOCK(&lk); 12538 return (0); 12539 } 12540 if (VFSTOUFS(vp->v_mount)->um_numindirdeps >= maxindirdeps || jlow) 12541 softdep_speedup(); 12542 stat_sync_limit_hit += 1; 12543 FREE_LOCK(&lk); 12544 if (DOINGSUJ(vp)) 12545 return (0); 12546 return (1); 12547} 12548 12549/* 12550 * Called by the allocation routines when they are about to fail 12551 * in the hope that we can free up the requested resource (inodes 12552 * or disk space). 12553 * 12554 * First check to see if the work list has anything on it. If it has, 12555 * clean up entries until we successfully free the requested resource. 12556 * Because this process holds inodes locked, we cannot handle any remove 12557 * requests that might block on a locked inode as that could lead to 12558 * deadlock. If the worklist yields none of the requested resource, 12559 * start syncing out vnodes to free up the needed space. 12560 */ 12561int 12562softdep_request_cleanup(fs, vp, cred, resource) 12563 struct fs *fs; 12564 struct vnode *vp; 12565 struct ucred *cred; 12566 int resource; 12567{ 12568 struct ufsmount *ump; 12569 struct mount *mp; 12570 struct vnode *lvp, *mvp; 12571 long starttime; 12572 ufs2_daddr_t needed; 12573 int error; 12574 12575 /* 12576 * If we are being called because of a process doing a 12577 * copy-on-write, then it is not safe to process any 12578 * worklist items as we will recurse into the copyonwrite 12579 * routine. This will result in an incoherent snapshot. 12580 * If the vnode that we hold is a snapshot, we must avoid 12581 * handling other resources that could cause deadlock. 12582 */ 12583 if ((curthread->td_pflags & TDP_COWINPROGRESS) || IS_SNAPSHOT(VTOI(vp))) 12584 return (0); 12585 12586 if (resource == FLUSH_BLOCKS_WAIT) 12587 stat_cleanup_blkrequests += 1; 12588 else 12589 stat_cleanup_inorequests += 1; 12590 12591 mp = vp->v_mount; 12592 ump = VFSTOUFS(mp); 12593 mtx_assert(UFS_MTX(ump), MA_OWNED); 12594 UFS_UNLOCK(ump); 12595 error = ffs_update(vp, 1); 12596 if (error != 0) { 12597 UFS_LOCK(ump); 12598 return (0); 12599 } 12600 /* 12601 * If we are in need of resources, consider pausing for 12602 * tickdelay to give ourselves some breathing room. 12603 */ 12604 ACQUIRE_LOCK(&lk); 12605 process_removes(vp); 12606 process_truncates(vp); 12607 request_cleanup(UFSTOVFS(ump), resource); 12608 FREE_LOCK(&lk); 12609 /* 12610 * Now clean up at least as many resources as we will need. 12611 * 12612 * When requested to clean up inodes, the number that are needed 12613 * is set by the number of simultaneous writers (mnt_writeopcount) 12614 * plus a bit of slop (2) in case some more writers show up while 12615 * we are cleaning. 12616 * 12617 * When requested to free up space, the amount of space that 12618 * we need is enough blocks to allocate a full-sized segment 12619 * (fs_contigsumsize). The number of such segments that will 12620 * be needed is set by the number of simultaneous writers 12621 * (mnt_writeopcount) plus a bit of slop (2) in case some more 12622 * writers show up while we are cleaning. 12623 * 12624 * Additionally, if we are unpriviledged and allocating space, 12625 * we need to ensure that we clean up enough blocks to get the 12626 * needed number of blocks over the threshhold of the minimum 12627 * number of blocks required to be kept free by the filesystem 12628 * (fs_minfree). 12629 */ 12630 if (resource == FLUSH_INODES_WAIT) { 12631 needed = vp->v_mount->mnt_writeopcount + 2; 12632 } else if (resource == FLUSH_BLOCKS_WAIT) { 12633 needed = (vp->v_mount->mnt_writeopcount + 2) * 12634 fs->fs_contigsumsize; 12635 if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE, 0)) 12636 needed += fragstoblks(fs, 12637 roundup((fs->fs_dsize * fs->fs_minfree / 100) - 12638 fs->fs_cstotal.cs_nffree, fs->fs_frag)); 12639 } else { 12640 UFS_LOCK(ump); 12641 printf("softdep_request_cleanup: Unknown resource type %d\n", 12642 resource); 12643 return (0); 12644 } 12645 starttime = time_second; 12646retry: 12647 if ((resource == FLUSH_BLOCKS_WAIT && ump->softdep_on_worklist > 0 && 12648 fs->fs_cstotal.cs_nbfree <= needed) || 12649 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 && 12650 fs->fs_cstotal.cs_nifree <= needed)) { 12651 ACQUIRE_LOCK(&lk); 12652 if (ump->softdep_on_worklist > 0 && 12653 process_worklist_item(UFSTOVFS(ump), 12654 ump->softdep_on_worklist, LK_NOWAIT) != 0) 12655 stat_worklist_push += 1; 12656 FREE_LOCK(&lk); 12657 } 12658 /* 12659 * If we still need resources and there are no more worklist 12660 * entries to process to obtain them, we have to start flushing 12661 * the dirty vnodes to force the release of additional requests 12662 * to the worklist that we can then process to reap addition 12663 * resources. We walk the vnodes associated with the mount point 12664 * until we get the needed worklist requests that we can reap. 12665 */ 12666 if ((resource == FLUSH_BLOCKS_WAIT && 12667 fs->fs_cstotal.cs_nbfree <= needed) || 12668 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 && 12669 fs->fs_cstotal.cs_nifree <= needed)) { 12670 MNT_VNODE_FOREACH_ALL(lvp, mp, mvp) { 12671 if (TAILQ_FIRST(&lvp->v_bufobj.bo_dirty.bv_hd) == 0) { 12672 VI_UNLOCK(lvp); 12673 continue; 12674 } 12675 if (vget(lvp, LK_EXCLUSIVE | LK_INTERLOCK | LK_NOWAIT, 12676 curthread)) 12677 continue; 12678 if (lvp->v_vflag & VV_NOSYNC) { /* unlinked */ 12679 vput(lvp); 12680 continue; 12681 } 12682 (void) ffs_syncvnode(lvp, MNT_NOWAIT, 0); 12683 vput(lvp); 12684 } 12685 lvp = ump->um_devvp; 12686 if (vn_lock(lvp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 12687 VOP_FSYNC(lvp, MNT_NOWAIT, curthread); 12688 VOP_UNLOCK(lvp, 0); 12689 } 12690 if (ump->softdep_on_worklist > 0) { 12691 stat_cleanup_retries += 1; 12692 goto retry; 12693 } 12694 stat_cleanup_failures += 1; 12695 } 12696 if (time_second - starttime > stat_cleanup_high_delay) 12697 stat_cleanup_high_delay = time_second - starttime; 12698 UFS_LOCK(ump); 12699 return (1); 12700} 12701 12702/* 12703 * If memory utilization has gotten too high, deliberately slow things 12704 * down and speed up the I/O processing. 12705 */ 12706extern struct thread *syncertd; 12707static int 12708request_cleanup(mp, resource) 12709 struct mount *mp; 12710 int resource; 12711{ 12712 struct thread *td = curthread; 12713 struct ufsmount *ump; 12714 12715 mtx_assert(&lk, MA_OWNED); 12716 /* 12717 * We never hold up the filesystem syncer or buf daemon. 12718 */ 12719 if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF)) 12720 return (0); 12721 ump = VFSTOUFS(mp); 12722 /* 12723 * First check to see if the work list has gotten backlogged. 12724 * If it has, co-opt this process to help clean up two entries. 12725 * Because this process may hold inodes locked, we cannot 12726 * handle any remove requests that might block on a locked 12727 * inode as that could lead to deadlock. We set TDP_SOFTDEP 12728 * to avoid recursively processing the worklist. 12729 */ 12730 if (ump->softdep_on_worklist > max_softdeps / 10) { 12731 td->td_pflags |= TDP_SOFTDEP; 12732 process_worklist_item(mp, 2, LK_NOWAIT); 12733 td->td_pflags &= ~TDP_SOFTDEP; 12734 stat_worklist_push += 2; 12735 return(1); 12736 } 12737 /* 12738 * Next, we attempt to speed up the syncer process. If that 12739 * is successful, then we allow the process to continue. 12740 */ 12741 if (softdep_speedup() && 12742 resource != FLUSH_BLOCKS_WAIT && 12743 resource != FLUSH_INODES_WAIT) 12744 return(0); 12745 /* 12746 * If we are resource constrained on inode dependencies, try 12747 * flushing some dirty inodes. Otherwise, we are constrained 12748 * by file deletions, so try accelerating flushes of directories 12749 * with removal dependencies. We would like to do the cleanup 12750 * here, but we probably hold an inode locked at this point and 12751 * that might deadlock against one that we try to clean. So, 12752 * the best that we can do is request the syncer daemon to do 12753 * the cleanup for us. 12754 */ 12755 switch (resource) { 12756 12757 case FLUSH_INODES: 12758 case FLUSH_INODES_WAIT: 12759 stat_ino_limit_push += 1; 12760 req_clear_inodedeps += 1; 12761 stat_countp = &stat_ino_limit_hit; 12762 break; 12763 12764 case FLUSH_BLOCKS: 12765 case FLUSH_BLOCKS_WAIT: 12766 stat_blk_limit_push += 1; 12767 req_clear_remove += 1; 12768 stat_countp = &stat_blk_limit_hit; 12769 break; 12770 12771 default: 12772 panic("request_cleanup: unknown type"); 12773 } 12774 /* 12775 * Hopefully the syncer daemon will catch up and awaken us. 12776 * We wait at most tickdelay before proceeding in any case. 12777 */ 12778 proc_waiting += 1; 12779 if (callout_pending(&softdep_callout) == FALSE) 12780 callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2, 12781 pause_timer, 0); 12782 12783 msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0); 12784 proc_waiting -= 1; 12785 return (1); 12786} 12787 12788/* 12789 * Awaken processes pausing in request_cleanup and clear proc_waiting 12790 * to indicate that there is no longer a timer running. 12791 */ 12792static void 12793pause_timer(arg) 12794 void *arg; 12795{ 12796 12797 /* 12798 * The callout_ API has acquired mtx and will hold it around this 12799 * function call. 12800 */ 12801 *stat_countp += 1; 12802 wakeup_one(&proc_waiting); 12803 if (proc_waiting > 0) 12804 callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2, 12805 pause_timer, 0); 12806} 12807 12808/* 12809 * Flush out a directory with at least one removal dependency in an effort to 12810 * reduce the number of dirrem, freefile, and freeblks dependency structures. 12811 */ 12812static void 12813clear_remove(void) 12814{ 12815 struct pagedep_hashhead *pagedephd; 12816 struct pagedep *pagedep; 12817 static int next = 0; 12818 struct mount *mp; 12819 struct vnode *vp; 12820 struct bufobj *bo; 12821 int error, cnt; 12822 ino_t ino; 12823 12824 mtx_assert(&lk, MA_OWNED); 12825 12826 for (cnt = 0; cnt < pagedep_hash; cnt++) { 12827 pagedephd = &pagedep_hashtbl[next++]; 12828 if (next >= pagedep_hash) 12829 next = 0; 12830 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 12831 if (LIST_EMPTY(&pagedep->pd_dirremhd)) 12832 continue; 12833 mp = pagedep->pd_list.wk_mp; 12834 ino = pagedep->pd_ino; 12835 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 12836 continue; 12837 FREE_LOCK(&lk); 12838 12839 /* 12840 * Let unmount clear deps 12841 */ 12842 error = vfs_busy(mp, MBF_NOWAIT); 12843 if (error != 0) 12844 goto finish_write; 12845 error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp, 12846 FFSV_FORCEINSMQ); 12847 vfs_unbusy(mp); 12848 if (error != 0) { 12849 softdep_error("clear_remove: vget", error); 12850 goto finish_write; 12851 } 12852 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0))) 12853 softdep_error("clear_remove: fsync", error); 12854 bo = &vp->v_bufobj; 12855 BO_LOCK(bo); 12856 drain_output(vp); 12857 BO_UNLOCK(bo); 12858 vput(vp); 12859 finish_write: 12860 vn_finished_write(mp); 12861 ACQUIRE_LOCK(&lk); 12862 return; 12863 } 12864 } 12865} 12866 12867/* 12868 * Clear out a block of dirty inodes in an effort to reduce 12869 * the number of inodedep dependency structures. 12870 */ 12871static void 12872clear_inodedeps(void) 12873{ 12874 struct inodedep_hashhead *inodedephd; 12875 struct inodedep *inodedep; 12876 static int next = 0; 12877 struct mount *mp; 12878 struct vnode *vp; 12879 struct fs *fs; 12880 int error, cnt; 12881 ino_t firstino, lastino, ino; 12882 12883 mtx_assert(&lk, MA_OWNED); 12884 /* 12885 * Pick a random inode dependency to be cleared. 12886 * We will then gather up all the inodes in its block 12887 * that have dependencies and flush them out. 12888 */ 12889 for (cnt = 0; cnt < inodedep_hash; cnt++) { 12890 inodedephd = &inodedep_hashtbl[next++]; 12891 if (next >= inodedep_hash) 12892 next = 0; 12893 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 12894 break; 12895 } 12896 if (inodedep == NULL) 12897 return; 12898 fs = inodedep->id_fs; 12899 mp = inodedep->id_list.wk_mp; 12900 /* 12901 * Find the last inode in the block with dependencies. 12902 */ 12903 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 12904 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 12905 if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0) 12906 break; 12907 /* 12908 * Asynchronously push all but the last inode with dependencies. 12909 * Synchronously push the last inode with dependencies to ensure 12910 * that the inode block gets written to free up the inodedeps. 12911 */ 12912 for (ino = firstino; ino <= lastino; ino++) { 12913 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 12914 continue; 12915 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 12916 continue; 12917 FREE_LOCK(&lk); 12918 error = vfs_busy(mp, MBF_NOWAIT); /* Let unmount clear deps */ 12919 if (error != 0) { 12920 vn_finished_write(mp); 12921 ACQUIRE_LOCK(&lk); 12922 return; 12923 } 12924 if ((error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp, 12925 FFSV_FORCEINSMQ)) != 0) { 12926 softdep_error("clear_inodedeps: vget", error); 12927 vfs_unbusy(mp); 12928 vn_finished_write(mp); 12929 ACQUIRE_LOCK(&lk); 12930 return; 12931 } 12932 vfs_unbusy(mp); 12933 if (ino == lastino) { 12934 if ((error = ffs_syncvnode(vp, MNT_WAIT, 0))) 12935 softdep_error("clear_inodedeps: fsync1", error); 12936 } else { 12937 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0))) 12938 softdep_error("clear_inodedeps: fsync2", error); 12939 BO_LOCK(&vp->v_bufobj); 12940 drain_output(vp); 12941 BO_UNLOCK(&vp->v_bufobj); 12942 } 12943 vput(vp); 12944 vn_finished_write(mp); 12945 ACQUIRE_LOCK(&lk); 12946 } 12947} 12948 12949void 12950softdep_buf_append(bp, wkhd) 12951 struct buf *bp; 12952 struct workhead *wkhd; 12953{ 12954 struct worklist *wk; 12955 12956 ACQUIRE_LOCK(&lk); 12957 while ((wk = LIST_FIRST(wkhd)) != NULL) { 12958 WORKLIST_REMOVE(wk); 12959 WORKLIST_INSERT(&bp->b_dep, wk); 12960 } 12961 FREE_LOCK(&lk); 12962 12963} 12964 12965void 12966softdep_inode_append(ip, cred, wkhd) 12967 struct inode *ip; 12968 struct ucred *cred; 12969 struct workhead *wkhd; 12970{ 12971 struct buf *bp; 12972 struct fs *fs; 12973 int error; 12974 12975 fs = ip->i_fs; 12976 error = bread(ip->i_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 12977 (int)fs->fs_bsize, cred, &bp); 12978 if (error) { 12979 softdep_freework(wkhd); 12980 return; 12981 } 12982 softdep_buf_append(bp, wkhd); 12983 bqrelse(bp); 12984} 12985 12986void 12987softdep_freework(wkhd) 12988 struct workhead *wkhd; 12989{ 12990 12991 ACQUIRE_LOCK(&lk); 12992 handle_jwork(wkhd); 12993 FREE_LOCK(&lk); 12994} 12995 12996/* 12997 * Function to determine if the buffer has outstanding dependencies 12998 * that will cause a roll-back if the buffer is written. If wantcount 12999 * is set, return number of dependencies, otherwise just yes or no. 13000 */ 13001static int 13002softdep_count_dependencies(bp, wantcount) 13003 struct buf *bp; 13004 int wantcount; 13005{ 13006 struct worklist *wk; 13007 struct bmsafemap *bmsafemap; 13008 struct freework *freework; 13009 struct inodedep *inodedep; 13010 struct indirdep *indirdep; 13011 struct freeblks *freeblks; 13012 struct allocindir *aip; 13013 struct pagedep *pagedep; 13014 struct dirrem *dirrem; 13015 struct newblk *newblk; 13016 struct mkdir *mkdir; 13017 struct diradd *dap; 13018 int i, retval; 13019 13020 retval = 0; 13021 ACQUIRE_LOCK(&lk); 13022 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 13023 switch (wk->wk_type) { 13024 13025 case D_INODEDEP: 13026 inodedep = WK_INODEDEP(wk); 13027 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 13028 /* bitmap allocation dependency */ 13029 retval += 1; 13030 if (!wantcount) 13031 goto out; 13032 } 13033 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 13034 /* direct block pointer dependency */ 13035 retval += 1; 13036 if (!wantcount) 13037 goto out; 13038 } 13039 if (TAILQ_FIRST(&inodedep->id_extupdt)) { 13040 /* direct block pointer dependency */ 13041 retval += 1; 13042 if (!wantcount) 13043 goto out; 13044 } 13045 if (TAILQ_FIRST(&inodedep->id_inoreflst)) { 13046 /* Add reference dependency. */ 13047 retval += 1; 13048 if (!wantcount) 13049 goto out; 13050 } 13051 continue; 13052 13053 case D_INDIRDEP: 13054 indirdep = WK_INDIRDEP(wk); 13055 13056 TAILQ_FOREACH(freework, &indirdep->ir_trunc, fw_next) { 13057 /* indirect truncation dependency */ 13058 retval += 1; 13059 if (!wantcount) 13060 goto out; 13061 } 13062 13063 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 13064 /* indirect block pointer dependency */ 13065 retval += 1; 13066 if (!wantcount) 13067 goto out; 13068 } 13069 continue; 13070 13071 case D_PAGEDEP: 13072 pagedep = WK_PAGEDEP(wk); 13073 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 13074 if (LIST_FIRST(&dirrem->dm_jremrefhd)) { 13075 /* Journal remove ref dependency. */ 13076 retval += 1; 13077 if (!wantcount) 13078 goto out; 13079 } 13080 } 13081 for (i = 0; i < DAHASHSZ; i++) { 13082 13083 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 13084 /* directory entry dependency */ 13085 retval += 1; 13086 if (!wantcount) 13087 goto out; 13088 } 13089 } 13090 continue; 13091 13092 case D_BMSAFEMAP: 13093 bmsafemap = WK_BMSAFEMAP(wk); 13094 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd)) { 13095 /* Add reference dependency. */ 13096 retval += 1; 13097 if (!wantcount) 13098 goto out; 13099 } 13100 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd)) { 13101 /* Allocate block dependency. */ 13102 retval += 1; 13103 if (!wantcount) 13104 goto out; 13105 } 13106 continue; 13107 13108 case D_FREEBLKS: 13109 freeblks = WK_FREEBLKS(wk); 13110 if (LIST_FIRST(&freeblks->fb_jblkdephd)) { 13111 /* Freeblk journal dependency. */ 13112 retval += 1; 13113 if (!wantcount) 13114 goto out; 13115 } 13116 continue; 13117 13118 case D_ALLOCDIRECT: 13119 case D_ALLOCINDIR: 13120 newblk = WK_NEWBLK(wk); 13121 if (newblk->nb_jnewblk) { 13122 /* Journal allocate dependency. */ 13123 retval += 1; 13124 if (!wantcount) 13125 goto out; 13126 } 13127 continue; 13128 13129 case D_MKDIR: 13130 mkdir = WK_MKDIR(wk); 13131 if (mkdir->md_jaddref) { 13132 /* Journal reference dependency. */ 13133 retval += 1; 13134 if (!wantcount) 13135 goto out; 13136 } 13137 continue; 13138 13139 case D_FREEWORK: 13140 case D_FREEDEP: 13141 case D_JSEGDEP: 13142 case D_JSEG: 13143 case D_SBDEP: 13144 /* never a dependency on these blocks */ 13145 continue; 13146 13147 default: 13148 panic("softdep_count_dependencies: Unexpected type %s", 13149 TYPENAME(wk->wk_type)); 13150 /* NOTREACHED */ 13151 } 13152 } 13153out: 13154 FREE_LOCK(&lk); 13155 return retval; 13156} 13157 13158/* 13159 * Acquire exclusive access to a buffer. 13160 * Must be called with a locked mtx parameter. 13161 * Return acquired buffer or NULL on failure. 13162 */ 13163static struct buf * 13164getdirtybuf(bp, mtx, waitfor) 13165 struct buf *bp; 13166 struct mtx *mtx; 13167 int waitfor; 13168{ 13169 int error; 13170 13171 mtx_assert(mtx, MA_OWNED); 13172 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) { 13173 if (waitfor != MNT_WAIT) 13174 return (NULL); 13175 error = BUF_LOCK(bp, 13176 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, mtx); 13177 /* 13178 * Even if we sucessfully acquire bp here, we have dropped 13179 * mtx, which may violates our guarantee. 13180 */ 13181 if (error == 0) 13182 BUF_UNLOCK(bp); 13183 else if (error != ENOLCK) 13184 panic("getdirtybuf: inconsistent lock: %d", error); 13185 mtx_lock(mtx); 13186 return (NULL); 13187 } 13188 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 13189 if (mtx == &lk && waitfor == MNT_WAIT) { 13190 mtx_unlock(mtx); 13191 BO_LOCK(bp->b_bufobj); 13192 BUF_UNLOCK(bp); 13193 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 13194 bp->b_vflags |= BV_BKGRDWAIT; 13195 msleep(&bp->b_xflags, BO_MTX(bp->b_bufobj), 13196 PRIBIO | PDROP, "getbuf", 0); 13197 } else 13198 BO_UNLOCK(bp->b_bufobj); 13199 mtx_lock(mtx); 13200 return (NULL); 13201 } 13202 BUF_UNLOCK(bp); 13203 if (waitfor != MNT_WAIT) 13204 return (NULL); 13205 /* 13206 * The mtx argument must be bp->b_vp's mutex in 13207 * this case. 13208 */ 13209#ifdef DEBUG_VFS_LOCKS 13210 if (bp->b_vp->v_type != VCHR) 13211 ASSERT_BO_LOCKED(bp->b_bufobj); 13212#endif 13213 bp->b_vflags |= BV_BKGRDWAIT; 13214 msleep(&bp->b_xflags, mtx, PRIBIO, "getbuf", 0); 13215 return (NULL); 13216 } 13217 if ((bp->b_flags & B_DELWRI) == 0) { 13218 BUF_UNLOCK(bp); 13219 return (NULL); 13220 } 13221 bremfree(bp); 13222 return (bp); 13223} 13224 13225 13226/* 13227 * Check if it is safe to suspend the file system now. On entry, 13228 * the vnode interlock for devvp should be held. Return 0 with 13229 * the mount interlock held if the file system can be suspended now, 13230 * otherwise return EAGAIN with the mount interlock held. 13231 */ 13232int 13233softdep_check_suspend(struct mount *mp, 13234 struct vnode *devvp, 13235 int softdep_deps, 13236 int softdep_accdeps, 13237 int secondary_writes, 13238 int secondary_accwrites) 13239{ 13240 struct bufobj *bo; 13241 struct ufsmount *ump; 13242 int error; 13243 13244 ump = VFSTOUFS(mp); 13245 bo = &devvp->v_bufobj; 13246 ASSERT_BO_LOCKED(bo); 13247 13248 for (;;) { 13249 if (!TRY_ACQUIRE_LOCK(&lk)) { 13250 BO_UNLOCK(bo); 13251 ACQUIRE_LOCK(&lk); 13252 FREE_LOCK(&lk); 13253 BO_LOCK(bo); 13254 continue; 13255 } 13256 MNT_ILOCK(mp); 13257 if (mp->mnt_secondary_writes != 0) { 13258 FREE_LOCK(&lk); 13259 BO_UNLOCK(bo); 13260 msleep(&mp->mnt_secondary_writes, 13261 MNT_MTX(mp), 13262 (PUSER - 1) | PDROP, "secwr", 0); 13263 BO_LOCK(bo); 13264 continue; 13265 } 13266 break; 13267 } 13268 13269 /* 13270 * Reasons for needing more work before suspend: 13271 * - Dirty buffers on devvp. 13272 * - Softdep activity occurred after start of vnode sync loop 13273 * - Secondary writes occurred after start of vnode sync loop 13274 */ 13275 error = 0; 13276 if (bo->bo_numoutput > 0 || 13277 bo->bo_dirty.bv_cnt > 0 || 13278 softdep_deps != 0 || 13279 ump->softdep_deps != 0 || 13280 softdep_accdeps != ump->softdep_accdeps || 13281 secondary_writes != 0 || 13282 mp->mnt_secondary_writes != 0 || 13283 secondary_accwrites != mp->mnt_secondary_accwrites) 13284 error = EAGAIN; 13285 FREE_LOCK(&lk); 13286 BO_UNLOCK(bo); 13287 return (error); 13288} 13289 13290 13291/* 13292 * Get the number of dependency structures for the file system, both 13293 * the current number and the total number allocated. These will 13294 * later be used to detect that softdep processing has occurred. 13295 */ 13296void 13297softdep_get_depcounts(struct mount *mp, 13298 int *softdep_depsp, 13299 int *softdep_accdepsp) 13300{ 13301 struct ufsmount *ump; 13302 13303 ump = VFSTOUFS(mp); 13304 ACQUIRE_LOCK(&lk); 13305 *softdep_depsp = ump->softdep_deps; 13306 *softdep_accdepsp = ump->softdep_accdeps; 13307 FREE_LOCK(&lk); 13308} 13309 13310/* 13311 * Wait for pending output on a vnode to complete. 13312 * Must be called with vnode lock and interlock locked. 13313 * 13314 * XXX: Should just be a call to bufobj_wwait(). 13315 */ 13316static void 13317drain_output(vp) 13318 struct vnode *vp; 13319{ 13320 struct bufobj *bo; 13321 13322 bo = &vp->v_bufobj; 13323 ASSERT_VOP_LOCKED(vp, "drain_output"); 13324 ASSERT_BO_LOCKED(bo); 13325 13326 while (bo->bo_numoutput) { 13327 bo->bo_flag |= BO_WWAIT; 13328 msleep((caddr_t)&bo->bo_numoutput, 13329 BO_MTX(bo), PRIBIO + 1, "drainvp", 0); 13330 } 13331} 13332 13333/* 13334 * Called whenever a buffer that is being invalidated or reallocated 13335 * contains dependencies. This should only happen if an I/O error has 13336 * occurred. The routine is called with the buffer locked. 13337 */ 13338static void 13339softdep_deallocate_dependencies(bp) 13340 struct buf *bp; 13341{ 13342 13343 if ((bp->b_ioflags & BIO_ERROR) == 0) 13344 panic("softdep_deallocate_dependencies: dangling deps"); 13345 if (bp->b_vp != NULL && bp->b_vp->v_mount != NULL) 13346 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); 13347 else 13348 printf("softdep_deallocate_dependencies: " 13349 "got error %d while accessing filesystem\n", bp->b_error); 13350 if (bp->b_error != ENXIO) 13351 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 13352} 13353 13354/* 13355 * Function to handle asynchronous write errors in the filesystem. 13356 */ 13357static void 13358softdep_error(func, error) 13359 char *func; 13360 int error; 13361{ 13362 13363 /* XXX should do something better! */ 13364 printf("%s: got error %d while accessing filesystem\n", func, error); 13365} 13366 13367#ifdef DDB 13368 13369static void 13370inodedep_print(struct inodedep *inodedep, int verbose) 13371{ 13372 db_printf("%p fs %p st %x ino %jd inoblk %jd delta %d nlink %d" 13373 " saveino %p\n", 13374 inodedep, inodedep->id_fs, inodedep->id_state, 13375 (intmax_t)inodedep->id_ino, 13376 (intmax_t)fsbtodb(inodedep->id_fs, 13377 ino_to_fsba(inodedep->id_fs, inodedep->id_ino)), 13378 inodedep->id_nlinkdelta, inodedep->id_savednlink, 13379 inodedep->id_savedino1); 13380 13381 if (verbose == 0) 13382 return; 13383 13384 db_printf("\tpendinghd %p, bufwait %p, inowait %p, inoreflst %p, " 13385 "mkdiradd %p\n", 13386 LIST_FIRST(&inodedep->id_pendinghd), 13387 LIST_FIRST(&inodedep->id_bufwait), 13388 LIST_FIRST(&inodedep->id_inowait), 13389 TAILQ_FIRST(&inodedep->id_inoreflst), 13390 inodedep->id_mkdiradd); 13391 db_printf("\tinoupdt %p, newinoupdt %p, extupdt %p, newextupdt %p\n", 13392 TAILQ_FIRST(&inodedep->id_inoupdt), 13393 TAILQ_FIRST(&inodedep->id_newinoupdt), 13394 TAILQ_FIRST(&inodedep->id_extupdt), 13395 TAILQ_FIRST(&inodedep->id_newextupdt)); 13396} 13397 13398DB_SHOW_COMMAND(inodedep, db_show_inodedep) 13399{ 13400 13401 if (have_addr == 0) { 13402 db_printf("Address required\n"); 13403 return; 13404 } 13405 inodedep_print((struct inodedep*)addr, 1); 13406} 13407 13408DB_SHOW_COMMAND(inodedeps, db_show_inodedeps) 13409{ 13410 struct inodedep_hashhead *inodedephd; 13411 struct inodedep *inodedep; 13412 struct fs *fs; 13413 int cnt; 13414 13415 fs = have_addr ? (struct fs *)addr : NULL; 13416 for (cnt = 0; cnt < inodedep_hash; cnt++) { 13417 inodedephd = &inodedep_hashtbl[cnt]; 13418 LIST_FOREACH(inodedep, inodedephd, id_hash) { 13419 if (fs != NULL && fs != inodedep->id_fs) 13420 continue; 13421 inodedep_print(inodedep, 0); 13422 } 13423 } 13424} 13425 13426DB_SHOW_COMMAND(worklist, db_show_worklist) 13427{ 13428 struct worklist *wk; 13429 13430 if (have_addr == 0) { 13431 db_printf("Address required\n"); 13432 return; 13433 } 13434 wk = (struct worklist *)addr; 13435 printf("worklist: %p type %s state 0x%X\n", 13436 wk, TYPENAME(wk->wk_type), wk->wk_state); 13437} 13438 13439DB_SHOW_COMMAND(workhead, db_show_workhead) 13440{ 13441 struct workhead *wkhd; 13442 struct worklist *wk; 13443 int i; 13444 13445 if (have_addr == 0) { 13446 db_printf("Address required\n"); 13447 return; 13448 } 13449 wkhd = (struct workhead *)addr; 13450 wk = LIST_FIRST(wkhd); 13451 for (i = 0; i < 100 && wk != NULL; i++, wk = LIST_NEXT(wk, wk_list)) 13452 db_printf("worklist: %p type %s state 0x%X", 13453 wk, TYPENAME(wk->wk_type), wk->wk_state); 13454 if (i == 100) 13455 db_printf("workhead overflow"); 13456 printf("\n"); 13457} 13458 13459 13460DB_SHOW_COMMAND(mkdirs, db_show_mkdirs) 13461{ 13462 struct jaddref *jaddref; 13463 struct diradd *diradd; 13464 struct mkdir *mkdir; 13465 13466 LIST_FOREACH(mkdir, &mkdirlisthd, md_mkdirs) { 13467 diradd = mkdir->md_diradd; 13468 db_printf("mkdir: %p state 0x%X dap %p state 0x%X", 13469 mkdir, mkdir->md_state, diradd, diradd->da_state); 13470 if ((jaddref = mkdir->md_jaddref) != NULL) 13471 db_printf(" jaddref %p jaddref state 0x%X", 13472 jaddref, jaddref->ja_state); 13473 db_printf("\n"); 13474 } 13475} 13476 13477#endif /* DDB */ 13478 13479#endif /* SOFTUPDATES */ 13480