ffs_softdep.c revision 144585
1219820Sjeff/*- 2219820Sjeff * Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved. 3219820Sjeff * 4219820Sjeff * The soft updates code is derived from the appendix of a University 5219820Sjeff * of Michigan technical report (Gregory R. Ganger and Yale N. Patt, 6219820Sjeff * "Soft Updates: A Solution to the Metadata Update Problem in File 7219820Sjeff * Systems", CSE-TR-254-95, August 1995). 8 * 9 * Further information about soft updates can be obtained from: 10 * 11 * Marshall Kirk McKusick http://www.mckusick.com/softdep/ 12 * 1614 Oxford Street mckusick@mckusick.com 13 * Berkeley, CA 94709-1608 +1-510-843-9542 14 * USA 15 * 16 * Redistribution and use in source and binary forms, with or without 17 * modification, are permitted provided that the following conditions 18 * are met: 19 * 20 * 1. Redistributions of source code must retain the above copyright 21 * notice, this list of conditions and the following disclaimer. 22 * 2. Redistributions in binary form must reproduce the above copyright 23 * notice, this list of conditions and the following disclaimer in the 24 * documentation and/or other materials provided with the distribution. 25 * 26 * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY 27 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 28 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 29 * DISCLAIMED. IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR 30 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * from: @(#)ffs_softdep.c 9.59 (McKusick) 6/21/00 39 */ 40 41#include <sys/cdefs.h> 42__FBSDID("$FreeBSD: head/sys/ufs/ffs/ffs_softdep.c 144585 2005-04-03 10:29:55Z jeff $"); 43 44/* 45 * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide. 46 */ 47#ifndef DIAGNOSTIC 48#define DIAGNOSTIC 49#endif 50#ifndef DEBUG 51#define DEBUG 52#endif 53 54#include <sys/param.h> 55#include <sys/kernel.h> 56#include <sys/systm.h> 57#include <sys/bio.h> 58#include <sys/buf.h> 59#include <sys/kdb.h> 60#include <sys/lock.h> 61#include <sys/malloc.h> 62#include <sys/mount.h> 63#include <sys/mutex.h> 64#include <sys/proc.h> 65#include <sys/stat.h> 66#include <sys/syslog.h> 67#include <sys/vnode.h> 68#include <sys/conf.h> 69#include <ufs/ufs/dir.h> 70#include <ufs/ufs/extattr.h> 71#include <ufs/ufs/quota.h> 72#include <ufs/ufs/inode.h> 73#include <ufs/ufs/ufsmount.h> 74#include <ufs/ffs/fs.h> 75#include <ufs/ffs/softdep.h> 76#include <ufs/ffs/ffs_extern.h> 77#include <ufs/ufs/ufs_extern.h> 78 79#include "opt_ffs.h" 80 81#ifndef SOFTUPDATES 82 83int 84softdep_flushfiles(oldmnt, flags, td) 85 struct mount *oldmnt; 86 int flags; 87 struct thread *td; 88{ 89 90 panic("softdep_flushfiles called"); 91} 92 93int 94softdep_mount(devvp, mp, fs, cred) 95 struct vnode *devvp; 96 struct mount *mp; 97 struct fs *fs; 98 struct ucred *cred; 99{ 100 101 return (0); 102} 103 104void 105softdep_initialize() 106{ 107 108 return; 109} 110 111void 112softdep_uninitialize() 113{ 114 115 return; 116} 117 118void 119softdep_setup_inomapdep(bp, ip, newinum) 120 struct buf *bp; 121 struct inode *ip; 122 ino_t newinum; 123{ 124 125 panic("softdep_setup_inomapdep called"); 126} 127 128void 129softdep_setup_blkmapdep(bp, fs, newblkno) 130 struct buf *bp; 131 struct fs *fs; 132 ufs2_daddr_t newblkno; 133{ 134 135 panic("softdep_setup_blkmapdep called"); 136} 137 138void 139softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 140 struct inode *ip; 141 ufs_lbn_t lbn; 142 ufs2_daddr_t newblkno; 143 ufs2_daddr_t oldblkno; 144 long newsize; 145 long oldsize; 146 struct buf *bp; 147{ 148 149 panic("softdep_setup_allocdirect called"); 150} 151 152void 153softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 154 struct inode *ip; 155 ufs_lbn_t lbn; 156 ufs2_daddr_t newblkno; 157 ufs2_daddr_t oldblkno; 158 long newsize; 159 long oldsize; 160 struct buf *bp; 161{ 162 163 panic("softdep_setup_allocdirect called"); 164} 165 166void 167softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 168 struct inode *ip; 169 ufs_lbn_t lbn; 170 struct buf *bp; 171 int ptrno; 172 ufs2_daddr_t newblkno; 173 ufs2_daddr_t oldblkno; 174 struct buf *nbp; 175{ 176 177 panic("softdep_setup_allocindir_page called"); 178} 179 180void 181softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 182 struct buf *nbp; 183 struct inode *ip; 184 struct buf *bp; 185 int ptrno; 186 ufs2_daddr_t newblkno; 187{ 188 189 panic("softdep_setup_allocindir_meta called"); 190} 191 192void 193softdep_setup_freeblocks(ip, length, flags) 194 struct inode *ip; 195 off_t length; 196 int flags; 197{ 198 199 panic("softdep_setup_freeblocks called"); 200} 201 202void 203softdep_freefile(pvp, ino, mode) 204 struct vnode *pvp; 205 ino_t ino; 206 int mode; 207{ 208 209 panic("softdep_freefile called"); 210} 211 212int 213softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 214 struct buf *bp; 215 struct inode *dp; 216 off_t diroffset; 217 ino_t newinum; 218 struct buf *newdirbp; 219 int isnewblk; 220{ 221 222 panic("softdep_setup_directory_add called"); 223} 224 225void 226softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize) 227 struct inode *dp; 228 caddr_t base; 229 caddr_t oldloc; 230 caddr_t newloc; 231 int entrysize; 232{ 233 234 panic("softdep_change_directoryentry_offset called"); 235} 236 237void 238softdep_setup_remove(bp, dp, ip, isrmdir) 239 struct buf *bp; 240 struct inode *dp; 241 struct inode *ip; 242 int isrmdir; 243{ 244 245 panic("softdep_setup_remove called"); 246} 247 248void 249softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 250 struct buf *bp; 251 struct inode *dp; 252 struct inode *ip; 253 ino_t newinum; 254 int isrmdir; 255{ 256 257 panic("softdep_setup_directory_change called"); 258} 259 260void 261softdep_change_linkcnt(ip) 262 struct inode *ip; 263{ 264 265 panic("softdep_change_linkcnt called"); 266} 267 268void 269softdep_load_inodeblock(ip) 270 struct inode *ip; 271{ 272 273 panic("softdep_load_inodeblock called"); 274} 275 276void 277softdep_update_inodeblock(ip, bp, waitfor) 278 struct inode *ip; 279 struct buf *bp; 280 int waitfor; 281{ 282 283 panic("softdep_update_inodeblock called"); 284} 285 286int 287softdep_fsync(vp) 288 struct vnode *vp; /* the "in_core" copy of the inode */ 289{ 290 291 return (0); 292} 293 294void 295softdep_fsync_mountdev(vp) 296 struct vnode *vp; 297{ 298 299 return; 300} 301 302int 303softdep_flushworklist(oldmnt, countp, td) 304 struct mount *oldmnt; 305 int *countp; 306 struct thread *td; 307{ 308 309 *countp = 0; 310 return (0); 311} 312 313int 314softdep_sync_metadata(struct vnode *vp) 315{ 316 317 return (0); 318} 319 320int 321softdep_slowdown(vp) 322 struct vnode *vp; 323{ 324 325 panic("softdep_slowdown called"); 326} 327 328void 329softdep_releasefile(ip) 330 struct inode *ip; /* inode with the zero effective link count */ 331{ 332 333 panic("softdep_releasefile called"); 334} 335 336int 337softdep_request_cleanup(fs, vp) 338 struct fs *fs; 339 struct vnode *vp; 340{ 341 342 return (0); 343} 344 345#else 346/* 347 * These definitions need to be adapted to the system to which 348 * this file is being ported. 349 */ 350/* 351 * malloc types defined for the softdep system. 352 */ 353static MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies"); 354static MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies"); 355static MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation"); 356static MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map"); 357static MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode"); 358static MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies"); 359static MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block"); 360static MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode"); 361static MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode"); 362static MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated"); 363static MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry"); 364static MALLOC_DEFINE(M_MKDIR, "mkdir","New directory"); 365static MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted"); 366static MALLOC_DEFINE(M_NEWDIRBLK, "newdirblk","Unclaimed new directory block"); 367static MALLOC_DEFINE(M_SAVEDINO, "savedino","Saved inodes"); 368 369#define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE) 370 371#define D_PAGEDEP 0 372#define D_INODEDEP 1 373#define D_NEWBLK 2 374#define D_BMSAFEMAP 3 375#define D_ALLOCDIRECT 4 376#define D_INDIRDEP 5 377#define D_ALLOCINDIR 6 378#define D_FREEFRAG 7 379#define D_FREEBLKS 8 380#define D_FREEFILE 9 381#define D_DIRADD 10 382#define D_MKDIR 11 383#define D_DIRREM 12 384#define D_NEWDIRBLK 13 385#define D_LAST D_NEWDIRBLK 386 387/* 388 * translate from workitem type to memory type 389 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 390 */ 391static struct malloc_type *memtype[] = { 392 M_PAGEDEP, 393 M_INODEDEP, 394 M_NEWBLK, 395 M_BMSAFEMAP, 396 M_ALLOCDIRECT, 397 M_INDIRDEP, 398 M_ALLOCINDIR, 399 M_FREEFRAG, 400 M_FREEBLKS, 401 M_FREEFILE, 402 M_DIRADD, 403 M_MKDIR, 404 M_DIRREM, 405 M_NEWDIRBLK 406}; 407 408#define DtoM(type) (memtype[type]) 409 410/* 411 * Names of malloc types. 412 */ 413#define TYPENAME(type) \ 414 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???") 415/* 416 * End system adaptaion definitions. 417 */ 418 419/* 420 * Forward declarations. 421 */ 422struct inodedep_hashhead; 423struct newblk_hashhead; 424struct pagedep_hashhead; 425 426/* 427 * Internal function prototypes. 428 */ 429static void softdep_error(char *, int); 430static void drain_output(struct vnode *); 431static struct buf *getdirtybuf(struct buf *, struct mtx *, int); 432static void clear_remove(struct thread *); 433static void clear_inodedeps(struct thread *); 434static int flush_pagedep_deps(struct vnode *, struct mount *, 435 struct diraddhd *); 436static int flush_inodedep_deps(struct fs *, ino_t); 437static int flush_deplist(struct allocdirectlst *, int, int *); 438static int handle_written_filepage(struct pagedep *, struct buf *); 439static void diradd_inode_written(struct diradd *, struct inodedep *); 440static int handle_written_inodeblock(struct inodedep *, struct buf *); 441static void handle_allocdirect_partdone(struct allocdirect *); 442static void handle_allocindir_partdone(struct allocindir *); 443static void initiate_write_filepage(struct pagedep *, struct buf *); 444static void handle_written_mkdir(struct mkdir *, int); 445static void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *); 446static void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *); 447static void handle_workitem_freefile(struct freefile *); 448static void handle_workitem_remove(struct dirrem *, struct vnode *); 449static struct dirrem *newdirrem(struct buf *, struct inode *, 450 struct inode *, int, struct dirrem **); 451static void free_diradd(struct diradd *); 452static void free_allocindir(struct allocindir *, struct inodedep *); 453static void free_newdirblk(struct newdirblk *); 454static int indir_trunc(struct freeblks *, ufs2_daddr_t, int, ufs_lbn_t, 455 ufs2_daddr_t *); 456static void deallocate_dependencies(struct buf *, struct inodedep *); 457static void free_allocdirect(struct allocdirectlst *, 458 struct allocdirect *, int); 459static int check_inode_unwritten(struct inodedep *); 460static int free_inodedep(struct inodedep *); 461static void handle_workitem_freeblocks(struct freeblks *, int); 462static void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *); 463static void setup_allocindir_phase2(struct buf *, struct inode *, 464 struct allocindir *); 465static struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t, 466 ufs2_daddr_t); 467static void handle_workitem_freefrag(struct freefrag *); 468static struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long); 469static void allocdirect_merge(struct allocdirectlst *, 470 struct allocdirect *, struct allocdirect *); 471static struct bmsafemap *bmsafemap_lookup(struct buf *); 472static int newblk_find(struct newblk_hashhead *, struct fs *, ufs2_daddr_t, 473 struct newblk **); 474static int newblk_lookup(struct fs *, ufs2_daddr_t, int, struct newblk **); 475static int inodedep_find(struct inodedep_hashhead *, struct fs *, ino_t, 476 struct inodedep **); 477static int inodedep_lookup(struct fs *, ino_t, int, struct inodedep **); 478static int pagedep_lookup(struct inode *, ufs_lbn_t, int, struct pagedep **); 479static int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t, 480 struct mount *mp, int, struct pagedep **); 481static void pause_timer(void *); 482static int request_cleanup(int); 483static int process_worklist_item(struct mount *, int); 484static void add_to_worklist(struct worklist *); 485 486/* 487 * Exported softdep operations. 488 */ 489static void softdep_disk_io_initiation(struct buf *); 490static void softdep_disk_write_complete(struct buf *); 491static void softdep_deallocate_dependencies(struct buf *); 492static int softdep_count_dependencies(struct buf *bp, int); 493 494static struct mtx lk; 495MTX_SYSINIT(softdep_lock, &lk, "Softdep Lock", MTX_DEF); 496 497#define ACQUIRE_LOCK(lk) mtx_lock(lk) 498#define FREE_LOCK(lk) mtx_unlock(lk) 499 500/* 501 * Worklist queue management. 502 * These routines require that the lock be held. 503 */ 504#ifndef /* NOT */ DEBUG 505#define WORKLIST_INSERT(head, item) do { \ 506 (item)->wk_state |= ONWORKLIST; \ 507 LIST_INSERT_HEAD(head, item, wk_list); \ 508} while (0) 509#define WORKLIST_REMOVE(item) do { \ 510 (item)->wk_state &= ~ONWORKLIST; \ 511 LIST_REMOVE(item, wk_list); \ 512} while (0) 513#define WORKITEM_FREE(item, type) FREE(item, DtoM(type)) 514 515#else /* DEBUG */ 516static void worklist_insert(struct workhead *, struct worklist *); 517static void worklist_remove(struct worklist *); 518static void workitem_free(struct worklist *, int); 519 520#define WORKLIST_INSERT(head, item) worklist_insert(head, item) 521#define WORKLIST_REMOVE(item) worklist_remove(item) 522#define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type) 523 524static void 525worklist_insert(head, item) 526 struct workhead *head; 527 struct worklist *item; 528{ 529 530 mtx_assert(&lk, MA_OWNED); 531 if (item->wk_state & ONWORKLIST) 532 panic("worklist_insert: already on list"); 533 item->wk_state |= ONWORKLIST; 534 LIST_INSERT_HEAD(head, item, wk_list); 535} 536 537static void 538worklist_remove(item) 539 struct worklist *item; 540{ 541 542 mtx_assert(&lk, MA_OWNED); 543 if ((item->wk_state & ONWORKLIST) == 0) 544 panic("worklist_remove: not on list"); 545 item->wk_state &= ~ONWORKLIST; 546 LIST_REMOVE(item, wk_list); 547} 548 549static void 550workitem_free(item, type) 551 struct worklist *item; 552 int type; 553{ 554 555 if (item->wk_state & ONWORKLIST) 556 panic("workitem_free: still on list"); 557 if (item->wk_type != type) 558 panic("workitem_free: type mismatch"); 559 FREE(item, DtoM(type)); 560} 561#endif /* DEBUG */ 562 563/* 564 * Workitem queue management 565 */ 566static struct workhead softdep_workitem_pending; 567static struct worklist *worklist_tail; 568static int num_on_worklist; /* number of worklist items to be processed */ 569static int softdep_worklist_busy; /* 1 => trying to do unmount */ 570static int softdep_worklist_req; /* serialized waiters */ 571static int max_softdeps; /* maximum number of structs before slowdown */ 572static int maxindirdeps = 50; /* max number of indirdeps before slowdown */ 573static int tickdelay = 2; /* number of ticks to pause during slowdown */ 574static int proc_waiting; /* tracks whether we have a timeout posted */ 575static int *stat_countp; /* statistic to count in proc_waiting timeout */ 576static struct callout_handle handle; /* handle on posted proc_waiting timeout */ 577static struct thread *filesys_syncer; /* proc of filesystem syncer process */ 578static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 579#define FLUSH_INODES 1 580static int req_clear_remove; /* syncer process flush some freeblks */ 581#define FLUSH_REMOVE 2 582#define FLUSH_REMOVE_WAIT 3 583/* 584 * runtime statistics 585 */ 586static int stat_worklist_push; /* number of worklist cleanups */ 587static int stat_blk_limit_push; /* number of times block limit neared */ 588static int stat_ino_limit_push; /* number of times inode limit neared */ 589static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 590static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 591static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 592static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 593static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 594static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 595static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 596#ifdef DEBUG 597#include <vm/vm.h> 598#include <sys/sysctl.h> 599SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, ""); 600SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, ""); 601SYSCTL_INT(_debug, OID_AUTO, maxindirdeps, CTLFLAG_RW, &maxindirdeps, 0, ""); 602SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,""); 603SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,""); 604SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,""); 605SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, ""); 606SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, ""); 607SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, ""); 608SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, ""); 609SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, ""); 610SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, ""); 611SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, ""); 612#endif /* DEBUG */ 613 614SYSCTL_DECL(_vfs_ffs); 615 616static int compute_summary_at_mount = 0; /* Whether to recompute the summary at mount time */ 617SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW, 618 &compute_summary_at_mount, 0, "Recompute summary at mount"); 619 620/* 621 * Add an item to the end of the work queue. 622 * This routine requires that the lock be held. 623 * This is the only routine that adds items to the list. 624 * The following routine is the only one that removes items 625 * and does so in order from first to last. 626 */ 627static void 628add_to_worklist(wk) 629 struct worklist *wk; 630{ 631 632 mtx_assert(&lk, MA_OWNED); 633 if (wk->wk_state & ONWORKLIST) 634 panic("add_to_worklist: already on list"); 635 wk->wk_state |= ONWORKLIST; 636 if (LIST_FIRST(&softdep_workitem_pending) == NULL) 637 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list); 638 else 639 LIST_INSERT_AFTER(worklist_tail, wk, wk_list); 640 worklist_tail = wk; 641 num_on_worklist += 1; 642} 643 644/* 645 * Process that runs once per second to handle items in the background queue. 646 * 647 * Note that we ensure that everything is done in the order in which they 648 * appear in the queue. The code below depends on this property to ensure 649 * that blocks of a file are freed before the inode itself is freed. This 650 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 651 * until all the old ones have been purged from the dependency lists. 652 */ 653int 654softdep_process_worklist(matchmnt) 655 struct mount *matchmnt; 656{ 657 struct thread *td = curthread; 658 int cnt, matchcnt, loopcount; 659 long starttime; 660 661 /* 662 * Record the process identifier of our caller so that we can give 663 * this process preferential treatment in request_cleanup below. 664 */ 665 filesys_syncer = td; 666 matchcnt = 0; 667 668 /* 669 * There is no danger of having multiple processes run this 670 * code, but we have to single-thread it when softdep_flushfiles() 671 * is in operation to get an accurate count of the number of items 672 * related to its mount point that are in the list. 673 */ 674 ACQUIRE_LOCK(&lk); 675 if (matchmnt == NULL) { 676 if (softdep_worklist_busy < 0) { 677 FREE_LOCK(&lk); 678 return(-1); 679 } 680 softdep_worklist_busy += 1; 681 } 682 683 /* 684 * If requested, try removing inode or removal dependencies. 685 */ 686 if (req_clear_inodedeps) { 687 clear_inodedeps(td); 688 req_clear_inodedeps -= 1; 689 wakeup_one(&proc_waiting); 690 } 691 if (req_clear_remove) { 692 clear_remove(td); 693 req_clear_remove -= 1; 694 wakeup_one(&proc_waiting); 695 } 696 loopcount = 1; 697 starttime = time_second; 698 while (num_on_worklist > 0) { 699 if ((cnt = process_worklist_item(matchmnt, 0)) == -1) 700 break; 701 else 702 matchcnt += cnt; 703 704 /* 705 * If a umount operation wants to run the worklist 706 * accurately, abort. 707 */ 708 if (softdep_worklist_req && matchmnt == NULL) { 709 matchcnt = -1; 710 break; 711 } 712 713 /* 714 * If requested, try removing inode or removal dependencies. 715 */ 716 if (req_clear_inodedeps) { 717 clear_inodedeps(td); 718 req_clear_inodedeps -= 1; 719 wakeup_one(&proc_waiting); 720 } 721 if (req_clear_remove) { 722 clear_remove(td); 723 req_clear_remove -= 1; 724 wakeup_one(&proc_waiting); 725 } 726 /* 727 * We do not generally want to stop for buffer space, but if 728 * we are really being a buffer hog, we will stop and wait. 729 */ 730 if (loopcount++ % 128 == 0) { 731 FREE_LOCK(&lk); 732 bwillwrite(); 733 ACQUIRE_LOCK(&lk); 734 } 735 /* 736 * Never allow processing to run for more than one 737 * second. Otherwise the other syncer tasks may get 738 * excessively backlogged. 739 */ 740 if (starttime != time_second && matchmnt == NULL) { 741 matchcnt = -1; 742 break; 743 } 744 } 745 if (matchmnt == NULL) { 746 softdep_worklist_busy -= 1; 747 if (softdep_worklist_req && softdep_worklist_busy == 0) 748 wakeup(&softdep_worklist_req); 749 } 750 FREE_LOCK(&lk); 751 return (matchcnt); 752} 753 754/* 755 * Process one item on the worklist. 756 */ 757static int 758process_worklist_item(matchmnt, flags) 759 struct mount *matchmnt; 760 int flags; 761{ 762 struct worklist *wk, *wkend; 763 struct mount *mp; 764 struct vnode *vp; 765 int matchcnt = 0; 766 767 mtx_assert(&lk, MA_OWNED); 768 /* 769 * If we are being called because of a process doing a 770 * copy-on-write, then it is not safe to write as we may 771 * recurse into the copy-on-write routine. 772 */ 773 if (curthread->td_pflags & TDP_COWINPROGRESS) 774 return (-1); 775 /* 776 * Normally we just process each item on the worklist in order. 777 * However, if we are in a situation where we cannot lock any 778 * inodes, we have to skip over any dirrem requests whose 779 * vnodes are resident and locked. 780 */ 781 vp = NULL; 782 LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) { 783 if (wk->wk_state & INPROGRESS) 784 continue; 785 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM) 786 break; 787 wk->wk_state |= INPROGRESS; 788 FREE_LOCK(&lk); 789 ffs_vget(WK_DIRREM(wk)->dm_mnt, WK_DIRREM(wk)->dm_oldinum, 790 LK_NOWAIT | LK_EXCLUSIVE, &vp); 791 ACQUIRE_LOCK(&lk); 792 wk->wk_state &= ~INPROGRESS; 793 if (vp != NULL) 794 break; 795 } 796 if (wk == 0) 797 return (-1); 798 /* 799 * Remove the item to be processed. If we are removing the last 800 * item on the list, we need to recalculate the tail pointer. 801 * As this happens rarely and usually when the list is short, 802 * we just run down the list to find it rather than tracking it 803 * in the above loop. 804 */ 805 WORKLIST_REMOVE(wk); 806 if (wk == worklist_tail) { 807 LIST_FOREACH(wkend, &softdep_workitem_pending, wk_list) 808 if (LIST_NEXT(wkend, wk_list) == NULL) 809 break; 810 worklist_tail = wkend; 811 } 812 num_on_worklist -= 1; 813 FREE_LOCK(&lk); 814 switch (wk->wk_type) { 815 816 case D_DIRREM: 817 /* removal of a directory entry */ 818 mp = WK_DIRREM(wk)->dm_mnt; 819 if (vn_write_suspend_wait(NULL, mp, V_NOWAIT)) 820 panic("%s: dirrem on suspended filesystem", 821 "process_worklist_item"); 822 if (mp == matchmnt) 823 matchcnt += 1; 824 handle_workitem_remove(WK_DIRREM(wk), vp); 825 break; 826 827 case D_FREEBLKS: 828 /* releasing blocks and/or fragments from a file */ 829 mp = WK_FREEBLKS(wk)->fb_mnt; 830 if (vn_write_suspend_wait(NULL, mp, V_NOWAIT)) 831 panic("%s: freeblks on suspended filesystem", 832 "process_worklist_item"); 833 if (mp == matchmnt) 834 matchcnt += 1; 835 handle_workitem_freeblocks(WK_FREEBLKS(wk), flags & LK_NOWAIT); 836 break; 837 838 case D_FREEFRAG: 839 /* releasing a fragment when replaced as a file grows */ 840 mp = WK_FREEFRAG(wk)->ff_mnt; 841 if (vn_write_suspend_wait(NULL, mp, V_NOWAIT)) 842 panic("%s: freefrag on suspended filesystem", 843 "process_worklist_item"); 844 if (mp == matchmnt) 845 matchcnt += 1; 846 handle_workitem_freefrag(WK_FREEFRAG(wk)); 847 break; 848 849 case D_FREEFILE: 850 /* releasing an inode when its link count drops to 0 */ 851 mp = WK_FREEFILE(wk)->fx_mnt; 852 if (vn_write_suspend_wait(NULL, mp, V_NOWAIT)) 853 panic("%s: freefile on suspended filesystem", 854 "process_worklist_item"); 855 if (mp == matchmnt) 856 matchcnt += 1; 857 handle_workitem_freefile(WK_FREEFILE(wk)); 858 break; 859 860 default: 861 panic("%s_process_worklist: Unknown type %s", 862 "softdep", TYPENAME(wk->wk_type)); 863 /* NOTREACHED */ 864 } 865 ACQUIRE_LOCK(&lk); 866 return (matchcnt); 867} 868 869/* 870 * Move dependencies from one buffer to another. 871 */ 872void 873softdep_move_dependencies(oldbp, newbp) 874 struct buf *oldbp; 875 struct buf *newbp; 876{ 877 struct worklist *wk, *wktail; 878 879 if (LIST_FIRST(&newbp->b_dep) != NULL) 880 panic("softdep_move_dependencies: need merge code"); 881 wktail = 0; 882 ACQUIRE_LOCK(&lk); 883 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 884 LIST_REMOVE(wk, wk_list); 885 if (wktail == 0) 886 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 887 else 888 LIST_INSERT_AFTER(wktail, wk, wk_list); 889 wktail = wk; 890 } 891 FREE_LOCK(&lk); 892} 893 894/* 895 * Purge the work list of all items associated with a particular mount point. 896 */ 897int 898softdep_flushworklist(oldmnt, countp, td) 899 struct mount *oldmnt; 900 int *countp; 901 struct thread *td; 902{ 903 struct vnode *devvp; 904 int count, error = 0; 905 906 /* 907 * Await our turn to clear out the queue, then serialize access. 908 */ 909 ACQUIRE_LOCK(&lk); 910 while (softdep_worklist_busy) { 911 softdep_worklist_req += 1; 912 msleep(&softdep_worklist_req, &lk, PRIBIO, "softflush", 0); 913 softdep_worklist_req -= 1; 914 } 915 softdep_worklist_busy = -1; 916 FREE_LOCK(&lk); 917 /* 918 * Alternately flush the block device associated with the mount 919 * point and process any dependencies that the flushing 920 * creates. We continue until no more worklist dependencies 921 * are found. 922 */ 923 *countp = 0; 924 devvp = VFSTOUFS(oldmnt)->um_devvp; 925 while ((count = softdep_process_worklist(oldmnt)) > 0) { 926 *countp += count; 927 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, td); 928 error = VOP_FSYNC(devvp, MNT_WAIT, td); 929 VOP_UNLOCK(devvp, 0, td); 930 if (error) 931 break; 932 } 933 ACQUIRE_LOCK(&lk); 934 softdep_worklist_busy = 0; 935 if (softdep_worklist_req) 936 wakeup(&softdep_worklist_req); 937 FREE_LOCK(&lk); 938 return (error); 939} 940 941/* 942 * Flush all vnodes and worklist items associated with a specified mount point. 943 */ 944int 945softdep_flushfiles(oldmnt, flags, td) 946 struct mount *oldmnt; 947 int flags; 948 struct thread *td; 949{ 950 int error, count, loopcnt; 951 952 error = 0; 953 954 /* 955 * Alternately flush the vnodes associated with the mount 956 * point and process any dependencies that the flushing 957 * creates. In theory, this loop can happen at most twice, 958 * but we give it a few extra just to be sure. 959 */ 960 for (loopcnt = 10; loopcnt > 0; loopcnt--) { 961 /* 962 * Do another flush in case any vnodes were brought in 963 * as part of the cleanup operations. 964 */ 965 if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0) 966 break; 967 if ((error = softdep_flushworklist(oldmnt, &count, td)) != 0 || 968 count == 0) 969 break; 970 } 971 /* 972 * If we are unmounting then it is an error to fail. If we 973 * are simply trying to downgrade to read-only, then filesystem 974 * activity can keep us busy forever, so we just fail with EBUSY. 975 */ 976 if (loopcnt == 0) { 977 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 978 panic("softdep_flushfiles: looping"); 979 error = EBUSY; 980 } 981 return (error); 982} 983 984/* 985 * Structure hashing. 986 * 987 * There are three types of structures that can be looked up: 988 * 1) pagedep structures identified by mount point, inode number, 989 * and logical block. 990 * 2) inodedep structures identified by mount point and inode number. 991 * 3) newblk structures identified by mount point and 992 * physical block number. 993 * 994 * The "pagedep" and "inodedep" dependency structures are hashed 995 * separately from the file blocks and inodes to which they correspond. 996 * This separation helps when the in-memory copy of an inode or 997 * file block must be replaced. It also obviates the need to access 998 * an inode or file page when simply updating (or de-allocating) 999 * dependency structures. Lookup of newblk structures is needed to 1000 * find newly allocated blocks when trying to associate them with 1001 * their allocdirect or allocindir structure. 1002 * 1003 * The lookup routines optionally create and hash a new instance when 1004 * an existing entry is not found. 1005 */ 1006#define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 1007#define NODELAY 0x0002 /* cannot do background work */ 1008 1009/* 1010 * Structures and routines associated with pagedep caching. 1011 */ 1012LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl; 1013u_long pagedep_hash; /* size of hash table - 1 */ 1014#define PAGEDEP_HASH(mp, inum, lbn) \ 1015 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \ 1016 pagedep_hash]) 1017 1018static int 1019pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp) 1020 struct pagedep_hashhead *pagedephd; 1021 ino_t ino; 1022 ufs_lbn_t lbn; 1023 struct mount *mp; 1024 int flags; 1025 struct pagedep **pagedeppp; 1026{ 1027 struct pagedep *pagedep; 1028 1029 LIST_FOREACH(pagedep, pagedephd, pd_hash) 1030 if (ino == pagedep->pd_ino && 1031 lbn == pagedep->pd_lbn && 1032 mp == pagedep->pd_mnt) 1033 break; 1034 if (pagedep) { 1035 *pagedeppp = pagedep; 1036 if ((flags & DEPALLOC) != 0 && 1037 (pagedep->pd_state & ONWORKLIST) == 0) 1038 return (0); 1039 return (1); 1040 } 1041 *pagedeppp = NULL; 1042 return (0); 1043} 1044/* 1045 * Look up a pagedep. Return 1 if found, 0 if not found or found 1046 * when asked to allocate but not associated with any buffer. 1047 * If not found, allocate if DEPALLOC flag is passed. 1048 * Found or allocated entry is returned in pagedeppp. 1049 * This routine must be called with splbio interrupts blocked. 1050 */ 1051static int 1052pagedep_lookup(ip, lbn, flags, pagedeppp) 1053 struct inode *ip; 1054 ufs_lbn_t lbn; 1055 int flags; 1056 struct pagedep **pagedeppp; 1057{ 1058 struct pagedep *pagedep; 1059 struct pagedep_hashhead *pagedephd; 1060 struct mount *mp; 1061 int ret; 1062 int i; 1063 1064 mtx_assert(&lk, MA_OWNED); 1065 mp = ITOV(ip)->v_mount; 1066 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn); 1067 1068 ret = pagedep_find(pagedephd, ip->i_number, lbn, mp, flags, pagedeppp); 1069 if (*pagedeppp || (flags & DEPALLOC) == 0) 1070 return (ret); 1071 FREE_LOCK(&lk); 1072 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), 1073 M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO); 1074 ACQUIRE_LOCK(&lk); 1075 ret = pagedep_find(pagedephd, ip->i_number, lbn, mp, flags, pagedeppp); 1076 if (*pagedeppp) { 1077 FREE(pagedep, M_PAGEDEP); 1078 return (ret); 1079 } 1080 pagedep->pd_list.wk_type = D_PAGEDEP; 1081 pagedep->pd_mnt = mp; 1082 pagedep->pd_ino = ip->i_number; 1083 pagedep->pd_lbn = lbn; 1084 LIST_INIT(&pagedep->pd_dirremhd); 1085 LIST_INIT(&pagedep->pd_pendinghd); 1086 for (i = 0; i < DAHASHSZ; i++) 1087 LIST_INIT(&pagedep->pd_diraddhd[i]); 1088 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 1089 *pagedeppp = pagedep; 1090 return (0); 1091} 1092 1093/* 1094 * Structures and routines associated with inodedep caching. 1095 */ 1096LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl; 1097static u_long inodedep_hash; /* size of hash table - 1 */ 1098static long num_inodedep; /* number of inodedep allocated */ 1099#define INODEDEP_HASH(fs, inum) \ 1100 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash]) 1101 1102static int 1103inodedep_find(inodedephd, fs, inum, inodedeppp) 1104 struct inodedep_hashhead *inodedephd; 1105 struct fs *fs; 1106 ino_t inum; 1107 struct inodedep **inodedeppp; 1108{ 1109 struct inodedep *inodedep; 1110 1111 LIST_FOREACH(inodedep, inodedephd, id_hash) 1112 if (inum == inodedep->id_ino && fs == inodedep->id_fs) 1113 break; 1114 if (inodedep) { 1115 *inodedeppp = inodedep; 1116 return (1); 1117 } 1118 *inodedeppp = NULL; 1119 1120 return (0); 1121} 1122/* 1123 * Look up an inodedep. Return 1 if found, 0 if not found. 1124 * If not found, allocate if DEPALLOC flag is passed. 1125 * Found or allocated entry is returned in inodedeppp. 1126 * This routine must be called with splbio interrupts blocked. 1127 */ 1128static int 1129inodedep_lookup(fs, inum, flags, inodedeppp) 1130 struct fs *fs; 1131 ino_t inum; 1132 int flags; 1133 struct inodedep **inodedeppp; 1134{ 1135 struct inodedep *inodedep; 1136 struct inodedep_hashhead *inodedephd; 1137 1138 mtx_assert(&lk, MA_OWNED); 1139 inodedephd = INODEDEP_HASH(fs, inum); 1140 1141 if (inodedep_find(inodedephd, fs, inum, inodedeppp)) 1142 return (1); 1143 if ((flags & DEPALLOC) == 0) 1144 return (0); 1145 /* 1146 * If we are over our limit, try to improve the situation. 1147 */ 1148 if (num_inodedep > max_softdeps && (flags & NODELAY) == 0) 1149 request_cleanup(FLUSH_INODES); 1150 FREE_LOCK(&lk); 1151 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep), 1152 M_INODEDEP, M_SOFTDEP_FLAGS); 1153 ACQUIRE_LOCK(&lk); 1154 if (inodedep_find(inodedephd, fs, inum, inodedeppp)) { 1155 FREE(inodedep, M_INODEDEP); 1156 return (1); 1157 } 1158 num_inodedep += 1; 1159 inodedep->id_list.wk_type = D_INODEDEP; 1160 inodedep->id_fs = fs; 1161 inodedep->id_ino = inum; 1162 inodedep->id_state = ALLCOMPLETE; 1163 inodedep->id_nlinkdelta = 0; 1164 inodedep->id_savedino1 = NULL; 1165 inodedep->id_savedsize = -1; 1166 inodedep->id_savedextsize = -1; 1167 inodedep->id_buf = NULL; 1168 LIST_INIT(&inodedep->id_pendinghd); 1169 LIST_INIT(&inodedep->id_inowait); 1170 LIST_INIT(&inodedep->id_bufwait); 1171 TAILQ_INIT(&inodedep->id_inoupdt); 1172 TAILQ_INIT(&inodedep->id_newinoupdt); 1173 TAILQ_INIT(&inodedep->id_extupdt); 1174 TAILQ_INIT(&inodedep->id_newextupdt); 1175 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 1176 *inodedeppp = inodedep; 1177 return (0); 1178} 1179 1180/* 1181 * Structures and routines associated with newblk caching. 1182 */ 1183LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl; 1184u_long newblk_hash; /* size of hash table - 1 */ 1185#define NEWBLK_HASH(fs, inum) \ 1186 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash]) 1187 1188static int 1189newblk_find(newblkhd, fs, newblkno, newblkpp) 1190 struct newblk_hashhead *newblkhd; 1191 struct fs *fs; 1192 ufs2_daddr_t newblkno; 1193 struct newblk **newblkpp; 1194{ 1195 struct newblk *newblk; 1196 1197 LIST_FOREACH(newblk, newblkhd, nb_hash) 1198 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs) 1199 break; 1200 if (newblk) { 1201 *newblkpp = newblk; 1202 return (1); 1203 } 1204 *newblkpp = NULL; 1205 return (0); 1206} 1207 1208/* 1209 * Look up a newblk. Return 1 if found, 0 if not found. 1210 * If not found, allocate if DEPALLOC flag is passed. 1211 * Found or allocated entry is returned in newblkpp. 1212 */ 1213static int 1214newblk_lookup(fs, newblkno, flags, newblkpp) 1215 struct fs *fs; 1216 ufs2_daddr_t newblkno; 1217 int flags; 1218 struct newblk **newblkpp; 1219{ 1220 struct newblk *newblk; 1221 struct newblk_hashhead *newblkhd; 1222 1223 newblkhd = NEWBLK_HASH(fs, newblkno); 1224 if (newblk_find(newblkhd, fs, newblkno, newblkpp)) 1225 return (1); 1226 if ((flags & DEPALLOC) == 0) 1227 return (0); 1228 FREE_LOCK(&lk); 1229 MALLOC(newblk, struct newblk *, sizeof(struct newblk), 1230 M_NEWBLK, M_SOFTDEP_FLAGS); 1231 ACQUIRE_LOCK(&lk); 1232 if (newblk_find(newblkhd, fs, newblkno, newblkpp)) { 1233 FREE(newblk, M_NEWBLK); 1234 return (1); 1235 } 1236 newblk->nb_state = 0; 1237 newblk->nb_fs = fs; 1238 newblk->nb_newblkno = newblkno; 1239 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 1240 *newblkpp = newblk; 1241 return (0); 1242} 1243 1244/* 1245 * Executed during filesystem system initialization before 1246 * mounting any filesystems. 1247 */ 1248void 1249softdep_initialize() 1250{ 1251 1252 LIST_INIT(&mkdirlisthd); 1253 LIST_INIT(&softdep_workitem_pending); 1254 max_softdeps = desiredvnodes * 4; 1255 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 1256 &pagedep_hash); 1257 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash); 1258 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash); 1259 1260 /* hooks through which the main kernel code calls us */ 1261 softdep_process_worklist_hook = softdep_process_worklist; 1262 1263 /* initialise bioops hack */ 1264 bioops.io_start = softdep_disk_io_initiation; 1265 bioops.io_complete = softdep_disk_write_complete; 1266 bioops.io_deallocate = softdep_deallocate_dependencies; 1267 bioops.io_countdeps = softdep_count_dependencies; 1268} 1269 1270/* 1271 * Executed after all filesystems have been unmounted during 1272 * filesystem module unload. 1273 */ 1274void 1275softdep_uninitialize() 1276{ 1277 1278 softdep_process_worklist_hook = NULL; 1279 hashdestroy(pagedep_hashtbl, M_PAGEDEP, pagedep_hash); 1280 hashdestroy(inodedep_hashtbl, M_INODEDEP, inodedep_hash); 1281 hashdestroy(newblk_hashtbl, M_NEWBLK, newblk_hash); 1282} 1283 1284/* 1285 * Called at mount time to notify the dependency code that a 1286 * filesystem wishes to use it. 1287 */ 1288int 1289softdep_mount(devvp, mp, fs, cred) 1290 struct vnode *devvp; 1291 struct mount *mp; 1292 struct fs *fs; 1293 struct ucred *cred; 1294{ 1295 struct csum_total cstotal; 1296 struct cg *cgp; 1297 struct buf *bp; 1298 int error, cyl; 1299 1300 mp->mnt_flag &= ~MNT_ASYNC; 1301 mp->mnt_flag |= MNT_SOFTDEP; 1302 /* 1303 * When doing soft updates, the counters in the 1304 * superblock may have gotten out of sync. Recomputation 1305 * can take a long time and can be deferred for background 1306 * fsck. However, the old behavior of scanning the cylinder 1307 * groups and recalculating them at mount time is available 1308 * by setting vfs.ffs.compute_summary_at_mount to one. 1309 */ 1310 if (compute_summary_at_mount == 0 || fs->fs_clean != 0) 1311 return (0); 1312 bzero(&cstotal, sizeof cstotal); 1313 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 1314 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 1315 fs->fs_cgsize, cred, &bp)) != 0) { 1316 brelse(bp); 1317 return (error); 1318 } 1319 cgp = (struct cg *)bp->b_data; 1320 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 1321 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 1322 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 1323 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 1324 fs->fs_cs(fs, cyl) = cgp->cg_cs; 1325 brelse(bp); 1326 } 1327#ifdef DEBUG 1328 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 1329 printf("%s: superblock summary recomputed\n", fs->fs_fsmnt); 1330#endif 1331 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 1332 return (0); 1333} 1334 1335/* 1336 * Protecting the freemaps (or bitmaps). 1337 * 1338 * To eliminate the need to execute fsck before mounting a filesystem 1339 * after a power failure, one must (conservatively) guarantee that the 1340 * on-disk copy of the bitmaps never indicate that a live inode or block is 1341 * free. So, when a block or inode is allocated, the bitmap should be 1342 * updated (on disk) before any new pointers. When a block or inode is 1343 * freed, the bitmap should not be updated until all pointers have been 1344 * reset. The latter dependency is handled by the delayed de-allocation 1345 * approach described below for block and inode de-allocation. The former 1346 * dependency is handled by calling the following procedure when a block or 1347 * inode is allocated. When an inode is allocated an "inodedep" is created 1348 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 1349 * Each "inodedep" is also inserted into the hash indexing structure so 1350 * that any additional link additions can be made dependent on the inode 1351 * allocation. 1352 * 1353 * The ufs filesystem maintains a number of free block counts (e.g., per 1354 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 1355 * in addition to the bitmaps. These counts are used to improve efficiency 1356 * during allocation and therefore must be consistent with the bitmaps. 1357 * There is no convenient way to guarantee post-crash consistency of these 1358 * counts with simple update ordering, for two main reasons: (1) The counts 1359 * and bitmaps for a single cylinder group block are not in the same disk 1360 * sector. If a disk write is interrupted (e.g., by power failure), one may 1361 * be written and the other not. (2) Some of the counts are located in the 1362 * superblock rather than the cylinder group block. So, we focus our soft 1363 * updates implementation on protecting the bitmaps. When mounting a 1364 * filesystem, we recompute the auxiliary counts from the bitmaps. 1365 */ 1366 1367/* 1368 * Called just after updating the cylinder group block to allocate an inode. 1369 */ 1370void 1371softdep_setup_inomapdep(bp, ip, newinum) 1372 struct buf *bp; /* buffer for cylgroup block with inode map */ 1373 struct inode *ip; /* inode related to allocation */ 1374 ino_t newinum; /* new inode number being allocated */ 1375{ 1376 struct inodedep *inodedep; 1377 struct bmsafemap *bmsafemap; 1378 1379 /* 1380 * Create a dependency for the newly allocated inode. 1381 * Panic if it already exists as something is seriously wrong. 1382 * Otherwise add it to the dependency list for the buffer holding 1383 * the cylinder group map from which it was allocated. 1384 */ 1385 ACQUIRE_LOCK(&lk); 1386 if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) 1387 panic("softdep_setup_inomapdep: found inode"); 1388 inodedep->id_buf = bp; 1389 inodedep->id_state &= ~DEPCOMPLETE; 1390 bmsafemap = bmsafemap_lookup(bp); 1391 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 1392 FREE_LOCK(&lk); 1393} 1394 1395/* 1396 * Called just after updating the cylinder group block to 1397 * allocate block or fragment. 1398 */ 1399void 1400softdep_setup_blkmapdep(bp, fs, newblkno) 1401 struct buf *bp; /* buffer for cylgroup block with block map */ 1402 struct fs *fs; /* filesystem doing allocation */ 1403 ufs2_daddr_t newblkno; /* number of newly allocated block */ 1404{ 1405 struct newblk *newblk; 1406 struct bmsafemap *bmsafemap; 1407 1408 /* 1409 * Create a dependency for the newly allocated block. 1410 * Add it to the dependency list for the buffer holding 1411 * the cylinder group map from which it was allocated. 1412 */ 1413 ACQUIRE_LOCK(&lk); 1414 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0) 1415 panic("softdep_setup_blkmapdep: found block"); 1416 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp); 1417 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 1418 FREE_LOCK(&lk); 1419} 1420 1421/* 1422 * Find the bmsafemap associated with a cylinder group buffer. 1423 * If none exists, create one. The buffer must be locked when 1424 * this routine is called and this routine must be called with 1425 * splbio interrupts blocked. 1426 */ 1427static struct bmsafemap * 1428bmsafemap_lookup(bp) 1429 struct buf *bp; 1430{ 1431 struct bmsafemap *bmsafemap; 1432 struct worklist *wk; 1433 1434 mtx_assert(&lk, MA_OWNED); 1435 LIST_FOREACH(wk, &bp->b_dep, wk_list) 1436 if (wk->wk_type == D_BMSAFEMAP) 1437 return (WK_BMSAFEMAP(wk)); 1438 FREE_LOCK(&lk); 1439 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap), 1440 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 1441 bmsafemap->sm_list.wk_type = D_BMSAFEMAP; 1442 bmsafemap->sm_list.wk_state = 0; 1443 bmsafemap->sm_buf = bp; 1444 LIST_INIT(&bmsafemap->sm_allocdirecthd); 1445 LIST_INIT(&bmsafemap->sm_allocindirhd); 1446 LIST_INIT(&bmsafemap->sm_inodedephd); 1447 LIST_INIT(&bmsafemap->sm_newblkhd); 1448 ACQUIRE_LOCK(&lk); 1449 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 1450 return (bmsafemap); 1451} 1452 1453/* 1454 * Direct block allocation dependencies. 1455 * 1456 * When a new block is allocated, the corresponding disk locations must be 1457 * initialized (with zeros or new data) before the on-disk inode points to 1458 * them. Also, the freemap from which the block was allocated must be 1459 * updated (on disk) before the inode's pointer. These two dependencies are 1460 * independent of each other and are needed for all file blocks and indirect 1461 * blocks that are pointed to directly by the inode. Just before the 1462 * "in-core" version of the inode is updated with a newly allocated block 1463 * number, a procedure (below) is called to setup allocation dependency 1464 * structures. These structures are removed when the corresponding 1465 * dependencies are satisfied or when the block allocation becomes obsolete 1466 * (i.e., the file is deleted, the block is de-allocated, or the block is a 1467 * fragment that gets upgraded). All of these cases are handled in 1468 * procedures described later. 1469 * 1470 * When a file extension causes a fragment to be upgraded, either to a larger 1471 * fragment or to a full block, the on-disk location may change (if the 1472 * previous fragment could not simply be extended). In this case, the old 1473 * fragment must be de-allocated, but not until after the inode's pointer has 1474 * been updated. In most cases, this is handled by later procedures, which 1475 * will construct a "freefrag" structure to be added to the workitem queue 1476 * when the inode update is complete (or obsolete). The main exception to 1477 * this is when an allocation occurs while a pending allocation dependency 1478 * (for the same block pointer) remains. This case is handled in the main 1479 * allocation dependency setup procedure by immediately freeing the 1480 * unreferenced fragments. 1481 */ 1482void 1483softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 1484 struct inode *ip; /* inode to which block is being added */ 1485 ufs_lbn_t lbn; /* block pointer within inode */ 1486 ufs2_daddr_t newblkno; /* disk block number being added */ 1487 ufs2_daddr_t oldblkno; /* previous block number, 0 unless frag */ 1488 long newsize; /* size of new block */ 1489 long oldsize; /* size of new block */ 1490 struct buf *bp; /* bp for allocated block */ 1491{ 1492 struct allocdirect *adp, *oldadp; 1493 struct allocdirectlst *adphead; 1494 struct bmsafemap *bmsafemap; 1495 struct inodedep *inodedep; 1496 struct pagedep *pagedep; 1497 struct newblk *newblk; 1498 1499 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect), 1500 M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO); 1501 adp->ad_list.wk_type = D_ALLOCDIRECT; 1502 adp->ad_lbn = lbn; 1503 adp->ad_newblkno = newblkno; 1504 adp->ad_oldblkno = oldblkno; 1505 adp->ad_newsize = newsize; 1506 adp->ad_oldsize = oldsize; 1507 adp->ad_state = ATTACHED; 1508 LIST_INIT(&adp->ad_newdirblk); 1509 if (newblkno == oldblkno) 1510 adp->ad_freefrag = NULL; 1511 else 1512 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1513 1514 ACQUIRE_LOCK(&lk); 1515 if (lbn >= NDADDR) { 1516 /* allocating an indirect block */ 1517 if (oldblkno != 0) 1518 panic("softdep_setup_allocdirect: non-zero indir"); 1519 } else { 1520 /* 1521 * Allocating a direct block. 1522 * 1523 * If we are allocating a directory block, then we must 1524 * allocate an associated pagedep to track additions and 1525 * deletions. 1526 */ 1527 if ((ip->i_mode & IFMT) == IFDIR && 1528 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1529 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 1530 } 1531 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1532 panic("softdep_setup_allocdirect: lost block"); 1533 if (newblk->nb_state == DEPCOMPLETE) { 1534 adp->ad_state |= DEPCOMPLETE; 1535 adp->ad_buf = NULL; 1536 } else { 1537 bmsafemap = newblk->nb_bmsafemap; 1538 adp->ad_buf = bmsafemap->sm_buf; 1539 LIST_REMOVE(newblk, nb_deps); 1540 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1541 } 1542 LIST_REMOVE(newblk, nb_hash); 1543 FREE(newblk, M_NEWBLK); 1544 1545 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep); 1546 adp->ad_inodedep = inodedep; 1547 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list); 1548 /* 1549 * The list of allocdirects must be kept in sorted and ascending 1550 * order so that the rollback routines can quickly determine the 1551 * first uncommitted block (the size of the file stored on disk 1552 * ends at the end of the lowest committed fragment, or if there 1553 * are no fragments, at the end of the highest committed block). 1554 * Since files generally grow, the typical case is that the new 1555 * block is to be added at the end of the list. We speed this 1556 * special case by checking against the last allocdirect in the 1557 * list before laboriously traversing the list looking for the 1558 * insertion point. 1559 */ 1560 adphead = &inodedep->id_newinoupdt; 1561 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1562 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1563 /* insert at end of list */ 1564 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1565 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1566 allocdirect_merge(adphead, adp, oldadp); 1567 FREE_LOCK(&lk); 1568 return; 1569 } 1570 TAILQ_FOREACH(oldadp, adphead, ad_next) { 1571 if (oldadp->ad_lbn >= lbn) 1572 break; 1573 } 1574 if (oldadp == NULL) 1575 panic("softdep_setup_allocdirect: lost entry"); 1576 /* insert in middle of list */ 1577 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1578 if (oldadp->ad_lbn == lbn) 1579 allocdirect_merge(adphead, adp, oldadp); 1580 FREE_LOCK(&lk); 1581} 1582 1583/* 1584 * Replace an old allocdirect dependency with a newer one. 1585 * This routine must be called with splbio interrupts blocked. 1586 */ 1587static void 1588allocdirect_merge(adphead, newadp, oldadp) 1589 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 1590 struct allocdirect *newadp; /* allocdirect being added */ 1591 struct allocdirect *oldadp; /* existing allocdirect being checked */ 1592{ 1593 struct worklist *wk; 1594 struct freefrag *freefrag; 1595 struct newdirblk *newdirblk; 1596 1597 mtx_assert(&lk, MA_OWNED); 1598 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 1599 newadp->ad_oldsize != oldadp->ad_newsize || 1600 newadp->ad_lbn >= NDADDR) 1601 panic("%s %jd != new %jd || old size %ld != new %ld", 1602 "allocdirect_merge: old blkno", 1603 (intmax_t)newadp->ad_oldblkno, 1604 (intmax_t)oldadp->ad_newblkno, 1605 newadp->ad_oldsize, oldadp->ad_newsize); 1606 newadp->ad_oldblkno = oldadp->ad_oldblkno; 1607 newadp->ad_oldsize = oldadp->ad_oldsize; 1608 /* 1609 * If the old dependency had a fragment to free or had never 1610 * previously had a block allocated, then the new dependency 1611 * can immediately post its freefrag and adopt the old freefrag. 1612 * This action is done by swapping the freefrag dependencies. 1613 * The new dependency gains the old one's freefrag, and the 1614 * old one gets the new one and then immediately puts it on 1615 * the worklist when it is freed by free_allocdirect. It is 1616 * not possible to do this swap when the old dependency had a 1617 * non-zero size but no previous fragment to free. This condition 1618 * arises when the new block is an extension of the old block. 1619 * Here, the first part of the fragment allocated to the new 1620 * dependency is part of the block currently claimed on disk by 1621 * the old dependency, so cannot legitimately be freed until the 1622 * conditions for the new dependency are fulfilled. 1623 */ 1624 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 1625 freefrag = newadp->ad_freefrag; 1626 newadp->ad_freefrag = oldadp->ad_freefrag; 1627 oldadp->ad_freefrag = freefrag; 1628 } 1629 /* 1630 * If we are tracking a new directory-block allocation, 1631 * move it from the old allocdirect to the new allocdirect. 1632 */ 1633 if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) { 1634 newdirblk = WK_NEWDIRBLK(wk); 1635 WORKLIST_REMOVE(&newdirblk->db_list); 1636 if (LIST_FIRST(&oldadp->ad_newdirblk) != NULL) 1637 panic("allocdirect_merge: extra newdirblk"); 1638 WORKLIST_INSERT(&newadp->ad_newdirblk, &newdirblk->db_list); 1639 } 1640 free_allocdirect(adphead, oldadp, 0); 1641} 1642 1643/* 1644 * Allocate a new freefrag structure if needed. 1645 */ 1646static struct freefrag * 1647newfreefrag(ip, blkno, size) 1648 struct inode *ip; 1649 ufs2_daddr_t blkno; 1650 long size; 1651{ 1652 struct freefrag *freefrag; 1653 struct fs *fs; 1654 1655 if (blkno == 0) 1656 return (NULL); 1657 fs = ip->i_fs; 1658 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 1659 panic("newfreefrag: frag size"); 1660 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag), 1661 M_FREEFRAG, M_SOFTDEP_FLAGS); 1662 freefrag->ff_list.wk_type = D_FREEFRAG; 1663 freefrag->ff_state = 0; 1664 freefrag->ff_inum = ip->i_number; 1665 freefrag->ff_mnt = ITOV(ip)->v_mount; 1666 freefrag->ff_blkno = blkno; 1667 freefrag->ff_fragsize = size; 1668 return (freefrag); 1669} 1670 1671/* 1672 * This workitem de-allocates fragments that were replaced during 1673 * file block allocation. 1674 */ 1675static void 1676handle_workitem_freefrag(freefrag) 1677 struct freefrag *freefrag; 1678{ 1679 struct ufsmount *ump = VFSTOUFS(freefrag->ff_mnt); 1680 1681 ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno, 1682 freefrag->ff_fragsize, freefrag->ff_inum); 1683 FREE(freefrag, M_FREEFRAG); 1684} 1685 1686/* 1687 * Set up a dependency structure for an external attributes data block. 1688 * This routine follows much of the structure of softdep_setup_allocdirect. 1689 * See the description of softdep_setup_allocdirect above for details. 1690 */ 1691void 1692softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 1693 struct inode *ip; 1694 ufs_lbn_t lbn; 1695 ufs2_daddr_t newblkno; 1696 ufs2_daddr_t oldblkno; 1697 long newsize; 1698 long oldsize; 1699 struct buf *bp; 1700{ 1701 struct allocdirect *adp, *oldadp; 1702 struct allocdirectlst *adphead; 1703 struct bmsafemap *bmsafemap; 1704 struct inodedep *inodedep; 1705 struct newblk *newblk; 1706 1707 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect), 1708 M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO); 1709 adp->ad_list.wk_type = D_ALLOCDIRECT; 1710 adp->ad_lbn = lbn; 1711 adp->ad_newblkno = newblkno; 1712 adp->ad_oldblkno = oldblkno; 1713 adp->ad_newsize = newsize; 1714 adp->ad_oldsize = oldsize; 1715 adp->ad_state = ATTACHED | EXTDATA; 1716 LIST_INIT(&adp->ad_newdirblk); 1717 if (newblkno == oldblkno) 1718 adp->ad_freefrag = NULL; 1719 else 1720 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1721 1722 ACQUIRE_LOCK(&lk); 1723 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1724 panic("softdep_setup_allocext: lost block"); 1725 1726 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep); 1727 adp->ad_inodedep = inodedep; 1728 1729 if (newblk->nb_state == DEPCOMPLETE) { 1730 adp->ad_state |= DEPCOMPLETE; 1731 adp->ad_buf = NULL; 1732 } else { 1733 bmsafemap = newblk->nb_bmsafemap; 1734 adp->ad_buf = bmsafemap->sm_buf; 1735 LIST_REMOVE(newblk, nb_deps); 1736 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1737 } 1738 LIST_REMOVE(newblk, nb_hash); 1739 FREE(newblk, M_NEWBLK); 1740 1741 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list); 1742 if (lbn >= NXADDR) 1743 panic("softdep_setup_allocext: lbn %lld > NXADDR", 1744 (long long)lbn); 1745 /* 1746 * The list of allocdirects must be kept in sorted and ascending 1747 * order so that the rollback routines can quickly determine the 1748 * first uncommitted block (the size of the file stored on disk 1749 * ends at the end of the lowest committed fragment, or if there 1750 * are no fragments, at the end of the highest committed block). 1751 * Since files generally grow, the typical case is that the new 1752 * block is to be added at the end of the list. We speed this 1753 * special case by checking against the last allocdirect in the 1754 * list before laboriously traversing the list looking for the 1755 * insertion point. 1756 */ 1757 adphead = &inodedep->id_newextupdt; 1758 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1759 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1760 /* insert at end of list */ 1761 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1762 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1763 allocdirect_merge(adphead, adp, oldadp); 1764 FREE_LOCK(&lk); 1765 return; 1766 } 1767 TAILQ_FOREACH(oldadp, adphead, ad_next) { 1768 if (oldadp->ad_lbn >= lbn) 1769 break; 1770 } 1771 if (oldadp == NULL) 1772 panic("softdep_setup_allocext: lost entry"); 1773 /* insert in middle of list */ 1774 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1775 if (oldadp->ad_lbn == lbn) 1776 allocdirect_merge(adphead, adp, oldadp); 1777 FREE_LOCK(&lk); 1778} 1779 1780/* 1781 * Indirect block allocation dependencies. 1782 * 1783 * The same dependencies that exist for a direct block also exist when 1784 * a new block is allocated and pointed to by an entry in a block of 1785 * indirect pointers. The undo/redo states described above are also 1786 * used here. Because an indirect block contains many pointers that 1787 * may have dependencies, a second copy of the entire in-memory indirect 1788 * block is kept. The buffer cache copy is always completely up-to-date. 1789 * The second copy, which is used only as a source for disk writes, 1790 * contains only the safe pointers (i.e., those that have no remaining 1791 * update dependencies). The second copy is freed when all pointers 1792 * are safe. The cache is not allowed to replace indirect blocks with 1793 * pending update dependencies. If a buffer containing an indirect 1794 * block with dependencies is written, these routines will mark it 1795 * dirty again. It can only be successfully written once all the 1796 * dependencies are removed. The ffs_fsync routine in conjunction with 1797 * softdep_sync_metadata work together to get all the dependencies 1798 * removed so that a file can be successfully written to disk. Three 1799 * procedures are used when setting up indirect block pointer 1800 * dependencies. The division is necessary because of the organization 1801 * of the "balloc" routine and because of the distinction between file 1802 * pages and file metadata blocks. 1803 */ 1804 1805/* 1806 * Allocate a new allocindir structure. 1807 */ 1808static struct allocindir * 1809newallocindir(ip, ptrno, newblkno, oldblkno) 1810 struct inode *ip; /* inode for file being extended */ 1811 int ptrno; /* offset of pointer in indirect block */ 1812 ufs2_daddr_t newblkno; /* disk block number being added */ 1813 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 1814{ 1815 struct allocindir *aip; 1816 1817 MALLOC(aip, struct allocindir *, sizeof(struct allocindir), 1818 M_ALLOCINDIR, M_SOFTDEP_FLAGS|M_ZERO); 1819 aip->ai_list.wk_type = D_ALLOCINDIR; 1820 aip->ai_state = ATTACHED; 1821 aip->ai_offset = ptrno; 1822 aip->ai_newblkno = newblkno; 1823 aip->ai_oldblkno = oldblkno; 1824 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize); 1825 return (aip); 1826} 1827 1828/* 1829 * Called just before setting an indirect block pointer 1830 * to a newly allocated file page. 1831 */ 1832void 1833softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 1834 struct inode *ip; /* inode for file being extended */ 1835 ufs_lbn_t lbn; /* allocated block number within file */ 1836 struct buf *bp; /* buffer with indirect blk referencing page */ 1837 int ptrno; /* offset of pointer in indirect block */ 1838 ufs2_daddr_t newblkno; /* disk block number being added */ 1839 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 1840 struct buf *nbp; /* buffer holding allocated page */ 1841{ 1842 struct allocindir *aip; 1843 struct pagedep *pagedep; 1844 1845 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page"); 1846 aip = newallocindir(ip, ptrno, newblkno, oldblkno); 1847 ACQUIRE_LOCK(&lk); 1848 /* 1849 * If we are allocating a directory page, then we must 1850 * allocate an associated pagedep to track additions and 1851 * deletions. 1852 */ 1853 if ((ip->i_mode & IFMT) == IFDIR && 1854 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1855 WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list); 1856 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1857 setup_allocindir_phase2(bp, ip, aip); 1858 FREE_LOCK(&lk); 1859} 1860 1861/* 1862 * Called just before setting an indirect block pointer to a 1863 * newly allocated indirect block. 1864 */ 1865void 1866softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 1867 struct buf *nbp; /* newly allocated indirect block */ 1868 struct inode *ip; /* inode for file being extended */ 1869 struct buf *bp; /* indirect block referencing allocated block */ 1870 int ptrno; /* offset of pointer in indirect block */ 1871 ufs2_daddr_t newblkno; /* disk block number being added */ 1872{ 1873 struct allocindir *aip; 1874 1875 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta"); 1876 aip = newallocindir(ip, ptrno, newblkno, 0); 1877 ACQUIRE_LOCK(&lk); 1878 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1879 setup_allocindir_phase2(bp, ip, aip); 1880 FREE_LOCK(&lk); 1881} 1882 1883/* 1884 * Called to finish the allocation of the "aip" allocated 1885 * by one of the two routines above. 1886 */ 1887static void 1888setup_allocindir_phase2(bp, ip, aip) 1889 struct buf *bp; /* in-memory copy of the indirect block */ 1890 struct inode *ip; /* inode for file being extended */ 1891 struct allocindir *aip; /* allocindir allocated by the above routines */ 1892{ 1893 struct worklist *wk; 1894 struct indirdep *indirdep, *newindirdep; 1895 struct bmsafemap *bmsafemap; 1896 struct allocindir *oldaip; 1897 struct freefrag *freefrag; 1898 struct newblk *newblk; 1899 ufs2_daddr_t blkno; 1900 1901 mtx_assert(&lk, MA_OWNED); 1902 if (bp->b_lblkno >= 0) 1903 panic("setup_allocindir_phase2: not indir blk"); 1904 for (indirdep = NULL, newindirdep = NULL; ; ) { 1905 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 1906 if (wk->wk_type != D_INDIRDEP) 1907 continue; 1908 indirdep = WK_INDIRDEP(wk); 1909 break; 1910 } 1911 if (indirdep == NULL && newindirdep) { 1912 indirdep = newindirdep; 1913 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 1914 newindirdep = NULL; 1915 } 1916 if (indirdep) { 1917 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0, 1918 &newblk) == 0) 1919 panic("setup_allocindir: lost block"); 1920 if (newblk->nb_state == DEPCOMPLETE) { 1921 aip->ai_state |= DEPCOMPLETE; 1922 aip->ai_buf = NULL; 1923 } else { 1924 bmsafemap = newblk->nb_bmsafemap; 1925 aip->ai_buf = bmsafemap->sm_buf; 1926 LIST_REMOVE(newblk, nb_deps); 1927 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd, 1928 aip, ai_deps); 1929 } 1930 LIST_REMOVE(newblk, nb_hash); 1931 FREE(newblk, M_NEWBLK); 1932 aip->ai_indirdep = indirdep; 1933 /* 1934 * Check to see if there is an existing dependency 1935 * for this block. If there is, merge the old 1936 * dependency into the new one. 1937 */ 1938 if (aip->ai_oldblkno == 0) 1939 oldaip = NULL; 1940 else 1941 1942 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) 1943 if (oldaip->ai_offset == aip->ai_offset) 1944 break; 1945 freefrag = NULL; 1946 if (oldaip != NULL) { 1947 if (oldaip->ai_newblkno != aip->ai_oldblkno) 1948 panic("setup_allocindir_phase2: blkno"); 1949 aip->ai_oldblkno = oldaip->ai_oldblkno; 1950 freefrag = aip->ai_freefrag; 1951 aip->ai_freefrag = oldaip->ai_freefrag; 1952 oldaip->ai_freefrag = NULL; 1953 free_allocindir(oldaip, NULL); 1954 } 1955 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 1956 if (ip->i_ump->um_fstype == UFS1) 1957 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data) 1958 [aip->ai_offset] = aip->ai_oldblkno; 1959 else 1960 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data) 1961 [aip->ai_offset] = aip->ai_oldblkno; 1962 FREE_LOCK(&lk); 1963 if (freefrag != NULL) 1964 handle_workitem_freefrag(freefrag); 1965 } else 1966 FREE_LOCK(&lk); 1967 if (newindirdep) { 1968 newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 1969 brelse(newindirdep->ir_savebp); 1970 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP); 1971 } 1972 if (indirdep) { 1973 ACQUIRE_LOCK(&lk); 1974 break; 1975 } 1976 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep), 1977 M_INDIRDEP, M_SOFTDEP_FLAGS); 1978 newindirdep->ir_list.wk_type = D_INDIRDEP; 1979 newindirdep->ir_state = ATTACHED; 1980 if (ip->i_ump->um_fstype == UFS1) 1981 newindirdep->ir_state |= UFS1FMT; 1982 LIST_INIT(&newindirdep->ir_deplisthd); 1983 LIST_INIT(&newindirdep->ir_donehd); 1984 if (bp->b_blkno == bp->b_lblkno) { 1985 ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp, 1986 NULL, NULL); 1987 bp->b_blkno = blkno; 1988 } 1989 newindirdep->ir_savebp = 1990 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0); 1991 BUF_KERNPROC(newindirdep->ir_savebp); 1992 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 1993 ACQUIRE_LOCK(&lk); 1994 } 1995} 1996 1997/* 1998 * Block de-allocation dependencies. 1999 * 2000 * When blocks are de-allocated, the on-disk pointers must be nullified before 2001 * the blocks are made available for use by other files. (The true 2002 * requirement is that old pointers must be nullified before new on-disk 2003 * pointers are set. We chose this slightly more stringent requirement to 2004 * reduce complexity.) Our implementation handles this dependency by updating 2005 * the inode (or indirect block) appropriately but delaying the actual block 2006 * de-allocation (i.e., freemap and free space count manipulation) until 2007 * after the updated versions reach stable storage. After the disk is 2008 * updated, the blocks can be safely de-allocated whenever it is convenient. 2009 * This implementation handles only the common case of reducing a file's 2010 * length to zero. Other cases are handled by the conventional synchronous 2011 * write approach. 2012 * 2013 * The ffs implementation with which we worked double-checks 2014 * the state of the block pointers and file size as it reduces 2015 * a file's length. Some of this code is replicated here in our 2016 * soft updates implementation. The freeblks->fb_chkcnt field is 2017 * used to transfer a part of this information to the procedure 2018 * that eventually de-allocates the blocks. 2019 * 2020 * This routine should be called from the routine that shortens 2021 * a file's length, before the inode's size or block pointers 2022 * are modified. It will save the block pointer information for 2023 * later release and zero the inode so that the calling routine 2024 * can release it. 2025 */ 2026void 2027softdep_setup_freeblocks(ip, length, flags) 2028 struct inode *ip; /* The inode whose length is to be reduced */ 2029 off_t length; /* The new length for the file */ 2030 int flags; /* IO_EXT and/or IO_NORMAL */ 2031{ 2032 struct freeblks *freeblks; 2033 struct inodedep *inodedep; 2034 struct allocdirect *adp; 2035 struct vnode *vp; 2036 struct buf *bp; 2037 struct fs *fs; 2038 ufs2_daddr_t extblocks, datablocks; 2039 int i, delay, error; 2040 2041 fs = ip->i_fs; 2042 if (length != 0) 2043 panic("softdep_setup_freeblocks: non-zero length"); 2044 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks), 2045 M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO); 2046 freeblks->fb_list.wk_type = D_FREEBLKS; 2047 freeblks->fb_uid = ip->i_uid; 2048 freeblks->fb_previousinum = ip->i_number; 2049 freeblks->fb_devvp = ip->i_devvp; 2050 freeblks->fb_mnt = ITOV(ip)->v_mount; 2051 extblocks = 0; 2052 if (fs->fs_magic == FS_UFS2_MAGIC) 2053 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 2054 datablocks = DIP(ip, i_blocks) - extblocks; 2055 if ((flags & IO_NORMAL) == 0) { 2056 freeblks->fb_oldsize = 0; 2057 freeblks->fb_chkcnt = 0; 2058 } else { 2059 freeblks->fb_oldsize = ip->i_size; 2060 ip->i_size = 0; 2061 DIP_SET(ip, i_size, 0); 2062 freeblks->fb_chkcnt = datablocks; 2063 for (i = 0; i < NDADDR; i++) { 2064 freeblks->fb_dblks[i] = DIP(ip, i_db[i]); 2065 DIP_SET(ip, i_db[i], 0); 2066 } 2067 for (i = 0; i < NIADDR; i++) { 2068 freeblks->fb_iblks[i] = DIP(ip, i_ib[i]); 2069 DIP_SET(ip, i_ib[i], 0); 2070 } 2071 /* 2072 * If the file was removed, then the space being freed was 2073 * accounted for then (see softdep_filereleased()). If the 2074 * file is merely being truncated, then we account for it now. 2075 */ 2076 if ((ip->i_flag & IN_SPACECOUNTED) == 0) { 2077 UFS_LOCK(ip->i_ump); 2078 fs->fs_pendingblocks += datablocks; 2079 UFS_UNLOCK(ip->i_ump); 2080 } 2081 } 2082 if ((flags & IO_EXT) == 0) { 2083 freeblks->fb_oldextsize = 0; 2084 } else { 2085 freeblks->fb_oldextsize = ip->i_din2->di_extsize; 2086 ip->i_din2->di_extsize = 0; 2087 freeblks->fb_chkcnt += extblocks; 2088 for (i = 0; i < NXADDR; i++) { 2089 freeblks->fb_eblks[i] = ip->i_din2->di_extb[i]; 2090 ip->i_din2->di_extb[i] = 0; 2091 } 2092 } 2093 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - freeblks->fb_chkcnt); 2094 /* 2095 * Push the zero'ed inode to to its disk buffer so that we are free 2096 * to delete its dependencies below. Once the dependencies are gone 2097 * the buffer can be safely released. 2098 */ 2099 if ((error = bread(ip->i_devvp, 2100 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 2101 (int)fs->fs_bsize, NOCRED, &bp)) != 0) { 2102 brelse(bp); 2103 softdep_error("softdep_setup_freeblocks", error); 2104 } 2105 if (ip->i_ump->um_fstype == UFS1) 2106 *((struct ufs1_dinode *)bp->b_data + 2107 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1; 2108 else 2109 *((struct ufs2_dinode *)bp->b_data + 2110 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2; 2111 /* 2112 * Find and eliminate any inode dependencies. 2113 */ 2114 ACQUIRE_LOCK(&lk); 2115 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep); 2116 if ((inodedep->id_state & IOSTARTED) != 0) 2117 panic("softdep_setup_freeblocks: inode busy"); 2118 /* 2119 * Add the freeblks structure to the list of operations that 2120 * must await the zero'ed inode being written to disk. If we 2121 * still have a bitmap dependency (delay == 0), then the inode 2122 * has never been written to disk, so we can process the 2123 * freeblks below once we have deleted the dependencies. 2124 */ 2125 delay = (inodedep->id_state & DEPCOMPLETE); 2126 if (delay) 2127 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list); 2128 /* 2129 * Because the file length has been truncated to zero, any 2130 * pending block allocation dependency structures associated 2131 * with this inode are obsolete and can simply be de-allocated. 2132 * We must first merge the two dependency lists to get rid of 2133 * any duplicate freefrag structures, then purge the merged list. 2134 * If we still have a bitmap dependency, then the inode has never 2135 * been written to disk, so we can free any fragments without delay. 2136 */ 2137 if (flags & IO_NORMAL) { 2138 merge_inode_lists(&inodedep->id_newinoupdt, 2139 &inodedep->id_inoupdt); 2140 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 2141 free_allocdirect(&inodedep->id_inoupdt, adp, delay); 2142 } 2143 if (flags & IO_EXT) { 2144 merge_inode_lists(&inodedep->id_newextupdt, 2145 &inodedep->id_extupdt); 2146 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0) 2147 free_allocdirect(&inodedep->id_extupdt, adp, delay); 2148 } 2149 FREE_LOCK(&lk); 2150 bdwrite(bp); 2151 /* 2152 * We must wait for any I/O in progress to finish so that 2153 * all potential buffers on the dirty list will be visible. 2154 * Once they are all there, walk the list and get rid of 2155 * any dependencies. 2156 */ 2157 vp = ITOV(ip); 2158 VI_LOCK(vp); 2159 drain_output(vp); 2160restart: 2161 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 2162 if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) || 2163 ((flags & IO_NORMAL) == 0 && 2164 (bp->b_xflags & BX_ALTDATA) == 0)) 2165 continue; 2166 if ((bp = getdirtybuf(bp, VI_MTX(vp), MNT_WAIT)) == NULL) 2167 goto restart; 2168 VI_UNLOCK(vp); 2169 ACQUIRE_LOCK(&lk); 2170 (void) inodedep_lookup(fs, ip->i_number, 0, &inodedep); 2171 deallocate_dependencies(bp, inodedep); 2172 FREE_LOCK(&lk); 2173 bp->b_flags |= B_INVAL | B_NOCACHE; 2174 brelse(bp); 2175 VI_LOCK(vp); 2176 goto restart; 2177 } 2178 VI_UNLOCK(vp); 2179 ACQUIRE_LOCK(&lk); 2180 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0) 2181 (void) free_inodedep(inodedep); 2182 FREE_LOCK(&lk); 2183 /* 2184 * If the inode has never been written to disk (delay == 0), 2185 * then we can process the freeblks now that we have deleted 2186 * the dependencies. 2187 */ 2188 if (!delay) 2189 handle_workitem_freeblocks(freeblks, 0); 2190} 2191 2192/* 2193 * Reclaim any dependency structures from a buffer that is about to 2194 * be reallocated to a new vnode. The buffer must be locked, thus, 2195 * no I/O completion operations can occur while we are manipulating 2196 * its associated dependencies. The mutex is held so that other I/O's 2197 * associated with related dependencies do not occur. 2198 */ 2199static void 2200deallocate_dependencies(bp, inodedep) 2201 struct buf *bp; 2202 struct inodedep *inodedep; 2203{ 2204 struct worklist *wk; 2205 struct indirdep *indirdep; 2206 struct allocindir *aip; 2207 struct pagedep *pagedep; 2208 struct dirrem *dirrem; 2209 struct diradd *dap; 2210 int i; 2211 2212 mtx_assert(&lk, MA_OWNED); 2213 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2214 switch (wk->wk_type) { 2215 2216 case D_INDIRDEP: 2217 indirdep = WK_INDIRDEP(wk); 2218 /* 2219 * None of the indirect pointers will ever be visible, 2220 * so they can simply be tossed. GOINGAWAY ensures 2221 * that allocated pointers will be saved in the buffer 2222 * cache until they are freed. Note that they will 2223 * only be able to be found by their physical address 2224 * since the inode mapping the logical address will 2225 * be gone. The save buffer used for the safe copy 2226 * was allocated in setup_allocindir_phase2 using 2227 * the physical address so it could be used for this 2228 * purpose. Hence we swap the safe copy with the real 2229 * copy, allowing the safe copy to be freed and holding 2230 * on to the real copy for later use in indir_trunc. 2231 */ 2232 if (indirdep->ir_state & GOINGAWAY) 2233 panic("deallocate_dependencies: already gone"); 2234 indirdep->ir_state |= GOINGAWAY; 2235 VFSTOUFS(bp->b_vp->v_mount)->um_numindirdeps += 1; 2236 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 2237 free_allocindir(aip, inodedep); 2238 if (bp->b_lblkno >= 0 || 2239 bp->b_blkno != indirdep->ir_savebp->b_lblkno) 2240 panic("deallocate_dependencies: not indir"); 2241 bcopy(bp->b_data, indirdep->ir_savebp->b_data, 2242 bp->b_bcount); 2243 WORKLIST_REMOVE(wk); 2244 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk); 2245 continue; 2246 2247 case D_PAGEDEP: 2248 pagedep = WK_PAGEDEP(wk); 2249 /* 2250 * None of the directory additions will ever be 2251 * visible, so they can simply be tossed. 2252 */ 2253 for (i = 0; i < DAHASHSZ; i++) 2254 while ((dap = 2255 LIST_FIRST(&pagedep->pd_diraddhd[i]))) 2256 free_diradd(dap); 2257 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0) 2258 free_diradd(dap); 2259 /* 2260 * Copy any directory remove dependencies to the list 2261 * to be processed after the zero'ed inode is written. 2262 * If the inode has already been written, then they 2263 * can be dumped directly onto the work list. 2264 */ 2265 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 2266 LIST_REMOVE(dirrem, dm_next); 2267 dirrem->dm_dirinum = pagedep->pd_ino; 2268 if (inodedep == NULL || 2269 (inodedep->id_state & ALLCOMPLETE) == 2270 ALLCOMPLETE) 2271 add_to_worklist(&dirrem->dm_list); 2272 else 2273 WORKLIST_INSERT(&inodedep->id_bufwait, 2274 &dirrem->dm_list); 2275 } 2276 if ((pagedep->pd_state & NEWBLOCK) != 0) { 2277 LIST_FOREACH(wk, &inodedep->id_bufwait, wk_list) 2278 if (wk->wk_type == D_NEWDIRBLK && 2279 WK_NEWDIRBLK(wk)->db_pagedep == 2280 pagedep) 2281 break; 2282 if (wk != NULL) { 2283 WORKLIST_REMOVE(wk); 2284 free_newdirblk(WK_NEWDIRBLK(wk)); 2285 } else 2286 panic("deallocate_dependencies: " 2287 "lost pagedep"); 2288 } 2289 WORKLIST_REMOVE(&pagedep->pd_list); 2290 LIST_REMOVE(pagedep, pd_hash); 2291 WORKITEM_FREE(pagedep, D_PAGEDEP); 2292 continue; 2293 2294 case D_ALLOCINDIR: 2295 free_allocindir(WK_ALLOCINDIR(wk), inodedep); 2296 continue; 2297 2298 case D_ALLOCDIRECT: 2299 case D_INODEDEP: 2300 panic("deallocate_dependencies: Unexpected type %s", 2301 TYPENAME(wk->wk_type)); 2302 /* NOTREACHED */ 2303 2304 default: 2305 panic("deallocate_dependencies: Unknown type %s", 2306 TYPENAME(wk->wk_type)); 2307 /* NOTREACHED */ 2308 } 2309 } 2310} 2311 2312/* 2313 * Free an allocdirect. Generate a new freefrag work request if appropriate. 2314 * This routine must be called with splbio interrupts blocked. 2315 */ 2316static void 2317free_allocdirect(adphead, adp, delay) 2318 struct allocdirectlst *adphead; 2319 struct allocdirect *adp; 2320 int delay; 2321{ 2322 struct newdirblk *newdirblk; 2323 struct worklist *wk; 2324 2325 mtx_assert(&lk, MA_OWNED); 2326 if ((adp->ad_state & DEPCOMPLETE) == 0) 2327 LIST_REMOVE(adp, ad_deps); 2328 TAILQ_REMOVE(adphead, adp, ad_next); 2329 if ((adp->ad_state & COMPLETE) == 0) 2330 WORKLIST_REMOVE(&adp->ad_list); 2331 if (adp->ad_freefrag != NULL) { 2332 if (delay) 2333 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 2334 &adp->ad_freefrag->ff_list); 2335 else 2336 add_to_worklist(&adp->ad_freefrag->ff_list); 2337 } 2338 if ((wk = LIST_FIRST(&adp->ad_newdirblk)) != NULL) { 2339 newdirblk = WK_NEWDIRBLK(wk); 2340 WORKLIST_REMOVE(&newdirblk->db_list); 2341 if (LIST_FIRST(&adp->ad_newdirblk) != NULL) 2342 panic("free_allocdirect: extra newdirblk"); 2343 if (delay) 2344 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 2345 &newdirblk->db_list); 2346 else 2347 free_newdirblk(newdirblk); 2348 } 2349 WORKITEM_FREE(adp, D_ALLOCDIRECT); 2350} 2351 2352/* 2353 * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep. 2354 * This routine must be called with splbio interrupts blocked. 2355 */ 2356static void 2357free_newdirblk(newdirblk) 2358 struct newdirblk *newdirblk; 2359{ 2360 struct pagedep *pagedep; 2361 struct diradd *dap; 2362 int i; 2363 2364 mtx_assert(&lk, MA_OWNED); 2365 /* 2366 * If the pagedep is still linked onto the directory buffer 2367 * dependency chain, then some of the entries on the 2368 * pd_pendinghd list may not be committed to disk yet. In 2369 * this case, we will simply clear the NEWBLOCK flag and 2370 * let the pd_pendinghd list be processed when the pagedep 2371 * is next written. If the pagedep is no longer on the buffer 2372 * dependency chain, then all the entries on the pd_pending 2373 * list are committed to disk and we can free them here. 2374 */ 2375 pagedep = newdirblk->db_pagedep; 2376 pagedep->pd_state &= ~NEWBLOCK; 2377 if ((pagedep->pd_state & ONWORKLIST) == 0) 2378 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 2379 free_diradd(dap); 2380 /* 2381 * If no dependencies remain, the pagedep will be freed. 2382 */ 2383 for (i = 0; i < DAHASHSZ; i++) 2384 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL) 2385 break; 2386 if (i == DAHASHSZ && (pagedep->pd_state & ONWORKLIST) == 0) { 2387 LIST_REMOVE(pagedep, pd_hash); 2388 WORKITEM_FREE(pagedep, D_PAGEDEP); 2389 } 2390 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 2391} 2392 2393/* 2394 * Prepare an inode to be freed. The actual free operation is not 2395 * done until the zero'ed inode has been written to disk. 2396 */ 2397void 2398softdep_freefile(pvp, ino, mode) 2399 struct vnode *pvp; 2400 ino_t ino; 2401 int mode; 2402{ 2403 struct inode *ip = VTOI(pvp); 2404 struct inodedep *inodedep; 2405 struct freefile *freefile; 2406 2407 /* 2408 * This sets up the inode de-allocation dependency. 2409 */ 2410 MALLOC(freefile, struct freefile *, sizeof(struct freefile), 2411 M_FREEFILE, M_SOFTDEP_FLAGS); 2412 freefile->fx_list.wk_type = D_FREEFILE; 2413 freefile->fx_list.wk_state = 0; 2414 freefile->fx_mode = mode; 2415 freefile->fx_oldinum = ino; 2416 freefile->fx_devvp = ip->i_devvp; 2417 freefile->fx_mnt = ITOV(ip)->v_mount; 2418 if ((ip->i_flag & IN_SPACECOUNTED) == 0) { 2419 UFS_LOCK(ip->i_ump); 2420 ip->i_fs->fs_pendinginodes += 1; 2421 UFS_UNLOCK(ip->i_ump); 2422 } 2423 2424 /* 2425 * If the inodedep does not exist, then the zero'ed inode has 2426 * been written to disk. If the allocated inode has never been 2427 * written to disk, then the on-disk inode is zero'ed. In either 2428 * case we can free the file immediately. 2429 */ 2430 ACQUIRE_LOCK(&lk); 2431 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 || 2432 check_inode_unwritten(inodedep)) { 2433 FREE_LOCK(&lk); 2434 handle_workitem_freefile(freefile); 2435 return; 2436 } 2437 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 2438 FREE_LOCK(&lk); 2439} 2440 2441/* 2442 * Check to see if an inode has never been written to disk. If 2443 * so free the inodedep and return success, otherwise return failure. 2444 * This routine must be called with splbio interrupts blocked. 2445 * 2446 * If we still have a bitmap dependency, then the inode has never 2447 * been written to disk. Drop the dependency as it is no longer 2448 * necessary since the inode is being deallocated. We set the 2449 * ALLCOMPLETE flags since the bitmap now properly shows that the 2450 * inode is not allocated. Even if the inode is actively being 2451 * written, it has been rolled back to its zero'ed state, so we 2452 * are ensured that a zero inode is what is on the disk. For short 2453 * lived files, this change will usually result in removing all the 2454 * dependencies from the inode so that it can be freed immediately. 2455 */ 2456static int 2457check_inode_unwritten(inodedep) 2458 struct inodedep *inodedep; 2459{ 2460 2461 mtx_assert(&lk, MA_OWNED); 2462 if ((inodedep->id_state & DEPCOMPLETE) != 0 || 2463 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2464 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2465 LIST_FIRST(&inodedep->id_inowait) != NULL || 2466 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2467 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2468 TAILQ_FIRST(&inodedep->id_extupdt) != NULL || 2469 TAILQ_FIRST(&inodedep->id_newextupdt) != NULL || 2470 inodedep->id_nlinkdelta != 0) 2471 return (0); 2472 inodedep->id_state |= ALLCOMPLETE; 2473 LIST_REMOVE(inodedep, id_deps); 2474 inodedep->id_buf = NULL; 2475 if (inodedep->id_state & ONWORKLIST) 2476 WORKLIST_REMOVE(&inodedep->id_list); 2477 if (inodedep->id_savedino1 != NULL) { 2478 FREE(inodedep->id_savedino1, M_SAVEDINO); 2479 inodedep->id_savedino1 = NULL; 2480 } 2481 if (free_inodedep(inodedep) == 0) 2482 panic("check_inode_unwritten: busy inode"); 2483 return (1); 2484} 2485 2486/* 2487 * Try to free an inodedep structure. Return 1 if it could be freed. 2488 */ 2489static int 2490free_inodedep(inodedep) 2491 struct inodedep *inodedep; 2492{ 2493 2494 mtx_assert(&lk, MA_OWNED); 2495 if ((inodedep->id_state & ONWORKLIST) != 0 || 2496 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 2497 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2498 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2499 LIST_FIRST(&inodedep->id_inowait) != NULL || 2500 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2501 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2502 TAILQ_FIRST(&inodedep->id_extupdt) != NULL || 2503 TAILQ_FIRST(&inodedep->id_newextupdt) != NULL || 2504 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino1 != NULL) 2505 return (0); 2506 LIST_REMOVE(inodedep, id_hash); 2507 WORKITEM_FREE(inodedep, D_INODEDEP); 2508 num_inodedep -= 1; 2509 return (1); 2510} 2511 2512/* 2513 * This workitem routine performs the block de-allocation. 2514 * The workitem is added to the pending list after the updated 2515 * inode block has been written to disk. As mentioned above, 2516 * checks regarding the number of blocks de-allocated (compared 2517 * to the number of blocks allocated for the file) are also 2518 * performed in this function. 2519 */ 2520static void 2521handle_workitem_freeblocks(freeblks, flags) 2522 struct freeblks *freeblks; 2523 int flags; 2524{ 2525 struct inode *ip; 2526 struct vnode *vp; 2527 struct fs *fs; 2528 struct ufsmount *ump; 2529 int i, nblocks, level, bsize; 2530 ufs2_daddr_t bn, blocksreleased = 0; 2531 int error, allerror = 0; 2532 ufs_lbn_t baselbns[NIADDR], tmpval; 2533 int fs_pendingblocks; 2534 2535 ump = VFSTOUFS(freeblks->fb_mnt); 2536 fs = ump->um_fs; 2537 fs_pendingblocks = 0; 2538 tmpval = 1; 2539 baselbns[0] = NDADDR; 2540 for (i = 1; i < NIADDR; i++) { 2541 tmpval *= NINDIR(fs); 2542 baselbns[i] = baselbns[i - 1] + tmpval; 2543 } 2544 nblocks = btodb(fs->fs_bsize); 2545 blocksreleased = 0; 2546 /* 2547 * Release all extended attribute blocks or frags. 2548 */ 2549 if (freeblks->fb_oldextsize > 0) { 2550 for (i = (NXADDR - 1); i >= 0; i--) { 2551 if ((bn = freeblks->fb_eblks[i]) == 0) 2552 continue; 2553 bsize = sblksize(fs, freeblks->fb_oldextsize, i); 2554 ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, bsize, 2555 freeblks->fb_previousinum); 2556 blocksreleased += btodb(bsize); 2557 } 2558 } 2559 /* 2560 * Release all data blocks or frags. 2561 */ 2562 if (freeblks->fb_oldsize > 0) { 2563 /* 2564 * Indirect blocks first. 2565 */ 2566 for (level = (NIADDR - 1); level >= 0; level--) { 2567 if ((bn = freeblks->fb_iblks[level]) == 0) 2568 continue; 2569 if ((error = indir_trunc(freeblks, fsbtodb(fs, bn), 2570 level, baselbns[level], &blocksreleased)) == 0) 2571 allerror = error; 2572 ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, 2573 fs->fs_bsize, freeblks->fb_previousinum); 2574 fs_pendingblocks += nblocks; 2575 blocksreleased += nblocks; 2576 } 2577 /* 2578 * All direct blocks or frags. 2579 */ 2580 for (i = (NDADDR - 1); i >= 0; i--) { 2581 if ((bn = freeblks->fb_dblks[i]) == 0) 2582 continue; 2583 bsize = sblksize(fs, freeblks->fb_oldsize, i); 2584 ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, bsize, 2585 freeblks->fb_previousinum); 2586 fs_pendingblocks += btodb(bsize); 2587 blocksreleased += btodb(bsize); 2588 } 2589 } 2590 UFS_LOCK(ump); 2591 fs->fs_pendingblocks -= fs_pendingblocks; 2592 UFS_UNLOCK(ump); 2593 /* 2594 * If we still have not finished background cleanup, then check 2595 * to see if the block count needs to be adjusted. 2596 */ 2597 if (freeblks->fb_chkcnt != blocksreleased && 2598 (fs->fs_flags & FS_UNCLEAN) != 0 && 2599 ffs_vget(freeblks->fb_mnt, freeblks->fb_previousinum, 2600 (flags & LK_NOWAIT) | LK_EXCLUSIVE, &vp) == 0) { 2601 ip = VTOI(vp); 2602 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + \ 2603 freeblks->fb_chkcnt - blocksreleased); 2604 ip->i_flag |= IN_CHANGE; 2605 vput(vp); 2606 } 2607 2608#ifdef DIAGNOSTIC 2609 if (freeblks->fb_chkcnt != blocksreleased && 2610 ((fs->fs_flags & FS_UNCLEAN) == 0 || (flags & LK_NOWAIT) != 0)) 2611 printf("handle_workitem_freeblocks: block count\n"); 2612 if (allerror) 2613 softdep_error("handle_workitem_freeblks", allerror); 2614#endif /* DIAGNOSTIC */ 2615 2616 WORKITEM_FREE(freeblks, D_FREEBLKS); 2617} 2618 2619/* 2620 * Release blocks associated with the inode ip and stored in the indirect 2621 * block dbn. If level is greater than SINGLE, the block is an indirect block 2622 * and recursive calls to indirtrunc must be used to cleanse other indirect 2623 * blocks. 2624 */ 2625static int 2626indir_trunc(freeblks, dbn, level, lbn, countp) 2627 struct freeblks *freeblks; 2628 ufs2_daddr_t dbn; 2629 int level; 2630 ufs_lbn_t lbn; 2631 ufs2_daddr_t *countp; 2632{ 2633 struct buf *bp; 2634 struct fs *fs; 2635 struct worklist *wk; 2636 struct indirdep *indirdep; 2637 struct ufsmount *ump; 2638 ufs1_daddr_t *bap1 = 0; 2639 ufs2_daddr_t nb, *bap2 = 0; 2640 ufs_lbn_t lbnadd; 2641 int i, nblocks, ufs1fmt; 2642 int error, allerror = 0; 2643 int fs_pendingblocks; 2644 2645 ump = VFSTOUFS(freeblks->fb_mnt); 2646 fs = ump->um_fs; 2647 fs_pendingblocks = 0; 2648 lbnadd = 1; 2649 for (i = level; i > 0; i--) 2650 lbnadd *= NINDIR(fs); 2651 /* 2652 * Get buffer of block pointers to be freed. This routine is not 2653 * called until the zero'ed inode has been written, so it is safe 2654 * to free blocks as they are encountered. Because the inode has 2655 * been zero'ed, calls to bmap on these blocks will fail. So, we 2656 * have to use the on-disk address and the block device for the 2657 * filesystem to look them up. If the file was deleted before its 2658 * indirect blocks were all written to disk, the routine that set 2659 * us up (deallocate_dependencies) will have arranged to leave 2660 * a complete copy of the indirect block in memory for our use. 2661 * Otherwise we have to read the blocks in from the disk. 2662 */ 2663#ifdef notyet 2664 bp = getblk(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 0, 0, 2665 GB_NOCREAT); 2666#else 2667 bp = incore(&freeblks->fb_devvp->v_bufobj, dbn); 2668#endif 2669 ACQUIRE_LOCK(&lk); 2670 if (bp != NULL && (wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2671 if (wk->wk_type != D_INDIRDEP || 2672 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp || 2673 (indirdep->ir_state & GOINGAWAY) == 0) 2674 panic("indir_trunc: lost indirdep"); 2675 WORKLIST_REMOVE(wk); 2676 WORKITEM_FREE(indirdep, D_INDIRDEP); 2677 if (LIST_FIRST(&bp->b_dep) != NULL) 2678 panic("indir_trunc: dangling dep"); 2679 VFSTOUFS(freeblks->fb_mnt)->um_numindirdeps -= 1; 2680 FREE_LOCK(&lk); 2681 } else { 2682#ifdef notyet 2683 if (bp) 2684 brelse(bp); 2685#endif 2686 FREE_LOCK(&lk); 2687 error = bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 2688 NOCRED, &bp); 2689 if (error) { 2690 brelse(bp); 2691 return (error); 2692 } 2693 } 2694 /* 2695 * Recursively free indirect blocks. 2696 */ 2697 if (VFSTOUFS(freeblks->fb_mnt)->um_fstype == UFS1) { 2698 ufs1fmt = 1; 2699 bap1 = (ufs1_daddr_t *)bp->b_data; 2700 } else { 2701 ufs1fmt = 0; 2702 bap2 = (ufs2_daddr_t *)bp->b_data; 2703 } 2704 nblocks = btodb(fs->fs_bsize); 2705 for (i = NINDIR(fs) - 1; i >= 0; i--) { 2706 if (ufs1fmt) 2707 nb = bap1[i]; 2708 else 2709 nb = bap2[i]; 2710 if (nb == 0) 2711 continue; 2712 if (level != 0) { 2713 if ((error = indir_trunc(freeblks, fsbtodb(fs, nb), 2714 level - 1, lbn + (i * lbnadd), countp)) != 0) 2715 allerror = error; 2716 } 2717 ffs_blkfree(ump, fs, freeblks->fb_devvp, nb, fs->fs_bsize, 2718 freeblks->fb_previousinum); 2719 fs_pendingblocks += nblocks; 2720 *countp += nblocks; 2721 } 2722 UFS_LOCK(ump); 2723 fs->fs_pendingblocks -= fs_pendingblocks; 2724 UFS_UNLOCK(ump); 2725 bp->b_flags |= B_INVAL | B_NOCACHE; 2726 brelse(bp); 2727 return (allerror); 2728} 2729 2730/* 2731 * Free an allocindir. 2732 * This routine must be called with splbio interrupts blocked. 2733 */ 2734static void 2735free_allocindir(aip, inodedep) 2736 struct allocindir *aip; 2737 struct inodedep *inodedep; 2738{ 2739 struct freefrag *freefrag; 2740 2741 mtx_assert(&lk, MA_OWNED); 2742 if ((aip->ai_state & DEPCOMPLETE) == 0) 2743 LIST_REMOVE(aip, ai_deps); 2744 if (aip->ai_state & ONWORKLIST) 2745 WORKLIST_REMOVE(&aip->ai_list); 2746 LIST_REMOVE(aip, ai_next); 2747 if ((freefrag = aip->ai_freefrag) != NULL) { 2748 if (inodedep == NULL) 2749 add_to_worklist(&freefrag->ff_list); 2750 else 2751 WORKLIST_INSERT(&inodedep->id_bufwait, 2752 &freefrag->ff_list); 2753 } 2754 WORKITEM_FREE(aip, D_ALLOCINDIR); 2755} 2756 2757/* 2758 * Directory entry addition dependencies. 2759 * 2760 * When adding a new directory entry, the inode (with its incremented link 2761 * count) must be written to disk before the directory entry's pointer to it. 2762 * Also, if the inode is newly allocated, the corresponding freemap must be 2763 * updated (on disk) before the directory entry's pointer. These requirements 2764 * are met via undo/redo on the directory entry's pointer, which consists 2765 * simply of the inode number. 2766 * 2767 * As directory entries are added and deleted, the free space within a 2768 * directory block can become fragmented. The ufs filesystem will compact 2769 * a fragmented directory block to make space for a new entry. When this 2770 * occurs, the offsets of previously added entries change. Any "diradd" 2771 * dependency structures corresponding to these entries must be updated with 2772 * the new offsets. 2773 */ 2774 2775/* 2776 * This routine is called after the in-memory inode's link 2777 * count has been incremented, but before the directory entry's 2778 * pointer to the inode has been set. 2779 */ 2780int 2781softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 2782 struct buf *bp; /* buffer containing directory block */ 2783 struct inode *dp; /* inode for directory */ 2784 off_t diroffset; /* offset of new entry in directory */ 2785 ino_t newinum; /* inode referenced by new directory entry */ 2786 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 2787 int isnewblk; /* entry is in a newly allocated block */ 2788{ 2789 int offset; /* offset of new entry within directory block */ 2790 ufs_lbn_t lbn; /* block in directory containing new entry */ 2791 struct fs *fs; 2792 struct diradd *dap; 2793 struct allocdirect *adp; 2794 struct pagedep *pagedep; 2795 struct inodedep *inodedep; 2796 struct newdirblk *newdirblk = 0; 2797 struct mkdir *mkdir1, *mkdir2; 2798 2799 /* 2800 * Whiteouts have no dependencies. 2801 */ 2802 if (newinum == WINO) { 2803 if (newdirbp != NULL) 2804 bdwrite(newdirbp); 2805 return (0); 2806 } 2807 2808 fs = dp->i_fs; 2809 lbn = lblkno(fs, diroffset); 2810 offset = blkoff(fs, diroffset); 2811 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, 2812 M_SOFTDEP_FLAGS|M_ZERO); 2813 dap->da_list.wk_type = D_DIRADD; 2814 dap->da_offset = offset; 2815 dap->da_newinum = newinum; 2816 dap->da_state = ATTACHED; 2817 if (isnewblk && lbn < NDADDR && fragoff(fs, diroffset) == 0) { 2818 MALLOC(newdirblk, struct newdirblk *, sizeof(struct newdirblk), 2819 M_NEWDIRBLK, M_SOFTDEP_FLAGS); 2820 newdirblk->db_list.wk_type = D_NEWDIRBLK; 2821 newdirblk->db_state = 0; 2822 } 2823 if (newdirbp == NULL) { 2824 dap->da_state |= DEPCOMPLETE; 2825 ACQUIRE_LOCK(&lk); 2826 } else { 2827 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 2828 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2829 M_SOFTDEP_FLAGS); 2830 mkdir1->md_list.wk_type = D_MKDIR; 2831 mkdir1->md_state = MKDIR_BODY; 2832 mkdir1->md_diradd = dap; 2833 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2834 M_SOFTDEP_FLAGS); 2835 mkdir2->md_list.wk_type = D_MKDIR; 2836 mkdir2->md_state = MKDIR_PARENT; 2837 mkdir2->md_diradd = dap; 2838 /* 2839 * Dependency on "." and ".." being written to disk. 2840 */ 2841 mkdir1->md_buf = newdirbp; 2842 ACQUIRE_LOCK(&lk); 2843 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 2844 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list); 2845 FREE_LOCK(&lk); 2846 bdwrite(newdirbp); 2847 /* 2848 * Dependency on link count increase for parent directory 2849 */ 2850 ACQUIRE_LOCK(&lk); 2851 if (inodedep_lookup(fs, dp->i_number, 0, &inodedep) == 0 2852 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2853 dap->da_state &= ~MKDIR_PARENT; 2854 WORKITEM_FREE(mkdir2, D_MKDIR); 2855 } else { 2856 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 2857 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 2858 } 2859 } 2860 /* 2861 * Link into parent directory pagedep to await its being written. 2862 */ 2863 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2864 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2865 dap->da_pagedep = pagedep; 2866 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 2867 da_pdlist); 2868 /* 2869 * Link into its inodedep. Put it on the id_bufwait list if the inode 2870 * is not yet written. If it is written, do the post-inode write 2871 * processing to put it on the id_pendinghd list. 2872 */ 2873 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep); 2874 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 2875 diradd_inode_written(dap, inodedep); 2876 else 2877 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2878 if (isnewblk) { 2879 /* 2880 * Directories growing into indirect blocks are rare 2881 * enough and the frequency of new block allocation 2882 * in those cases even more rare, that we choose not 2883 * to bother tracking them. Rather we simply force the 2884 * new directory entry to disk. 2885 */ 2886 if (lbn >= NDADDR) { 2887 FREE_LOCK(&lk); 2888 /* 2889 * We only have a new allocation when at the 2890 * beginning of a new block, not when we are 2891 * expanding into an existing block. 2892 */ 2893 if (blkoff(fs, diroffset) == 0) 2894 return (1); 2895 return (0); 2896 } 2897 /* 2898 * We only have a new allocation when at the beginning 2899 * of a new fragment, not when we are expanding into an 2900 * existing fragment. Also, there is nothing to do if we 2901 * are already tracking this block. 2902 */ 2903 if (fragoff(fs, diroffset) != 0) { 2904 FREE_LOCK(&lk); 2905 return (0); 2906 } 2907 if ((pagedep->pd_state & NEWBLOCK) != 0) { 2908 FREE_LOCK(&lk); 2909 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 2910 return (0); 2911 } 2912 /* 2913 * Find our associated allocdirect and have it track us. 2914 */ 2915 if (inodedep_lookup(fs, dp->i_number, 0, &inodedep) == 0) 2916 panic("softdep_setup_directory_add: lost inodedep"); 2917 adp = TAILQ_LAST(&inodedep->id_newinoupdt, allocdirectlst); 2918 if (adp == NULL || adp->ad_lbn != lbn) 2919 panic("softdep_setup_directory_add: lost entry"); 2920 pagedep->pd_state |= NEWBLOCK; 2921 newdirblk->db_pagedep = pagedep; 2922 WORKLIST_INSERT(&adp->ad_newdirblk, &newdirblk->db_list); 2923 } 2924 FREE_LOCK(&lk); 2925 return (0); 2926} 2927 2928/* 2929 * This procedure is called to change the offset of a directory 2930 * entry when compacting a directory block which must be owned 2931 * exclusively by the caller. Note that the actual entry movement 2932 * must be done in this procedure to ensure that no I/O completions 2933 * occur while the move is in progress. 2934 */ 2935void 2936softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize) 2937 struct inode *dp; /* inode for directory */ 2938 caddr_t base; /* address of dp->i_offset */ 2939 caddr_t oldloc; /* address of old directory location */ 2940 caddr_t newloc; /* address of new directory location */ 2941 int entrysize; /* size of directory entry */ 2942{ 2943 int offset, oldoffset, newoffset; 2944 struct pagedep *pagedep; 2945 struct diradd *dap; 2946 ufs_lbn_t lbn; 2947 2948 ACQUIRE_LOCK(&lk); 2949 lbn = lblkno(dp->i_fs, dp->i_offset); 2950 offset = blkoff(dp->i_fs, dp->i_offset); 2951 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) 2952 goto done; 2953 oldoffset = offset + (oldloc - base); 2954 newoffset = offset + (newloc - base); 2955 2956 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) { 2957 if (dap->da_offset != oldoffset) 2958 continue; 2959 dap->da_offset = newoffset; 2960 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) 2961 break; 2962 LIST_REMOVE(dap, da_pdlist); 2963 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], 2964 dap, da_pdlist); 2965 break; 2966 } 2967 if (dap == NULL) { 2968 2969 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) { 2970 if (dap->da_offset == oldoffset) { 2971 dap->da_offset = newoffset; 2972 break; 2973 } 2974 } 2975 } 2976done: 2977 bcopy(oldloc, newloc, entrysize); 2978 FREE_LOCK(&lk); 2979} 2980 2981/* 2982 * Free a diradd dependency structure. This routine must be called 2983 * with splbio interrupts blocked. 2984 */ 2985static void 2986free_diradd(dap) 2987 struct diradd *dap; 2988{ 2989 struct dirrem *dirrem; 2990 struct pagedep *pagedep; 2991 struct inodedep *inodedep; 2992 struct mkdir *mkdir, *nextmd; 2993 2994 mtx_assert(&lk, MA_OWNED); 2995 WORKLIST_REMOVE(&dap->da_list); 2996 LIST_REMOVE(dap, da_pdlist); 2997 if ((dap->da_state & DIRCHG) == 0) { 2998 pagedep = dap->da_pagedep; 2999 } else { 3000 dirrem = dap->da_previous; 3001 pagedep = dirrem->dm_pagedep; 3002 dirrem->dm_dirinum = pagedep->pd_ino; 3003 add_to_worklist(&dirrem->dm_list); 3004 } 3005 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum, 3006 0, &inodedep) != 0) 3007 (void) free_inodedep(inodedep); 3008 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 3009 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 3010 nextmd = LIST_NEXT(mkdir, md_mkdirs); 3011 if (mkdir->md_diradd != dap) 3012 continue; 3013 dap->da_state &= ~mkdir->md_state; 3014 WORKLIST_REMOVE(&mkdir->md_list); 3015 LIST_REMOVE(mkdir, md_mkdirs); 3016 WORKITEM_FREE(mkdir, D_MKDIR); 3017 } 3018 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 3019 panic("free_diradd: unfound ref"); 3020 } 3021 WORKITEM_FREE(dap, D_DIRADD); 3022} 3023 3024/* 3025 * Directory entry removal dependencies. 3026 * 3027 * When removing a directory entry, the entry's inode pointer must be 3028 * zero'ed on disk before the corresponding inode's link count is decremented 3029 * (possibly freeing the inode for re-use). This dependency is handled by 3030 * updating the directory entry but delaying the inode count reduction until 3031 * after the directory block has been written to disk. After this point, the 3032 * inode count can be decremented whenever it is convenient. 3033 */ 3034 3035/* 3036 * This routine should be called immediately after removing 3037 * a directory entry. The inode's link count should not be 3038 * decremented by the calling procedure -- the soft updates 3039 * code will do this task when it is safe. 3040 */ 3041void 3042softdep_setup_remove(bp, dp, ip, isrmdir) 3043 struct buf *bp; /* buffer containing directory block */ 3044 struct inode *dp; /* inode for the directory being modified */ 3045 struct inode *ip; /* inode for directory entry being removed */ 3046 int isrmdir; /* indicates if doing RMDIR */ 3047{ 3048 struct dirrem *dirrem, *prevdirrem; 3049 3050 /* 3051 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 3052 */ 3053 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 3054 3055 /* 3056 * If the COMPLETE flag is clear, then there were no active 3057 * entries and we want to roll back to a zeroed entry until 3058 * the new inode is committed to disk. If the COMPLETE flag is 3059 * set then we have deleted an entry that never made it to 3060 * disk. If the entry we deleted resulted from a name change, 3061 * then the old name still resides on disk. We cannot delete 3062 * its inode (returned to us in prevdirrem) until the zeroed 3063 * directory entry gets to disk. The new inode has never been 3064 * referenced on the disk, so can be deleted immediately. 3065 */ 3066 if ((dirrem->dm_state & COMPLETE) == 0) { 3067 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 3068 dm_next); 3069 FREE_LOCK(&lk); 3070 } else { 3071 if (prevdirrem != NULL) 3072 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 3073 prevdirrem, dm_next); 3074 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 3075 FREE_LOCK(&lk); 3076 handle_workitem_remove(dirrem, NULL); 3077 } 3078} 3079 3080/* 3081 * Allocate a new dirrem if appropriate and return it along with 3082 * its associated pagedep. Called without a lock, returns with lock. 3083 */ 3084static long num_dirrem; /* number of dirrem allocated */ 3085static struct dirrem * 3086newdirrem(bp, dp, ip, isrmdir, prevdirremp) 3087 struct buf *bp; /* buffer containing directory block */ 3088 struct inode *dp; /* inode for the directory being modified */ 3089 struct inode *ip; /* inode for directory entry being removed */ 3090 int isrmdir; /* indicates if doing RMDIR */ 3091 struct dirrem **prevdirremp; /* previously referenced inode, if any */ 3092{ 3093 int offset; 3094 ufs_lbn_t lbn; 3095 struct diradd *dap; 3096 struct dirrem *dirrem; 3097 struct pagedep *pagedep; 3098 3099 /* 3100 * Whiteouts have no deletion dependencies. 3101 */ 3102 if (ip == NULL) 3103 panic("newdirrem: whiteout"); 3104 /* 3105 * If we are over our limit, try to improve the situation. 3106 * Limiting the number of dirrem structures will also limit 3107 * the number of freefile and freeblks structures. 3108 */ 3109 ACQUIRE_LOCK(&lk); 3110 if (num_dirrem > max_softdeps / 2) 3111 (void) request_cleanup(FLUSH_REMOVE); 3112 num_dirrem += 1; 3113 FREE_LOCK(&lk); 3114 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem), 3115 M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO); 3116 dirrem->dm_list.wk_type = D_DIRREM; 3117 dirrem->dm_state = isrmdir ? RMDIR : 0; 3118 dirrem->dm_mnt = ITOV(ip)->v_mount; 3119 dirrem->dm_oldinum = ip->i_number; 3120 *prevdirremp = NULL; 3121 3122 ACQUIRE_LOCK(&lk); 3123 lbn = lblkno(dp->i_fs, dp->i_offset); 3124 offset = blkoff(dp->i_fs, dp->i_offset); 3125 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 3126 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 3127 dirrem->dm_pagedep = pagedep; 3128 /* 3129 * Check for a diradd dependency for the same directory entry. 3130 * If present, then both dependencies become obsolete and can 3131 * be de-allocated. Check for an entry on both the pd_dirraddhd 3132 * list and the pd_pendinghd list. 3133 */ 3134 3135 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 3136 if (dap->da_offset == offset) 3137 break; 3138 if (dap == NULL) { 3139 3140 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 3141 if (dap->da_offset == offset) 3142 break; 3143 if (dap == NULL) 3144 return (dirrem); 3145 } 3146 /* 3147 * Must be ATTACHED at this point. 3148 */ 3149 if ((dap->da_state & ATTACHED) == 0) 3150 panic("newdirrem: not ATTACHED"); 3151 if (dap->da_newinum != ip->i_number) 3152 panic("newdirrem: inum %d should be %d", 3153 ip->i_number, dap->da_newinum); 3154 /* 3155 * If we are deleting a changed name that never made it to disk, 3156 * then return the dirrem describing the previous inode (which 3157 * represents the inode currently referenced from this entry on disk). 3158 */ 3159 if ((dap->da_state & DIRCHG) != 0) { 3160 *prevdirremp = dap->da_previous; 3161 dap->da_state &= ~DIRCHG; 3162 dap->da_pagedep = pagedep; 3163 } 3164 /* 3165 * We are deleting an entry that never made it to disk. 3166 * Mark it COMPLETE so we can delete its inode immediately. 3167 */ 3168 dirrem->dm_state |= COMPLETE; 3169 free_diradd(dap); 3170 return (dirrem); 3171} 3172 3173/* 3174 * Directory entry change dependencies. 3175 * 3176 * Changing an existing directory entry requires that an add operation 3177 * be completed first followed by a deletion. The semantics for the addition 3178 * are identical to the description of adding a new entry above except 3179 * that the rollback is to the old inode number rather than zero. Once 3180 * the addition dependency is completed, the removal is done as described 3181 * in the removal routine above. 3182 */ 3183 3184/* 3185 * This routine should be called immediately after changing 3186 * a directory entry. The inode's link count should not be 3187 * decremented by the calling procedure -- the soft updates 3188 * code will perform this task when it is safe. 3189 */ 3190void 3191softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 3192 struct buf *bp; /* buffer containing directory block */ 3193 struct inode *dp; /* inode for the directory being modified */ 3194 struct inode *ip; /* inode for directory entry being removed */ 3195 ino_t newinum; /* new inode number for changed entry */ 3196 int isrmdir; /* indicates if doing RMDIR */ 3197{ 3198 int offset; 3199 struct diradd *dap = NULL; 3200 struct dirrem *dirrem, *prevdirrem; 3201 struct pagedep *pagedep; 3202 struct inodedep *inodedep; 3203 3204 offset = blkoff(dp->i_fs, dp->i_offset); 3205 3206 /* 3207 * Whiteouts do not need diradd dependencies. 3208 */ 3209 if (newinum != WINO) { 3210 MALLOC(dap, struct diradd *, sizeof(struct diradd), 3211 M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO); 3212 dap->da_list.wk_type = D_DIRADD; 3213 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 3214 dap->da_offset = offset; 3215 dap->da_newinum = newinum; 3216 } 3217 3218 /* 3219 * Allocate a new dirrem and ACQUIRE_LOCK. 3220 */ 3221 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 3222 pagedep = dirrem->dm_pagedep; 3223 /* 3224 * The possible values for isrmdir: 3225 * 0 - non-directory file rename 3226 * 1 - directory rename within same directory 3227 * inum - directory rename to new directory of given inode number 3228 * When renaming to a new directory, we are both deleting and 3229 * creating a new directory entry, so the link count on the new 3230 * directory should not change. Thus we do not need the followup 3231 * dirrem which is usually done in handle_workitem_remove. We set 3232 * the DIRCHG flag to tell handle_workitem_remove to skip the 3233 * followup dirrem. 3234 */ 3235 if (isrmdir > 1) 3236 dirrem->dm_state |= DIRCHG; 3237 3238 /* 3239 * Whiteouts have no additional dependencies, 3240 * so just put the dirrem on the correct list. 3241 */ 3242 if (newinum == WINO) { 3243 if ((dirrem->dm_state & COMPLETE) == 0) { 3244 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 3245 dm_next); 3246 } else { 3247 dirrem->dm_dirinum = pagedep->pd_ino; 3248 add_to_worklist(&dirrem->dm_list); 3249 } 3250 FREE_LOCK(&lk); 3251 return; 3252 } 3253 3254 /* 3255 * If the COMPLETE flag is clear, then there were no active 3256 * entries and we want to roll back to the previous inode until 3257 * the new inode is committed to disk. If the COMPLETE flag is 3258 * set, then we have deleted an entry that never made it to disk. 3259 * If the entry we deleted resulted from a name change, then the old 3260 * inode reference still resides on disk. Any rollback that we do 3261 * needs to be to that old inode (returned to us in prevdirrem). If 3262 * the entry we deleted resulted from a create, then there is 3263 * no entry on the disk, so we want to roll back to zero rather 3264 * than the uncommitted inode. In either of the COMPLETE cases we 3265 * want to immediately free the unwritten and unreferenced inode. 3266 */ 3267 if ((dirrem->dm_state & COMPLETE) == 0) { 3268 dap->da_previous = dirrem; 3269 } else { 3270 if (prevdirrem != NULL) { 3271 dap->da_previous = prevdirrem; 3272 } else { 3273 dap->da_state &= ~DIRCHG; 3274 dap->da_pagedep = pagedep; 3275 } 3276 dirrem->dm_dirinum = pagedep->pd_ino; 3277 add_to_worklist(&dirrem->dm_list); 3278 } 3279 /* 3280 * Link into its inodedep. Put it on the id_bufwait list if the inode 3281 * is not yet written. If it is written, do the post-inode write 3282 * processing to put it on the id_pendinghd list. 3283 */ 3284 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 || 3285 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 3286 dap->da_state |= COMPLETE; 3287 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3288 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 3289 } else { 3290 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 3291 dap, da_pdlist); 3292 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 3293 } 3294 FREE_LOCK(&lk); 3295} 3296 3297/* 3298 * Called whenever the link count on an inode is changed. 3299 * It creates an inode dependency so that the new reference(s) 3300 * to the inode cannot be committed to disk until the updated 3301 * inode has been written. 3302 */ 3303void 3304softdep_change_linkcnt(ip) 3305 struct inode *ip; /* the inode with the increased link count */ 3306{ 3307 struct inodedep *inodedep; 3308 3309 ACQUIRE_LOCK(&lk); 3310 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 3311 if (ip->i_nlink < ip->i_effnlink) 3312 panic("softdep_change_linkcnt: bad delta"); 3313 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3314 FREE_LOCK(&lk); 3315} 3316 3317/* 3318 * Called when the effective link count and the reference count 3319 * on an inode drops to zero. At this point there are no names 3320 * referencing the file in the filesystem and no active file 3321 * references. The space associated with the file will be freed 3322 * as soon as the necessary soft dependencies are cleared. 3323 */ 3324void 3325softdep_releasefile(ip) 3326 struct inode *ip; /* inode with the zero effective link count */ 3327{ 3328 struct inodedep *inodedep; 3329 struct fs *fs; 3330 int extblocks; 3331 3332 if (ip->i_effnlink > 0) 3333 panic("softdep_filerelease: file still referenced"); 3334 /* 3335 * We may be called several times as the real reference count 3336 * drops to zero. We only want to account for the space once. 3337 */ 3338 if (ip->i_flag & IN_SPACECOUNTED) 3339 return; 3340 /* 3341 * We have to deactivate a snapshot otherwise copyonwrites may 3342 * add blocks and the cleanup may remove blocks after we have 3343 * tried to account for them. 3344 */ 3345 if ((ip->i_flags & SF_SNAPSHOT) != 0) 3346 ffs_snapremove(ITOV(ip)); 3347 /* 3348 * If we are tracking an nlinkdelta, we have to also remember 3349 * whether we accounted for the freed space yet. 3350 */ 3351 ACQUIRE_LOCK(&lk); 3352 if ((inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep))) 3353 inodedep->id_state |= SPACECOUNTED; 3354 FREE_LOCK(&lk); 3355 fs = ip->i_fs; 3356 extblocks = 0; 3357 if (fs->fs_magic == FS_UFS2_MAGIC) 3358 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 3359 UFS_LOCK(ip->i_ump); 3360 ip->i_fs->fs_pendingblocks += DIP(ip, i_blocks) - extblocks; 3361 ip->i_fs->fs_pendinginodes += 1; 3362 UFS_UNLOCK(ip->i_ump); 3363 ip->i_flag |= IN_SPACECOUNTED; 3364} 3365 3366/* 3367 * This workitem decrements the inode's link count. 3368 * If the link count reaches zero, the file is removed. 3369 */ 3370static void 3371handle_workitem_remove(dirrem, xp) 3372 struct dirrem *dirrem; 3373 struct vnode *xp; 3374{ 3375 struct thread *td = curthread; 3376 struct inodedep *inodedep; 3377 struct vnode *vp; 3378 struct inode *ip; 3379 ino_t oldinum; 3380 int error; 3381 3382 if ((vp = xp) == NULL && 3383 (error = ffs_vget(dirrem->dm_mnt, dirrem->dm_oldinum, LK_EXCLUSIVE, 3384 &vp)) != 0) { 3385 softdep_error("handle_workitem_remove: vget", error); 3386 return; 3387 } 3388 ip = VTOI(vp); 3389 ACQUIRE_LOCK(&lk); 3390 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0) 3391 panic("handle_workitem_remove: lost inodedep"); 3392 /* 3393 * Normal file deletion. 3394 */ 3395 if ((dirrem->dm_state & RMDIR) == 0) { 3396 ip->i_nlink--; 3397 DIP_SET(ip, i_nlink, ip->i_nlink); 3398 ip->i_flag |= IN_CHANGE; 3399 if (ip->i_nlink < ip->i_effnlink) 3400 panic("handle_workitem_remove: bad file delta"); 3401 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3402 num_dirrem -= 1; 3403 FREE_LOCK(&lk); 3404 vput(vp); 3405 WORKITEM_FREE(dirrem, D_DIRREM); 3406 return; 3407 } 3408 /* 3409 * Directory deletion. Decrement reference count for both the 3410 * just deleted parent directory entry and the reference for ".". 3411 * Next truncate the directory to length zero. When the 3412 * truncation completes, arrange to have the reference count on 3413 * the parent decremented to account for the loss of "..". 3414 */ 3415 ip->i_nlink -= 2; 3416 DIP_SET(ip, i_nlink, ip->i_nlink); 3417 ip->i_flag |= IN_CHANGE; 3418 if (ip->i_nlink < ip->i_effnlink) 3419 panic("handle_workitem_remove: bad dir delta"); 3420 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3421 FREE_LOCK(&lk); 3422 if ((error = ffs_truncate(vp, (off_t)0, 0, td->td_ucred, td)) != 0) 3423 softdep_error("handle_workitem_remove: truncate", error); 3424 ACQUIRE_LOCK(&lk); 3425 /* 3426 * Rename a directory to a new parent. Since, we are both deleting 3427 * and creating a new directory entry, the link count on the new 3428 * directory should not change. Thus we skip the followup dirrem. 3429 */ 3430 if (dirrem->dm_state & DIRCHG) { 3431 num_dirrem -= 1; 3432 FREE_LOCK(&lk); 3433 vput(vp); 3434 WORKITEM_FREE(dirrem, D_DIRREM); 3435 return; 3436 } 3437 /* 3438 * If the inodedep does not exist, then the zero'ed inode has 3439 * been written to disk. If the allocated inode has never been 3440 * written to disk, then the on-disk inode is zero'ed. In either 3441 * case we can remove the file immediately. 3442 */ 3443 dirrem->dm_state = 0; 3444 oldinum = dirrem->dm_oldinum; 3445 dirrem->dm_oldinum = dirrem->dm_dirinum; 3446 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 || 3447 check_inode_unwritten(inodedep)) { 3448 FREE_LOCK(&lk); 3449 vput(vp); 3450 handle_workitem_remove(dirrem, NULL); 3451 return; 3452 } 3453 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 3454 FREE_LOCK(&lk); 3455 vput(vp); 3456} 3457 3458/* 3459 * Inode de-allocation dependencies. 3460 * 3461 * When an inode's link count is reduced to zero, it can be de-allocated. We 3462 * found it convenient to postpone de-allocation until after the inode is 3463 * written to disk with its new link count (zero). At this point, all of the 3464 * on-disk inode's block pointers are nullified and, with careful dependency 3465 * list ordering, all dependencies related to the inode will be satisfied and 3466 * the corresponding dependency structures de-allocated. So, if/when the 3467 * inode is reused, there will be no mixing of old dependencies with new 3468 * ones. This artificial dependency is set up by the block de-allocation 3469 * procedure above (softdep_setup_freeblocks) and completed by the 3470 * following procedure. 3471 */ 3472static void 3473handle_workitem_freefile(freefile) 3474 struct freefile *freefile; 3475{ 3476 struct fs *fs; 3477 struct inodedep *idp; 3478 struct ufsmount *ump; 3479 int error; 3480 3481 ump = VFSTOUFS(freefile->fx_mnt); 3482 fs = ump->um_fs; 3483#ifdef DEBUG 3484 ACQUIRE_LOCK(&lk); 3485 error = inodedep_lookup(fs, freefile->fx_oldinum, 0, &idp); 3486 FREE_LOCK(&lk); 3487 if (error) 3488 panic("handle_workitem_freefile: inodedep survived"); 3489#endif 3490 UFS_LOCK(ump); 3491 fs->fs_pendinginodes -= 1; 3492 UFS_UNLOCK(ump); 3493 if ((error = ffs_freefile(VFSTOUFS(freefile->fx_mnt), fs, 3494 freefile->fx_devvp, freefile->fx_oldinum, freefile->fx_mode)) != 0) 3495 softdep_error("handle_workitem_freefile", error); 3496 WORKITEM_FREE(freefile, D_FREEFILE); 3497} 3498 3499/* 3500 * Disk writes. 3501 * 3502 * The dependency structures constructed above are most actively used when file 3503 * system blocks are written to disk. No constraints are placed on when a 3504 * block can be written, but unsatisfied update dependencies are made safe by 3505 * modifying (or replacing) the source memory for the duration of the disk 3506 * write. When the disk write completes, the memory block is again brought 3507 * up-to-date. 3508 * 3509 * In-core inode structure reclamation. 3510 * 3511 * Because there are a finite number of "in-core" inode structures, they are 3512 * reused regularly. By transferring all inode-related dependencies to the 3513 * in-memory inode block and indexing them separately (via "inodedep"s), we 3514 * can allow "in-core" inode structures to be reused at any time and avoid 3515 * any increase in contention. 3516 * 3517 * Called just before entering the device driver to initiate a new disk I/O. 3518 * The buffer must be locked, thus, no I/O completion operations can occur 3519 * while we are manipulating its associated dependencies. 3520 */ 3521static void 3522softdep_disk_io_initiation(bp) 3523 struct buf *bp; /* structure describing disk write to occur */ 3524{ 3525 struct worklist *wk, *nextwk; 3526 struct indirdep *indirdep; 3527 struct inodedep *inodedep; 3528 3529 /* 3530 * We only care about write operations. There should never 3531 * be dependencies for reads. 3532 */ 3533 if (bp->b_iocmd != BIO_WRITE) 3534 panic("softdep_disk_io_initiation: not write"); 3535 ACQUIRE_LOCK(&lk); 3536 /* 3537 * Do any necessary pre-I/O processing. 3538 */ 3539 LIST_FOREACH_SAFE(wk, &bp->b_dep, wk_list, nextwk) { 3540 switch (wk->wk_type) { 3541 3542 case D_PAGEDEP: 3543 initiate_write_filepage(WK_PAGEDEP(wk), bp); 3544 continue; 3545 3546 case D_INODEDEP: 3547 inodedep = WK_INODEDEP(wk); 3548 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) 3549 initiate_write_inodeblock_ufs1(inodedep, bp); 3550 else 3551 initiate_write_inodeblock_ufs2(inodedep, bp); 3552 continue; 3553 3554 case D_INDIRDEP: 3555 indirdep = WK_INDIRDEP(wk); 3556 if (indirdep->ir_state & GOINGAWAY) 3557 panic("disk_io_initiation: indirdep gone"); 3558 /* 3559 * If there are no remaining dependencies, this 3560 * will be writing the real pointers, so the 3561 * dependency can be freed. 3562 */ 3563 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) { 3564 struct buf *bp; 3565 3566 bp = indirdep->ir_savebp; 3567 bp->b_flags |= B_INVAL | B_NOCACHE; 3568 /* inline expand WORKLIST_REMOVE(wk); */ 3569 wk->wk_state &= ~ONWORKLIST; 3570 LIST_REMOVE(wk, wk_list); 3571 WORKITEM_FREE(indirdep, D_INDIRDEP); 3572 FREE_LOCK(&lk); 3573 brelse(bp); 3574 ACQUIRE_LOCK(&lk); 3575 continue; 3576 } 3577 /* 3578 * Replace up-to-date version with safe version. 3579 */ 3580 FREE_LOCK(&lk); 3581 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount, 3582 M_INDIRDEP, M_SOFTDEP_FLAGS); 3583 ACQUIRE_LOCK(&lk); 3584 indirdep->ir_state &= ~ATTACHED; 3585 indirdep->ir_state |= UNDONE; 3586 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 3587 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 3588 bp->b_bcount); 3589 continue; 3590 3591 case D_MKDIR: 3592 case D_BMSAFEMAP: 3593 case D_ALLOCDIRECT: 3594 case D_ALLOCINDIR: 3595 continue; 3596 3597 default: 3598 panic("handle_disk_io_initiation: Unexpected type %s", 3599 TYPENAME(wk->wk_type)); 3600 /* NOTREACHED */ 3601 } 3602 } 3603 FREE_LOCK(&lk); 3604} 3605 3606/* 3607 * Called from within the procedure above to deal with unsatisfied 3608 * allocation dependencies in a directory. The buffer must be locked, 3609 * thus, no I/O completion operations can occur while we are 3610 * manipulating its associated dependencies. 3611 */ 3612static void 3613initiate_write_filepage(pagedep, bp) 3614 struct pagedep *pagedep; 3615 struct buf *bp; 3616{ 3617 struct diradd *dap; 3618 struct direct *ep; 3619 int i; 3620 3621 if (pagedep->pd_state & IOSTARTED) { 3622 /* 3623 * This can only happen if there is a driver that does not 3624 * understand chaining. Here biodone will reissue the call 3625 * to strategy for the incomplete buffers. 3626 */ 3627 printf("initiate_write_filepage: already started\n"); 3628 return; 3629 } 3630 pagedep->pd_state |= IOSTARTED; 3631 for (i = 0; i < DAHASHSZ; i++) { 3632 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 3633 ep = (struct direct *) 3634 ((char *)bp->b_data + dap->da_offset); 3635 if (ep->d_ino != dap->da_newinum) 3636 panic("%s: dir inum %d != new %d", 3637 "initiate_write_filepage", 3638 ep->d_ino, dap->da_newinum); 3639 if (dap->da_state & DIRCHG) 3640 ep->d_ino = dap->da_previous->dm_oldinum; 3641 else 3642 ep->d_ino = 0; 3643 dap->da_state &= ~ATTACHED; 3644 dap->da_state |= UNDONE; 3645 } 3646 } 3647} 3648 3649/* 3650 * Version of initiate_write_inodeblock that handles UFS1 dinodes. 3651 * Note that any bug fixes made to this routine must be done in the 3652 * version found below. 3653 * 3654 * Called from within the procedure above to deal with unsatisfied 3655 * allocation dependencies in an inodeblock. The buffer must be 3656 * locked, thus, no I/O completion operations can occur while we 3657 * are manipulating its associated dependencies. 3658 */ 3659static void 3660initiate_write_inodeblock_ufs1(inodedep, bp) 3661 struct inodedep *inodedep; 3662 struct buf *bp; /* The inode block */ 3663{ 3664 struct allocdirect *adp, *lastadp; 3665 struct ufs1_dinode *dp; 3666 struct fs *fs; 3667 ufs_lbn_t i, prevlbn = 0; 3668 int deplist; 3669 3670 if (inodedep->id_state & IOSTARTED) 3671 panic("initiate_write_inodeblock_ufs1: already started"); 3672 inodedep->id_state |= IOSTARTED; 3673 fs = inodedep->id_fs; 3674 dp = (struct ufs1_dinode *)bp->b_data + 3675 ino_to_fsbo(fs, inodedep->id_ino); 3676 /* 3677 * If the bitmap is not yet written, then the allocated 3678 * inode cannot be written to disk. 3679 */ 3680 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 3681 if (inodedep->id_savedino1 != NULL) 3682 panic("initiate_write_inodeblock_ufs1: I/O underway"); 3683 FREE_LOCK(&lk); 3684 MALLOC(inodedep->id_savedino1, struct ufs1_dinode *, 3685 sizeof(struct ufs1_dinode), M_SAVEDINO, M_SOFTDEP_FLAGS); 3686 ACQUIRE_LOCK(&lk); 3687 *inodedep->id_savedino1 = *dp; 3688 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 3689 return; 3690 } 3691 /* 3692 * If no dependencies, then there is nothing to roll back. 3693 */ 3694 inodedep->id_savedsize = dp->di_size; 3695 inodedep->id_savedextsize = 0; 3696 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 3697 return; 3698 /* 3699 * Set the dependencies to busy. 3700 */ 3701 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3702 adp = TAILQ_NEXT(adp, ad_next)) { 3703#ifdef DIAGNOSTIC 3704 if (deplist != 0 && prevlbn >= adp->ad_lbn) 3705 panic("softdep_write_inodeblock: lbn order"); 3706 prevlbn = adp->ad_lbn; 3707 if (adp->ad_lbn < NDADDR && 3708 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) 3709 panic("%s: direct pointer #%jd mismatch %d != %jd", 3710 "softdep_write_inodeblock", 3711 (intmax_t)adp->ad_lbn, 3712 dp->di_db[adp->ad_lbn], 3713 (intmax_t)adp->ad_newblkno); 3714 if (adp->ad_lbn >= NDADDR && 3715 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) 3716 panic("%s: indirect pointer #%jd mismatch %d != %jd", 3717 "softdep_write_inodeblock", 3718 (intmax_t)adp->ad_lbn - NDADDR, 3719 dp->di_ib[adp->ad_lbn - NDADDR], 3720 (intmax_t)adp->ad_newblkno); 3721 deplist |= 1 << adp->ad_lbn; 3722 if ((adp->ad_state & ATTACHED) == 0) 3723 panic("softdep_write_inodeblock: Unknown state 0x%x", 3724 adp->ad_state); 3725#endif /* DIAGNOSTIC */ 3726 adp->ad_state &= ~ATTACHED; 3727 adp->ad_state |= UNDONE; 3728 } 3729 /* 3730 * The on-disk inode cannot claim to be any larger than the last 3731 * fragment that has been written. Otherwise, the on-disk inode 3732 * might have fragments that were not the last block in the file 3733 * which would corrupt the filesystem. 3734 */ 3735 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3736 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 3737 if (adp->ad_lbn >= NDADDR) 3738 break; 3739 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 3740 /* keep going until hitting a rollback to a frag */ 3741 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 3742 continue; 3743 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 3744 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 3745#ifdef DIAGNOSTIC 3746 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 3747 panic("softdep_write_inodeblock: lost dep1"); 3748#endif /* DIAGNOSTIC */ 3749 dp->di_db[i] = 0; 3750 } 3751 for (i = 0; i < NIADDR; i++) { 3752#ifdef DIAGNOSTIC 3753 if (dp->di_ib[i] != 0 && 3754 (deplist & ((1 << NDADDR) << i)) == 0) 3755 panic("softdep_write_inodeblock: lost dep2"); 3756#endif /* DIAGNOSTIC */ 3757 dp->di_ib[i] = 0; 3758 } 3759 return; 3760 } 3761 /* 3762 * If we have zero'ed out the last allocated block of the file, 3763 * roll back the size to the last currently allocated block. 3764 * We know that this last allocated block is a full-sized as 3765 * we already checked for fragments in the loop above. 3766 */ 3767 if (lastadp != NULL && 3768 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 3769 for (i = lastadp->ad_lbn; i >= 0; i--) 3770 if (dp->di_db[i] != 0) 3771 break; 3772 dp->di_size = (i + 1) * fs->fs_bsize; 3773 } 3774 /* 3775 * The only dependencies are for indirect blocks. 3776 * 3777 * The file size for indirect block additions is not guaranteed. 3778 * Such a guarantee would be non-trivial to achieve. The conventional 3779 * synchronous write implementation also does not make this guarantee. 3780 * Fsck should catch and fix discrepancies. Arguably, the file size 3781 * can be over-estimated without destroying integrity when the file 3782 * moves into the indirect blocks (i.e., is large). If we want to 3783 * postpone fsck, we are stuck with this argument. 3784 */ 3785 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 3786 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 3787} 3788 3789/* 3790 * Version of initiate_write_inodeblock that handles UFS2 dinodes. 3791 * Note that any bug fixes made to this routine must be done in the 3792 * version found above. 3793 * 3794 * Called from within the procedure above to deal with unsatisfied 3795 * allocation dependencies in an inodeblock. The buffer must be 3796 * locked, thus, no I/O completion operations can occur while we 3797 * are manipulating its associated dependencies. 3798 */ 3799static void 3800initiate_write_inodeblock_ufs2(inodedep, bp) 3801 struct inodedep *inodedep; 3802 struct buf *bp; /* The inode block */ 3803{ 3804 struct allocdirect *adp, *lastadp; 3805 struct ufs2_dinode *dp; 3806 struct fs *fs; 3807 ufs_lbn_t i, prevlbn = 0; 3808 int deplist; 3809 3810 if (inodedep->id_state & IOSTARTED) 3811 panic("initiate_write_inodeblock_ufs2: already started"); 3812 inodedep->id_state |= IOSTARTED; 3813 fs = inodedep->id_fs; 3814 dp = (struct ufs2_dinode *)bp->b_data + 3815 ino_to_fsbo(fs, inodedep->id_ino); 3816 /* 3817 * If the bitmap is not yet written, then the allocated 3818 * inode cannot be written to disk. 3819 */ 3820 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 3821 if (inodedep->id_savedino2 != NULL) 3822 panic("initiate_write_inodeblock_ufs2: I/O underway"); 3823 FREE_LOCK(&lk); 3824 MALLOC(inodedep->id_savedino2, struct ufs2_dinode *, 3825 sizeof(struct ufs2_dinode), M_SAVEDINO, M_SOFTDEP_FLAGS); 3826 ACQUIRE_LOCK(&lk); 3827 *inodedep->id_savedino2 = *dp; 3828 bzero((caddr_t)dp, sizeof(struct ufs2_dinode)); 3829 return; 3830 } 3831 /* 3832 * If no dependencies, then there is nothing to roll back. 3833 */ 3834 inodedep->id_savedsize = dp->di_size; 3835 inodedep->id_savedextsize = dp->di_extsize; 3836 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL && 3837 TAILQ_FIRST(&inodedep->id_extupdt) == NULL) 3838 return; 3839 /* 3840 * Set the ext data dependencies to busy. 3841 */ 3842 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 3843 adp = TAILQ_NEXT(adp, ad_next)) { 3844#ifdef DIAGNOSTIC 3845 if (deplist != 0 && prevlbn >= adp->ad_lbn) 3846 panic("softdep_write_inodeblock: lbn order"); 3847 prevlbn = adp->ad_lbn; 3848 if (dp->di_extb[adp->ad_lbn] != adp->ad_newblkno) 3849 panic("%s: direct pointer #%jd mismatch %jd != %jd", 3850 "softdep_write_inodeblock", 3851 (intmax_t)adp->ad_lbn, 3852 (intmax_t)dp->di_extb[adp->ad_lbn], 3853 (intmax_t)adp->ad_newblkno); 3854 deplist |= 1 << adp->ad_lbn; 3855 if ((adp->ad_state & ATTACHED) == 0) 3856 panic("softdep_write_inodeblock: Unknown state 0x%x", 3857 adp->ad_state); 3858#endif /* DIAGNOSTIC */ 3859 adp->ad_state &= ~ATTACHED; 3860 adp->ad_state |= UNDONE; 3861 } 3862 /* 3863 * The on-disk inode cannot claim to be any larger than the last 3864 * fragment that has been written. Otherwise, the on-disk inode 3865 * might have fragments that were not the last block in the ext 3866 * data which would corrupt the filesystem. 3867 */ 3868 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 3869 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 3870 dp->di_extb[adp->ad_lbn] = adp->ad_oldblkno; 3871 /* keep going until hitting a rollback to a frag */ 3872 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 3873 continue; 3874 dp->di_extsize = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 3875 for (i = adp->ad_lbn + 1; i < NXADDR; i++) { 3876#ifdef DIAGNOSTIC 3877 if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0) 3878 panic("softdep_write_inodeblock: lost dep1"); 3879#endif /* DIAGNOSTIC */ 3880 dp->di_extb[i] = 0; 3881 } 3882 lastadp = NULL; 3883 break; 3884 } 3885 /* 3886 * If we have zero'ed out the last allocated block of the ext 3887 * data, roll back the size to the last currently allocated block. 3888 * We know that this last allocated block is a full-sized as 3889 * we already checked for fragments in the loop above. 3890 */ 3891 if (lastadp != NULL && 3892 dp->di_extsize <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 3893 for (i = lastadp->ad_lbn; i >= 0; i--) 3894 if (dp->di_extb[i] != 0) 3895 break; 3896 dp->di_extsize = (i + 1) * fs->fs_bsize; 3897 } 3898 /* 3899 * Set the file data dependencies to busy. 3900 */ 3901 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3902 adp = TAILQ_NEXT(adp, ad_next)) { 3903#ifdef DIAGNOSTIC 3904 if (deplist != 0 && prevlbn >= adp->ad_lbn) 3905 panic("softdep_write_inodeblock: lbn order"); 3906 prevlbn = adp->ad_lbn; 3907 if (adp->ad_lbn < NDADDR && 3908 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) 3909 panic("%s: direct pointer #%jd mismatch %jd != %jd", 3910 "softdep_write_inodeblock", 3911 (intmax_t)adp->ad_lbn, 3912 (intmax_t)dp->di_db[adp->ad_lbn], 3913 (intmax_t)adp->ad_newblkno); 3914 if (adp->ad_lbn >= NDADDR && 3915 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) 3916 panic("%s indirect pointer #%jd mismatch %jd != %jd", 3917 "softdep_write_inodeblock:", 3918 (intmax_t)adp->ad_lbn - NDADDR, 3919 (intmax_t)dp->di_ib[adp->ad_lbn - NDADDR], 3920 (intmax_t)adp->ad_newblkno); 3921 deplist |= 1 << adp->ad_lbn; 3922 if ((adp->ad_state & ATTACHED) == 0) 3923 panic("softdep_write_inodeblock: Unknown state 0x%x", 3924 adp->ad_state); 3925#endif /* DIAGNOSTIC */ 3926 adp->ad_state &= ~ATTACHED; 3927 adp->ad_state |= UNDONE; 3928 } 3929 /* 3930 * The on-disk inode cannot claim to be any larger than the last 3931 * fragment that has been written. Otherwise, the on-disk inode 3932 * might have fragments that were not the last block in the file 3933 * which would corrupt the filesystem. 3934 */ 3935 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3936 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 3937 if (adp->ad_lbn >= NDADDR) 3938 break; 3939 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 3940 /* keep going until hitting a rollback to a frag */ 3941 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 3942 continue; 3943 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 3944 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 3945#ifdef DIAGNOSTIC 3946 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 3947 panic("softdep_write_inodeblock: lost dep2"); 3948#endif /* DIAGNOSTIC */ 3949 dp->di_db[i] = 0; 3950 } 3951 for (i = 0; i < NIADDR; i++) { 3952#ifdef DIAGNOSTIC 3953 if (dp->di_ib[i] != 0 && 3954 (deplist & ((1 << NDADDR) << i)) == 0) 3955 panic("softdep_write_inodeblock: lost dep3"); 3956#endif /* DIAGNOSTIC */ 3957 dp->di_ib[i] = 0; 3958 } 3959 return; 3960 } 3961 /* 3962 * If we have zero'ed out the last allocated block of the file, 3963 * roll back the size to the last currently allocated block. 3964 * We know that this last allocated block is a full-sized as 3965 * we already checked for fragments in the loop above. 3966 */ 3967 if (lastadp != NULL && 3968 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 3969 for (i = lastadp->ad_lbn; i >= 0; i--) 3970 if (dp->di_db[i] != 0) 3971 break; 3972 dp->di_size = (i + 1) * fs->fs_bsize; 3973 } 3974 /* 3975 * The only dependencies are for indirect blocks. 3976 * 3977 * The file size for indirect block additions is not guaranteed. 3978 * Such a guarantee would be non-trivial to achieve. The conventional 3979 * synchronous write implementation also does not make this guarantee. 3980 * Fsck should catch and fix discrepancies. Arguably, the file size 3981 * can be over-estimated without destroying integrity when the file 3982 * moves into the indirect blocks (i.e., is large). If we want to 3983 * postpone fsck, we are stuck with this argument. 3984 */ 3985 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 3986 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 3987} 3988 3989/* 3990 * This routine is called during the completion interrupt 3991 * service routine for a disk write (from the procedure called 3992 * by the device driver to inform the filesystem caches of 3993 * a request completion). It should be called early in this 3994 * procedure, before the block is made available to other 3995 * processes or other routines are called. 3996 */ 3997static void 3998softdep_disk_write_complete(bp) 3999 struct buf *bp; /* describes the completed disk write */ 4000{ 4001 struct worklist *wk; 4002 struct worklist *owk; 4003 struct workhead reattach; 4004 struct newblk *newblk; 4005 struct allocindir *aip; 4006 struct allocdirect *adp; 4007 struct indirdep *indirdep; 4008 struct inodedep *inodedep; 4009 struct bmsafemap *bmsafemap; 4010 4011 /* 4012 * If an error occurred while doing the write, then the data 4013 * has not hit the disk and the dependencies cannot be unrolled. 4014 */ 4015 if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0) 4016 return; 4017 LIST_INIT(&reattach); 4018 /* 4019 * This lock must not be released anywhere in this code segment. 4020 */ 4021 ACQUIRE_LOCK(&lk); 4022 owk = NULL; 4023 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 4024 WORKLIST_REMOVE(wk); 4025 if (wk == owk) 4026 panic("duplicate worklist: %p\n", wk); 4027 owk = wk; 4028 switch (wk->wk_type) { 4029 4030 case D_PAGEDEP: 4031 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 4032 WORKLIST_INSERT(&reattach, wk); 4033 continue; 4034 4035 case D_INODEDEP: 4036 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 4037 WORKLIST_INSERT(&reattach, wk); 4038 continue; 4039 4040 case D_BMSAFEMAP: 4041 bmsafemap = WK_BMSAFEMAP(wk); 4042 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) { 4043 newblk->nb_state |= DEPCOMPLETE; 4044 newblk->nb_bmsafemap = NULL; 4045 LIST_REMOVE(newblk, nb_deps); 4046 } 4047 while ((adp = 4048 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) { 4049 adp->ad_state |= DEPCOMPLETE; 4050 adp->ad_buf = NULL; 4051 LIST_REMOVE(adp, ad_deps); 4052 handle_allocdirect_partdone(adp); 4053 } 4054 while ((aip = 4055 LIST_FIRST(&bmsafemap->sm_allocindirhd))) { 4056 aip->ai_state |= DEPCOMPLETE; 4057 aip->ai_buf = NULL; 4058 LIST_REMOVE(aip, ai_deps); 4059 handle_allocindir_partdone(aip); 4060 } 4061 while ((inodedep = 4062 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { 4063 inodedep->id_state |= DEPCOMPLETE; 4064 LIST_REMOVE(inodedep, id_deps); 4065 inodedep->id_buf = NULL; 4066 } 4067 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 4068 continue; 4069 4070 case D_MKDIR: 4071 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 4072 continue; 4073 4074 case D_ALLOCDIRECT: 4075 adp = WK_ALLOCDIRECT(wk); 4076 adp->ad_state |= COMPLETE; 4077 handle_allocdirect_partdone(adp); 4078 continue; 4079 4080 case D_ALLOCINDIR: 4081 aip = WK_ALLOCINDIR(wk); 4082 aip->ai_state |= COMPLETE; 4083 handle_allocindir_partdone(aip); 4084 continue; 4085 4086 case D_INDIRDEP: 4087 indirdep = WK_INDIRDEP(wk); 4088 if (indirdep->ir_state & GOINGAWAY) 4089 panic("disk_write_complete: indirdep gone"); 4090 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 4091 FREE(indirdep->ir_saveddata, M_INDIRDEP); 4092 indirdep->ir_saveddata = 0; 4093 indirdep->ir_state &= ~UNDONE; 4094 indirdep->ir_state |= ATTACHED; 4095 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 4096 handle_allocindir_partdone(aip); 4097 if (aip == LIST_FIRST(&indirdep->ir_donehd)) 4098 panic("disk_write_complete: not gone"); 4099 } 4100 WORKLIST_INSERT(&reattach, wk); 4101 if ((bp->b_flags & B_DELWRI) == 0) 4102 stat_indir_blk_ptrs++; 4103 bdirty(bp); 4104 continue; 4105 4106 default: 4107 panic("handle_disk_write_complete: Unknown type %s", 4108 TYPENAME(wk->wk_type)); 4109 /* NOTREACHED */ 4110 } 4111 } 4112 /* 4113 * Reattach any requests that must be redone. 4114 */ 4115 while ((wk = LIST_FIRST(&reattach)) != NULL) { 4116 WORKLIST_REMOVE(wk); 4117 WORKLIST_INSERT(&bp->b_dep, wk); 4118 } 4119 FREE_LOCK(&lk); 4120} 4121 4122/* 4123 * Called from within softdep_disk_write_complete above. Note that 4124 * this routine is always called from interrupt level with further 4125 * splbio interrupts blocked. 4126 */ 4127static void 4128handle_allocdirect_partdone(adp) 4129 struct allocdirect *adp; /* the completed allocdirect */ 4130{ 4131 struct allocdirectlst *listhead; 4132 struct allocdirect *listadp; 4133 struct inodedep *inodedep; 4134 long bsize, delay; 4135 4136 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 4137 return; 4138 if (adp->ad_buf != NULL) 4139 panic("handle_allocdirect_partdone: dangling dep"); 4140 /* 4141 * The on-disk inode cannot claim to be any larger than the last 4142 * fragment that has been written. Otherwise, the on-disk inode 4143 * might have fragments that were not the last block in the file 4144 * which would corrupt the filesystem. Thus, we cannot free any 4145 * allocdirects after one whose ad_oldblkno claims a fragment as 4146 * these blocks must be rolled back to zero before writing the inode. 4147 * We check the currently active set of allocdirects in id_inoupdt 4148 * or id_extupdt as appropriate. 4149 */ 4150 inodedep = adp->ad_inodedep; 4151 bsize = inodedep->id_fs->fs_bsize; 4152 if (adp->ad_state & EXTDATA) 4153 listhead = &inodedep->id_extupdt; 4154 else 4155 listhead = &inodedep->id_inoupdt; 4156 TAILQ_FOREACH(listadp, listhead, ad_next) { 4157 /* found our block */ 4158 if (listadp == adp) 4159 break; 4160 /* continue if ad_oldlbn is not a fragment */ 4161 if (listadp->ad_oldsize == 0 || 4162 listadp->ad_oldsize == bsize) 4163 continue; 4164 /* hit a fragment */ 4165 return; 4166 } 4167 /* 4168 * If we have reached the end of the current list without 4169 * finding the just finished dependency, then it must be 4170 * on the future dependency list. Future dependencies cannot 4171 * be freed until they are moved to the current list. 4172 */ 4173 if (listadp == NULL) { 4174#ifdef DEBUG 4175 if (adp->ad_state & EXTDATA) 4176 listhead = &inodedep->id_newextupdt; 4177 else 4178 listhead = &inodedep->id_newinoupdt; 4179 TAILQ_FOREACH(listadp, listhead, ad_next) 4180 /* found our block */ 4181 if (listadp == adp) 4182 break; 4183 if (listadp == NULL) 4184 panic("handle_allocdirect_partdone: lost dep"); 4185#endif /* DEBUG */ 4186 return; 4187 } 4188 /* 4189 * If we have found the just finished dependency, then free 4190 * it along with anything that follows it that is complete. 4191 * If the inode still has a bitmap dependency, then it has 4192 * never been written to disk, hence the on-disk inode cannot 4193 * reference the old fragment so we can free it without delay. 4194 */ 4195 delay = (inodedep->id_state & DEPCOMPLETE); 4196 for (; adp; adp = listadp) { 4197 listadp = TAILQ_NEXT(adp, ad_next); 4198 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 4199 return; 4200 free_allocdirect(listhead, adp, delay); 4201 } 4202} 4203 4204/* 4205 * Called from within softdep_disk_write_complete above. Note that 4206 * this routine is always called from interrupt level with further 4207 * splbio interrupts blocked. 4208 */ 4209static void 4210handle_allocindir_partdone(aip) 4211 struct allocindir *aip; /* the completed allocindir */ 4212{ 4213 struct indirdep *indirdep; 4214 4215 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 4216 return; 4217 if (aip->ai_buf != NULL) 4218 panic("handle_allocindir_partdone: dangling dependency"); 4219 indirdep = aip->ai_indirdep; 4220 if (indirdep->ir_state & UNDONE) { 4221 LIST_REMOVE(aip, ai_next); 4222 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 4223 return; 4224 } 4225 if (indirdep->ir_state & UFS1FMT) 4226 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 4227 aip->ai_newblkno; 4228 else 4229 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 4230 aip->ai_newblkno; 4231 LIST_REMOVE(aip, ai_next); 4232 if (aip->ai_freefrag != NULL) 4233 add_to_worklist(&aip->ai_freefrag->ff_list); 4234 WORKITEM_FREE(aip, D_ALLOCINDIR); 4235} 4236 4237/* 4238 * Called from within softdep_disk_write_complete above to restore 4239 * in-memory inode block contents to their most up-to-date state. Note 4240 * that this routine is always called from interrupt level with further 4241 * splbio interrupts blocked. 4242 */ 4243static int 4244handle_written_inodeblock(inodedep, bp) 4245 struct inodedep *inodedep; 4246 struct buf *bp; /* buffer containing the inode block */ 4247{ 4248 struct worklist *wk, *filefree; 4249 struct allocdirect *adp, *nextadp; 4250 struct ufs1_dinode *dp1 = NULL; 4251 struct ufs2_dinode *dp2 = NULL; 4252 int hadchanges, fstype; 4253 4254 if ((inodedep->id_state & IOSTARTED) == 0) 4255 panic("handle_written_inodeblock: not started"); 4256 inodedep->id_state &= ~IOSTARTED; 4257 inodedep->id_state |= COMPLETE; 4258 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) { 4259 fstype = UFS1; 4260 dp1 = (struct ufs1_dinode *)bp->b_data + 4261 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 4262 } else { 4263 fstype = UFS2; 4264 dp2 = (struct ufs2_dinode *)bp->b_data + 4265 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 4266 } 4267 /* 4268 * If we had to rollback the inode allocation because of 4269 * bitmaps being incomplete, then simply restore it. 4270 * Keep the block dirty so that it will not be reclaimed until 4271 * all associated dependencies have been cleared and the 4272 * corresponding updates written to disk. 4273 */ 4274 if (inodedep->id_savedino1 != NULL) { 4275 if (fstype == UFS1) 4276 *dp1 = *inodedep->id_savedino1; 4277 else 4278 *dp2 = *inodedep->id_savedino2; 4279 FREE(inodedep->id_savedino1, M_SAVEDINO); 4280 inodedep->id_savedino1 = NULL; 4281 if ((bp->b_flags & B_DELWRI) == 0) 4282 stat_inode_bitmap++; 4283 bdirty(bp); 4284 return (1); 4285 } 4286 /* 4287 * Roll forward anything that had to be rolled back before 4288 * the inode could be updated. 4289 */ 4290 hadchanges = 0; 4291 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 4292 nextadp = TAILQ_NEXT(adp, ad_next); 4293 if (adp->ad_state & ATTACHED) 4294 panic("handle_written_inodeblock: new entry"); 4295 if (fstype == UFS1) { 4296 if (adp->ad_lbn < NDADDR) { 4297 if (dp1->di_db[adp->ad_lbn]!=adp->ad_oldblkno) 4298 panic("%s %s #%jd mismatch %d != %jd", 4299 "handle_written_inodeblock:", 4300 "direct pointer", 4301 (intmax_t)adp->ad_lbn, 4302 dp1->di_db[adp->ad_lbn], 4303 (intmax_t)adp->ad_oldblkno); 4304 dp1->di_db[adp->ad_lbn] = adp->ad_newblkno; 4305 } else { 4306 if (dp1->di_ib[adp->ad_lbn - NDADDR] != 0) 4307 panic("%s: %s #%jd allocated as %d", 4308 "handle_written_inodeblock", 4309 "indirect pointer", 4310 (intmax_t)adp->ad_lbn - NDADDR, 4311 dp1->di_ib[adp->ad_lbn - NDADDR]); 4312 dp1->di_ib[adp->ad_lbn - NDADDR] = 4313 adp->ad_newblkno; 4314 } 4315 } else { 4316 if (adp->ad_lbn < NDADDR) { 4317 if (dp2->di_db[adp->ad_lbn]!=adp->ad_oldblkno) 4318 panic("%s: %s #%jd %s %jd != %jd", 4319 "handle_written_inodeblock", 4320 "direct pointer", 4321 (intmax_t)adp->ad_lbn, "mismatch", 4322 (intmax_t)dp2->di_db[adp->ad_lbn], 4323 (intmax_t)adp->ad_oldblkno); 4324 dp2->di_db[adp->ad_lbn] = adp->ad_newblkno; 4325 } else { 4326 if (dp2->di_ib[adp->ad_lbn - NDADDR] != 0) 4327 panic("%s: %s #%jd allocated as %jd", 4328 "handle_written_inodeblock", 4329 "indirect pointer", 4330 (intmax_t)adp->ad_lbn - NDADDR, 4331 (intmax_t) 4332 dp2->di_ib[adp->ad_lbn - NDADDR]); 4333 dp2->di_ib[adp->ad_lbn - NDADDR] = 4334 adp->ad_newblkno; 4335 } 4336 } 4337 adp->ad_state &= ~UNDONE; 4338 adp->ad_state |= ATTACHED; 4339 hadchanges = 1; 4340 } 4341 for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) { 4342 nextadp = TAILQ_NEXT(adp, ad_next); 4343 if (adp->ad_state & ATTACHED) 4344 panic("handle_written_inodeblock: new entry"); 4345 if (dp2->di_extb[adp->ad_lbn] != adp->ad_oldblkno) 4346 panic("%s: direct pointers #%jd %s %jd != %jd", 4347 "handle_written_inodeblock", 4348 (intmax_t)adp->ad_lbn, "mismatch", 4349 (intmax_t)dp2->di_extb[adp->ad_lbn], 4350 (intmax_t)adp->ad_oldblkno); 4351 dp2->di_extb[adp->ad_lbn] = adp->ad_newblkno; 4352 adp->ad_state &= ~UNDONE; 4353 adp->ad_state |= ATTACHED; 4354 hadchanges = 1; 4355 } 4356 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 4357 stat_direct_blk_ptrs++; 4358 /* 4359 * Reset the file size to its most up-to-date value. 4360 */ 4361 if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1) 4362 panic("handle_written_inodeblock: bad size"); 4363 if (fstype == UFS1) { 4364 if (dp1->di_size != inodedep->id_savedsize) { 4365 dp1->di_size = inodedep->id_savedsize; 4366 hadchanges = 1; 4367 } 4368 } else { 4369 if (dp2->di_size != inodedep->id_savedsize) { 4370 dp2->di_size = inodedep->id_savedsize; 4371 hadchanges = 1; 4372 } 4373 if (dp2->di_extsize != inodedep->id_savedextsize) { 4374 dp2->di_extsize = inodedep->id_savedextsize; 4375 hadchanges = 1; 4376 } 4377 } 4378 inodedep->id_savedsize = -1; 4379 inodedep->id_savedextsize = -1; 4380 /* 4381 * If there were any rollbacks in the inode block, then it must be 4382 * marked dirty so that its will eventually get written back in 4383 * its correct form. 4384 */ 4385 if (hadchanges) 4386 bdirty(bp); 4387 /* 4388 * Process any allocdirects that completed during the update. 4389 */ 4390 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 4391 handle_allocdirect_partdone(adp); 4392 if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 4393 handle_allocdirect_partdone(adp); 4394 /* 4395 * Process deallocations that were held pending until the 4396 * inode had been written to disk. Freeing of the inode 4397 * is delayed until after all blocks have been freed to 4398 * avoid creation of new <vfsid, inum, lbn> triples 4399 * before the old ones have been deleted. 4400 */ 4401 filefree = NULL; 4402 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 4403 WORKLIST_REMOVE(wk); 4404 switch (wk->wk_type) { 4405 4406 case D_FREEFILE: 4407 /* 4408 * We defer adding filefree to the worklist until 4409 * all other additions have been made to ensure 4410 * that it will be done after all the old blocks 4411 * have been freed. 4412 */ 4413 if (filefree != NULL) 4414 panic("handle_written_inodeblock: filefree"); 4415 filefree = wk; 4416 continue; 4417 4418 case D_MKDIR: 4419 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 4420 continue; 4421 4422 case D_DIRADD: 4423 diradd_inode_written(WK_DIRADD(wk), inodedep); 4424 continue; 4425 4426 case D_FREEBLKS: 4427 case D_FREEFRAG: 4428 case D_DIRREM: 4429 add_to_worklist(wk); 4430 continue; 4431 4432 case D_NEWDIRBLK: 4433 free_newdirblk(WK_NEWDIRBLK(wk)); 4434 continue; 4435 4436 default: 4437 panic("handle_written_inodeblock: Unknown type %s", 4438 TYPENAME(wk->wk_type)); 4439 /* NOTREACHED */ 4440 } 4441 } 4442 if (filefree != NULL) { 4443 if (free_inodedep(inodedep) == 0) 4444 panic("handle_written_inodeblock: live inodedep"); 4445 add_to_worklist(filefree); 4446 return (0); 4447 } 4448 4449 /* 4450 * If no outstanding dependencies, free it. 4451 */ 4452 if (free_inodedep(inodedep) || 4453 (TAILQ_FIRST(&inodedep->id_inoupdt) == 0 && 4454 TAILQ_FIRST(&inodedep->id_extupdt) == 0)) 4455 return (0); 4456 return (hadchanges); 4457} 4458 4459/* 4460 * Process a diradd entry after its dependent inode has been written. 4461 * This routine must be called with splbio interrupts blocked. 4462 */ 4463static void 4464diradd_inode_written(dap, inodedep) 4465 struct diradd *dap; 4466 struct inodedep *inodedep; 4467{ 4468 struct pagedep *pagedep; 4469 4470 dap->da_state |= COMPLETE; 4471 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 4472 if (dap->da_state & DIRCHG) 4473 pagedep = dap->da_previous->dm_pagedep; 4474 else 4475 pagedep = dap->da_pagedep; 4476 LIST_REMOVE(dap, da_pdlist); 4477 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 4478 } 4479 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 4480} 4481 4482/* 4483 * Handle the completion of a mkdir dependency. 4484 */ 4485static void 4486handle_written_mkdir(mkdir, type) 4487 struct mkdir *mkdir; 4488 int type; 4489{ 4490 struct diradd *dap; 4491 struct pagedep *pagedep; 4492 4493 if (mkdir->md_state != type) 4494 panic("handle_written_mkdir: bad type"); 4495 dap = mkdir->md_diradd; 4496 dap->da_state &= ~type; 4497 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 4498 dap->da_state |= DEPCOMPLETE; 4499 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 4500 if (dap->da_state & DIRCHG) 4501 pagedep = dap->da_previous->dm_pagedep; 4502 else 4503 pagedep = dap->da_pagedep; 4504 LIST_REMOVE(dap, da_pdlist); 4505 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 4506 } 4507 LIST_REMOVE(mkdir, md_mkdirs); 4508 WORKITEM_FREE(mkdir, D_MKDIR); 4509} 4510 4511/* 4512 * Called from within softdep_disk_write_complete above. 4513 * A write operation was just completed. Removed inodes can 4514 * now be freed and associated block pointers may be committed. 4515 * Note that this routine is always called from interrupt level 4516 * with further splbio interrupts blocked. 4517 */ 4518static int 4519handle_written_filepage(pagedep, bp) 4520 struct pagedep *pagedep; 4521 struct buf *bp; /* buffer containing the written page */ 4522{ 4523 struct dirrem *dirrem; 4524 struct diradd *dap, *nextdap; 4525 struct direct *ep; 4526 int i, chgs; 4527 4528 if ((pagedep->pd_state & IOSTARTED) == 0) 4529 panic("handle_written_filepage: not started"); 4530 pagedep->pd_state &= ~IOSTARTED; 4531 /* 4532 * Process any directory removals that have been committed. 4533 */ 4534 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 4535 LIST_REMOVE(dirrem, dm_next); 4536 dirrem->dm_dirinum = pagedep->pd_ino; 4537 add_to_worklist(&dirrem->dm_list); 4538 } 4539 /* 4540 * Free any directory additions that have been committed. 4541 * If it is a newly allocated block, we have to wait until 4542 * the on-disk directory inode claims the new block. 4543 */ 4544 if ((pagedep->pd_state & NEWBLOCK) == 0) 4545 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 4546 free_diradd(dap); 4547 /* 4548 * Uncommitted directory entries must be restored. 4549 */ 4550 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 4551 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 4552 dap = nextdap) { 4553 nextdap = LIST_NEXT(dap, da_pdlist); 4554 if (dap->da_state & ATTACHED) 4555 panic("handle_written_filepage: attached"); 4556 ep = (struct direct *) 4557 ((char *)bp->b_data + dap->da_offset); 4558 ep->d_ino = dap->da_newinum; 4559 dap->da_state &= ~UNDONE; 4560 dap->da_state |= ATTACHED; 4561 chgs = 1; 4562 /* 4563 * If the inode referenced by the directory has 4564 * been written out, then the dependency can be 4565 * moved to the pending list. 4566 */ 4567 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 4568 LIST_REMOVE(dap, da_pdlist); 4569 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 4570 da_pdlist); 4571 } 4572 } 4573 } 4574 /* 4575 * If there were any rollbacks in the directory, then it must be 4576 * marked dirty so that its will eventually get written back in 4577 * its correct form. 4578 */ 4579 if (chgs) { 4580 if ((bp->b_flags & B_DELWRI) == 0) 4581 stat_dir_entry++; 4582 bdirty(bp); 4583 return (1); 4584 } 4585 /* 4586 * If we are not waiting for a new directory block to be 4587 * claimed by its inode, then the pagedep will be freed. 4588 * Otherwise it will remain to track any new entries on 4589 * the page in case they are fsync'ed. 4590 */ 4591 if ((pagedep->pd_state & NEWBLOCK) == 0) { 4592 LIST_REMOVE(pagedep, pd_hash); 4593 WORKITEM_FREE(pagedep, D_PAGEDEP); 4594 } 4595 return (0); 4596} 4597 4598/* 4599 * Writing back in-core inode structures. 4600 * 4601 * The filesystem only accesses an inode's contents when it occupies an 4602 * "in-core" inode structure. These "in-core" structures are separate from 4603 * the page frames used to cache inode blocks. Only the latter are 4604 * transferred to/from the disk. So, when the updated contents of the 4605 * "in-core" inode structure are copied to the corresponding in-memory inode 4606 * block, the dependencies are also transferred. The following procedure is 4607 * called when copying a dirty "in-core" inode to a cached inode block. 4608 */ 4609 4610/* 4611 * Called when an inode is loaded from disk. If the effective link count 4612 * differed from the actual link count when it was last flushed, then we 4613 * need to ensure that the correct effective link count is put back. 4614 */ 4615void 4616softdep_load_inodeblock(ip) 4617 struct inode *ip; /* the "in_core" copy of the inode */ 4618{ 4619 struct inodedep *inodedep; 4620 4621 /* 4622 * Check for alternate nlink count. 4623 */ 4624 ip->i_effnlink = ip->i_nlink; 4625 ACQUIRE_LOCK(&lk); 4626 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 4627 FREE_LOCK(&lk); 4628 return; 4629 } 4630 ip->i_effnlink -= inodedep->id_nlinkdelta; 4631 if (inodedep->id_state & SPACECOUNTED) 4632 ip->i_flag |= IN_SPACECOUNTED; 4633 FREE_LOCK(&lk); 4634} 4635 4636/* 4637 * This routine is called just before the "in-core" inode 4638 * information is to be copied to the in-memory inode block. 4639 * Recall that an inode block contains several inodes. If 4640 * the force flag is set, then the dependencies will be 4641 * cleared so that the update can always be made. Note that 4642 * the buffer is locked when this routine is called, so we 4643 * will never be in the middle of writing the inode block 4644 * to disk. 4645 */ 4646void 4647softdep_update_inodeblock(ip, bp, waitfor) 4648 struct inode *ip; /* the "in_core" copy of the inode */ 4649 struct buf *bp; /* the buffer containing the inode block */ 4650 int waitfor; /* nonzero => update must be allowed */ 4651{ 4652 struct inodedep *inodedep; 4653 struct worklist *wk; 4654 struct buf *ibp; 4655 int error; 4656 4657 /* 4658 * If the effective link count is not equal to the actual link 4659 * count, then we must track the difference in an inodedep while 4660 * the inode is (potentially) tossed out of the cache. Otherwise, 4661 * if there is no existing inodedep, then there are no dependencies 4662 * to track. 4663 */ 4664 ACQUIRE_LOCK(&lk); 4665 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 4666 FREE_LOCK(&lk); 4667 if (ip->i_effnlink != ip->i_nlink) 4668 panic("softdep_update_inodeblock: bad link count"); 4669 return; 4670 } 4671 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) 4672 panic("softdep_update_inodeblock: bad delta"); 4673 /* 4674 * Changes have been initiated. Anything depending on these 4675 * changes cannot occur until this inode has been written. 4676 */ 4677 inodedep->id_state &= ~COMPLETE; 4678 if ((inodedep->id_state & ONWORKLIST) == 0) 4679 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 4680 /* 4681 * Any new dependencies associated with the incore inode must 4682 * now be moved to the list associated with the buffer holding 4683 * the in-memory copy of the inode. Once merged process any 4684 * allocdirects that are completed by the merger. 4685 */ 4686 merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt); 4687 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL) 4688 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); 4689 merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt); 4690 if (TAILQ_FIRST(&inodedep->id_extupdt) != NULL) 4691 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt)); 4692 /* 4693 * Now that the inode has been pushed into the buffer, the 4694 * operations dependent on the inode being written to disk 4695 * can be moved to the id_bufwait so that they will be 4696 * processed when the buffer I/O completes. 4697 */ 4698 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 4699 WORKLIST_REMOVE(wk); 4700 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 4701 } 4702 /* 4703 * Newly allocated inodes cannot be written until the bitmap 4704 * that allocates them have been written (indicated by 4705 * DEPCOMPLETE being set in id_state). If we are doing a 4706 * forced sync (e.g., an fsync on a file), we force the bitmap 4707 * to be written so that the update can be done. 4708 */ 4709 if (waitfor == 0) { 4710 FREE_LOCK(&lk); 4711 return; 4712 } 4713retry: 4714 if ((inodedep->id_state & DEPCOMPLETE) != 0) { 4715 FREE_LOCK(&lk); 4716 return; 4717 } 4718 ibp = inodedep->id_buf; 4719 ibp = getdirtybuf(ibp, &lk, MNT_WAIT); 4720 if (ibp == NULL) { 4721 /* 4722 * If ibp came back as NULL, the dependency could have been 4723 * freed while we slept. Look it up again, and check to see 4724 * that it has completed. 4725 */ 4726 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) != 0) 4727 goto retry; 4728 FREE_LOCK(&lk); 4729 return; 4730 } 4731 FREE_LOCK(&lk); 4732 if ((error = bwrite(ibp)) != 0) 4733 softdep_error("softdep_update_inodeblock: bwrite", error); 4734} 4735 4736/* 4737 * Merge the a new inode dependency list (such as id_newinoupdt) into an 4738 * old inode dependency list (such as id_inoupdt). This routine must be 4739 * called with splbio interrupts blocked. 4740 */ 4741static void 4742merge_inode_lists(newlisthead, oldlisthead) 4743 struct allocdirectlst *newlisthead; 4744 struct allocdirectlst *oldlisthead; 4745{ 4746 struct allocdirect *listadp, *newadp; 4747 4748 newadp = TAILQ_FIRST(newlisthead); 4749 for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) { 4750 if (listadp->ad_lbn < newadp->ad_lbn) { 4751 listadp = TAILQ_NEXT(listadp, ad_next); 4752 continue; 4753 } 4754 TAILQ_REMOVE(newlisthead, newadp, ad_next); 4755 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 4756 if (listadp->ad_lbn == newadp->ad_lbn) { 4757 allocdirect_merge(oldlisthead, newadp, 4758 listadp); 4759 listadp = newadp; 4760 } 4761 newadp = TAILQ_FIRST(newlisthead); 4762 } 4763 while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) { 4764 TAILQ_REMOVE(newlisthead, newadp, ad_next); 4765 TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next); 4766 } 4767} 4768 4769/* 4770 * If we are doing an fsync, then we must ensure that any directory 4771 * entries for the inode have been written after the inode gets to disk. 4772 */ 4773int 4774softdep_fsync(vp) 4775 struct vnode *vp; /* the "in_core" copy of the inode */ 4776{ 4777 struct inodedep *inodedep; 4778 struct pagedep *pagedep; 4779 struct worklist *wk; 4780 struct diradd *dap; 4781 struct mount *mnt; 4782 struct vnode *pvp; 4783 struct inode *ip; 4784 struct buf *bp; 4785 struct fs *fs; 4786 struct thread *td = curthread; 4787 int error, flushparent; 4788 ino_t parentino; 4789 ufs_lbn_t lbn; 4790 4791 ip = VTOI(vp); 4792 fs = ip->i_fs; 4793 ACQUIRE_LOCK(&lk); 4794 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) { 4795 FREE_LOCK(&lk); 4796 return (0); 4797 } 4798 if (LIST_FIRST(&inodedep->id_inowait) != NULL || 4799 LIST_FIRST(&inodedep->id_bufwait) != NULL || 4800 TAILQ_FIRST(&inodedep->id_extupdt) != NULL || 4801 TAILQ_FIRST(&inodedep->id_newextupdt) != NULL || 4802 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 4803 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) 4804 panic("softdep_fsync: pending ops"); 4805 for (error = 0, flushparent = 0; ; ) { 4806 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 4807 break; 4808 if (wk->wk_type != D_DIRADD) 4809 panic("softdep_fsync: Unexpected type %s", 4810 TYPENAME(wk->wk_type)); 4811 dap = WK_DIRADD(wk); 4812 /* 4813 * Flush our parent if this directory entry has a MKDIR_PARENT 4814 * dependency or is contained in a newly allocated block. 4815 */ 4816 if (dap->da_state & DIRCHG) 4817 pagedep = dap->da_previous->dm_pagedep; 4818 else 4819 pagedep = dap->da_pagedep; 4820 mnt = pagedep->pd_mnt; 4821 parentino = pagedep->pd_ino; 4822 lbn = pagedep->pd_lbn; 4823 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) 4824 panic("softdep_fsync: dirty"); 4825 if ((dap->da_state & MKDIR_PARENT) || 4826 (pagedep->pd_state & NEWBLOCK)) 4827 flushparent = 1; 4828 else 4829 flushparent = 0; 4830 /* 4831 * If we are being fsync'ed as part of vgone'ing this vnode, 4832 * then we will not be able to release and recover the 4833 * vnode below, so we just have to give up on writing its 4834 * directory entry out. It will eventually be written, just 4835 * not now, but then the user was not asking to have it 4836 * written, so we are not breaking any promises. 4837 */ 4838 if (vp->v_iflag & VI_DOOMED) 4839 break; 4840 /* 4841 * We prevent deadlock by always fetching inodes from the 4842 * root, moving down the directory tree. Thus, when fetching 4843 * our parent directory, we first try to get the lock. If 4844 * that fails, we must unlock ourselves before requesting 4845 * the lock on our parent. See the comment in ufs_lookup 4846 * for details on possible races. 4847 */ 4848 FREE_LOCK(&lk); 4849 if (ffs_vget(mnt, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp)) { 4850 VOP_UNLOCK(vp, 0, td); 4851 error = ffs_vget(mnt, parentino, LK_EXCLUSIVE, &pvp); 4852 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 4853 if (error != 0) 4854 return (error); 4855 } 4856 /* 4857 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps 4858 * that are contained in direct blocks will be resolved by 4859 * doing a ffs_update. Pagedeps contained in indirect blocks 4860 * may require a complete sync'ing of the directory. So, we 4861 * try the cheap and fast ffs_update first, and if that fails, 4862 * then we do the slower ffs_syncvnode of the directory. 4863 */ 4864 if (flushparent) { 4865 if ((error = ffs_update(pvp, 1)) != 0) { 4866 vput(pvp); 4867 return (error); 4868 } 4869 if ((pagedep->pd_state & NEWBLOCK) && 4870 (error = ffs_syncvnode(pvp, MNT_WAIT))) { 4871 vput(pvp); 4872 return (error); 4873 } 4874 } 4875 /* 4876 * Flush directory page containing the inode's name. 4877 */ 4878 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred, 4879 &bp); 4880 if (error == 0) 4881 error = bwrite(bp); 4882 else 4883 brelse(bp); 4884 vput(pvp); 4885 if (error != 0) 4886 return (error); 4887 ACQUIRE_LOCK(&lk); 4888 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) 4889 break; 4890 } 4891 FREE_LOCK(&lk); 4892 return (0); 4893} 4894 4895/* 4896 * Flush all the dirty bitmaps associated with the block device 4897 * before flushing the rest of the dirty blocks so as to reduce 4898 * the number of dependencies that will have to be rolled back. 4899 */ 4900void 4901softdep_fsync_mountdev(vp) 4902 struct vnode *vp; 4903{ 4904 struct buf *bp, *nbp; 4905 struct worklist *wk; 4906 4907 if (!vn_isdisk(vp, NULL)) 4908 panic("softdep_fsync_mountdev: vnode not a disk"); 4909 ACQUIRE_LOCK(&lk); 4910 VI_LOCK(vp); 4911 TAILQ_FOREACH_SAFE(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs, nbp) { 4912 /* 4913 * If it is already scheduled, skip to the next buffer. 4914 */ 4915 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 4916 continue; 4917 4918 if ((bp->b_flags & B_DELWRI) == 0) 4919 panic("softdep_fsync_mountdev: not dirty"); 4920 /* 4921 * We are only interested in bitmaps with outstanding 4922 * dependencies. 4923 */ 4924 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 4925 wk->wk_type != D_BMSAFEMAP || 4926 (bp->b_vflags & BV_BKGRDINPROG)) { 4927 BUF_UNLOCK(bp); 4928 continue; 4929 } 4930 VI_UNLOCK(vp); 4931 FREE_LOCK(&lk); 4932 bremfree(bp); 4933 (void) bawrite(bp); 4934 ACQUIRE_LOCK(&lk); 4935 /* 4936 * Since we may have slept during the I/O, we need 4937 * to start from a known point. 4938 */ 4939 VI_LOCK(vp); 4940 nbp = TAILQ_FIRST(&vp->v_bufobj.bo_dirty.bv_hd); 4941 } 4942 FREE_LOCK(&lk); 4943 drain_output(vp); 4944 VI_UNLOCK(vp); 4945} 4946 4947/* 4948 * This routine is called when we are trying to synchronously flush a 4949 * file. This routine must eliminate any filesystem metadata dependencies 4950 * so that the syncing routine can succeed by pushing the dirty blocks 4951 * associated with the file. If any I/O errors occur, they are returned. 4952 */ 4953int 4954softdep_sync_metadata(struct vnode *vp) 4955{ 4956 struct pagedep *pagedep; 4957 struct allocdirect *adp; 4958 struct allocindir *aip; 4959 struct buf *bp, *nbp; 4960 struct worklist *wk; 4961 int i, error, waitfor; 4962 4963 if (!DOINGSOFTDEP(vp)) 4964 return (0); 4965 /* 4966 * Ensure that any direct block dependencies have been cleared. 4967 */ 4968 ACQUIRE_LOCK(&lk); 4969 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) { 4970 FREE_LOCK(&lk); 4971 return (error); 4972 } 4973 FREE_LOCK(&lk); 4974 /* 4975 * For most files, the only metadata dependencies are the 4976 * cylinder group maps that allocate their inode or blocks. 4977 * The block allocation dependencies can be found by traversing 4978 * the dependency lists for any buffers that remain on their 4979 * dirty buffer list. The inode allocation dependency will 4980 * be resolved when the inode is updated with MNT_WAIT. 4981 * This work is done in two passes. The first pass grabs most 4982 * of the buffers and begins asynchronously writing them. The 4983 * only way to wait for these asynchronous writes is to sleep 4984 * on the filesystem vnode which may stay busy for a long time 4985 * if the filesystem is active. So, instead, we make a second 4986 * pass over the dependencies blocking on each write. In the 4987 * usual case we will be blocking against a write that we 4988 * initiated, so when it is done the dependency will have been 4989 * resolved. Thus the second pass is expected to end quickly. 4990 */ 4991 waitfor = MNT_NOWAIT; 4992 4993top: 4994 /* 4995 * We must wait for any I/O in progress to finish so that 4996 * all potential buffers on the dirty list will be visible. 4997 */ 4998 VI_LOCK(vp); 4999 drain_output(vp); 5000 while ((bp = TAILQ_FIRST(&vp->v_bufobj.bo_dirty.bv_hd)) != NULL) { 5001 bp = getdirtybuf(bp, VI_MTX(vp), MNT_WAIT); 5002 if (bp) 5003 break; 5004 } 5005 VI_UNLOCK(vp); 5006 if (bp == NULL) 5007 return (0); 5008loop: 5009 /* While syncing snapshots, we must allow recursive lookups */ 5010 bp->b_lock.lk_flags |= LK_CANRECURSE; 5011 ACQUIRE_LOCK(&lk); 5012 /* 5013 * As we hold the buffer locked, none of its dependencies 5014 * will disappear. 5015 */ 5016 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5017 switch (wk->wk_type) { 5018 5019 case D_ALLOCDIRECT: 5020 adp = WK_ALLOCDIRECT(wk); 5021 if (adp->ad_state & DEPCOMPLETE) 5022 continue; 5023 nbp = adp->ad_buf; 5024 nbp = getdirtybuf(nbp, &lk, waitfor); 5025 if (nbp == NULL) 5026 continue; 5027 FREE_LOCK(&lk); 5028 if (waitfor == MNT_NOWAIT) { 5029 bawrite(nbp); 5030 } else if ((error = bwrite(nbp)) != 0) { 5031 break; 5032 } 5033 ACQUIRE_LOCK(&lk); 5034 continue; 5035 5036 case D_ALLOCINDIR: 5037 aip = WK_ALLOCINDIR(wk); 5038 if (aip->ai_state & DEPCOMPLETE) 5039 continue; 5040 nbp = aip->ai_buf; 5041 nbp = getdirtybuf(nbp, &lk, waitfor); 5042 if (nbp == NULL) 5043 continue; 5044 FREE_LOCK(&lk); 5045 if (waitfor == MNT_NOWAIT) { 5046 bawrite(nbp); 5047 } else if ((error = bwrite(nbp)) != 0) { 5048 break; 5049 } 5050 ACQUIRE_LOCK(&lk); 5051 continue; 5052 5053 case D_INDIRDEP: 5054 restart: 5055 5056 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) { 5057 if (aip->ai_state & DEPCOMPLETE) 5058 continue; 5059 nbp = aip->ai_buf; 5060 nbp = getdirtybuf(nbp, &lk, MNT_WAIT); 5061 if (nbp == NULL) 5062 goto restart; 5063 FREE_LOCK(&lk); 5064 if ((error = bwrite(nbp)) != 0) { 5065 break; 5066 } 5067 ACQUIRE_LOCK(&lk); 5068 goto restart; 5069 } 5070 continue; 5071 5072 case D_INODEDEP: 5073 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs, 5074 WK_INODEDEP(wk)->id_ino)) != 0) { 5075 FREE_LOCK(&lk); 5076 break; 5077 } 5078 continue; 5079 5080 case D_PAGEDEP: 5081 /* 5082 * We are trying to sync a directory that may 5083 * have dependencies on both its own metadata 5084 * and/or dependencies on the inodes of any 5085 * recently allocated files. We walk its diradd 5086 * lists pushing out the associated inode. 5087 */ 5088 pagedep = WK_PAGEDEP(wk); 5089 for (i = 0; i < DAHASHSZ; i++) { 5090 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 5091 continue; 5092 if ((error = 5093 flush_pagedep_deps(vp, pagedep->pd_mnt, 5094 &pagedep->pd_diraddhd[i]))) { 5095 FREE_LOCK(&lk); 5096 break; 5097 } 5098 } 5099 continue; 5100 5101 case D_MKDIR: 5102 /* 5103 * This case should never happen if the vnode has 5104 * been properly sync'ed. However, if this function 5105 * is used at a place where the vnode has not yet 5106 * been sync'ed, this dependency can show up. So, 5107 * rather than panic, just flush it. 5108 */ 5109 nbp = WK_MKDIR(wk)->md_buf; 5110 nbp = getdirtybuf(nbp, &lk, waitfor); 5111 if (nbp == NULL) 5112 continue; 5113 FREE_LOCK(&lk); 5114 if (waitfor == MNT_NOWAIT) { 5115 bawrite(nbp); 5116 } else if ((error = bwrite(nbp)) != 0) { 5117 break; 5118 } 5119 ACQUIRE_LOCK(&lk); 5120 continue; 5121 5122 case D_BMSAFEMAP: 5123 /* 5124 * This case should never happen if the vnode has 5125 * been properly sync'ed. However, if this function 5126 * is used at a place where the vnode has not yet 5127 * been sync'ed, this dependency can show up. So, 5128 * rather than panic, just flush it. 5129 */ 5130 nbp = WK_BMSAFEMAP(wk)->sm_buf; 5131 nbp = getdirtybuf(nbp, &lk, waitfor); 5132 if (nbp == NULL) 5133 continue; 5134 FREE_LOCK(&lk); 5135 if (waitfor == MNT_NOWAIT) { 5136 bawrite(nbp); 5137 } else if ((error = bwrite(nbp)) != 0) { 5138 break; 5139 } 5140 ACQUIRE_LOCK(&lk); 5141 continue; 5142 5143 default: 5144 panic("softdep_sync_metadata: Unknown type %s", 5145 TYPENAME(wk->wk_type)); 5146 /* NOTREACHED */ 5147 } 5148 /* We reach here only in error and unlocked */ 5149 if (error == 0) 5150 panic("softdep_sync_metadata: zero error"); 5151 bp->b_lock.lk_flags &= ~LK_CANRECURSE; 5152 bawrite(bp); 5153 return (error); 5154 } 5155 FREE_LOCK(&lk); 5156 VI_LOCK(vp); 5157 while ((nbp = TAILQ_NEXT(bp, b_bobufs)) != NULL) { 5158 nbp = getdirtybuf(nbp, VI_MTX(vp), MNT_WAIT); 5159 if (nbp) 5160 break; 5161 } 5162 VI_UNLOCK(vp); 5163 bp->b_lock.lk_flags &= ~LK_CANRECURSE; 5164 bawrite(bp); 5165 if (nbp != NULL) { 5166 bp = nbp; 5167 goto loop; 5168 } 5169 /* 5170 * The brief unlock is to allow any pent up dependency 5171 * processing to be done. Then proceed with the second pass. 5172 */ 5173 if (waitfor == MNT_NOWAIT) { 5174 waitfor = MNT_WAIT; 5175 goto top; 5176 } 5177 5178 /* 5179 * If we have managed to get rid of all the dirty buffers, 5180 * then we are done. For certain directories and block 5181 * devices, we may need to do further work. 5182 * 5183 * We must wait for any I/O in progress to finish so that 5184 * all potential buffers on the dirty list will be visible. 5185 */ 5186 VI_LOCK(vp); 5187 drain_output(vp); 5188 VI_UNLOCK(vp); 5189 return (0); 5190} 5191 5192/* 5193 * Flush the dependencies associated with an inodedep. 5194 * Called with splbio blocked. 5195 */ 5196static int 5197flush_inodedep_deps(fs, ino) 5198 struct fs *fs; 5199 ino_t ino; 5200{ 5201 struct inodedep *inodedep; 5202 int error, waitfor; 5203 5204 /* 5205 * This work is done in two passes. The first pass grabs most 5206 * of the buffers and begins asynchronously writing them. The 5207 * only way to wait for these asynchronous writes is to sleep 5208 * on the filesystem vnode which may stay busy for a long time 5209 * if the filesystem is active. So, instead, we make a second 5210 * pass over the dependencies blocking on each write. In the 5211 * usual case we will be blocking against a write that we 5212 * initiated, so when it is done the dependency will have been 5213 * resolved. Thus the second pass is expected to end quickly. 5214 * We give a brief window at the top of the loop to allow 5215 * any pending I/O to complete. 5216 */ 5217 for (error = 0, waitfor = MNT_NOWAIT; ; ) { 5218 if (error) 5219 return (error); 5220 FREE_LOCK(&lk); 5221 ACQUIRE_LOCK(&lk); 5222 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 5223 return (0); 5224 if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) || 5225 flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) || 5226 flush_deplist(&inodedep->id_extupdt, waitfor, &error) || 5227 flush_deplist(&inodedep->id_newextupdt, waitfor, &error)) 5228 continue; 5229 /* 5230 * If pass2, we are done, otherwise do pass 2. 5231 */ 5232 if (waitfor == MNT_WAIT) 5233 break; 5234 waitfor = MNT_WAIT; 5235 } 5236 /* 5237 * Try freeing inodedep in case all dependencies have been removed. 5238 */ 5239 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0) 5240 (void) free_inodedep(inodedep); 5241 return (0); 5242} 5243 5244/* 5245 * Flush an inode dependency list. 5246 * Called with splbio blocked. 5247 */ 5248static int 5249flush_deplist(listhead, waitfor, errorp) 5250 struct allocdirectlst *listhead; 5251 int waitfor; 5252 int *errorp; 5253{ 5254 struct allocdirect *adp; 5255 struct buf *bp; 5256 5257 mtx_assert(&lk, MA_OWNED); 5258 TAILQ_FOREACH(adp, listhead, ad_next) { 5259 if (adp->ad_state & DEPCOMPLETE) 5260 continue; 5261 bp = adp->ad_buf; 5262 bp = getdirtybuf(bp, &lk, waitfor); 5263 if (bp == NULL) { 5264 if (waitfor == MNT_NOWAIT) 5265 continue; 5266 return (1); 5267 } 5268 FREE_LOCK(&lk); 5269 if (waitfor == MNT_NOWAIT) { 5270 bawrite(bp); 5271 } else if ((*errorp = bwrite(bp)) != 0) { 5272 ACQUIRE_LOCK(&lk); 5273 return (1); 5274 } 5275 ACQUIRE_LOCK(&lk); 5276 return (1); 5277 } 5278 return (0); 5279} 5280 5281/* 5282 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 5283 * Called with splbio blocked. 5284 */ 5285static int 5286flush_pagedep_deps(pvp, mp, diraddhdp) 5287 struct vnode *pvp; 5288 struct mount *mp; 5289 struct diraddhd *diraddhdp; 5290{ 5291 struct inodedep *inodedep; 5292 struct ufsmount *ump; 5293 struct diradd *dap; 5294 struct vnode *vp; 5295 int error = 0; 5296 struct buf *bp; 5297 ino_t inum; 5298 5299 ump = VFSTOUFS(mp); 5300 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 5301 /* 5302 * Flush ourselves if this directory entry 5303 * has a MKDIR_PARENT dependency. 5304 */ 5305 if (dap->da_state & MKDIR_PARENT) { 5306 FREE_LOCK(&lk); 5307 if ((error = ffs_update(pvp, 1)) != 0) 5308 break; 5309 ACQUIRE_LOCK(&lk); 5310 /* 5311 * If that cleared dependencies, go on to next. 5312 */ 5313 if (dap != LIST_FIRST(diraddhdp)) 5314 continue; 5315 if (dap->da_state & MKDIR_PARENT) 5316 panic("flush_pagedep_deps: MKDIR_PARENT"); 5317 } 5318 /* 5319 * A newly allocated directory must have its "." and 5320 * ".." entries written out before its name can be 5321 * committed in its parent. We do not want or need 5322 * the full semantics of a synchronous ffs_syncvnode as 5323 * that may end up here again, once for each directory 5324 * level in the filesystem. Instead, we push the blocks 5325 * and wait for them to clear. We have to fsync twice 5326 * because the first call may choose to defer blocks 5327 * that still have dependencies, but deferral will 5328 * happen at most once. 5329 */ 5330 inum = dap->da_newinum; 5331 if (dap->da_state & MKDIR_BODY) { 5332 FREE_LOCK(&lk); 5333 if ((error = ffs_vget(mp, inum, LK_EXCLUSIVE, &vp))) 5334 break; 5335 if ((error=ffs_syncvnode(vp, MNT_NOWAIT)) || 5336 (error=ffs_syncvnode(vp, MNT_NOWAIT))) { 5337 vput(vp); 5338 break; 5339 } 5340 VI_LOCK(vp); 5341 drain_output(vp); 5342 VI_UNLOCK(vp); 5343 vput(vp); 5344 ACQUIRE_LOCK(&lk); 5345 /* 5346 * If that cleared dependencies, go on to next. 5347 */ 5348 if (dap != LIST_FIRST(diraddhdp)) 5349 continue; 5350 if (dap->da_state & MKDIR_BODY) 5351 panic("flush_pagedep_deps: MKDIR_BODY"); 5352 } 5353 /* 5354 * Flush the inode on which the directory entry depends. 5355 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 5356 * the only remaining dependency is that the updated inode 5357 * count must get pushed to disk. The inode has already 5358 * been pushed into its inode buffer (via VOP_UPDATE) at 5359 * the time of the reference count change. So we need only 5360 * locate that buffer, ensure that there will be no rollback 5361 * caused by a bitmap dependency, then write the inode buffer. 5362 */ 5363retry: 5364 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) 5365 panic("flush_pagedep_deps: lost inode"); 5366 /* 5367 * If the inode still has bitmap dependencies, 5368 * push them to disk. 5369 */ 5370 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 5371 bp = inodedep->id_buf; 5372 bp = getdirtybuf(bp, &lk, MNT_WAIT); 5373 if (bp == NULL) 5374 goto retry; 5375 FREE_LOCK(&lk); 5376 if ((error = bwrite(bp)) != 0) 5377 break; 5378 ACQUIRE_LOCK(&lk); 5379 if (dap != LIST_FIRST(diraddhdp)) 5380 continue; 5381 } 5382 /* 5383 * If the inode is still sitting in a buffer waiting 5384 * to be written, push it to disk. 5385 */ 5386 FREE_LOCK(&lk); 5387 if ((error = bread(ump->um_devvp, 5388 fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), 5389 (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0) { 5390 brelse(bp); 5391 break; 5392 } 5393 if ((error = bwrite(bp)) != 0) 5394 break; 5395 ACQUIRE_LOCK(&lk); 5396 /* 5397 * If we have failed to get rid of all the dependencies 5398 * then something is seriously wrong. 5399 */ 5400 if (dap == LIST_FIRST(diraddhdp)) 5401 panic("flush_pagedep_deps: flush failed"); 5402 } 5403 if (error) 5404 ACQUIRE_LOCK(&lk); 5405 return (error); 5406} 5407 5408/* 5409 * A large burst of file addition or deletion activity can drive the 5410 * memory load excessively high. First attempt to slow things down 5411 * using the techniques below. If that fails, this routine requests 5412 * the offending operations to fall back to running synchronously 5413 * until the memory load returns to a reasonable level. 5414 */ 5415int 5416softdep_slowdown(vp) 5417 struct vnode *vp; 5418{ 5419 int max_softdeps_hard; 5420 5421 max_softdeps_hard = max_softdeps * 11 / 10; 5422 if (num_dirrem < max_softdeps_hard / 2 && 5423 num_inodedep < max_softdeps_hard && 5424 VFSTOUFS(vp->v_mount)->um_numindirdeps < maxindirdeps) 5425 return (0); 5426 if (VFSTOUFS(vp->v_mount)->um_numindirdeps >= maxindirdeps) 5427 speedup_syncer(); 5428 stat_sync_limit_hit += 1; 5429 return (1); 5430} 5431 5432/* 5433 * Called by the allocation routines when they are about to fail 5434 * in the hope that we can free up some disk space. 5435 * 5436 * First check to see if the work list has anything on it. If it has, 5437 * clean up entries until we successfully free some space. Because this 5438 * process holds inodes locked, we cannot handle any remove requests 5439 * that might block on a locked inode as that could lead to deadlock. 5440 * If the worklist yields no free space, encourage the syncer daemon 5441 * to help us. In no event will we try for longer than tickdelay seconds. 5442 */ 5443int 5444softdep_request_cleanup(fs, vp) 5445 struct fs *fs; 5446 struct vnode *vp; 5447{ 5448 struct ufsmount *ump; 5449 long starttime; 5450 ufs2_daddr_t needed; 5451 int error; 5452 5453 ump = VTOI(vp)->i_ump; 5454 mtx_assert(UFS_MTX(ump), MA_OWNED); 5455 needed = fs->fs_cstotal.cs_nbfree + fs->fs_contigsumsize; 5456 starttime = time_second + tickdelay; 5457 /* 5458 * If we are being called because of a process doing a 5459 * copy-on-write, then it is not safe to update the vnode 5460 * as we may recurse into the copy-on-write routine. 5461 */ 5462 if (!(curthread->td_pflags & TDP_COWINPROGRESS)) { 5463 UFS_UNLOCK(ump); 5464 error = ffs_update(vp, 1); 5465 UFS_LOCK(ump); 5466 if (error != 0) 5467 return (0); 5468 } 5469 while (fs->fs_pendingblocks > 0 && fs->fs_cstotal.cs_nbfree <= needed) { 5470 if (time_second > starttime) 5471 return (0); 5472 UFS_UNLOCK(ump); 5473 ACQUIRE_LOCK(&lk); 5474 if (num_on_worklist > 0 && 5475 process_worklist_item(NULL, LK_NOWAIT) != -1) { 5476 stat_worklist_push += 1; 5477 FREE_LOCK(&lk); 5478 UFS_LOCK(ump); 5479 continue; 5480 } 5481 request_cleanup(FLUSH_REMOVE_WAIT); 5482 FREE_LOCK(&lk); 5483 UFS_LOCK(ump); 5484 } 5485 return (1); 5486} 5487 5488/* 5489 * If memory utilization has gotten too high, deliberately slow things 5490 * down and speed up the I/O processing. 5491 */ 5492static int 5493request_cleanup(resource) 5494 int resource; 5495{ 5496 struct thread *td = curthread; 5497 5498 mtx_assert(&lk, MA_OWNED); 5499 /* 5500 * We never hold up the filesystem syncer process. 5501 */ 5502 if (td == filesys_syncer || (td->td_pflags & TDP_SOFTDEP)) 5503 return (0); 5504 /* 5505 * First check to see if the work list has gotten backlogged. 5506 * If it has, co-opt this process to help clean up two entries. 5507 * Because this process may hold inodes locked, we cannot 5508 * handle any remove requests that might block on a locked 5509 * inode as that could lead to deadlock. We set TDP_SOFTDEP 5510 * to avoid recursively processing the worklist. 5511 */ 5512 if (num_on_worklist > max_softdeps / 10) { 5513 td->td_pflags |= TDP_SOFTDEP; 5514 process_worklist_item(NULL, LK_NOWAIT); 5515 process_worklist_item(NULL, LK_NOWAIT); 5516 td->td_pflags &= ~TDP_SOFTDEP; 5517 stat_worklist_push += 2; 5518 return(1); 5519 } 5520 /* 5521 * Next, we attempt to speed up the syncer process. If that 5522 * is successful, then we allow the process to continue. 5523 */ 5524 if (speedup_syncer() && resource != FLUSH_REMOVE_WAIT) 5525 return(0); 5526 /* 5527 * If we are resource constrained on inode dependencies, try 5528 * flushing some dirty inodes. Otherwise, we are constrained 5529 * by file deletions, so try accelerating flushes of directories 5530 * with removal dependencies. We would like to do the cleanup 5531 * here, but we probably hold an inode locked at this point and 5532 * that might deadlock against one that we try to clean. So, 5533 * the best that we can do is request the syncer daemon to do 5534 * the cleanup for us. 5535 */ 5536 switch (resource) { 5537 5538 case FLUSH_INODES: 5539 stat_ino_limit_push += 1; 5540 req_clear_inodedeps += 1; 5541 stat_countp = &stat_ino_limit_hit; 5542 break; 5543 5544 case FLUSH_REMOVE: 5545 case FLUSH_REMOVE_WAIT: 5546 stat_blk_limit_push += 1; 5547 req_clear_remove += 1; 5548 stat_countp = &stat_blk_limit_hit; 5549 break; 5550 5551 default: 5552 panic("request_cleanup: unknown type"); 5553 } 5554 /* 5555 * Hopefully the syncer daemon will catch up and awaken us. 5556 * We wait at most tickdelay before proceeding in any case. 5557 */ 5558 proc_waiting += 1; 5559 if (handle.callout == NULL) 5560 handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2); 5561 msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0); 5562 proc_waiting -= 1; 5563 return (1); 5564} 5565 5566/* 5567 * Awaken processes pausing in request_cleanup and clear proc_waiting 5568 * to indicate that there is no longer a timer running. 5569 */ 5570static void 5571pause_timer(arg) 5572 void *arg; 5573{ 5574 5575 ACQUIRE_LOCK(&lk); 5576 *stat_countp += 1; 5577 wakeup_one(&proc_waiting); 5578 if (proc_waiting > 0) 5579 handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2); 5580 else 5581 handle.callout = NULL; 5582 FREE_LOCK(&lk); 5583} 5584 5585/* 5586 * Flush out a directory with at least one removal dependency in an effort to 5587 * reduce the number of dirrem, freefile, and freeblks dependency structures. 5588 */ 5589static void 5590clear_remove(td) 5591 struct thread *td; 5592{ 5593 struct pagedep_hashhead *pagedephd; 5594 struct pagedep *pagedep; 5595 static int next = 0; 5596 struct mount *mp; 5597 struct vnode *vp; 5598 int error, cnt; 5599 ino_t ino; 5600 5601 mtx_assert(&lk, MA_OWNED); 5602 5603 for (cnt = 0; cnt < pagedep_hash; cnt++) { 5604 pagedephd = &pagedep_hashtbl[next++]; 5605 if (next >= pagedep_hash) 5606 next = 0; 5607 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 5608 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL) 5609 continue; 5610 mp = pagedep->pd_mnt; 5611 ino = pagedep->pd_ino; 5612 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 5613 continue; 5614 FREE_LOCK(&lk); 5615 if ((error = ffs_vget(mp, ino, LK_EXCLUSIVE, &vp))) { 5616 softdep_error("clear_remove: vget", error); 5617 vn_finished_write(mp); 5618 ACQUIRE_LOCK(&lk); 5619 return; 5620 } 5621 if ((error = ffs_syncvnode(vp, MNT_NOWAIT))) 5622 softdep_error("clear_remove: fsync", error); 5623 VI_LOCK(vp); 5624 drain_output(vp); 5625 VI_UNLOCK(vp); 5626 vput(vp); 5627 vn_finished_write(mp); 5628 ACQUIRE_LOCK(&lk); 5629 return; 5630 } 5631 } 5632} 5633 5634/* 5635 * Clear out a block of dirty inodes in an effort to reduce 5636 * the number of inodedep dependency structures. 5637 */ 5638static void 5639clear_inodedeps(td) 5640 struct thread *td; 5641{ 5642 struct inodedep_hashhead *inodedephd; 5643 struct inodedep *inodedep; 5644 static int next = 0; 5645 struct mount *mp; 5646 struct vnode *vp; 5647 struct fs *fs; 5648 int error, cnt; 5649 ino_t firstino, lastino, ino; 5650 5651 mtx_assert(&lk, MA_OWNED); 5652 /* 5653 * Pick a random inode dependency to be cleared. 5654 * We will then gather up all the inodes in its block 5655 * that have dependencies and flush them out. 5656 */ 5657 for (cnt = 0; cnt < inodedep_hash; cnt++) { 5658 inodedephd = &inodedep_hashtbl[next++]; 5659 if (next >= inodedep_hash) 5660 next = 0; 5661 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 5662 break; 5663 } 5664 if (inodedep == NULL) 5665 return; 5666 /* 5667 * Ugly code to find mount point given pointer to superblock. 5668 */ 5669 fs = inodedep->id_fs; 5670 TAILQ_FOREACH(mp, &mountlist, mnt_list) 5671 if ((mp->mnt_flag & MNT_SOFTDEP) && fs == VFSTOUFS(mp)->um_fs) 5672 break; 5673 /* 5674 * Find the last inode in the block with dependencies. 5675 */ 5676 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 5677 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 5678 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0) 5679 break; 5680 /* 5681 * Asynchronously push all but the last inode with dependencies. 5682 * Synchronously push the last inode with dependencies to ensure 5683 * that the inode block gets written to free up the inodedeps. 5684 */ 5685 for (ino = firstino; ino <= lastino; ino++) { 5686 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 5687 continue; 5688 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 5689 continue; 5690 FREE_LOCK(&lk); 5691 if ((error = ffs_vget(mp, ino, LK_EXCLUSIVE, &vp)) != 0) { 5692 softdep_error("clear_inodedeps: vget", error); 5693 vn_finished_write(mp); 5694 ACQUIRE_LOCK(&lk); 5695 return; 5696 } 5697 if (ino == lastino) { 5698 if ((error = ffs_syncvnode(vp, MNT_WAIT))) 5699 softdep_error("clear_inodedeps: fsync1", error); 5700 } else { 5701 if ((error = ffs_syncvnode(vp, MNT_NOWAIT))) 5702 softdep_error("clear_inodedeps: fsync2", error); 5703 VI_LOCK(vp); 5704 drain_output(vp); 5705 VI_UNLOCK(vp); 5706 } 5707 vput(vp); 5708 vn_finished_write(mp); 5709 ACQUIRE_LOCK(&lk); 5710 } 5711} 5712 5713/* 5714 * Function to determine if the buffer has outstanding dependencies 5715 * that will cause a roll-back if the buffer is written. If wantcount 5716 * is set, return number of dependencies, otherwise just yes or no. 5717 */ 5718static int 5719softdep_count_dependencies(bp, wantcount) 5720 struct buf *bp; 5721 int wantcount; 5722{ 5723 struct worklist *wk; 5724 struct inodedep *inodedep; 5725 struct indirdep *indirdep; 5726 struct allocindir *aip; 5727 struct pagedep *pagedep; 5728 struct diradd *dap; 5729 int i, retval; 5730 5731 retval = 0; 5732 ACQUIRE_LOCK(&lk); 5733 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5734 switch (wk->wk_type) { 5735 5736 case D_INODEDEP: 5737 inodedep = WK_INODEDEP(wk); 5738 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 5739 /* bitmap allocation dependency */ 5740 retval += 1; 5741 if (!wantcount) 5742 goto out; 5743 } 5744 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 5745 /* direct block pointer dependency */ 5746 retval += 1; 5747 if (!wantcount) 5748 goto out; 5749 } 5750 if (TAILQ_FIRST(&inodedep->id_extupdt)) { 5751 /* direct block pointer dependency */ 5752 retval += 1; 5753 if (!wantcount) 5754 goto out; 5755 } 5756 continue; 5757 5758 case D_INDIRDEP: 5759 indirdep = WK_INDIRDEP(wk); 5760 5761 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 5762 /* indirect block pointer dependency */ 5763 retval += 1; 5764 if (!wantcount) 5765 goto out; 5766 } 5767 continue; 5768 5769 case D_PAGEDEP: 5770 pagedep = WK_PAGEDEP(wk); 5771 for (i = 0; i < DAHASHSZ; i++) { 5772 5773 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 5774 /* directory entry dependency */ 5775 retval += 1; 5776 if (!wantcount) 5777 goto out; 5778 } 5779 } 5780 continue; 5781 5782 case D_BMSAFEMAP: 5783 case D_ALLOCDIRECT: 5784 case D_ALLOCINDIR: 5785 case D_MKDIR: 5786 /* never a dependency on these blocks */ 5787 continue; 5788 5789 default: 5790 panic("softdep_check_for_rollback: Unexpected type %s", 5791 TYPENAME(wk->wk_type)); 5792 /* NOTREACHED */ 5793 } 5794 } 5795out: 5796 FREE_LOCK(&lk); 5797 return retval; 5798} 5799 5800/* 5801 * Acquire exclusive access to a buffer. 5802 * Must be called with a locked mtx parameter. 5803 * Return acquired buffer or NULL on failure. 5804 */ 5805static struct buf * 5806getdirtybuf(bp, mtx, waitfor) 5807 struct buf *bp; 5808 struct mtx *mtx; 5809 int waitfor; 5810{ 5811 int error; 5812 5813 mtx_assert(mtx, MA_OWNED); 5814 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) { 5815 if (waitfor != MNT_WAIT) 5816 return (NULL); 5817 error = BUF_LOCK(bp, 5818 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, mtx); 5819 /* 5820 * Even if we sucessfully acquire bp here, we have dropped 5821 * mtx, which may violates our guarantee. 5822 */ 5823 if (error == 0) 5824 BUF_UNLOCK(bp); 5825 else if (error != ENOLCK) 5826 panic("getdirtybuf: inconsistent lock: %d", error); 5827 mtx_lock(mtx); 5828 return (NULL); 5829 } 5830 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 5831 BUF_UNLOCK(bp); 5832 if (waitfor != MNT_WAIT) 5833 return (NULL); 5834 /* 5835 * The mtx argument must be bp->b_vp's mutex in 5836 * this case. 5837 */ 5838#ifdef DEBUG_VFS_LOCKS 5839 if (bp->b_vp->v_type != VCHR) 5840 ASSERT_VI_LOCKED(bp->b_vp, "getdirtybuf"); 5841#endif 5842 bp->b_vflags |= BV_BKGRDWAIT; 5843 msleep(&bp->b_xflags, mtx, PRIBIO, "getbuf", 0); 5844 return (NULL); 5845 } 5846 if ((bp->b_flags & B_DELWRI) == 0) { 5847 BUF_UNLOCK(bp); 5848 return (NULL); 5849 } 5850 bremfree(bp); 5851 return (bp); 5852} 5853 5854/* 5855 * Wait for pending output on a vnode to complete. 5856 * Must be called with vnode lock and interlock locked. 5857 * 5858 * XXX: Should just be a call to bufobj_wwait(). 5859 */ 5860static void 5861drain_output(vp) 5862 struct vnode *vp; 5863{ 5864 ASSERT_VOP_LOCKED(vp, "drain_output"); 5865 ASSERT_VI_LOCKED(vp, "drain_output"); 5866 5867 while (vp->v_bufobj.bo_numoutput) { 5868 vp->v_bufobj.bo_flag |= BO_WWAIT; 5869 msleep((caddr_t)&vp->v_bufobj.bo_numoutput, 5870 VI_MTX(vp), PRIBIO + 1, "drainvp", 0); 5871 } 5872} 5873 5874/* 5875 * Called whenever a buffer that is being invalidated or reallocated 5876 * contains dependencies. This should only happen if an I/O error has 5877 * occurred. The routine is called with the buffer locked. 5878 */ 5879static void 5880softdep_deallocate_dependencies(bp) 5881 struct buf *bp; 5882{ 5883 5884 if ((bp->b_ioflags & BIO_ERROR) == 0) 5885 panic("softdep_deallocate_dependencies: dangling deps"); 5886 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); 5887 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 5888} 5889 5890/* 5891 * Function to handle asynchronous write errors in the filesystem. 5892 */ 5893static void 5894softdep_error(func, error) 5895 char *func; 5896 int error; 5897{ 5898 5899 /* XXX should do something better! */ 5900 printf("%s: got error %d while accessing filesystem\n", func, error); 5901} 5902 5903#endif /* SOFTUPDATES */ 5904