1 /* 2 * Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved. 3 * 4 * The soft updates code is derived from the appendix of a University 5 * of Michigan technical report (Gregory R. Ganger and Yale N. Patt, 6 * "Soft Updates: A Solution to the Metadata Update Problem in File 7 * 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 * $FreeBSD: src/sys/ufs/ffs/ffs_softdep.c,v 1.57.2.11 2002/02/05 18:46:53 dillon Exp $ 40 */ 41 42 /* 43 * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide. 44 */ 45 #ifndef DIAGNOSTIC 46 #define DIAGNOSTIC 47 #endif 48 #ifndef DEBUG 49 #define DEBUG 50 #endif 51 52 #include <sys/param.h> 53 #include <sys/kernel.h> 54 #include <sys/systm.h> 55 #include <sys/buf.h> 56 #include <sys/malloc.h> 57 #include <sys/mount.h> 58 #include <sys/proc.h> 59 #include <sys/syslog.h> 60 #include <sys/vnode.h> 61 #include <sys/conf.h> 62 #include <machine/inttypes.h> 63 #include "dir.h" 64 #include "quota.h" 65 #include "inode.h" 66 #include "ufsmount.h" 67 #include "fs.h" 68 #include "softdep.h" 69 #include "ffs_extern.h" 70 #include "ufs_extern.h" 71 72 #include <sys/buf2.h> 73 #include <sys/thread2.h> 74 #include <sys/lock.h> 75 76 /* 77 * These definitions need to be adapted to the system to which 78 * this file is being ported. 79 */ 80 /* 81 * malloc types defined for the softdep system. 82 */ 83 MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies"); 84 MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies"); 85 MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation"); 86 MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map"); 87 MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode"); 88 MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies"); 89 MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block"); 90 MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode"); 91 MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode"); 92 MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated"); 93 MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry"); 94 MALLOC_DEFINE(M_MKDIR, "mkdir","New directory"); 95 MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted"); 96 97 #define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE) 98 99 #define D_PAGEDEP 0 100 #define D_INODEDEP 1 101 #define D_NEWBLK 2 102 #define D_BMSAFEMAP 3 103 #define D_ALLOCDIRECT 4 104 #define D_INDIRDEP 5 105 #define D_ALLOCINDIR 6 106 #define D_FREEFRAG 7 107 #define D_FREEBLKS 8 108 #define D_FREEFILE 9 109 #define D_DIRADD 10 110 #define D_MKDIR 11 111 #define D_DIRREM 12 112 #define D_LAST D_DIRREM 113 114 /* 115 * translate from workitem type to memory type 116 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 117 */ 118 static struct malloc_type *memtype[] = { 119 M_PAGEDEP, 120 M_INODEDEP, 121 M_NEWBLK, 122 M_BMSAFEMAP, 123 M_ALLOCDIRECT, 124 M_INDIRDEP, 125 M_ALLOCINDIR, 126 M_FREEFRAG, 127 M_FREEBLKS, 128 M_FREEFILE, 129 M_DIRADD, 130 M_MKDIR, 131 M_DIRREM 132 }; 133 134 #define DtoM(type) (memtype[type]) 135 136 /* 137 * Names of malloc types. 138 */ 139 #define TYPENAME(type) \ 140 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???") 141 /* 142 * End system adaptaion definitions. 143 */ 144 145 /* 146 * Internal function prototypes. 147 */ 148 static void softdep_error(char *, int); 149 static void drain_output(struct vnode *, int); 150 static int getdirtybuf(struct buf **, int); 151 static void clear_remove(struct thread *); 152 static void clear_inodedeps(struct thread *); 153 static int flush_pagedep_deps(struct vnode *, struct mount *, 154 struct diraddhd *); 155 static int flush_inodedep_deps(struct fs *, ino_t); 156 static int handle_written_filepage(struct pagedep *, struct buf *); 157 static void diradd_inode_written(struct diradd *, struct inodedep *); 158 static int handle_written_inodeblock(struct inodedep *, struct buf *); 159 static void handle_allocdirect_partdone(struct allocdirect *); 160 static void handle_allocindir_partdone(struct allocindir *); 161 static void initiate_write_filepage(struct pagedep *, struct buf *); 162 static void handle_written_mkdir(struct mkdir *, int); 163 static void initiate_write_inodeblock(struct inodedep *, struct buf *); 164 static void handle_workitem_freefile(struct freefile *); 165 static void handle_workitem_remove(struct dirrem *); 166 static struct dirrem *newdirrem(struct buf *, struct inode *, 167 struct inode *, int, struct dirrem **); 168 static void free_diradd(struct diradd *); 169 static void free_allocindir(struct allocindir *, struct inodedep *); 170 static int indir_trunc (struct inode *, off_t, int, ufs_lbn_t, long *); 171 static void deallocate_dependencies(struct buf *, struct inodedep *); 172 static void free_allocdirect(struct allocdirectlst *, 173 struct allocdirect *, int); 174 static int check_inode_unwritten(struct inodedep *); 175 static int free_inodedep(struct inodedep *); 176 static void handle_workitem_freeblocks(struct freeblks *); 177 static void merge_inode_lists(struct inodedep *); 178 static void setup_allocindir_phase2(struct buf *, struct inode *, 179 struct allocindir *); 180 static struct allocindir *newallocindir(struct inode *, int, ufs_daddr_t, 181 ufs_daddr_t); 182 static void handle_workitem_freefrag(struct freefrag *); 183 static struct freefrag *newfreefrag(struct inode *, ufs_daddr_t, long); 184 static void allocdirect_merge(struct allocdirectlst *, 185 struct allocdirect *, struct allocdirect *); 186 static struct bmsafemap *bmsafemap_lookup(struct buf *); 187 static int newblk_lookup(struct fs *, ufs_daddr_t, int, 188 struct newblk **); 189 static int inodedep_lookup(struct fs *, ino_t, int, struct inodedep **); 190 static int pagedep_lookup(struct inode *, ufs_lbn_t, int, 191 struct pagedep **); 192 static int request_cleanup(int); 193 static int process_worklist_item(struct mount *, int); 194 static void add_to_worklist(struct worklist *); 195 196 /* 197 * Exported softdep operations. 198 */ 199 static void softdep_disk_io_initiation(struct buf *); 200 static void softdep_disk_write_complete(struct buf *); 201 static void softdep_deallocate_dependencies(struct buf *); 202 static int softdep_fsync(struct vnode *); 203 static int softdep_process_worklist(struct mount *); 204 static void softdep_move_dependencies(struct buf *, struct buf *); 205 static int softdep_count_dependencies(struct buf *bp, int); 206 static int softdep_checkread(struct buf *bp); 207 static int softdep_checkwrite(struct buf *bp); 208 209 static struct bio_ops softdep_bioops = { 210 .io_start = softdep_disk_io_initiation, 211 .io_complete = softdep_disk_write_complete, 212 .io_deallocate = softdep_deallocate_dependencies, 213 .io_fsync = softdep_fsync, 214 .io_sync = softdep_process_worklist, 215 .io_movedeps = softdep_move_dependencies, 216 .io_countdeps = softdep_count_dependencies, 217 .io_checkread = softdep_checkread, 218 .io_checkwrite = softdep_checkwrite 219 }; 220 221 /* 222 * Locking primitives. 223 */ 224 static void acquire_lock(struct lock *); 225 static void free_lock(struct lock *); 226 #ifdef INVARIANTS 227 static int lock_held(struct lock *); 228 #endif 229 230 static struct lock lk; 231 232 #define ACQUIRE_LOCK(lkp) acquire_lock(lkp) 233 #define FREE_LOCK(lkp) free_lock(lkp) 234 235 static void 236 acquire_lock(struct lock *lkp) 237 { 238 lockmgr(lkp, LK_EXCLUSIVE); 239 } 240 241 static void 242 free_lock(struct lock *lkp) 243 { 244 lockmgr(lkp, LK_RELEASE); 245 } 246 247 #ifdef INVARIANTS 248 static int 249 lock_held(struct lock *lkp) 250 { 251 return lockcountnb(lkp); 252 } 253 #endif 254 255 /* 256 * Place holder for real semaphores. 257 */ 258 struct sema { 259 int value; 260 thread_t holder; 261 char *name; 262 int timo; 263 struct spinlock spin; 264 }; 265 static void sema_init(struct sema *, char *, int); 266 static int sema_get(struct sema *, struct lock *); 267 static void sema_release(struct sema *, struct lock *); 268 269 #define NOHOLDER ((struct thread *) -1) 270 271 static void 272 sema_init(struct sema *semap, char *name, int timo) 273 { 274 semap->holder = NOHOLDER; 275 semap->value = 0; 276 semap->name = name; 277 semap->timo = timo; 278 spin_init(&semap->spin); 279 } 280 281 /* 282 * Obtain exclusive access, semaphore is protected by the interlock. 283 * If interlock is NULL we must protect the semaphore ourselves. 284 */ 285 static int 286 sema_get(struct sema *semap, struct lock *interlock) 287 { 288 int rv; 289 290 if (interlock) { 291 if (semap->value > 0) { 292 ++semap->value; /* serves as wakeup flag */ 293 lksleep(semap, interlock, 0, 294 semap->name, semap->timo); 295 rv = 0; 296 } else { 297 semap->value = 1; /* serves as owned flag */ 298 semap->holder = curthread; 299 rv = 1; 300 } 301 } else { 302 spin_lock(&semap->spin); 303 if (semap->value > 0) { 304 ++semap->value; /* serves as wakeup flag */ 305 ssleep(semap, &semap->spin, 0, 306 semap->name, semap->timo); 307 spin_unlock(&semap->spin); 308 rv = 0; 309 } else { 310 semap->value = 1; /* serves as owned flag */ 311 semap->holder = curthread; 312 spin_unlock(&semap->spin); 313 rv = 1; 314 } 315 } 316 return (rv); 317 } 318 319 static void 320 sema_release(struct sema *semap, struct lock *lk) 321 { 322 if (semap->value <= 0 || semap->holder != curthread) 323 panic("sema_release: not held"); 324 if (lk) { 325 semap->holder = NOHOLDER; 326 if (--semap->value > 0) { 327 semap->value = 0; 328 wakeup(semap); 329 } 330 } else { 331 spin_lock(&semap->spin); 332 semap->holder = NOHOLDER; 333 if (--semap->value > 0) { 334 semap->value = 0; 335 spin_unlock(&semap->spin); 336 wakeup(semap); 337 } else { 338 spin_unlock(&semap->spin); 339 } 340 } 341 } 342 343 /* 344 * Worklist queue management. 345 * These routines require that the lock be held. 346 */ 347 static void worklist_insert(struct workhead *, struct worklist *); 348 static void worklist_remove(struct worklist *); 349 static void workitem_free(struct worklist *, int); 350 351 #define WORKLIST_INSERT_BP(bp, item) do { \ 352 (bp)->b_ops = &softdep_bioops; \ 353 worklist_insert(&(bp)->b_dep, item); \ 354 } while (0) 355 356 #define WORKLIST_INSERT(head, item) worklist_insert(head, item) 357 #define WORKLIST_REMOVE(item) worklist_remove(item) 358 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type) 359 360 static void 361 worklist_insert(struct workhead *head, struct worklist *item) 362 { 363 KKASSERT(lock_held(&lk) > 0); 364 365 if (item->wk_state & ONWORKLIST) { 366 panic("worklist_insert: already on list"); 367 } 368 item->wk_state |= ONWORKLIST; 369 LIST_INSERT_HEAD(head, item, wk_list); 370 } 371 372 static void 373 worklist_remove(struct worklist *item) 374 { 375 376 KKASSERT(lock_held(&lk)); 377 if ((item->wk_state & ONWORKLIST) == 0) 378 panic("worklist_remove: not on list"); 379 380 item->wk_state &= ~ONWORKLIST; 381 LIST_REMOVE(item, wk_list); 382 } 383 384 static void 385 workitem_free(struct worklist *item, int type) 386 { 387 388 if (item->wk_state & ONWORKLIST) 389 panic("workitem_free: still on list"); 390 if (item->wk_type != type) 391 panic("workitem_free: type mismatch"); 392 393 kfree(item, DtoM(type)); 394 } 395 396 /* 397 * Workitem queue management 398 */ 399 static struct workhead softdep_workitem_pending; 400 static int num_on_worklist; /* number of worklist items to be processed */ 401 static int softdep_worklist_busy; /* 1 => trying to do unmount */ 402 static int softdep_worklist_req; /* serialized waiters */ 403 static int max_softdeps; /* maximum number of structs before slowdown */ 404 static int tickdelay = 2; /* number of ticks to pause during slowdown */ 405 static int *stat_countp; /* statistic to count in proc_waiting timeout */ 406 static int proc_waiting; /* tracks whether we have a timeout posted */ 407 static struct thread *filesys_syncer; /* proc of filesystem syncer process */ 408 static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 409 #define FLUSH_INODES 1 410 static int req_clear_remove; /* syncer process flush some freeblks */ 411 #define FLUSH_REMOVE 2 412 /* 413 * runtime statistics 414 */ 415 static int stat_worklist_push; /* number of worklist cleanups */ 416 static int stat_blk_limit_push; /* number of times block limit neared */ 417 static int stat_ino_limit_push; /* number of times inode limit neared */ 418 static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 419 static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 420 static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 421 static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 422 static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 423 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 424 static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 425 #ifdef DEBUG 426 #include <vm/vm.h> 427 #include <sys/sysctl.h> 428 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, 429 "Maximum soft dependencies before slowdown occurs"); 430 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, 431 "Ticks to delay before allocating during slowdown"); 432 SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0, 433 "Number of worklist cleanups"); 434 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0, 435 "Number of times block limit neared"); 436 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0, 437 "Number of times inode limit neared"); 438 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, 439 "Number of times block slowdown imposed"); 440 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, 441 "Number of times inode slowdown imposed "); 442 SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, 443 "Number of synchronous slowdowns imposed"); 444 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, 445 "Bufs redirtied as indir ptrs not written"); 446 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, 447 "Bufs redirtied as inode bitmap not written"); 448 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, 449 "Bufs redirtied as direct ptrs not written"); 450 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, 451 "Bufs redirtied as dir entry cannot write"); 452 #endif /* DEBUG */ 453 454 /* 455 * Add an item to the end of the work queue. 456 * This routine requires that the lock be held. 457 * This is the only routine that adds items to the list. 458 * The following routine is the only one that removes items 459 * and does so in order from first to last. 460 */ 461 static void 462 add_to_worklist(struct worklist *wk) 463 { 464 static struct worklist *worklist_tail; 465 466 if (wk->wk_state & ONWORKLIST) { 467 panic("add_to_worklist: already on list"); 468 } 469 wk->wk_state |= ONWORKLIST; 470 if (LIST_FIRST(&softdep_workitem_pending) == NULL) 471 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list); 472 else 473 LIST_INSERT_AFTER(worklist_tail, wk, wk_list); 474 worklist_tail = wk; 475 num_on_worklist += 1; 476 } 477 478 /* 479 * Process that runs once per second to handle items in the background queue. 480 * 481 * Note that we ensure that everything is done in the order in which they 482 * appear in the queue. The code below depends on this property to ensure 483 * that blocks of a file are freed before the inode itself is freed. This 484 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 485 * until all the old ones have been purged from the dependency lists. 486 * 487 * bioops callback - hold io_token 488 */ 489 static int 490 softdep_process_worklist(struct mount *matchmnt) 491 { 492 thread_t td = curthread; 493 int matchcnt, loopcount; 494 int starttime; 495 496 ACQUIRE_LOCK(&lk); 497 498 /* 499 * Record the process identifier of our caller so that we can give 500 * this process preferential treatment in request_cleanup below. 501 */ 502 filesys_syncer = td; 503 matchcnt = 0; 504 505 /* 506 * There is no danger of having multiple processes run this 507 * code, but we have to single-thread it when softdep_flushfiles() 508 * is in operation to get an accurate count of the number of items 509 * related to its mount point that are in the list. 510 */ 511 if (matchmnt == NULL) { 512 if (softdep_worklist_busy < 0) { 513 matchcnt = -1; 514 goto done; 515 } 516 softdep_worklist_busy += 1; 517 } 518 519 /* 520 * If requested, try removing inode or removal dependencies. 521 */ 522 if (req_clear_inodedeps) { 523 clear_inodedeps(td); 524 req_clear_inodedeps -= 1; 525 wakeup_one(&proc_waiting); 526 } 527 if (req_clear_remove) { 528 clear_remove(td); 529 req_clear_remove -= 1; 530 wakeup_one(&proc_waiting); 531 } 532 loopcount = 1; 533 starttime = ticks; 534 while (num_on_worklist > 0) { 535 matchcnt += process_worklist_item(matchmnt, 0); 536 537 /* 538 * If a umount operation wants to run the worklist 539 * accurately, abort. 540 */ 541 if (softdep_worklist_req && matchmnt == NULL) { 542 matchcnt = -1; 543 break; 544 } 545 546 /* 547 * If requested, try removing inode or removal dependencies. 548 */ 549 if (req_clear_inodedeps) { 550 clear_inodedeps(td); 551 req_clear_inodedeps -= 1; 552 wakeup_one(&proc_waiting); 553 } 554 if (req_clear_remove) { 555 clear_remove(td); 556 req_clear_remove -= 1; 557 wakeup_one(&proc_waiting); 558 } 559 /* 560 * We do not generally want to stop for buffer space, but if 561 * we are really being a buffer hog, we will stop and wait. 562 */ 563 if (loopcount++ % 128 == 0) { 564 FREE_LOCK(&lk); 565 bwillinode(1); 566 ACQUIRE_LOCK(&lk); 567 } 568 569 /* 570 * Never allow processing to run for more than one 571 * second. Otherwise the other syncer tasks may get 572 * excessively backlogged. 573 * 574 * Use ticks to avoid boundary condition w/time_second or 575 * time_uptime. 576 */ 577 if ((ticks - starttime) > hz && matchmnt == NULL) { 578 matchcnt = -1; 579 break; 580 } 581 } 582 if (matchmnt == NULL) { 583 --softdep_worklist_busy; 584 if (softdep_worklist_req && softdep_worklist_busy == 0) 585 wakeup(&softdep_worklist_req); 586 } 587 done: 588 FREE_LOCK(&lk); 589 return (matchcnt); 590 } 591 592 /* 593 * Process one item on the worklist. 594 */ 595 static int 596 process_worklist_item(struct mount *matchmnt, int flags) 597 { 598 struct ufsmount *ump; 599 struct worklist *wk; 600 struct dirrem *dirrem; 601 struct fs *matchfs; 602 struct vnode *vp; 603 int matchcnt = 0; 604 605 KKASSERT(lock_held(&lk) > 0); 606 607 matchfs = NULL; 608 if (matchmnt != NULL) 609 matchfs = VFSTOUFS(matchmnt)->um_fs; 610 611 /* 612 * Normally we just process each item on the worklist in order. 613 * However, if we are in a situation where we cannot lock any 614 * inodes, we have to skip over any dirrem requests whose 615 * vnodes are resident and locked. 616 */ 617 LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) { 618 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM) 619 break; 620 dirrem = WK_DIRREM(wk); 621 ump = VFSTOUFS(dirrem->dm_mnt); 622 lwkt_gettoken(&ump->um_mountp->mnt_token); 623 vp = ufs_ihashlookup(ump, ump->um_dev, dirrem->dm_oldinum); 624 lwkt_reltoken(&ump->um_mountp->mnt_token); 625 if (vp == NULL || !vn_islocked(vp)) 626 break; 627 } 628 if (wk == NULL) { 629 return (0); 630 } 631 WORKLIST_REMOVE(wk); 632 num_on_worklist -= 1; 633 FREE_LOCK(&lk); 634 switch (wk->wk_type) { 635 case D_DIRREM: 636 /* removal of a directory entry */ 637 if (WK_DIRREM(wk)->dm_mnt == matchmnt) 638 matchcnt += 1; 639 handle_workitem_remove(WK_DIRREM(wk)); 640 break; 641 642 case D_FREEBLKS: 643 /* releasing blocks and/or fragments from a file */ 644 if (WK_FREEBLKS(wk)->fb_fs == matchfs) 645 matchcnt += 1; 646 handle_workitem_freeblocks(WK_FREEBLKS(wk)); 647 break; 648 649 case D_FREEFRAG: 650 /* releasing a fragment when replaced as a file grows */ 651 if (WK_FREEFRAG(wk)->ff_fs == matchfs) 652 matchcnt += 1; 653 handle_workitem_freefrag(WK_FREEFRAG(wk)); 654 break; 655 656 case D_FREEFILE: 657 /* releasing an inode when its link count drops to 0 */ 658 if (WK_FREEFILE(wk)->fx_fs == matchfs) 659 matchcnt += 1; 660 handle_workitem_freefile(WK_FREEFILE(wk)); 661 break; 662 663 default: 664 panic("%s_process_worklist: Unknown type %s", 665 "softdep", TYPENAME(wk->wk_type)); 666 /* NOTREACHED */ 667 } 668 ACQUIRE_LOCK(&lk); 669 return (matchcnt); 670 } 671 672 /* 673 * Move dependencies from one buffer to another. 674 * 675 * bioops callback - hold io_token 676 */ 677 static void 678 softdep_move_dependencies(struct buf *oldbp, struct buf *newbp) 679 { 680 struct worklist *wk, *wktail; 681 682 if (LIST_FIRST(&newbp->b_dep) != NULL) 683 panic("softdep_move_dependencies: need merge code"); 684 wktail = NULL; 685 ACQUIRE_LOCK(&lk); 686 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 687 LIST_REMOVE(wk, wk_list); 688 if (wktail == NULL) 689 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 690 else 691 LIST_INSERT_AFTER(wktail, wk, wk_list); 692 wktail = wk; 693 newbp->b_ops = &softdep_bioops; 694 } 695 FREE_LOCK(&lk); 696 } 697 698 /* 699 * Purge the work list of all items associated with a particular mount point. 700 */ 701 int 702 softdep_flushfiles(struct mount *oldmnt, int flags) 703 { 704 struct vnode *devvp; 705 int error, loopcnt; 706 707 /* 708 * Await our turn to clear out the queue, then serialize access. 709 */ 710 ACQUIRE_LOCK(&lk); 711 while (softdep_worklist_busy != 0) { 712 softdep_worklist_req += 1; 713 lksleep(&softdep_worklist_req, &lk, 0, "softflush", 0); 714 softdep_worklist_req -= 1; 715 } 716 softdep_worklist_busy = -1; 717 FREE_LOCK(&lk); 718 719 if ((error = ffs_flushfiles(oldmnt, flags)) != 0) { 720 softdep_worklist_busy = 0; 721 if (softdep_worklist_req) 722 wakeup(&softdep_worklist_req); 723 return (error); 724 } 725 /* 726 * Alternately flush the block device associated with the mount 727 * point and process any dependencies that the flushing 728 * creates. In theory, this loop can happen at most twice, 729 * but we give it a few extra just to be sure. 730 */ 731 devvp = VFSTOUFS(oldmnt)->um_devvp; 732 for (loopcnt = 10; loopcnt > 0; ) { 733 if (softdep_process_worklist(oldmnt) == 0) { 734 loopcnt--; 735 /* 736 * Do another flush in case any vnodes were brought in 737 * as part of the cleanup operations. 738 */ 739 if ((error = ffs_flushfiles(oldmnt, flags)) != 0) 740 break; 741 /* 742 * If we still found nothing to do, we are really done. 743 */ 744 if (softdep_process_worklist(oldmnt) == 0) 745 break; 746 } 747 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 748 error = VOP_FSYNC(devvp, MNT_WAIT, 0); 749 vn_unlock(devvp); 750 if (error) 751 break; 752 } 753 ACQUIRE_LOCK(&lk); 754 softdep_worklist_busy = 0; 755 if (softdep_worklist_req) 756 wakeup(&softdep_worklist_req); 757 FREE_LOCK(&lk); 758 759 /* 760 * If we are unmounting then it is an error to fail. If we 761 * are simply trying to downgrade to read-only, then filesystem 762 * activity can keep us busy forever, so we just fail with EBUSY. 763 */ 764 if (loopcnt == 0) { 765 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 766 panic("softdep_flushfiles: looping"); 767 error = EBUSY; 768 } 769 return (error); 770 } 771 772 /* 773 * Structure hashing. 774 * 775 * There are three types of structures that can be looked up: 776 * 1) pagedep structures identified by mount point, inode number, 777 * and logical block. 778 * 2) inodedep structures identified by mount point and inode number. 779 * 3) newblk structures identified by mount point and 780 * physical block number. 781 * 782 * The "pagedep" and "inodedep" dependency structures are hashed 783 * separately from the file blocks and inodes to which they correspond. 784 * This separation helps when the in-memory copy of an inode or 785 * file block must be replaced. It also obviates the need to access 786 * an inode or file page when simply updating (or de-allocating) 787 * dependency structures. Lookup of newblk structures is needed to 788 * find newly allocated blocks when trying to associate them with 789 * their allocdirect or allocindir structure. 790 * 791 * The lookup routines optionally create and hash a new instance when 792 * an existing entry is not found. 793 */ 794 #define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 795 #define NODELAY 0x0002 /* cannot do background work */ 796 797 /* 798 * Structures and routines associated with pagedep caching. 799 */ 800 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl; 801 u_long pagedep_hash; /* size of hash table - 1 */ 802 #define PAGEDEP_HASH(mp, inum, lbn) \ 803 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \ 804 pagedep_hash]) 805 static struct sema pagedep_in_progress; 806 807 /* 808 * Helper routine for pagedep_lookup() 809 */ 810 static __inline 811 struct pagedep * 812 pagedep_find(struct pagedep_hashhead *pagedephd, ino_t ino, ufs_lbn_t lbn, 813 struct mount *mp) 814 { 815 struct pagedep *pagedep; 816 817 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 818 if (ino == pagedep->pd_ino && 819 lbn == pagedep->pd_lbn && 820 mp == pagedep->pd_mnt) { 821 return (pagedep); 822 } 823 } 824 return(NULL); 825 } 826 827 /* 828 * Look up a pagedep. Return 1 if found, 0 if not found. 829 * If not found, allocate if DEPALLOC flag is passed. 830 * Found or allocated entry is returned in pagedeppp. 831 * This routine must be called with splbio interrupts blocked. 832 */ 833 static int 834 pagedep_lookup(struct inode *ip, ufs_lbn_t lbn, int flags, 835 struct pagedep **pagedeppp) 836 { 837 struct pagedep *pagedep; 838 struct pagedep_hashhead *pagedephd; 839 struct mount *mp; 840 int i; 841 842 KKASSERT(lock_held(&lk) > 0); 843 844 mp = ITOV(ip)->v_mount; 845 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn); 846 top: 847 *pagedeppp = pagedep_find(pagedephd, ip->i_number, lbn, mp); 848 if (*pagedeppp) 849 return(1); 850 if ((flags & DEPALLOC) == 0) 851 return (0); 852 if (sema_get(&pagedep_in_progress, &lk) == 0) 853 goto top; 854 855 FREE_LOCK(&lk); 856 pagedep = kmalloc(sizeof(struct pagedep), M_PAGEDEP, 857 M_SOFTDEP_FLAGS | M_ZERO); 858 ACQUIRE_LOCK(&lk); 859 if (pagedep_find(pagedephd, ip->i_number, lbn, mp)) { 860 kprintf("pagedep_lookup: blocking race avoided\n"); 861 sema_release(&pagedep_in_progress, &lk); 862 kfree(pagedep, M_PAGEDEP); 863 goto top; 864 } 865 866 pagedep->pd_list.wk_type = D_PAGEDEP; 867 pagedep->pd_mnt = mp; 868 pagedep->pd_ino = ip->i_number; 869 pagedep->pd_lbn = lbn; 870 LIST_INIT(&pagedep->pd_dirremhd); 871 LIST_INIT(&pagedep->pd_pendinghd); 872 for (i = 0; i < DAHASHSZ; i++) 873 LIST_INIT(&pagedep->pd_diraddhd[i]); 874 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 875 sema_release(&pagedep_in_progress, &lk); 876 *pagedeppp = pagedep; 877 return (0); 878 } 879 880 /* 881 * Structures and routines associated with inodedep caching. 882 */ 883 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl; 884 static u_long inodedep_hash; /* size of hash table - 1 */ 885 static long num_inodedep; /* number of inodedep allocated */ 886 #define INODEDEP_HASH(fs, inum) \ 887 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash]) 888 static struct sema inodedep_in_progress; 889 890 /* 891 * Helper routine for inodedep_lookup() 892 */ 893 static __inline 894 struct inodedep * 895 inodedep_find(struct inodedep_hashhead *inodedephd, struct fs *fs, ino_t inum) 896 { 897 struct inodedep *inodedep; 898 899 LIST_FOREACH(inodedep, inodedephd, id_hash) { 900 if (inum == inodedep->id_ino && fs == inodedep->id_fs) 901 return(inodedep); 902 } 903 return (NULL); 904 } 905 906 /* 907 * Look up a inodedep. Return 1 if found, 0 if not found. 908 * If not found, allocate if DEPALLOC flag is passed. 909 * Found or allocated entry is returned in inodedeppp. 910 * This routine must be called with splbio interrupts blocked. 911 */ 912 static int 913 inodedep_lookup(struct fs *fs, ino_t inum, int flags, 914 struct inodedep **inodedeppp) 915 { 916 struct inodedep *inodedep; 917 struct inodedep_hashhead *inodedephd; 918 919 KKASSERT(lock_held(&lk) > 0); 920 921 inodedephd = INODEDEP_HASH(fs, inum); 922 top: 923 *inodedeppp = inodedep_find(inodedephd, fs, inum); 924 if (*inodedeppp) 925 return (1); 926 if ((flags & DEPALLOC) == 0) 927 return (0); 928 929 /* 930 * If we are over our limit, try to improve the situation. 931 */ 932 if (num_inodedep > max_softdeps / 2) 933 speedup_syncer(NULL); 934 if (num_inodedep > max_softdeps && 935 (flags & NODELAY) == 0 && 936 request_cleanup(FLUSH_INODES)) { 937 goto top; 938 } 939 if (sema_get(&inodedep_in_progress, &lk) == 0) 940 goto top; 941 942 FREE_LOCK(&lk); 943 inodedep = kmalloc(sizeof(struct inodedep), M_INODEDEP, 944 M_SOFTDEP_FLAGS | M_ZERO); 945 ACQUIRE_LOCK(&lk); 946 if (inodedep_find(inodedephd, fs, inum)) { 947 kprintf("inodedep_lookup: blocking race avoided\n"); 948 sema_release(&inodedep_in_progress, &lk); 949 kfree(inodedep, M_INODEDEP); 950 goto top; 951 } 952 inodedep->id_list.wk_type = D_INODEDEP; 953 inodedep->id_fs = fs; 954 inodedep->id_ino = inum; 955 inodedep->id_state = ALLCOMPLETE; 956 inodedep->id_nlinkdelta = 0; 957 inodedep->id_savedino = NULL; 958 inodedep->id_savedsize = -1; 959 inodedep->id_buf = NULL; 960 LIST_INIT(&inodedep->id_pendinghd); 961 LIST_INIT(&inodedep->id_inowait); 962 LIST_INIT(&inodedep->id_bufwait); 963 TAILQ_INIT(&inodedep->id_inoupdt); 964 TAILQ_INIT(&inodedep->id_newinoupdt); 965 num_inodedep += 1; 966 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 967 sema_release(&inodedep_in_progress, &lk); 968 *inodedeppp = inodedep; 969 return (0); 970 } 971 972 /* 973 * Structures and routines associated with newblk caching. 974 */ 975 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl; 976 u_long newblk_hash; /* size of hash table - 1 */ 977 #define NEWBLK_HASH(fs, inum) \ 978 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash]) 979 static struct sema newblk_in_progress; 980 981 /* 982 * Helper routine for newblk_lookup() 983 */ 984 static __inline 985 struct newblk * 986 newblk_find(struct newblk_hashhead *newblkhd, struct fs *fs, 987 ufs_daddr_t newblkno) 988 { 989 struct newblk *newblk; 990 991 LIST_FOREACH(newblk, newblkhd, nb_hash) { 992 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs) 993 return (newblk); 994 } 995 return(NULL); 996 } 997 998 /* 999 * Look up a newblk. Return 1 if found, 0 if not found. 1000 * If not found, allocate if DEPALLOC flag is passed. 1001 * Found or allocated entry is returned in newblkpp. 1002 */ 1003 static int 1004 newblk_lookup(struct fs *fs, ufs_daddr_t newblkno, int flags, 1005 struct newblk **newblkpp) 1006 { 1007 struct newblk *newblk; 1008 struct newblk_hashhead *newblkhd; 1009 1010 newblkhd = NEWBLK_HASH(fs, newblkno); 1011 top: 1012 *newblkpp = newblk_find(newblkhd, fs, newblkno); 1013 if (*newblkpp) 1014 return(1); 1015 if ((flags & DEPALLOC) == 0) 1016 return (0); 1017 if (sema_get(&newblk_in_progress, NULL) == 0) 1018 goto top; 1019 1020 newblk = kmalloc(sizeof(struct newblk), M_NEWBLK, 1021 M_SOFTDEP_FLAGS | M_ZERO); 1022 1023 if (newblk_find(newblkhd, fs, newblkno)) { 1024 kprintf("newblk_lookup: blocking race avoided\n"); 1025 sema_release(&pagedep_in_progress, NULL); 1026 kfree(newblk, M_NEWBLK); 1027 goto top; 1028 } 1029 newblk->nb_state = 0; 1030 newblk->nb_fs = fs; 1031 newblk->nb_newblkno = newblkno; 1032 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 1033 sema_release(&newblk_in_progress, NULL); 1034 *newblkpp = newblk; 1035 return (0); 1036 } 1037 1038 /* 1039 * Executed during filesystem system initialization before 1040 * mounting any filesystems. 1041 */ 1042 void 1043 softdep_initialize(void) 1044 { 1045 LIST_INIT(&mkdirlisthd); 1046 LIST_INIT(&softdep_workitem_pending); 1047 max_softdeps = min(desiredvnodes * 8, 1048 M_INODEDEP->ks_limit / (2 * sizeof(struct inodedep))); 1049 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 1050 &pagedep_hash); 1051 lockinit(&lk, "ffs_softdep", 0, LK_CANRECURSE); 1052 sema_init(&pagedep_in_progress, "pagedep", 0); 1053 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash); 1054 sema_init(&inodedep_in_progress, "inodedep", 0); 1055 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash); 1056 sema_init(&newblk_in_progress, "newblk", 0); 1057 add_bio_ops(&softdep_bioops); 1058 } 1059 1060 /* 1061 * Called at mount time to notify the dependency code that a 1062 * filesystem wishes to use it. 1063 */ 1064 int 1065 softdep_mount(struct vnode *devvp, struct mount *mp, struct fs *fs) 1066 { 1067 struct csum cstotal; 1068 struct cg *cgp; 1069 struct buf *bp; 1070 int error, cyl; 1071 1072 mp->mnt_flag &= ~MNT_ASYNC; 1073 mp->mnt_flag |= MNT_SOFTDEP; 1074 mp->mnt_bioops = &softdep_bioops; 1075 /* 1076 * When doing soft updates, the counters in the 1077 * superblock may have gotten out of sync, so we have 1078 * to scan the cylinder groups and recalculate them. 1079 */ 1080 if (fs->fs_clean != 0) 1081 return (0); 1082 bzero(&cstotal, sizeof cstotal); 1083 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 1084 if ((error = bread(devvp, fsbtodoff(fs, cgtod(fs, cyl)), 1085 fs->fs_cgsize, &bp)) != 0) { 1086 brelse(bp); 1087 return (error); 1088 } 1089 cgp = (struct cg *)bp->b_data; 1090 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 1091 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 1092 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 1093 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 1094 fs->fs_cs(fs, cyl) = cgp->cg_cs; 1095 brelse(bp); 1096 } 1097 #ifdef DEBUG 1098 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 1099 kprintf("ffs_mountfs: superblock updated for soft updates\n"); 1100 #endif 1101 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 1102 return (0); 1103 } 1104 1105 /* 1106 * Protecting the freemaps (or bitmaps). 1107 * 1108 * To eliminate the need to execute fsck before mounting a filesystem 1109 * after a power failure, one must (conservatively) guarantee that the 1110 * on-disk copy of the bitmaps never indicate that a live inode or block is 1111 * free. So, when a block or inode is allocated, the bitmap should be 1112 * updated (on disk) before any new pointers. When a block or inode is 1113 * freed, the bitmap should not be updated until all pointers have been 1114 * reset. The latter dependency is handled by the delayed de-allocation 1115 * approach described below for block and inode de-allocation. The former 1116 * dependency is handled by calling the following procedure when a block or 1117 * inode is allocated. When an inode is allocated an "inodedep" is created 1118 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 1119 * Each "inodedep" is also inserted into the hash indexing structure so 1120 * that any additional link additions can be made dependent on the inode 1121 * allocation. 1122 * 1123 * The ufs filesystem maintains a number of free block counts (e.g., per 1124 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 1125 * in addition to the bitmaps. These counts are used to improve efficiency 1126 * during allocation and therefore must be consistent with the bitmaps. 1127 * There is no convenient way to guarantee post-crash consistency of these 1128 * counts with simple update ordering, for two main reasons: (1) The counts 1129 * and bitmaps for a single cylinder group block are not in the same disk 1130 * sector. If a disk write is interrupted (e.g., by power failure), one may 1131 * be written and the other not. (2) Some of the counts are located in the 1132 * superblock rather than the cylinder group block. So, we focus our soft 1133 * updates implementation on protecting the bitmaps. When mounting a 1134 * filesystem, we recompute the auxiliary counts from the bitmaps. 1135 */ 1136 1137 /* 1138 * Called just after updating the cylinder group block to allocate an inode. 1139 * 1140 * Parameters: 1141 * bp: buffer for cylgroup block with inode map 1142 * ip: inode related to allocation 1143 * newinum: new inode number being allocated 1144 */ 1145 void 1146 softdep_setup_inomapdep(struct buf *bp, struct inode *ip, ino_t newinum) 1147 { 1148 struct inodedep *inodedep; 1149 struct bmsafemap *bmsafemap; 1150 1151 /* 1152 * Create a dependency for the newly allocated inode. 1153 * Panic if it already exists as something is seriously wrong. 1154 * Otherwise add it to the dependency list for the buffer holding 1155 * the cylinder group map from which it was allocated. 1156 */ 1157 ACQUIRE_LOCK(&lk); 1158 if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) { 1159 panic("softdep_setup_inomapdep: found inode"); 1160 } 1161 inodedep->id_buf = bp; 1162 inodedep->id_state &= ~DEPCOMPLETE; 1163 bmsafemap = bmsafemap_lookup(bp); 1164 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 1165 FREE_LOCK(&lk); 1166 } 1167 1168 /* 1169 * Called just after updating the cylinder group block to 1170 * allocate block or fragment. 1171 * 1172 * Parameters: 1173 * bp: buffer for cylgroup block with block map 1174 * fs: filesystem doing allocation 1175 * newblkno: number of newly allocated block 1176 */ 1177 void 1178 softdep_setup_blkmapdep(struct buf *bp, struct fs *fs, 1179 ufs_daddr_t newblkno) 1180 { 1181 struct newblk *newblk; 1182 struct bmsafemap *bmsafemap; 1183 1184 /* 1185 * Create a dependency for the newly allocated block. 1186 * Add it to the dependency list for the buffer holding 1187 * the cylinder group map from which it was allocated. 1188 */ 1189 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0) 1190 panic("softdep_setup_blkmapdep: found block"); 1191 ACQUIRE_LOCK(&lk); 1192 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp); 1193 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 1194 FREE_LOCK(&lk); 1195 } 1196 1197 /* 1198 * Find the bmsafemap associated with a cylinder group buffer. 1199 * If none exists, create one. The buffer must be locked when 1200 * this routine is called and this routine must be called with 1201 * splbio interrupts blocked. 1202 */ 1203 static struct bmsafemap * 1204 bmsafemap_lookup(struct buf *bp) 1205 { 1206 struct bmsafemap *bmsafemap; 1207 struct worklist *wk; 1208 1209 KKASSERT(lock_held(&lk) > 0); 1210 1211 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 1212 if (wk->wk_type == D_BMSAFEMAP) 1213 return (WK_BMSAFEMAP(wk)); 1214 } 1215 FREE_LOCK(&lk); 1216 bmsafemap = kmalloc(sizeof(struct bmsafemap), M_BMSAFEMAP, 1217 M_SOFTDEP_FLAGS); 1218 bmsafemap->sm_list.wk_type = D_BMSAFEMAP; 1219 bmsafemap->sm_list.wk_state = 0; 1220 bmsafemap->sm_buf = bp; 1221 LIST_INIT(&bmsafemap->sm_allocdirecthd); 1222 LIST_INIT(&bmsafemap->sm_allocindirhd); 1223 LIST_INIT(&bmsafemap->sm_inodedephd); 1224 LIST_INIT(&bmsafemap->sm_newblkhd); 1225 ACQUIRE_LOCK(&lk); 1226 WORKLIST_INSERT_BP(bp, &bmsafemap->sm_list); 1227 return (bmsafemap); 1228 } 1229 1230 /* 1231 * Direct block allocation dependencies. 1232 * 1233 * When a new block is allocated, the corresponding disk locations must be 1234 * initialized (with zeros or new data) before the on-disk inode points to 1235 * them. Also, the freemap from which the block was allocated must be 1236 * updated (on disk) before the inode's pointer. These two dependencies are 1237 * independent of each other and are needed for all file blocks and indirect 1238 * blocks that are pointed to directly by the inode. Just before the 1239 * "in-core" version of the inode is updated with a newly allocated block 1240 * number, a procedure (below) is called to setup allocation dependency 1241 * structures. These structures are removed when the corresponding 1242 * dependencies are satisfied or when the block allocation becomes obsolete 1243 * (i.e., the file is deleted, the block is de-allocated, or the block is a 1244 * fragment that gets upgraded). All of these cases are handled in 1245 * procedures described later. 1246 * 1247 * When a file extension causes a fragment to be upgraded, either to a larger 1248 * fragment or to a full block, the on-disk location may change (if the 1249 * previous fragment could not simply be extended). In this case, the old 1250 * fragment must be de-allocated, but not until after the inode's pointer has 1251 * been updated. In most cases, this is handled by later procedures, which 1252 * will construct a "freefrag" structure to be added to the workitem queue 1253 * when the inode update is complete (or obsolete). The main exception to 1254 * this is when an allocation occurs while a pending allocation dependency 1255 * (for the same block pointer) remains. This case is handled in the main 1256 * allocation dependency setup procedure by immediately freeing the 1257 * unreferenced fragments. 1258 * 1259 * Parameters: 1260 * ip: inode to which block is being added 1261 * lbn: block pointer within inode 1262 * newblkno: disk block number being added 1263 * oldblkno: previous block number, 0 unless frag 1264 * newsize: size of new block 1265 * oldsize: size of new block 1266 * bp: bp for allocated block 1267 */ 1268 void 1269 softdep_setup_allocdirect(struct inode *ip, ufs_lbn_t lbn, ufs_daddr_t newblkno, 1270 ufs_daddr_t oldblkno, long newsize, long oldsize, 1271 struct buf *bp) 1272 { 1273 struct allocdirect *adp, *oldadp; 1274 struct allocdirectlst *adphead; 1275 struct bmsafemap *bmsafemap; 1276 struct inodedep *inodedep; 1277 struct pagedep *pagedep; 1278 struct newblk *newblk; 1279 1280 adp = kmalloc(sizeof(struct allocdirect), M_ALLOCDIRECT, 1281 M_SOFTDEP_FLAGS | M_ZERO); 1282 adp->ad_list.wk_type = D_ALLOCDIRECT; 1283 adp->ad_lbn = lbn; 1284 adp->ad_newblkno = newblkno; 1285 adp->ad_oldblkno = oldblkno; 1286 adp->ad_newsize = newsize; 1287 adp->ad_oldsize = oldsize; 1288 adp->ad_state = ATTACHED; 1289 if (newblkno == oldblkno) 1290 adp->ad_freefrag = NULL; 1291 else 1292 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1293 1294 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1295 panic("softdep_setup_allocdirect: lost block"); 1296 1297 ACQUIRE_LOCK(&lk); 1298 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep); 1299 adp->ad_inodedep = inodedep; 1300 1301 if (newblk->nb_state == DEPCOMPLETE) { 1302 adp->ad_state |= DEPCOMPLETE; 1303 adp->ad_buf = NULL; 1304 } else { 1305 bmsafemap = newblk->nb_bmsafemap; 1306 adp->ad_buf = bmsafemap->sm_buf; 1307 LIST_REMOVE(newblk, nb_deps); 1308 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1309 } 1310 LIST_REMOVE(newblk, nb_hash); 1311 kfree(newblk, M_NEWBLK); 1312 1313 WORKLIST_INSERT_BP(bp, &adp->ad_list); 1314 if (lbn >= NDADDR) { 1315 /* allocating an indirect block */ 1316 if (oldblkno != 0) { 1317 panic("softdep_setup_allocdirect: non-zero indir"); 1318 } 1319 } else { 1320 /* 1321 * Allocating a direct block. 1322 * 1323 * If we are allocating a directory block, then we must 1324 * allocate an associated pagedep to track additions and 1325 * deletions. 1326 */ 1327 if ((ip->i_mode & IFMT) == IFDIR && 1328 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) { 1329 WORKLIST_INSERT_BP(bp, &pagedep->pd_list); 1330 } 1331 } 1332 /* 1333 * The list of allocdirects must be kept in sorted and ascending 1334 * order so that the rollback routines can quickly determine the 1335 * first uncommitted block (the size of the file stored on disk 1336 * ends at the end of the lowest committed fragment, or if there 1337 * are no fragments, at the end of the highest committed block). 1338 * Since files generally grow, the typical case is that the new 1339 * block is to be added at the end of the list. We speed this 1340 * special case by checking against the last allocdirect in the 1341 * list before laboriously traversing the list looking for the 1342 * insertion point. 1343 */ 1344 adphead = &inodedep->id_newinoupdt; 1345 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1346 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1347 /* insert at end of list */ 1348 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1349 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1350 allocdirect_merge(adphead, adp, oldadp); 1351 FREE_LOCK(&lk); 1352 return; 1353 } 1354 TAILQ_FOREACH(oldadp, adphead, ad_next) { 1355 if (oldadp->ad_lbn >= lbn) 1356 break; 1357 } 1358 if (oldadp == NULL) { 1359 panic("softdep_setup_allocdirect: lost entry"); 1360 } 1361 /* insert in middle of list */ 1362 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1363 if (oldadp->ad_lbn == lbn) 1364 allocdirect_merge(adphead, adp, oldadp); 1365 FREE_LOCK(&lk); 1366 } 1367 1368 /* 1369 * Replace an old allocdirect dependency with a newer one. 1370 * This routine must be called with splbio interrupts blocked. 1371 * 1372 * Parameters: 1373 * adphead: head of list holding allocdirects 1374 * newadp: allocdirect being added 1375 * oldadp: existing allocdirect being checked 1376 */ 1377 static void 1378 allocdirect_merge(struct allocdirectlst *adphead, 1379 struct allocdirect *newadp, 1380 struct allocdirect *oldadp) 1381 { 1382 struct freefrag *freefrag; 1383 1384 KKASSERT(lock_held(&lk) > 0); 1385 1386 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 1387 newadp->ad_oldsize != oldadp->ad_newsize || 1388 newadp->ad_lbn >= NDADDR) { 1389 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d", 1390 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn, 1391 NDADDR); 1392 } 1393 newadp->ad_oldblkno = oldadp->ad_oldblkno; 1394 newadp->ad_oldsize = oldadp->ad_oldsize; 1395 /* 1396 * If the old dependency had a fragment to free or had never 1397 * previously had a block allocated, then the new dependency 1398 * can immediately post its freefrag and adopt the old freefrag. 1399 * This action is done by swapping the freefrag dependencies. 1400 * The new dependency gains the old one's freefrag, and the 1401 * old one gets the new one and then immediately puts it on 1402 * the worklist when it is freed by free_allocdirect. It is 1403 * not possible to do this swap when the old dependency had a 1404 * non-zero size but no previous fragment to free. This condition 1405 * arises when the new block is an extension of the old block. 1406 * Here, the first part of the fragment allocated to the new 1407 * dependency is part of the block currently claimed on disk by 1408 * the old dependency, so cannot legitimately be freed until the 1409 * conditions for the new dependency are fulfilled. 1410 */ 1411 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 1412 freefrag = newadp->ad_freefrag; 1413 newadp->ad_freefrag = oldadp->ad_freefrag; 1414 oldadp->ad_freefrag = freefrag; 1415 } 1416 free_allocdirect(adphead, oldadp, 0); 1417 } 1418 1419 /* 1420 * Allocate a new freefrag structure if needed. 1421 */ 1422 static struct freefrag * 1423 newfreefrag(struct inode *ip, ufs_daddr_t blkno, long size) 1424 { 1425 struct freefrag *freefrag; 1426 struct fs *fs; 1427 1428 if (blkno == 0) 1429 return (NULL); 1430 fs = ip->i_fs; 1431 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 1432 panic("newfreefrag: frag size"); 1433 freefrag = kmalloc(sizeof(struct freefrag), M_FREEFRAG, 1434 M_SOFTDEP_FLAGS); 1435 freefrag->ff_list.wk_type = D_FREEFRAG; 1436 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */ 1437 freefrag->ff_inum = ip->i_number; 1438 freefrag->ff_fs = fs; 1439 freefrag->ff_devvp = ip->i_devvp; 1440 freefrag->ff_blkno = blkno; 1441 freefrag->ff_fragsize = size; 1442 return (freefrag); 1443 } 1444 1445 /* 1446 * This workitem de-allocates fragments that were replaced during 1447 * file block allocation. 1448 */ 1449 static void 1450 handle_workitem_freefrag(struct freefrag *freefrag) 1451 { 1452 struct inode tip; 1453 1454 tip.i_fs = freefrag->ff_fs; 1455 tip.i_devvp = freefrag->ff_devvp; 1456 tip.i_dev = freefrag->ff_devvp->v_rdev; 1457 tip.i_number = freefrag->ff_inum; 1458 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */ 1459 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize); 1460 kfree(freefrag, M_FREEFRAG); 1461 } 1462 1463 /* 1464 * Indirect block allocation dependencies. 1465 * 1466 * The same dependencies that exist for a direct block also exist when 1467 * a new block is allocated and pointed to by an entry in a block of 1468 * indirect pointers. The undo/redo states described above are also 1469 * used here. Because an indirect block contains many pointers that 1470 * may have dependencies, a second copy of the entire in-memory indirect 1471 * block is kept. The buffer cache copy is always completely up-to-date. 1472 * The second copy, which is used only as a source for disk writes, 1473 * contains only the safe pointers (i.e., those that have no remaining 1474 * update dependencies). The second copy is freed when all pointers 1475 * are safe. The cache is not allowed to replace indirect blocks with 1476 * pending update dependencies. If a buffer containing an indirect 1477 * block with dependencies is written, these routines will mark it 1478 * dirty again. It can only be successfully written once all the 1479 * dependencies are removed. The ffs_fsync routine in conjunction with 1480 * softdep_sync_metadata work together to get all the dependencies 1481 * removed so that a file can be successfully written to disk. Three 1482 * procedures are used when setting up indirect block pointer 1483 * dependencies. The division is necessary because of the organization 1484 * of the "balloc" routine and because of the distinction between file 1485 * pages and file metadata blocks. 1486 */ 1487 1488 /* 1489 * Allocate a new allocindir structure. 1490 * 1491 * Parameters: 1492 * ip: inode for file being extended 1493 * ptrno: offset of pointer in indirect block 1494 * newblkno: disk block number being added 1495 * oldblkno: previous block number, 0 if none 1496 */ 1497 static struct allocindir * 1498 newallocindir(struct inode *ip, int ptrno, ufs_daddr_t newblkno, 1499 ufs_daddr_t oldblkno) 1500 { 1501 struct allocindir *aip; 1502 1503 aip = kmalloc(sizeof(struct allocindir), M_ALLOCINDIR, 1504 M_SOFTDEP_FLAGS | M_ZERO); 1505 aip->ai_list.wk_type = D_ALLOCINDIR; 1506 aip->ai_state = ATTACHED; 1507 aip->ai_offset = ptrno; 1508 aip->ai_newblkno = newblkno; 1509 aip->ai_oldblkno = oldblkno; 1510 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize); 1511 return (aip); 1512 } 1513 1514 /* 1515 * Called just before setting an indirect block pointer 1516 * to a newly allocated file page. 1517 * 1518 * Parameters: 1519 * ip: inode for file being extended 1520 * lbn: allocated block number within file 1521 * bp: buffer with indirect blk referencing page 1522 * ptrno: offset of pointer in indirect block 1523 * newblkno: disk block number being added 1524 * oldblkno: previous block number, 0 if none 1525 * nbp: buffer holding allocated page 1526 */ 1527 void 1528 softdep_setup_allocindir_page(struct inode *ip, ufs_lbn_t lbn, 1529 struct buf *bp, int ptrno, 1530 ufs_daddr_t newblkno, ufs_daddr_t oldblkno, 1531 struct buf *nbp) 1532 { 1533 struct allocindir *aip; 1534 struct pagedep *pagedep; 1535 1536 aip = newallocindir(ip, ptrno, newblkno, oldblkno); 1537 ACQUIRE_LOCK(&lk); 1538 /* 1539 * If we are allocating a directory page, then we must 1540 * allocate an associated pagedep to track additions and 1541 * deletions. 1542 */ 1543 if ((ip->i_mode & IFMT) == IFDIR && 1544 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1545 WORKLIST_INSERT_BP(nbp, &pagedep->pd_list); 1546 WORKLIST_INSERT_BP(nbp, &aip->ai_list); 1547 FREE_LOCK(&lk); 1548 setup_allocindir_phase2(bp, ip, aip); 1549 } 1550 1551 /* 1552 * Called just before setting an indirect block pointer to a 1553 * newly allocated indirect block. 1554 * Parameters: 1555 * nbp: newly allocated indirect block 1556 * ip: inode for file being extended 1557 * bp: indirect block referencing allocated block 1558 * ptrno: offset of pointer in indirect block 1559 * newblkno: disk block number being added 1560 */ 1561 void 1562 softdep_setup_allocindir_meta(struct buf *nbp, struct inode *ip, 1563 struct buf *bp, int ptrno, 1564 ufs_daddr_t newblkno) 1565 { 1566 struct allocindir *aip; 1567 1568 aip = newallocindir(ip, ptrno, newblkno, 0); 1569 ACQUIRE_LOCK(&lk); 1570 WORKLIST_INSERT_BP(nbp, &aip->ai_list); 1571 FREE_LOCK(&lk); 1572 setup_allocindir_phase2(bp, ip, aip); 1573 } 1574 1575 /* 1576 * Called to finish the allocation of the "aip" allocated 1577 * by one of the two routines above. 1578 * 1579 * Parameters: 1580 * bp: in-memory copy of the indirect block 1581 * ip: inode for file being extended 1582 * aip: allocindir allocated by the above routines 1583 */ 1584 static void 1585 setup_allocindir_phase2(struct buf *bp, struct inode *ip, 1586 struct allocindir *aip) 1587 { 1588 struct worklist *wk; 1589 struct indirdep *indirdep, *newindirdep; 1590 struct bmsafemap *bmsafemap; 1591 struct allocindir *oldaip; 1592 struct freefrag *freefrag; 1593 struct newblk *newblk; 1594 1595 if (bp->b_loffset >= 0) 1596 panic("setup_allocindir_phase2: not indir blk"); 1597 for (indirdep = NULL, newindirdep = NULL; ; ) { 1598 ACQUIRE_LOCK(&lk); 1599 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 1600 if (wk->wk_type != D_INDIRDEP) 1601 continue; 1602 indirdep = WK_INDIRDEP(wk); 1603 break; 1604 } 1605 if (indirdep == NULL && newindirdep) { 1606 indirdep = newindirdep; 1607 WORKLIST_INSERT_BP(bp, &indirdep->ir_list); 1608 newindirdep = NULL; 1609 } 1610 FREE_LOCK(&lk); 1611 if (indirdep) { 1612 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0, 1613 &newblk) == 0) 1614 panic("setup_allocindir: lost block"); 1615 ACQUIRE_LOCK(&lk); 1616 if (newblk->nb_state == DEPCOMPLETE) { 1617 aip->ai_state |= DEPCOMPLETE; 1618 aip->ai_buf = NULL; 1619 } else { 1620 bmsafemap = newblk->nb_bmsafemap; 1621 aip->ai_buf = bmsafemap->sm_buf; 1622 LIST_REMOVE(newblk, nb_deps); 1623 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd, 1624 aip, ai_deps); 1625 } 1626 LIST_REMOVE(newblk, nb_hash); 1627 kfree(newblk, M_NEWBLK); 1628 aip->ai_indirdep = indirdep; 1629 /* 1630 * Check to see if there is an existing dependency 1631 * for this block. If there is, merge the old 1632 * dependency into the new one. 1633 */ 1634 if (aip->ai_oldblkno == 0) 1635 oldaip = NULL; 1636 else 1637 1638 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) 1639 if (oldaip->ai_offset == aip->ai_offset) 1640 break; 1641 if (oldaip != NULL) { 1642 if (oldaip->ai_newblkno != aip->ai_oldblkno) { 1643 panic("setup_allocindir_phase2: blkno"); 1644 } 1645 aip->ai_oldblkno = oldaip->ai_oldblkno; 1646 freefrag = oldaip->ai_freefrag; 1647 oldaip->ai_freefrag = aip->ai_freefrag; 1648 aip->ai_freefrag = freefrag; 1649 free_allocindir(oldaip, NULL); 1650 } 1651 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 1652 ((ufs_daddr_t *)indirdep->ir_savebp->b_data) 1653 [aip->ai_offset] = aip->ai_oldblkno; 1654 FREE_LOCK(&lk); 1655 } 1656 if (newindirdep) { 1657 /* 1658 * Avoid any possibility of data corruption by 1659 * ensuring that our old version is thrown away. 1660 */ 1661 newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 1662 brelse(newindirdep->ir_savebp); 1663 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP); 1664 } 1665 if (indirdep) 1666 break; 1667 newindirdep = kmalloc(sizeof(struct indirdep), M_INDIRDEP, 1668 M_SOFTDEP_FLAGS); 1669 newindirdep->ir_list.wk_type = D_INDIRDEP; 1670 newindirdep->ir_state = ATTACHED; 1671 LIST_INIT(&newindirdep->ir_deplisthd); 1672 LIST_INIT(&newindirdep->ir_donehd); 1673 if (bp->b_bio2.bio_offset == NOOFFSET) { 1674 VOP_BMAP(bp->b_vp, bp->b_bio1.bio_offset, 1675 &bp->b_bio2.bio_offset, NULL, NULL, 1676 BUF_CMD_WRITE); 1677 } 1678 KKASSERT(bp->b_bio2.bio_offset != NOOFFSET); 1679 newindirdep->ir_savebp = getblk(ip->i_devvp, 1680 bp->b_bio2.bio_offset, 1681 bp->b_bcount, 0, 0); 1682 BUF_KERNPROC(newindirdep->ir_savebp); 1683 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 1684 } 1685 } 1686 1687 /* 1688 * Block de-allocation dependencies. 1689 * 1690 * When blocks are de-allocated, the on-disk pointers must be nullified before 1691 * the blocks are made available for use by other files. (The true 1692 * requirement is that old pointers must be nullified before new on-disk 1693 * pointers are set. We chose this slightly more stringent requirement to 1694 * reduce complexity.) Our implementation handles this dependency by updating 1695 * the inode (or indirect block) appropriately but delaying the actual block 1696 * de-allocation (i.e., freemap and free space count manipulation) until 1697 * after the updated versions reach stable storage. After the disk is 1698 * updated, the blocks can be safely de-allocated whenever it is convenient. 1699 * This implementation handles only the common case of reducing a file's 1700 * length to zero. Other cases are handled by the conventional synchronous 1701 * write approach. 1702 * 1703 * The ffs implementation with which we worked double-checks 1704 * the state of the block pointers and file size as it reduces 1705 * a file's length. Some of this code is replicated here in our 1706 * soft updates implementation. The freeblks->fb_chkcnt field is 1707 * used to transfer a part of this information to the procedure 1708 * that eventually de-allocates the blocks. 1709 * 1710 * This routine should be called from the routine that shortens 1711 * a file's length, before the inode's size or block pointers 1712 * are modified. It will save the block pointer information for 1713 * later release and zero the inode so that the calling routine 1714 * can release it. 1715 */ 1716 struct softdep_setup_freeblocks_info { 1717 struct fs *fs; 1718 struct inode *ip; 1719 }; 1720 1721 static int softdep_setup_freeblocks_bp(struct buf *bp, void *data); 1722 1723 /* 1724 * Parameters: 1725 * ip: The inode whose length is to be reduced 1726 * length: The new length for the file 1727 */ 1728 void 1729 softdep_setup_freeblocks(struct inode *ip, off_t length) 1730 { 1731 struct softdep_setup_freeblocks_info info; 1732 struct freeblks *freeblks; 1733 struct inodedep *inodedep; 1734 struct allocdirect *adp; 1735 struct vnode *vp; 1736 struct buf *bp; 1737 struct fs *fs; 1738 int i, error, delay; 1739 int count; 1740 1741 fs = ip->i_fs; 1742 if (length != 0) 1743 panic("softde_setup_freeblocks: non-zero length"); 1744 freeblks = kmalloc(sizeof(struct freeblks), M_FREEBLKS, 1745 M_SOFTDEP_FLAGS | M_ZERO); 1746 freeblks->fb_list.wk_type = D_FREEBLKS; 1747 freeblks->fb_state = ATTACHED; 1748 freeblks->fb_uid = ip->i_uid; 1749 freeblks->fb_previousinum = ip->i_number; 1750 freeblks->fb_devvp = ip->i_devvp; 1751 freeblks->fb_fs = fs; 1752 freeblks->fb_oldsize = ip->i_size; 1753 freeblks->fb_newsize = length; 1754 freeblks->fb_chkcnt = ip->i_blocks; 1755 for (i = 0; i < NDADDR; i++) { 1756 freeblks->fb_dblks[i] = ip->i_db[i]; 1757 ip->i_db[i] = 0; 1758 } 1759 for (i = 0; i < NIADDR; i++) { 1760 freeblks->fb_iblks[i] = ip->i_ib[i]; 1761 ip->i_ib[i] = 0; 1762 } 1763 ip->i_blocks = 0; 1764 ip->i_size = 0; 1765 /* 1766 * Push the zero'ed inode to to its disk buffer so that we are free 1767 * to delete its dependencies below. Once the dependencies are gone 1768 * the buffer can be safely released. 1769 */ 1770 if ((error = bread(ip->i_devvp, 1771 fsbtodoff(fs, ino_to_fsba(fs, ip->i_number)), 1772 (int)fs->fs_bsize, &bp)) != 0) 1773 softdep_error("softdep_setup_freeblocks", error); 1774 *((struct ufs1_dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) = 1775 ip->i_din; 1776 /* 1777 * Find and eliminate any inode dependencies. 1778 */ 1779 ACQUIRE_LOCK(&lk); 1780 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep); 1781 if ((inodedep->id_state & IOSTARTED) != 0) { 1782 panic("softdep_setup_freeblocks: inode busy"); 1783 } 1784 /* 1785 * Add the freeblks structure to the list of operations that 1786 * must await the zero'ed inode being written to disk. If we 1787 * still have a bitmap dependency (delay == 0), then the inode 1788 * has never been written to disk, so we can process the 1789 * freeblks below once we have deleted the dependencies. 1790 */ 1791 delay = (inodedep->id_state & DEPCOMPLETE); 1792 if (delay) 1793 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list); 1794 /* 1795 * Because the file length has been truncated to zero, any 1796 * pending block allocation dependency structures associated 1797 * with this inode are obsolete and can simply be de-allocated. 1798 * We must first merge the two dependency lists to get rid of 1799 * any duplicate freefrag structures, then purge the merged list. 1800 */ 1801 merge_inode_lists(inodedep); 1802 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 1803 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 1804 FREE_LOCK(&lk); 1805 bdwrite(bp); 1806 /* 1807 * We must wait for any I/O in progress to finish so that 1808 * all potential buffers on the dirty list will be visible. 1809 * Once they are all there, walk the list and get rid of 1810 * any dependencies. 1811 */ 1812 vp = ITOV(ip); 1813 ACQUIRE_LOCK(&lk); 1814 drain_output(vp, 1); 1815 1816 info.fs = fs; 1817 info.ip = ip; 1818 lwkt_gettoken(&vp->v_token); 1819 do { 1820 count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 1821 softdep_setup_freeblocks_bp, &info); 1822 } while (count != 0); 1823 lwkt_reltoken(&vp->v_token); 1824 1825 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0) 1826 (void)free_inodedep(inodedep); 1827 1828 if (delay) { 1829 freeblks->fb_state |= DEPCOMPLETE; 1830 /* 1831 * If the inode with zeroed block pointers is now on disk 1832 * we can start freeing blocks. Add freeblks to the worklist 1833 * instead of calling handle_workitem_freeblocks directly as 1834 * it is more likely that additional IO is needed to complete 1835 * the request here than in the !delay case. 1836 */ 1837 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 1838 add_to_worklist(&freeblks->fb_list); 1839 } 1840 1841 FREE_LOCK(&lk); 1842 /* 1843 * If the inode has never been written to disk (delay == 0), 1844 * then we can process the freeblks now that we have deleted 1845 * the dependencies. 1846 */ 1847 if (!delay) 1848 handle_workitem_freeblocks(freeblks); 1849 } 1850 1851 static int 1852 softdep_setup_freeblocks_bp(struct buf *bp, void *data) 1853 { 1854 struct softdep_setup_freeblocks_info *info = data; 1855 struct inodedep *inodedep; 1856 1857 if (getdirtybuf(&bp, MNT_WAIT) == 0) { 1858 kprintf("softdep_setup_freeblocks_bp(1): caught bp %p going away\n", bp); 1859 return(-1); 1860 } 1861 if (bp->b_vp != ITOV(info->ip) || (bp->b_flags & B_DELWRI) == 0) { 1862 kprintf("softdep_setup_freeblocks_bp(2): caught bp %p going away\n", bp); 1863 BUF_UNLOCK(bp); 1864 return(-1); 1865 } 1866 (void) inodedep_lookup(info->fs, info->ip->i_number, 0, &inodedep); 1867 deallocate_dependencies(bp, inodedep); 1868 bp->b_flags |= B_INVAL | B_NOCACHE; 1869 FREE_LOCK(&lk); 1870 brelse(bp); 1871 ACQUIRE_LOCK(&lk); 1872 return(1); 1873 } 1874 1875 /* 1876 * Reclaim any dependency structures from a buffer that is about to 1877 * be reallocated to a new vnode. The buffer must be locked, thus, 1878 * no I/O completion operations can occur while we are manipulating 1879 * its associated dependencies. The mutex is held so that other I/O's 1880 * associated with related dependencies do not occur. 1881 */ 1882 static void 1883 deallocate_dependencies(struct buf *bp, struct inodedep *inodedep) 1884 { 1885 struct worklist *wk; 1886 struct indirdep *indirdep; 1887 struct allocindir *aip; 1888 struct pagedep *pagedep; 1889 struct dirrem *dirrem; 1890 struct diradd *dap; 1891 int i; 1892 1893 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 1894 switch (wk->wk_type) { 1895 1896 case D_INDIRDEP: 1897 indirdep = WK_INDIRDEP(wk); 1898 /* 1899 * None of the indirect pointers will ever be visible, 1900 * so they can simply be tossed. GOINGAWAY ensures 1901 * that allocated pointers will be saved in the buffer 1902 * cache until they are freed. Note that they will 1903 * only be able to be found by their physical address 1904 * since the inode mapping the logical address will 1905 * be gone. The save buffer used for the safe copy 1906 * was allocated in setup_allocindir_phase2 using 1907 * the physical address so it could be used for this 1908 * purpose. Hence we swap the safe copy with the real 1909 * copy, allowing the safe copy to be freed and holding 1910 * on to the real copy for later use in indir_trunc. 1911 * 1912 * NOTE: ir_savebp is relative to the block device 1913 * so b_bio1 contains the device block number. 1914 */ 1915 if (indirdep->ir_state & GOINGAWAY) { 1916 panic("deallocate_dependencies: already gone"); 1917 } 1918 indirdep->ir_state |= GOINGAWAY; 1919 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != NULL) 1920 free_allocindir(aip, inodedep); 1921 if (bp->b_bio1.bio_offset >= 0 || 1922 bp->b_bio2.bio_offset != indirdep->ir_savebp->b_bio1.bio_offset) { 1923 panic("deallocate_dependencies: not indir"); 1924 } 1925 bcopy(bp->b_data, indirdep->ir_savebp->b_data, 1926 bp->b_bcount); 1927 WORKLIST_REMOVE(wk); 1928 WORKLIST_INSERT_BP(indirdep->ir_savebp, wk); 1929 continue; 1930 1931 case D_PAGEDEP: 1932 pagedep = WK_PAGEDEP(wk); 1933 /* 1934 * None of the directory additions will ever be 1935 * visible, so they can simply be tossed. 1936 */ 1937 for (i = 0; i < DAHASHSZ; i++) 1938 while ((dap = 1939 LIST_FIRST(&pagedep->pd_diraddhd[i]))) 1940 free_diradd(dap); 1941 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 1942 free_diradd(dap); 1943 /* 1944 * Copy any directory remove dependencies to the list 1945 * to be processed after the zero'ed inode is written. 1946 * If the inode has already been written, then they 1947 * can be dumped directly onto the work list. 1948 */ 1949 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 1950 LIST_REMOVE(dirrem, dm_next); 1951 dirrem->dm_dirinum = pagedep->pd_ino; 1952 if (inodedep == NULL || 1953 (inodedep->id_state & ALLCOMPLETE) == 1954 ALLCOMPLETE) 1955 add_to_worklist(&dirrem->dm_list); 1956 else 1957 WORKLIST_INSERT(&inodedep->id_bufwait, 1958 &dirrem->dm_list); 1959 } 1960 WORKLIST_REMOVE(&pagedep->pd_list); 1961 LIST_REMOVE(pagedep, pd_hash); 1962 WORKITEM_FREE(pagedep, D_PAGEDEP); 1963 continue; 1964 1965 case D_ALLOCINDIR: 1966 free_allocindir(WK_ALLOCINDIR(wk), inodedep); 1967 continue; 1968 1969 case D_ALLOCDIRECT: 1970 case D_INODEDEP: 1971 panic("deallocate_dependencies: Unexpected type %s", 1972 TYPENAME(wk->wk_type)); 1973 /* NOTREACHED */ 1974 1975 default: 1976 panic("deallocate_dependencies: Unknown type %s", 1977 TYPENAME(wk->wk_type)); 1978 /* NOTREACHED */ 1979 } 1980 } 1981 } 1982 1983 /* 1984 * Free an allocdirect. Generate a new freefrag work request if appropriate. 1985 * This routine must be called with splbio interrupts blocked. 1986 */ 1987 static void 1988 free_allocdirect(struct allocdirectlst *adphead, 1989 struct allocdirect *adp, int delay) 1990 { 1991 KKASSERT(lock_held(&lk) > 0); 1992 1993 if ((adp->ad_state & DEPCOMPLETE) == 0) 1994 LIST_REMOVE(adp, ad_deps); 1995 TAILQ_REMOVE(adphead, adp, ad_next); 1996 if ((adp->ad_state & COMPLETE) == 0) 1997 WORKLIST_REMOVE(&adp->ad_list); 1998 if (adp->ad_freefrag != NULL) { 1999 if (delay) 2000 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 2001 &adp->ad_freefrag->ff_list); 2002 else 2003 add_to_worklist(&adp->ad_freefrag->ff_list); 2004 } 2005 WORKITEM_FREE(adp, D_ALLOCDIRECT); 2006 } 2007 2008 /* 2009 * Prepare an inode to be freed. The actual free operation is not 2010 * done until the zero'ed inode has been written to disk. 2011 */ 2012 void 2013 softdep_freefile(struct vnode *pvp, ino_t ino, int mode) 2014 { 2015 struct inode *ip = VTOI(pvp); 2016 struct inodedep *inodedep; 2017 struct freefile *freefile; 2018 2019 /* 2020 * This sets up the inode de-allocation dependency. 2021 */ 2022 freefile = kmalloc(sizeof(struct freefile), M_FREEFILE, 2023 M_SOFTDEP_FLAGS); 2024 freefile->fx_list.wk_type = D_FREEFILE; 2025 freefile->fx_list.wk_state = 0; 2026 freefile->fx_mode = mode; 2027 freefile->fx_oldinum = ino; 2028 freefile->fx_devvp = ip->i_devvp; 2029 freefile->fx_fs = ip->i_fs; 2030 2031 /* 2032 * If the inodedep does not exist, then the zero'ed inode has 2033 * been written to disk. If the allocated inode has never been 2034 * written to disk, then the on-disk inode is zero'ed. In either 2035 * case we can free the file immediately. 2036 */ 2037 ACQUIRE_LOCK(&lk); 2038 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 || 2039 check_inode_unwritten(inodedep)) { 2040 FREE_LOCK(&lk); 2041 handle_workitem_freefile(freefile); 2042 return; 2043 } 2044 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 2045 FREE_LOCK(&lk); 2046 } 2047 2048 /* 2049 * Check to see if an inode has never been written to disk. If 2050 * so free the inodedep and return success, otherwise return failure. 2051 * This routine must be called with splbio interrupts blocked. 2052 * 2053 * If we still have a bitmap dependency, then the inode has never 2054 * been written to disk. Drop the dependency as it is no longer 2055 * necessary since the inode is being deallocated. We set the 2056 * ALLCOMPLETE flags since the bitmap now properly shows that the 2057 * inode is not allocated. Even if the inode is actively being 2058 * written, it has been rolled back to its zero'ed state, so we 2059 * are ensured that a zero inode is what is on the disk. For short 2060 * lived files, this change will usually result in removing all the 2061 * dependencies from the inode so that it can be freed immediately. 2062 */ 2063 static int 2064 check_inode_unwritten(struct inodedep *inodedep) 2065 { 2066 2067 if ((inodedep->id_state & DEPCOMPLETE) != 0 || 2068 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2069 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2070 LIST_FIRST(&inodedep->id_inowait) != NULL || 2071 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2072 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2073 inodedep->id_nlinkdelta != 0) 2074 return (0); 2075 2076 /* 2077 * Another process might be in initiate_write_inodeblock 2078 * trying to allocate memory without holding "Softdep Lock". 2079 */ 2080 if ((inodedep->id_state & IOSTARTED) != 0 && 2081 inodedep->id_savedino == NULL) 2082 return(0); 2083 2084 inodedep->id_state |= ALLCOMPLETE; 2085 LIST_REMOVE(inodedep, id_deps); 2086 inodedep->id_buf = NULL; 2087 if (inodedep->id_state & ONWORKLIST) 2088 WORKLIST_REMOVE(&inodedep->id_list); 2089 if (inodedep->id_savedino != NULL) { 2090 kfree(inodedep->id_savedino, M_INODEDEP); 2091 inodedep->id_savedino = NULL; 2092 } 2093 if (free_inodedep(inodedep) == 0) { 2094 panic("check_inode_unwritten: busy inode"); 2095 } 2096 return (1); 2097 } 2098 2099 /* 2100 * Try to free an inodedep structure. Return 1 if it could be freed. 2101 */ 2102 static int 2103 free_inodedep(struct inodedep *inodedep) 2104 { 2105 2106 if ((inodedep->id_state & ONWORKLIST) != 0 || 2107 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 2108 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2109 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2110 LIST_FIRST(&inodedep->id_inowait) != NULL || 2111 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2112 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2113 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL) 2114 return (0); 2115 LIST_REMOVE(inodedep, id_hash); 2116 WORKITEM_FREE(inodedep, D_INODEDEP); 2117 num_inodedep -= 1; 2118 return (1); 2119 } 2120 2121 /* 2122 * This workitem routine performs the block de-allocation. 2123 * The workitem is added to the pending list after the updated 2124 * inode block has been written to disk. As mentioned above, 2125 * checks regarding the number of blocks de-allocated (compared 2126 * to the number of blocks allocated for the file) are also 2127 * performed in this function. 2128 */ 2129 static void 2130 handle_workitem_freeblocks(struct freeblks *freeblks) 2131 { 2132 struct inode tip; 2133 ufs_daddr_t bn; 2134 struct fs *fs; 2135 int i, level, bsize; 2136 long nblocks, blocksreleased = 0; 2137 int error, allerror = 0; 2138 ufs_lbn_t baselbns[NIADDR], tmpval; 2139 2140 tip.i_number = freeblks->fb_previousinum; 2141 tip.i_devvp = freeblks->fb_devvp; 2142 tip.i_dev = freeblks->fb_devvp->v_rdev; 2143 tip.i_fs = freeblks->fb_fs; 2144 tip.i_size = freeblks->fb_oldsize; 2145 tip.i_uid = freeblks->fb_uid; 2146 fs = freeblks->fb_fs; 2147 tmpval = 1; 2148 baselbns[0] = NDADDR; 2149 for (i = 1; i < NIADDR; i++) { 2150 tmpval *= NINDIR(fs); 2151 baselbns[i] = baselbns[i - 1] + tmpval; 2152 } 2153 nblocks = btodb(fs->fs_bsize); 2154 blocksreleased = 0; 2155 /* 2156 * Indirect blocks first. 2157 */ 2158 for (level = (NIADDR - 1); level >= 0; level--) { 2159 if ((bn = freeblks->fb_iblks[level]) == 0) 2160 continue; 2161 if ((error = indir_trunc(&tip, fsbtodoff(fs, bn), level, 2162 baselbns[level], &blocksreleased)) == 0) 2163 allerror = error; 2164 ffs_blkfree(&tip, bn, fs->fs_bsize); 2165 blocksreleased += nblocks; 2166 } 2167 /* 2168 * All direct blocks or frags. 2169 */ 2170 for (i = (NDADDR - 1); i >= 0; i--) { 2171 if ((bn = freeblks->fb_dblks[i]) == 0) 2172 continue; 2173 bsize = blksize(fs, &tip, i); 2174 ffs_blkfree(&tip, bn, bsize); 2175 blocksreleased += btodb(bsize); 2176 } 2177 2178 #ifdef DIAGNOSTIC 2179 if (freeblks->fb_chkcnt != blocksreleased) 2180 kprintf("handle_workitem_freeblocks: block count\n"); 2181 if (allerror) 2182 softdep_error("handle_workitem_freeblks", allerror); 2183 #endif /* DIAGNOSTIC */ 2184 WORKITEM_FREE(freeblks, D_FREEBLKS); 2185 } 2186 2187 /* 2188 * Release blocks associated with the inode ip and stored in the indirect 2189 * block at doffset. If level is greater than SINGLE, the block is an 2190 * indirect block and recursive calls to indirtrunc must be used to 2191 * cleanse other indirect blocks. 2192 */ 2193 static int 2194 indir_trunc(struct inode *ip, off_t doffset, int level, ufs_lbn_t lbn, 2195 long *countp) 2196 { 2197 struct buf *bp; 2198 ufs_daddr_t *bap; 2199 ufs_daddr_t nb; 2200 struct fs *fs; 2201 struct worklist *wk; 2202 struct indirdep *indirdep; 2203 int i, lbnadd, nblocks; 2204 int error, allerror = 0; 2205 2206 fs = ip->i_fs; 2207 lbnadd = 1; 2208 for (i = level; i > 0; i--) 2209 lbnadd *= NINDIR(fs); 2210 /* 2211 * Get buffer of block pointers to be freed. This routine is not 2212 * called until the zero'ed inode has been written, so it is safe 2213 * to free blocks as they are encountered. Because the inode has 2214 * been zero'ed, calls to bmap on these blocks will fail. So, we 2215 * have to use the on-disk address and the block device for the 2216 * filesystem to look them up. If the file was deleted before its 2217 * indirect blocks were all written to disk, the routine that set 2218 * us up (deallocate_dependencies) will have arranged to leave 2219 * a complete copy of the indirect block in memory for our use. 2220 * Otherwise we have to read the blocks in from the disk. 2221 */ 2222 ACQUIRE_LOCK(&lk); 2223 if ((bp = findblk(ip->i_devvp, doffset, FINDBLK_TEST)) != NULL && 2224 (wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2225 /* 2226 * bp must be ir_savebp, which is held locked for our use. 2227 */ 2228 if (wk->wk_type != D_INDIRDEP || 2229 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp || 2230 (indirdep->ir_state & GOINGAWAY) == 0) { 2231 panic("indir_trunc: lost indirdep"); 2232 } 2233 WORKLIST_REMOVE(wk); 2234 WORKITEM_FREE(indirdep, D_INDIRDEP); 2235 if (LIST_FIRST(&bp->b_dep) != NULL) { 2236 panic("indir_trunc: dangling dep"); 2237 } 2238 FREE_LOCK(&lk); 2239 } else { 2240 FREE_LOCK(&lk); 2241 error = bread(ip->i_devvp, doffset, (int)fs->fs_bsize, &bp); 2242 if (error) 2243 return (error); 2244 } 2245 /* 2246 * Recursively free indirect blocks. 2247 */ 2248 bap = (ufs_daddr_t *)bp->b_data; 2249 nblocks = btodb(fs->fs_bsize); 2250 for (i = NINDIR(fs) - 1; i >= 0; i--) { 2251 if ((nb = bap[i]) == 0) 2252 continue; 2253 if (level != 0) { 2254 if ((error = indir_trunc(ip, fsbtodoff(fs, nb), 2255 level - 1, lbn + (i * lbnadd), countp)) != 0) 2256 allerror = error; 2257 } 2258 ffs_blkfree(ip, nb, fs->fs_bsize); 2259 *countp += nblocks; 2260 } 2261 bp->b_flags |= B_INVAL | B_NOCACHE; 2262 brelse(bp); 2263 return (allerror); 2264 } 2265 2266 /* 2267 * Free an allocindir. 2268 * This routine must be called with splbio interrupts blocked. 2269 */ 2270 static void 2271 free_allocindir(struct allocindir *aip, struct inodedep *inodedep) 2272 { 2273 struct freefrag *freefrag; 2274 2275 KKASSERT(lock_held(&lk) > 0); 2276 2277 if ((aip->ai_state & DEPCOMPLETE) == 0) 2278 LIST_REMOVE(aip, ai_deps); 2279 if (aip->ai_state & ONWORKLIST) 2280 WORKLIST_REMOVE(&aip->ai_list); 2281 LIST_REMOVE(aip, ai_next); 2282 if ((freefrag = aip->ai_freefrag) != NULL) { 2283 if (inodedep == NULL) 2284 add_to_worklist(&freefrag->ff_list); 2285 else 2286 WORKLIST_INSERT(&inodedep->id_bufwait, 2287 &freefrag->ff_list); 2288 } 2289 WORKITEM_FREE(aip, D_ALLOCINDIR); 2290 } 2291 2292 /* 2293 * Directory entry addition dependencies. 2294 * 2295 * When adding a new directory entry, the inode (with its incremented link 2296 * count) must be written to disk before the directory entry's pointer to it. 2297 * Also, if the inode is newly allocated, the corresponding freemap must be 2298 * updated (on disk) before the directory entry's pointer. These requirements 2299 * are met via undo/redo on the directory entry's pointer, which consists 2300 * simply of the inode number. 2301 * 2302 * As directory entries are added and deleted, the free space within a 2303 * directory block can become fragmented. The ufs filesystem will compact 2304 * a fragmented directory block to make space for a new entry. When this 2305 * occurs, the offsets of previously added entries change. Any "diradd" 2306 * dependency structures corresponding to these entries must be updated with 2307 * the new offsets. 2308 */ 2309 2310 /* 2311 * This routine is called after the in-memory inode's link 2312 * count has been incremented, but before the directory entry's 2313 * pointer to the inode has been set. 2314 * 2315 * Parameters: 2316 * bp: buffer containing directory block 2317 * dp: inode for directory 2318 * diroffset: offset of new entry in directory 2319 * newinum: inode referenced by new directory entry 2320 * newdirbp: non-NULL => contents of new mkdir 2321 */ 2322 void 2323 softdep_setup_directory_add(struct buf *bp, struct inode *dp, off_t diroffset, 2324 ino_t newinum, struct buf *newdirbp) 2325 { 2326 int offset; /* offset of new entry within directory block */ 2327 ufs_lbn_t lbn; /* block in directory containing new entry */ 2328 struct fs *fs; 2329 struct diradd *dap; 2330 struct pagedep *pagedep; 2331 struct inodedep *inodedep; 2332 struct mkdir *mkdir1, *mkdir2; 2333 2334 /* 2335 * Whiteouts have no dependencies. 2336 */ 2337 if (newinum == WINO) { 2338 if (newdirbp != NULL) 2339 bdwrite(newdirbp); 2340 return; 2341 } 2342 2343 fs = dp->i_fs; 2344 lbn = lblkno(fs, diroffset); 2345 offset = blkoff(fs, diroffset); 2346 dap = kmalloc(sizeof(struct diradd), M_DIRADD, 2347 M_SOFTDEP_FLAGS | M_ZERO); 2348 dap->da_list.wk_type = D_DIRADD; 2349 dap->da_offset = offset; 2350 dap->da_newinum = newinum; 2351 dap->da_state = ATTACHED; 2352 if (newdirbp == NULL) { 2353 dap->da_state |= DEPCOMPLETE; 2354 ACQUIRE_LOCK(&lk); 2355 } else { 2356 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 2357 mkdir1 = kmalloc(sizeof(struct mkdir), M_MKDIR, 2358 M_SOFTDEP_FLAGS); 2359 mkdir1->md_list.wk_type = D_MKDIR; 2360 mkdir1->md_state = MKDIR_BODY; 2361 mkdir1->md_diradd = dap; 2362 mkdir2 = kmalloc(sizeof(struct mkdir), M_MKDIR, 2363 M_SOFTDEP_FLAGS); 2364 mkdir2->md_list.wk_type = D_MKDIR; 2365 mkdir2->md_state = MKDIR_PARENT; 2366 mkdir2->md_diradd = dap; 2367 /* 2368 * Dependency on "." and ".." being written to disk. 2369 */ 2370 mkdir1->md_buf = newdirbp; 2371 ACQUIRE_LOCK(&lk); 2372 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 2373 WORKLIST_INSERT_BP(newdirbp, &mkdir1->md_list); 2374 FREE_LOCK(&lk); 2375 bdwrite(newdirbp); 2376 /* 2377 * Dependency on link count increase for parent directory 2378 */ 2379 ACQUIRE_LOCK(&lk); 2380 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0 2381 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2382 dap->da_state &= ~MKDIR_PARENT; 2383 WORKITEM_FREE(mkdir2, D_MKDIR); 2384 } else { 2385 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 2386 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 2387 } 2388 } 2389 /* 2390 * Link into parent directory pagedep to await its being written. 2391 */ 2392 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2393 WORKLIST_INSERT_BP(bp, &pagedep->pd_list); 2394 dap->da_pagedep = pagedep; 2395 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 2396 da_pdlist); 2397 /* 2398 * Link into its inodedep. Put it on the id_bufwait list if the inode 2399 * is not yet written. If it is written, do the post-inode write 2400 * processing to put it on the id_pendinghd list. 2401 */ 2402 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep); 2403 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 2404 diradd_inode_written(dap, inodedep); 2405 else 2406 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2407 FREE_LOCK(&lk); 2408 } 2409 2410 /* 2411 * This procedure is called to change the offset of a directory 2412 * entry when compacting a directory block which must be owned 2413 * exclusively by the caller. Note that the actual entry movement 2414 * must be done in this procedure to ensure that no I/O completions 2415 * occur while the move is in progress. 2416 * 2417 * Parameters: 2418 * dp: inode for directory 2419 * base: address of dp->i_offset 2420 * oldloc: address of old directory location 2421 * newloc: address of new directory location 2422 * entrysize: size of directory entry 2423 */ 2424 void 2425 softdep_change_directoryentry_offset(struct inode *dp, caddr_t base, 2426 caddr_t oldloc, caddr_t newloc, 2427 int entrysize) 2428 { 2429 int offset, oldoffset, newoffset; 2430 struct pagedep *pagedep; 2431 struct diradd *dap; 2432 ufs_lbn_t lbn; 2433 2434 ACQUIRE_LOCK(&lk); 2435 lbn = lblkno(dp->i_fs, dp->i_offset); 2436 offset = blkoff(dp->i_fs, dp->i_offset); 2437 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) 2438 goto done; 2439 oldoffset = offset + (oldloc - base); 2440 newoffset = offset + (newloc - base); 2441 2442 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) { 2443 if (dap->da_offset != oldoffset) 2444 continue; 2445 dap->da_offset = newoffset; 2446 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) 2447 break; 2448 LIST_REMOVE(dap, da_pdlist); 2449 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], 2450 dap, da_pdlist); 2451 break; 2452 } 2453 if (dap == NULL) { 2454 2455 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) { 2456 if (dap->da_offset == oldoffset) { 2457 dap->da_offset = newoffset; 2458 break; 2459 } 2460 } 2461 } 2462 done: 2463 bcopy(oldloc, newloc, entrysize); 2464 FREE_LOCK(&lk); 2465 } 2466 2467 /* 2468 * Free a diradd dependency structure. This routine must be called 2469 * with splbio interrupts blocked. 2470 */ 2471 static void 2472 free_diradd(struct diradd *dap) 2473 { 2474 struct dirrem *dirrem; 2475 struct pagedep *pagedep; 2476 struct inodedep *inodedep; 2477 struct mkdir *mkdir, *nextmd; 2478 2479 KKASSERT(lock_held(&lk) > 0); 2480 2481 WORKLIST_REMOVE(&dap->da_list); 2482 LIST_REMOVE(dap, da_pdlist); 2483 if ((dap->da_state & DIRCHG) == 0) { 2484 pagedep = dap->da_pagedep; 2485 } else { 2486 dirrem = dap->da_previous; 2487 pagedep = dirrem->dm_pagedep; 2488 dirrem->dm_dirinum = pagedep->pd_ino; 2489 add_to_worklist(&dirrem->dm_list); 2490 } 2491 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum, 2492 0, &inodedep) != 0) 2493 (void) free_inodedep(inodedep); 2494 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2495 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 2496 nextmd = LIST_NEXT(mkdir, md_mkdirs); 2497 if (mkdir->md_diradd != dap) 2498 continue; 2499 dap->da_state &= ~mkdir->md_state; 2500 WORKLIST_REMOVE(&mkdir->md_list); 2501 LIST_REMOVE(mkdir, md_mkdirs); 2502 WORKITEM_FREE(mkdir, D_MKDIR); 2503 } 2504 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2505 panic("free_diradd: unfound ref"); 2506 } 2507 } 2508 WORKITEM_FREE(dap, D_DIRADD); 2509 } 2510 2511 /* 2512 * Directory entry removal dependencies. 2513 * 2514 * When removing a directory entry, the entry's inode pointer must be 2515 * zero'ed on disk before the corresponding inode's link count is decremented 2516 * (possibly freeing the inode for re-use). This dependency is handled by 2517 * updating the directory entry but delaying the inode count reduction until 2518 * after the directory block has been written to disk. After this point, the 2519 * inode count can be decremented whenever it is convenient. 2520 */ 2521 2522 /* 2523 * This routine should be called immediately after removing 2524 * a directory entry. The inode's link count should not be 2525 * decremented by the calling procedure -- the soft updates 2526 * code will do this task when it is safe. 2527 * 2528 * Parameters: 2529 * bp: buffer containing directory block 2530 * dp: inode for the directory being modified 2531 * ip: inode for directory entry being removed 2532 * isrmdir: indicates if doing RMDIR 2533 */ 2534 void 2535 softdep_setup_remove(struct buf *bp, struct inode *dp, struct inode *ip, 2536 int isrmdir) 2537 { 2538 struct dirrem *dirrem, *prevdirrem; 2539 2540 /* 2541 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 2542 */ 2543 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2544 2545 /* 2546 * If the COMPLETE flag is clear, then there were no active 2547 * entries and we want to roll back to a zeroed entry until 2548 * the new inode is committed to disk. If the COMPLETE flag is 2549 * set then we have deleted an entry that never made it to 2550 * disk. If the entry we deleted resulted from a name change, 2551 * then the old name still resides on disk. We cannot delete 2552 * its inode (returned to us in prevdirrem) until the zeroed 2553 * directory entry gets to disk. The new inode has never been 2554 * referenced on the disk, so can be deleted immediately. 2555 */ 2556 if ((dirrem->dm_state & COMPLETE) == 0) { 2557 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 2558 dm_next); 2559 FREE_LOCK(&lk); 2560 } else { 2561 if (prevdirrem != NULL) 2562 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 2563 prevdirrem, dm_next); 2564 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 2565 FREE_LOCK(&lk); 2566 handle_workitem_remove(dirrem); 2567 } 2568 } 2569 2570 /* 2571 * Allocate a new dirrem if appropriate and return it along with 2572 * its associated pagedep. Called without a lock, returns with lock. 2573 */ 2574 static long num_dirrem; /* number of dirrem allocated */ 2575 2576 /* 2577 * Parameters: 2578 * bp: buffer containing directory block 2579 * dp: inode for the directory being modified 2580 * ip: inode for directory entry being removed 2581 * isrmdir: indicates if doing RMDIR 2582 * prevdirremp: previously referenced inode, if any 2583 */ 2584 static struct dirrem * 2585 newdirrem(struct buf *bp, struct inode *dp, struct inode *ip, 2586 int isrmdir, struct dirrem **prevdirremp) 2587 { 2588 int offset; 2589 ufs_lbn_t lbn; 2590 struct diradd *dap; 2591 struct dirrem *dirrem; 2592 struct pagedep *pagedep; 2593 2594 /* 2595 * Whiteouts have no deletion dependencies. 2596 */ 2597 if (ip == NULL) 2598 panic("newdirrem: whiteout"); 2599 /* 2600 * If we are over our limit, try to improve the situation. 2601 * Limiting the number of dirrem structures will also limit 2602 * the number of freefile and freeblks structures. 2603 */ 2604 if (num_dirrem > max_softdeps / 4) 2605 speedup_syncer(NULL); 2606 if (num_dirrem > max_softdeps / 2) { 2607 ACQUIRE_LOCK(&lk); 2608 request_cleanup(FLUSH_REMOVE); 2609 FREE_LOCK(&lk); 2610 } 2611 2612 num_dirrem += 1; 2613 dirrem = kmalloc(sizeof(struct dirrem), M_DIRREM, 2614 M_SOFTDEP_FLAGS | M_ZERO); 2615 dirrem->dm_list.wk_type = D_DIRREM; 2616 dirrem->dm_state = isrmdir ? RMDIR : 0; 2617 dirrem->dm_mnt = ITOV(ip)->v_mount; 2618 dirrem->dm_oldinum = ip->i_number; 2619 *prevdirremp = NULL; 2620 2621 ACQUIRE_LOCK(&lk); 2622 lbn = lblkno(dp->i_fs, dp->i_offset); 2623 offset = blkoff(dp->i_fs, dp->i_offset); 2624 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2625 WORKLIST_INSERT_BP(bp, &pagedep->pd_list); 2626 dirrem->dm_pagedep = pagedep; 2627 /* 2628 * Check for a diradd dependency for the same directory entry. 2629 * If present, then both dependencies become obsolete and can 2630 * be de-allocated. Check for an entry on both the pd_dirraddhd 2631 * list and the pd_pendinghd list. 2632 */ 2633 2634 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 2635 if (dap->da_offset == offset) 2636 break; 2637 if (dap == NULL) { 2638 2639 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 2640 if (dap->da_offset == offset) 2641 break; 2642 if (dap == NULL) 2643 return (dirrem); 2644 } 2645 /* 2646 * Must be ATTACHED at this point. 2647 */ 2648 if ((dap->da_state & ATTACHED) == 0) { 2649 panic("newdirrem: not ATTACHED"); 2650 } 2651 if (dap->da_newinum != ip->i_number) { 2652 panic("newdirrem: inum %"PRId64" should be %"PRId64, 2653 ip->i_number, dap->da_newinum); 2654 } 2655 /* 2656 * If we are deleting a changed name that never made it to disk, 2657 * then return the dirrem describing the previous inode (which 2658 * represents the inode currently referenced from this entry on disk). 2659 */ 2660 if ((dap->da_state & DIRCHG) != 0) { 2661 *prevdirremp = dap->da_previous; 2662 dap->da_state &= ~DIRCHG; 2663 dap->da_pagedep = pagedep; 2664 } 2665 /* 2666 * We are deleting an entry that never made it to disk. 2667 * Mark it COMPLETE so we can delete its inode immediately. 2668 */ 2669 dirrem->dm_state |= COMPLETE; 2670 free_diradd(dap); 2671 return (dirrem); 2672 } 2673 2674 /* 2675 * Directory entry change dependencies. 2676 * 2677 * Changing an existing directory entry requires that an add operation 2678 * be completed first followed by a deletion. The semantics for the addition 2679 * are identical to the description of adding a new entry above except 2680 * that the rollback is to the old inode number rather than zero. Once 2681 * the addition dependency is completed, the removal is done as described 2682 * in the removal routine above. 2683 */ 2684 2685 /* 2686 * This routine should be called immediately after changing 2687 * a directory entry. The inode's link count should not be 2688 * decremented by the calling procedure -- the soft updates 2689 * code will perform this task when it is safe. 2690 * 2691 * Parameters: 2692 * bp: buffer containing directory block 2693 * dp: inode for the directory being modified 2694 * ip: inode for directory entry being removed 2695 * newinum: new inode number for changed entry 2696 * isrmdir: indicates if doing RMDIR 2697 */ 2698 void 2699 softdep_setup_directory_change(struct buf *bp, struct inode *dp, 2700 struct inode *ip, ino_t newinum, 2701 int isrmdir) 2702 { 2703 int offset; 2704 struct diradd *dap = NULL; 2705 struct dirrem *dirrem, *prevdirrem; 2706 struct pagedep *pagedep; 2707 struct inodedep *inodedep; 2708 2709 offset = blkoff(dp->i_fs, dp->i_offset); 2710 2711 /* 2712 * Whiteouts do not need diradd dependencies. 2713 */ 2714 if (newinum != WINO) { 2715 dap = kmalloc(sizeof(struct diradd), M_DIRADD, 2716 M_SOFTDEP_FLAGS | M_ZERO); 2717 dap->da_list.wk_type = D_DIRADD; 2718 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 2719 dap->da_offset = offset; 2720 dap->da_newinum = newinum; 2721 } 2722 2723 /* 2724 * Allocate a new dirrem and ACQUIRE_LOCK. 2725 */ 2726 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2727 pagedep = dirrem->dm_pagedep; 2728 /* 2729 * The possible values for isrmdir: 2730 * 0 - non-directory file rename 2731 * 1 - directory rename within same directory 2732 * inum - directory rename to new directory of given inode number 2733 * When renaming to a new directory, we are both deleting and 2734 * creating a new directory entry, so the link count on the new 2735 * directory should not change. Thus we do not need the followup 2736 * dirrem which is usually done in handle_workitem_remove. We set 2737 * the DIRCHG flag to tell handle_workitem_remove to skip the 2738 * followup dirrem. 2739 */ 2740 if (isrmdir > 1) 2741 dirrem->dm_state |= DIRCHG; 2742 2743 /* 2744 * Whiteouts have no additional dependencies, 2745 * so just put the dirrem on the correct list. 2746 */ 2747 if (newinum == WINO) { 2748 if ((dirrem->dm_state & COMPLETE) == 0) { 2749 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 2750 dm_next); 2751 } else { 2752 dirrem->dm_dirinum = pagedep->pd_ino; 2753 add_to_worklist(&dirrem->dm_list); 2754 } 2755 FREE_LOCK(&lk); 2756 return; 2757 } 2758 2759 /* 2760 * If the COMPLETE flag is clear, then there were no active 2761 * entries and we want to roll back to the previous inode until 2762 * the new inode is committed to disk. If the COMPLETE flag is 2763 * set, then we have deleted an entry that never made it to disk. 2764 * If the entry we deleted resulted from a name change, then the old 2765 * inode reference still resides on disk. Any rollback that we do 2766 * needs to be to that old inode (returned to us in prevdirrem). If 2767 * the entry we deleted resulted from a create, then there is 2768 * no entry on the disk, so we want to roll back to zero rather 2769 * than the uncommitted inode. In either of the COMPLETE cases we 2770 * want to immediately free the unwritten and unreferenced inode. 2771 */ 2772 if ((dirrem->dm_state & COMPLETE) == 0) { 2773 dap->da_previous = dirrem; 2774 } else { 2775 if (prevdirrem != NULL) { 2776 dap->da_previous = prevdirrem; 2777 } else { 2778 dap->da_state &= ~DIRCHG; 2779 dap->da_pagedep = pagedep; 2780 } 2781 dirrem->dm_dirinum = pagedep->pd_ino; 2782 add_to_worklist(&dirrem->dm_list); 2783 } 2784 /* 2785 * Link into its inodedep. Put it on the id_bufwait list if the inode 2786 * is not yet written. If it is written, do the post-inode write 2787 * processing to put it on the id_pendinghd list. 2788 */ 2789 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 || 2790 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2791 dap->da_state |= COMPLETE; 2792 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 2793 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 2794 } else { 2795 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 2796 dap, da_pdlist); 2797 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2798 } 2799 FREE_LOCK(&lk); 2800 } 2801 2802 /* 2803 * Called whenever the link count on an inode is changed. 2804 * It creates an inode dependency so that the new reference(s) 2805 * to the inode cannot be committed to disk until the updated 2806 * inode has been written. 2807 * 2808 * Parameters: 2809 * ip: the inode with the increased link count 2810 */ 2811 void 2812 softdep_change_linkcnt(struct inode *ip) 2813 { 2814 struct inodedep *inodedep; 2815 2816 ACQUIRE_LOCK(&lk); 2817 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 2818 if (ip->i_nlink < ip->i_effnlink) { 2819 panic("softdep_change_linkcnt: bad delta"); 2820 } 2821 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2822 FREE_LOCK(&lk); 2823 } 2824 2825 /* 2826 * This workitem decrements the inode's link count. 2827 * If the link count reaches zero, the file is removed. 2828 */ 2829 static void 2830 handle_workitem_remove(struct dirrem *dirrem) 2831 { 2832 struct inodedep *inodedep; 2833 struct vnode *vp; 2834 struct inode *ip; 2835 ino_t oldinum; 2836 int error; 2837 2838 error = VFS_VGET(dirrem->dm_mnt, NULL, dirrem->dm_oldinum, &vp); 2839 if (error) { 2840 softdep_error("handle_workitem_remove: vget", error); 2841 return; 2842 } 2843 ip = VTOI(vp); 2844 ACQUIRE_LOCK(&lk); 2845 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){ 2846 panic("handle_workitem_remove: lost inodedep"); 2847 } 2848 /* 2849 * Normal file deletion. 2850 */ 2851 if ((dirrem->dm_state & RMDIR) == 0) { 2852 ip->i_nlink--; 2853 ip->i_flag |= IN_CHANGE; 2854 if (ip->i_nlink < ip->i_effnlink) { 2855 panic("handle_workitem_remove: bad file delta"); 2856 } 2857 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2858 FREE_LOCK(&lk); 2859 vput(vp); 2860 num_dirrem -= 1; 2861 WORKITEM_FREE(dirrem, D_DIRREM); 2862 return; 2863 } 2864 /* 2865 * Directory deletion. Decrement reference count for both the 2866 * just deleted parent directory entry and the reference for ".". 2867 * Next truncate the directory to length zero. When the 2868 * truncation completes, arrange to have the reference count on 2869 * the parent decremented to account for the loss of "..". 2870 */ 2871 ip->i_nlink -= 2; 2872 ip->i_flag |= IN_CHANGE; 2873 if (ip->i_nlink < ip->i_effnlink) { 2874 panic("handle_workitem_remove: bad dir delta"); 2875 } 2876 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2877 FREE_LOCK(&lk); 2878 if ((error = ffs_truncate(vp, (off_t)0, 0, proc0.p_ucred)) != 0) 2879 softdep_error("handle_workitem_remove: truncate", error); 2880 /* 2881 * Rename a directory to a new parent. Since, we are both deleting 2882 * and creating a new directory entry, the link count on the new 2883 * directory should not change. Thus we skip the followup dirrem. 2884 */ 2885 if (dirrem->dm_state & DIRCHG) { 2886 vput(vp); 2887 num_dirrem -= 1; 2888 WORKITEM_FREE(dirrem, D_DIRREM); 2889 return; 2890 } 2891 /* 2892 * If the inodedep does not exist, then the zero'ed inode has 2893 * been written to disk. If the allocated inode has never been 2894 * written to disk, then the on-disk inode is zero'ed. In either 2895 * case we can remove the file immediately. 2896 */ 2897 ACQUIRE_LOCK(&lk); 2898 dirrem->dm_state = 0; 2899 oldinum = dirrem->dm_oldinum; 2900 dirrem->dm_oldinum = dirrem->dm_dirinum; 2901 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 || 2902 check_inode_unwritten(inodedep)) { 2903 FREE_LOCK(&lk); 2904 vput(vp); 2905 handle_workitem_remove(dirrem); 2906 return; 2907 } 2908 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 2909 FREE_LOCK(&lk); 2910 ip->i_flag |= IN_CHANGE; 2911 ffs_update(vp, 0); 2912 vput(vp); 2913 } 2914 2915 /* 2916 * Inode de-allocation dependencies. 2917 * 2918 * When an inode's link count is reduced to zero, it can be de-allocated. We 2919 * found it convenient to postpone de-allocation until after the inode is 2920 * written to disk with its new link count (zero). At this point, all of the 2921 * on-disk inode's block pointers are nullified and, with careful dependency 2922 * list ordering, all dependencies related to the inode will be satisfied and 2923 * the corresponding dependency structures de-allocated. So, if/when the 2924 * inode is reused, there will be no mixing of old dependencies with new 2925 * ones. This artificial dependency is set up by the block de-allocation 2926 * procedure above (softdep_setup_freeblocks) and completed by the 2927 * following procedure. 2928 */ 2929 static void 2930 handle_workitem_freefile(struct freefile *freefile) 2931 { 2932 struct vnode vp; 2933 struct inode tip; 2934 struct inodedep *idp; 2935 int error; 2936 2937 #ifdef DEBUG 2938 ACQUIRE_LOCK(&lk); 2939 error = inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp); 2940 FREE_LOCK(&lk); 2941 if (error) 2942 panic("handle_workitem_freefile: inodedep survived"); 2943 #endif 2944 tip.i_devvp = freefile->fx_devvp; 2945 tip.i_dev = freefile->fx_devvp->v_rdev; 2946 tip.i_fs = freefile->fx_fs; 2947 vp.v_data = &tip; 2948 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0) 2949 softdep_error("handle_workitem_freefile", error); 2950 WORKITEM_FREE(freefile, D_FREEFILE); 2951 } 2952 2953 /* 2954 * Helper function which unlinks marker element from work list and returns 2955 * the next element on the list. 2956 */ 2957 static __inline struct worklist * 2958 markernext(struct worklist *marker) 2959 { 2960 struct worklist *next; 2961 2962 next = LIST_NEXT(marker, wk_list); 2963 LIST_REMOVE(marker, wk_list); 2964 return next; 2965 } 2966 2967 /* 2968 * checkread, checkwrite 2969 * 2970 * bioops callback - hold io_token 2971 */ 2972 static int 2973 softdep_checkread(struct buf *bp) 2974 { 2975 /* nothing to do, mp lock not needed */ 2976 return(0); 2977 } 2978 2979 /* 2980 * bioops callback - hold io_token 2981 */ 2982 static int 2983 softdep_checkwrite(struct buf *bp) 2984 { 2985 /* nothing to do, mp lock not needed */ 2986 return(0); 2987 } 2988 2989 /* 2990 * Disk writes. 2991 * 2992 * The dependency structures constructed above are most actively used when file 2993 * system blocks are written to disk. No constraints are placed on when a 2994 * block can be written, but unsatisfied update dependencies are made safe by 2995 * modifying (or replacing) the source memory for the duration of the disk 2996 * write. When the disk write completes, the memory block is again brought 2997 * up-to-date. 2998 * 2999 * In-core inode structure reclamation. 3000 * 3001 * Because there are a finite number of "in-core" inode structures, they are 3002 * reused regularly. By transferring all inode-related dependencies to the 3003 * in-memory inode block and indexing them separately (via "inodedep"s), we 3004 * can allow "in-core" inode structures to be reused at any time and avoid 3005 * any increase in contention. 3006 * 3007 * Called just before entering the device driver to initiate a new disk I/O. 3008 * The buffer must be locked, thus, no I/O completion operations can occur 3009 * while we are manipulating its associated dependencies. 3010 * 3011 * bioops callback - hold io_token 3012 * 3013 * Parameters: 3014 * bp: structure describing disk write to occur 3015 */ 3016 static void 3017 softdep_disk_io_initiation(struct buf *bp) 3018 { 3019 struct worklist *wk; 3020 struct worklist marker; 3021 struct indirdep *indirdep; 3022 3023 /* 3024 * We only care about write operations. There should never 3025 * be dependencies for reads. 3026 */ 3027 if (bp->b_cmd == BUF_CMD_READ) 3028 panic("softdep_disk_io_initiation: read"); 3029 3030 ACQUIRE_LOCK(&lk); 3031 marker.wk_type = D_LAST + 1; /* Not a normal workitem */ 3032 3033 /* 3034 * Do any necessary pre-I/O processing. 3035 */ 3036 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = markernext(&marker)) { 3037 LIST_INSERT_AFTER(wk, &marker, wk_list); 3038 3039 switch (wk->wk_type) { 3040 case D_PAGEDEP: 3041 initiate_write_filepage(WK_PAGEDEP(wk), bp); 3042 continue; 3043 3044 case D_INODEDEP: 3045 initiate_write_inodeblock(WK_INODEDEP(wk), bp); 3046 continue; 3047 3048 case D_INDIRDEP: 3049 indirdep = WK_INDIRDEP(wk); 3050 if (indirdep->ir_state & GOINGAWAY) 3051 panic("disk_io_initiation: indirdep gone"); 3052 /* 3053 * If there are no remaining dependencies, this 3054 * will be writing the real pointers, so the 3055 * dependency can be freed. 3056 */ 3057 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) { 3058 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 3059 brelse(indirdep->ir_savebp); 3060 /* inline expand WORKLIST_REMOVE(wk); */ 3061 wk->wk_state &= ~ONWORKLIST; 3062 LIST_REMOVE(wk, wk_list); 3063 WORKITEM_FREE(indirdep, D_INDIRDEP); 3064 continue; 3065 } 3066 /* 3067 * Replace up-to-date version with safe version. 3068 */ 3069 indirdep->ir_saveddata = kmalloc(bp->b_bcount, 3070 M_INDIRDEP, 3071 M_SOFTDEP_FLAGS); 3072 ACQUIRE_LOCK(&lk); 3073 indirdep->ir_state &= ~ATTACHED; 3074 indirdep->ir_state |= UNDONE; 3075 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 3076 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 3077 bp->b_bcount); 3078 FREE_LOCK(&lk); 3079 continue; 3080 3081 case D_MKDIR: 3082 case D_BMSAFEMAP: 3083 case D_ALLOCDIRECT: 3084 case D_ALLOCINDIR: 3085 continue; 3086 3087 default: 3088 panic("handle_disk_io_initiation: Unexpected type %s", 3089 TYPENAME(wk->wk_type)); 3090 /* NOTREACHED */ 3091 } 3092 } 3093 FREE_LOCK(&lk); 3094 } 3095 3096 /* 3097 * Called from within the procedure above to deal with unsatisfied 3098 * allocation dependencies in a directory. The buffer must be locked, 3099 * thus, no I/O completion operations can occur while we are 3100 * manipulating its associated dependencies. 3101 */ 3102 static void 3103 initiate_write_filepage(struct pagedep *pagedep, struct buf *bp) 3104 { 3105 struct diradd *dap; 3106 struct direct *ep; 3107 int i; 3108 3109 if (pagedep->pd_state & IOSTARTED) { 3110 /* 3111 * This can only happen if there is a driver that does not 3112 * understand chaining. Here biodone will reissue the call 3113 * to strategy for the incomplete buffers. 3114 */ 3115 kprintf("initiate_write_filepage: already started\n"); 3116 return; 3117 } 3118 pagedep->pd_state |= IOSTARTED; 3119 ACQUIRE_LOCK(&lk); 3120 for (i = 0; i < DAHASHSZ; i++) { 3121 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 3122 ep = (struct direct *) 3123 ((char *)bp->b_data + dap->da_offset); 3124 if (ep->d_ino != dap->da_newinum) { 3125 panic("%s: dir inum %d != new %"PRId64, 3126 "initiate_write_filepage", 3127 ep->d_ino, dap->da_newinum); 3128 } 3129 if (dap->da_state & DIRCHG) 3130 ep->d_ino = dap->da_previous->dm_oldinum; 3131 else 3132 ep->d_ino = 0; 3133 dap->da_state &= ~ATTACHED; 3134 dap->da_state |= UNDONE; 3135 } 3136 } 3137 FREE_LOCK(&lk); 3138 } 3139 3140 /* 3141 * Called from within the procedure above to deal with unsatisfied 3142 * allocation dependencies in an inodeblock. The buffer must be 3143 * locked, thus, no I/O completion operations can occur while we 3144 * are manipulating its associated dependencies. 3145 * 3146 * Parameters: 3147 * bp: The inode block 3148 */ 3149 static void 3150 initiate_write_inodeblock(struct inodedep *inodedep, struct buf *bp) 3151 { 3152 struct allocdirect *adp, *lastadp; 3153 struct ufs1_dinode *dp; 3154 struct ufs1_dinode *sip; 3155 struct fs *fs; 3156 ufs_lbn_t prevlbn = 0; 3157 int i, deplist; 3158 3159 if (inodedep->id_state & IOSTARTED) 3160 panic("initiate_write_inodeblock: already started"); 3161 inodedep->id_state |= IOSTARTED; 3162 fs = inodedep->id_fs; 3163 dp = (struct ufs1_dinode *)bp->b_data + 3164 ino_to_fsbo(fs, inodedep->id_ino); 3165 /* 3166 * If the bitmap is not yet written, then the allocated 3167 * inode cannot be written to disk. 3168 */ 3169 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 3170 if (inodedep->id_savedino != NULL) 3171 panic("initiate_write_inodeblock: already doing I/O"); 3172 sip = kmalloc(sizeof(struct ufs1_dinode), M_INODEDEP, 3173 M_SOFTDEP_FLAGS); 3174 inodedep->id_savedino = sip; 3175 *inodedep->id_savedino = *dp; 3176 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 3177 dp->di_gen = inodedep->id_savedino->di_gen; 3178 return; 3179 } 3180 /* 3181 * If no dependencies, then there is nothing to roll back. 3182 */ 3183 inodedep->id_savedsize = dp->di_size; 3184 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 3185 return; 3186 /* 3187 * Set the dependencies to busy. 3188 */ 3189 ACQUIRE_LOCK(&lk); 3190 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3191 adp = TAILQ_NEXT(adp, ad_next)) { 3192 #ifdef DIAGNOSTIC 3193 if (deplist != 0 && prevlbn >= adp->ad_lbn) { 3194 panic("softdep_write_inodeblock: lbn order"); 3195 } 3196 prevlbn = adp->ad_lbn; 3197 if (adp->ad_lbn < NDADDR && 3198 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) { 3199 panic("%s: direct pointer #%ld mismatch %d != %d", 3200 "softdep_write_inodeblock", adp->ad_lbn, 3201 dp->di_db[adp->ad_lbn], adp->ad_newblkno); 3202 } 3203 if (adp->ad_lbn >= NDADDR && 3204 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) { 3205 panic("%s: indirect pointer #%ld mismatch %d != %d", 3206 "softdep_write_inodeblock", adp->ad_lbn - NDADDR, 3207 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno); 3208 } 3209 deplist |= 1 << adp->ad_lbn; 3210 if ((adp->ad_state & ATTACHED) == 0) { 3211 panic("softdep_write_inodeblock: Unknown state 0x%x", 3212 adp->ad_state); 3213 } 3214 #endif /* DIAGNOSTIC */ 3215 adp->ad_state &= ~ATTACHED; 3216 adp->ad_state |= UNDONE; 3217 } 3218 /* 3219 * The on-disk inode cannot claim to be any larger than the last 3220 * fragment that has been written. Otherwise, the on-disk inode 3221 * might have fragments that were not the last block in the file 3222 * which would corrupt the filesystem. 3223 */ 3224 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3225 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 3226 if (adp->ad_lbn >= NDADDR) 3227 break; 3228 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 3229 /* keep going until hitting a rollback to a frag */ 3230 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 3231 continue; 3232 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 3233 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 3234 #ifdef DIAGNOSTIC 3235 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) { 3236 panic("softdep_write_inodeblock: lost dep1"); 3237 } 3238 #endif /* DIAGNOSTIC */ 3239 dp->di_db[i] = 0; 3240 } 3241 for (i = 0; i < NIADDR; i++) { 3242 #ifdef DIAGNOSTIC 3243 if (dp->di_ib[i] != 0 && 3244 (deplist & ((1 << NDADDR) << i)) == 0) { 3245 panic("softdep_write_inodeblock: lost dep2"); 3246 } 3247 #endif /* DIAGNOSTIC */ 3248 dp->di_ib[i] = 0; 3249 } 3250 FREE_LOCK(&lk); 3251 return; 3252 } 3253 /* 3254 * If we have zero'ed out the last allocated block of the file, 3255 * roll back the size to the last currently allocated block. 3256 * We know that this last allocated block is a full-sized as 3257 * we already checked for fragments in the loop above. 3258 */ 3259 if (lastadp != NULL && 3260 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 3261 for (i = lastadp->ad_lbn; i >= 0; i--) 3262 if (dp->di_db[i] != 0) 3263 break; 3264 dp->di_size = (i + 1) * fs->fs_bsize; 3265 } 3266 /* 3267 * The only dependencies are for indirect blocks. 3268 * 3269 * The file size for indirect block additions is not guaranteed. 3270 * Such a guarantee would be non-trivial to achieve. The conventional 3271 * synchronous write implementation also does not make this guarantee. 3272 * Fsck should catch and fix discrepancies. Arguably, the file size 3273 * can be over-estimated without destroying integrity when the file 3274 * moves into the indirect blocks (i.e., is large). If we want to 3275 * postpone fsck, we are stuck with this argument. 3276 */ 3277 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 3278 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 3279 FREE_LOCK(&lk); 3280 } 3281 3282 /* 3283 * This routine is called during the completion interrupt 3284 * service routine for a disk write (from the procedure called 3285 * by the device driver to inform the filesystem caches of 3286 * a request completion). It should be called early in this 3287 * procedure, before the block is made available to other 3288 * processes or other routines are called. 3289 * 3290 * bioops callback - hold io_token 3291 * 3292 * Parameters: 3293 * bp: describes the completed disk write 3294 */ 3295 static void 3296 softdep_disk_write_complete(struct buf *bp) 3297 { 3298 struct worklist *wk; 3299 struct workhead reattach; 3300 struct newblk *newblk; 3301 struct allocindir *aip; 3302 struct allocdirect *adp; 3303 struct indirdep *indirdep; 3304 struct inodedep *inodedep; 3305 struct bmsafemap *bmsafemap; 3306 3307 ACQUIRE_LOCK(&lk); 3308 3309 LIST_INIT(&reattach); 3310 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 3311 WORKLIST_REMOVE(wk); 3312 switch (wk->wk_type) { 3313 3314 case D_PAGEDEP: 3315 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 3316 WORKLIST_INSERT(&reattach, wk); 3317 continue; 3318 3319 case D_INODEDEP: 3320 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 3321 WORKLIST_INSERT(&reattach, wk); 3322 continue; 3323 3324 case D_BMSAFEMAP: 3325 bmsafemap = WK_BMSAFEMAP(wk); 3326 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) { 3327 newblk->nb_state |= DEPCOMPLETE; 3328 newblk->nb_bmsafemap = NULL; 3329 LIST_REMOVE(newblk, nb_deps); 3330 } 3331 while ((adp = 3332 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) { 3333 adp->ad_state |= DEPCOMPLETE; 3334 adp->ad_buf = NULL; 3335 LIST_REMOVE(adp, ad_deps); 3336 handle_allocdirect_partdone(adp); 3337 } 3338 while ((aip = 3339 LIST_FIRST(&bmsafemap->sm_allocindirhd))) { 3340 aip->ai_state |= DEPCOMPLETE; 3341 aip->ai_buf = NULL; 3342 LIST_REMOVE(aip, ai_deps); 3343 handle_allocindir_partdone(aip); 3344 } 3345 while ((inodedep = 3346 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { 3347 inodedep->id_state |= DEPCOMPLETE; 3348 LIST_REMOVE(inodedep, id_deps); 3349 inodedep->id_buf = NULL; 3350 } 3351 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 3352 continue; 3353 3354 case D_MKDIR: 3355 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 3356 continue; 3357 3358 case D_ALLOCDIRECT: 3359 adp = WK_ALLOCDIRECT(wk); 3360 adp->ad_state |= COMPLETE; 3361 handle_allocdirect_partdone(adp); 3362 continue; 3363 3364 case D_ALLOCINDIR: 3365 aip = WK_ALLOCINDIR(wk); 3366 aip->ai_state |= COMPLETE; 3367 handle_allocindir_partdone(aip); 3368 continue; 3369 3370 case D_INDIRDEP: 3371 indirdep = WK_INDIRDEP(wk); 3372 if (indirdep->ir_state & GOINGAWAY) { 3373 panic("disk_write_complete: indirdep gone"); 3374 } 3375 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 3376 kfree(indirdep->ir_saveddata, M_INDIRDEP); 3377 indirdep->ir_saveddata = NULL; 3378 indirdep->ir_state &= ~UNDONE; 3379 indirdep->ir_state |= ATTACHED; 3380 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != NULL) { 3381 handle_allocindir_partdone(aip); 3382 if (aip == LIST_FIRST(&indirdep->ir_donehd)) { 3383 panic("disk_write_complete: not gone"); 3384 } 3385 } 3386 WORKLIST_INSERT(&reattach, wk); 3387 if ((bp->b_flags & B_DELWRI) == 0) 3388 stat_indir_blk_ptrs++; 3389 bdirty(bp); 3390 continue; 3391 3392 default: 3393 panic("handle_disk_write_complete: Unknown type %s", 3394 TYPENAME(wk->wk_type)); 3395 /* NOTREACHED */ 3396 } 3397 } 3398 /* 3399 * Reattach any requests that must be redone. 3400 */ 3401 while ((wk = LIST_FIRST(&reattach)) != NULL) { 3402 WORKLIST_REMOVE(wk); 3403 WORKLIST_INSERT_BP(bp, wk); 3404 } 3405 3406 FREE_LOCK(&lk); 3407 } 3408 3409 /* 3410 * Called from within softdep_disk_write_complete above. Note that 3411 * this routine is always called from interrupt level with further 3412 * splbio interrupts blocked. 3413 * 3414 * Parameters: 3415 * adp: the completed allocdirect 3416 */ 3417 static void 3418 handle_allocdirect_partdone(struct allocdirect *adp) 3419 { 3420 struct allocdirect *listadp; 3421 struct inodedep *inodedep; 3422 long bsize; 3423 3424 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3425 return; 3426 if (adp->ad_buf != NULL) 3427 panic("handle_allocdirect_partdone: dangling dep"); 3428 3429 /* 3430 * The on-disk inode cannot claim to be any larger than the last 3431 * fragment that has been written. Otherwise, the on-disk inode 3432 * might have fragments that were not the last block in the file 3433 * which would corrupt the filesystem. Thus, we cannot free any 3434 * allocdirects after one whose ad_oldblkno claims a fragment as 3435 * these blocks must be rolled back to zero before writing the inode. 3436 * We check the currently active set of allocdirects in id_inoupdt. 3437 */ 3438 inodedep = adp->ad_inodedep; 3439 bsize = inodedep->id_fs->fs_bsize; 3440 TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) { 3441 /* found our block */ 3442 if (listadp == adp) 3443 break; 3444 /* continue if ad_oldlbn is not a fragment */ 3445 if (listadp->ad_oldsize == 0 || 3446 listadp->ad_oldsize == bsize) 3447 continue; 3448 /* hit a fragment */ 3449 return; 3450 } 3451 /* 3452 * If we have reached the end of the current list without 3453 * finding the just finished dependency, then it must be 3454 * on the future dependency list. Future dependencies cannot 3455 * be freed until they are moved to the current list. 3456 */ 3457 if (listadp == NULL) { 3458 #ifdef DEBUG 3459 TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next) 3460 /* found our block */ 3461 if (listadp == adp) 3462 break; 3463 if (listadp == NULL) 3464 panic("handle_allocdirect_partdone: lost dep"); 3465 #endif /* DEBUG */ 3466 return; 3467 } 3468 /* 3469 * If we have found the just finished dependency, then free 3470 * it along with anything that follows it that is complete. 3471 */ 3472 for (; adp; adp = listadp) { 3473 listadp = TAILQ_NEXT(adp, ad_next); 3474 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3475 return; 3476 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 3477 } 3478 } 3479 3480 /* 3481 * Called from within softdep_disk_write_complete above. Note that 3482 * this routine is always called from interrupt level with further 3483 * splbio interrupts blocked. 3484 * 3485 * Parameters: 3486 * aip: the completed allocindir 3487 */ 3488 static void 3489 handle_allocindir_partdone(struct allocindir *aip) 3490 { 3491 struct indirdep *indirdep; 3492 3493 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 3494 return; 3495 if (aip->ai_buf != NULL) 3496 panic("handle_allocindir_partdone: dangling dependency"); 3497 3498 indirdep = aip->ai_indirdep; 3499 if (indirdep->ir_state & UNDONE) { 3500 LIST_REMOVE(aip, ai_next); 3501 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 3502 return; 3503 } 3504 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 3505 aip->ai_newblkno; 3506 LIST_REMOVE(aip, ai_next); 3507 if (aip->ai_freefrag != NULL) 3508 add_to_worklist(&aip->ai_freefrag->ff_list); 3509 WORKITEM_FREE(aip, D_ALLOCINDIR); 3510 } 3511 3512 /* 3513 * Called from within softdep_disk_write_complete above to restore 3514 * in-memory inode block contents to their most up-to-date state. Note 3515 * that this routine is always called from interrupt level with further 3516 * splbio interrupts blocked. 3517 * 3518 * Parameters: 3519 * bp: buffer containing the inode block 3520 */ 3521 static int 3522 handle_written_inodeblock(struct inodedep *inodedep, struct buf *bp) 3523 { 3524 struct worklist *wk, *filefree; 3525 struct allocdirect *adp, *nextadp; 3526 struct ufs1_dinode *dp; 3527 int hadchanges; 3528 3529 if ((inodedep->id_state & IOSTARTED) == 0) 3530 panic("handle_written_inodeblock: not started"); 3531 3532 inodedep->id_state &= ~IOSTARTED; 3533 dp = (struct ufs1_dinode *)bp->b_data + 3534 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 3535 /* 3536 * If we had to rollback the inode allocation because of 3537 * bitmaps being incomplete, then simply restore it. 3538 * Keep the block dirty so that it will not be reclaimed until 3539 * all associated dependencies have been cleared and the 3540 * corresponding updates written to disk. 3541 */ 3542 if (inodedep->id_savedino != NULL) { 3543 *dp = *inodedep->id_savedino; 3544 kfree(inodedep->id_savedino, M_INODEDEP); 3545 inodedep->id_savedino = NULL; 3546 if ((bp->b_flags & B_DELWRI) == 0) 3547 stat_inode_bitmap++; 3548 bdirty(bp); 3549 return (1); 3550 } 3551 inodedep->id_state |= COMPLETE; 3552 /* 3553 * Roll forward anything that had to be rolled back before 3554 * the inode could be updated. 3555 */ 3556 hadchanges = 0; 3557 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 3558 nextadp = TAILQ_NEXT(adp, ad_next); 3559 if (adp->ad_state & ATTACHED) 3560 panic("handle_written_inodeblock: new entry"); 3561 3562 if (adp->ad_lbn < NDADDR) { 3563 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) { 3564 panic("%s: %s #%ld mismatch %d != %d", 3565 "handle_written_inodeblock", 3566 "direct pointer", adp->ad_lbn, 3567 dp->di_db[adp->ad_lbn], adp->ad_oldblkno); 3568 } 3569 dp->di_db[adp->ad_lbn] = adp->ad_newblkno; 3570 } else { 3571 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) { 3572 panic("%s: %s #%ld allocated as %d", 3573 "handle_written_inodeblock", 3574 "indirect pointer", adp->ad_lbn - NDADDR, 3575 dp->di_ib[adp->ad_lbn - NDADDR]); 3576 } 3577 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno; 3578 } 3579 adp->ad_state &= ~UNDONE; 3580 adp->ad_state |= ATTACHED; 3581 hadchanges = 1; 3582 } 3583 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 3584 stat_direct_blk_ptrs++; 3585 /* 3586 * Reset the file size to its most up-to-date value. 3587 */ 3588 if (inodedep->id_savedsize == -1) { 3589 panic("handle_written_inodeblock: bad size"); 3590 } 3591 if (dp->di_size != inodedep->id_savedsize) { 3592 dp->di_size = inodedep->id_savedsize; 3593 hadchanges = 1; 3594 } 3595 inodedep->id_savedsize = -1; 3596 /* 3597 * If there were any rollbacks in the inode block, then it must be 3598 * marked dirty so that its will eventually get written back in 3599 * its correct form. 3600 */ 3601 if (hadchanges) 3602 bdirty(bp); 3603 /* 3604 * Process any allocdirects that completed during the update. 3605 */ 3606 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 3607 handle_allocdirect_partdone(adp); 3608 /* 3609 * Process deallocations that were held pending until the 3610 * inode had been written to disk. Freeing of the inode 3611 * is delayed until after all blocks have been freed to 3612 * avoid creation of new <vfsid, inum, lbn> triples 3613 * before the old ones have been deleted. 3614 */ 3615 filefree = NULL; 3616 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 3617 WORKLIST_REMOVE(wk); 3618 switch (wk->wk_type) { 3619 3620 case D_FREEFILE: 3621 /* 3622 * We defer adding filefree to the worklist until 3623 * all other additions have been made to ensure 3624 * that it will be done after all the old blocks 3625 * have been freed. 3626 */ 3627 if (filefree != NULL) { 3628 panic("handle_written_inodeblock: filefree"); 3629 } 3630 filefree = wk; 3631 continue; 3632 3633 case D_MKDIR: 3634 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 3635 continue; 3636 3637 case D_DIRADD: 3638 diradd_inode_written(WK_DIRADD(wk), inodedep); 3639 continue; 3640 3641 case D_FREEBLKS: 3642 wk->wk_state |= COMPLETE; 3643 if ((wk->wk_state & ALLCOMPLETE) != ALLCOMPLETE) 3644 continue; 3645 /* -- fall through -- */ 3646 case D_FREEFRAG: 3647 case D_DIRREM: 3648 add_to_worklist(wk); 3649 continue; 3650 3651 default: 3652 panic("handle_written_inodeblock: Unknown type %s", 3653 TYPENAME(wk->wk_type)); 3654 /* NOTREACHED */ 3655 } 3656 } 3657 if (filefree != NULL) { 3658 if (free_inodedep(inodedep) == 0) { 3659 panic("handle_written_inodeblock: live inodedep"); 3660 } 3661 add_to_worklist(filefree); 3662 return (0); 3663 } 3664 3665 /* 3666 * If no outstanding dependencies, free it. 3667 */ 3668 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 3669 return (0); 3670 return (hadchanges); 3671 } 3672 3673 /* 3674 * Process a diradd entry after its dependent inode has been written. 3675 * This routine must be called with splbio interrupts blocked. 3676 */ 3677 static void 3678 diradd_inode_written(struct diradd *dap, struct inodedep *inodedep) 3679 { 3680 struct pagedep *pagedep; 3681 3682 dap->da_state |= COMPLETE; 3683 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3684 if (dap->da_state & DIRCHG) 3685 pagedep = dap->da_previous->dm_pagedep; 3686 else 3687 pagedep = dap->da_pagedep; 3688 LIST_REMOVE(dap, da_pdlist); 3689 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3690 } 3691 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 3692 } 3693 3694 /* 3695 * Handle the completion of a mkdir dependency. 3696 */ 3697 static void 3698 handle_written_mkdir(struct mkdir *mkdir, int type) 3699 { 3700 struct diradd *dap; 3701 struct pagedep *pagedep; 3702 3703 if (mkdir->md_state != type) { 3704 panic("handle_written_mkdir: bad type"); 3705 } 3706 dap = mkdir->md_diradd; 3707 dap->da_state &= ~type; 3708 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 3709 dap->da_state |= DEPCOMPLETE; 3710 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3711 if (dap->da_state & DIRCHG) 3712 pagedep = dap->da_previous->dm_pagedep; 3713 else 3714 pagedep = dap->da_pagedep; 3715 LIST_REMOVE(dap, da_pdlist); 3716 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3717 } 3718 LIST_REMOVE(mkdir, md_mkdirs); 3719 WORKITEM_FREE(mkdir, D_MKDIR); 3720 } 3721 3722 /* 3723 * Called from within softdep_disk_write_complete above. 3724 * A write operation was just completed. Removed inodes can 3725 * now be freed and associated block pointers may be committed. 3726 * Note that this routine is always called from interrupt level 3727 * with further splbio interrupts blocked. 3728 * 3729 * Parameters: 3730 * bp: buffer containing the written page 3731 */ 3732 static int 3733 handle_written_filepage(struct pagedep *pagedep, struct buf *bp) 3734 { 3735 struct dirrem *dirrem; 3736 struct diradd *dap, *nextdap; 3737 struct direct *ep; 3738 int i, chgs; 3739 3740 if ((pagedep->pd_state & IOSTARTED) == 0) { 3741 panic("handle_written_filepage: not started"); 3742 } 3743 pagedep->pd_state &= ~IOSTARTED; 3744 /* 3745 * Process any directory removals that have been committed. 3746 */ 3747 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 3748 LIST_REMOVE(dirrem, dm_next); 3749 dirrem->dm_dirinum = pagedep->pd_ino; 3750 add_to_worklist(&dirrem->dm_list); 3751 } 3752 /* 3753 * Free any directory additions that have been committed. 3754 */ 3755 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 3756 free_diradd(dap); 3757 /* 3758 * Uncommitted directory entries must be restored. 3759 */ 3760 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 3761 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 3762 dap = nextdap) { 3763 nextdap = LIST_NEXT(dap, da_pdlist); 3764 if (dap->da_state & ATTACHED) { 3765 panic("handle_written_filepage: attached"); 3766 } 3767 ep = (struct direct *) 3768 ((char *)bp->b_data + dap->da_offset); 3769 ep->d_ino = dap->da_newinum; 3770 dap->da_state &= ~UNDONE; 3771 dap->da_state |= ATTACHED; 3772 chgs = 1; 3773 /* 3774 * If the inode referenced by the directory has 3775 * been written out, then the dependency can be 3776 * moved to the pending list. 3777 */ 3778 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3779 LIST_REMOVE(dap, da_pdlist); 3780 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 3781 da_pdlist); 3782 } 3783 } 3784 } 3785 /* 3786 * If there were any rollbacks in the directory, then it must be 3787 * marked dirty so that its will eventually get written back in 3788 * its correct form. 3789 */ 3790 if (chgs) { 3791 if ((bp->b_flags & B_DELWRI) == 0) 3792 stat_dir_entry++; 3793 bdirty(bp); 3794 } 3795 /* 3796 * If no dependencies remain, the pagedep will be freed. 3797 * Otherwise it will remain to update the page before it 3798 * is written back to disk. 3799 */ 3800 if (LIST_FIRST(&pagedep->pd_pendinghd) == NULL) { 3801 for (i = 0; i < DAHASHSZ; i++) 3802 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL) 3803 break; 3804 if (i == DAHASHSZ) { 3805 LIST_REMOVE(pagedep, pd_hash); 3806 WORKITEM_FREE(pagedep, D_PAGEDEP); 3807 return (0); 3808 } 3809 } 3810 return (1); 3811 } 3812 3813 /* 3814 * Writing back in-core inode structures. 3815 * 3816 * The filesystem only accesses an inode's contents when it occupies an 3817 * "in-core" inode structure. These "in-core" structures are separate from 3818 * the page frames used to cache inode blocks. Only the latter are 3819 * transferred to/from the disk. So, when the updated contents of the 3820 * "in-core" inode structure are copied to the corresponding in-memory inode 3821 * block, the dependencies are also transferred. The following procedure is 3822 * called when copying a dirty "in-core" inode to a cached inode block. 3823 */ 3824 3825 /* 3826 * Called when an inode is loaded from disk. If the effective link count 3827 * differed from the actual link count when it was last flushed, then we 3828 * need to ensure that the correct effective link count is put back. 3829 * 3830 * Parameters: 3831 * ip: the "in_core" copy of the inode 3832 */ 3833 void 3834 softdep_load_inodeblock(struct inode *ip) 3835 { 3836 struct inodedep *inodedep; 3837 3838 /* 3839 * Check for alternate nlink count. 3840 */ 3841 ip->i_effnlink = ip->i_nlink; 3842 ACQUIRE_LOCK(&lk); 3843 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3844 FREE_LOCK(&lk); 3845 return; 3846 } 3847 ip->i_effnlink -= inodedep->id_nlinkdelta; 3848 FREE_LOCK(&lk); 3849 } 3850 3851 /* 3852 * This routine is called just before the "in-core" inode 3853 * information is to be copied to the in-memory inode block. 3854 * Recall that an inode block contains several inodes. If 3855 * the force flag is set, then the dependencies will be 3856 * cleared so that the update can always be made. Note that 3857 * the buffer is locked when this routine is called, so we 3858 * will never be in the middle of writing the inode block 3859 * to disk. 3860 * 3861 * Parameters: 3862 * ip: the "in_core" copy of the inode 3863 * bp: the buffer containing the inode block 3864 * waitfor: nonzero => update must be allowed 3865 */ 3866 void 3867 softdep_update_inodeblock(struct inode *ip, struct buf *bp, 3868 int waitfor) 3869 { 3870 struct inodedep *inodedep; 3871 struct worklist *wk; 3872 struct buf *ibp; 3873 int error, gotit; 3874 3875 /* 3876 * If the effective link count is not equal to the actual link 3877 * count, then we must track the difference in an inodedep while 3878 * the inode is (potentially) tossed out of the cache. Otherwise, 3879 * if there is no existing inodedep, then there are no dependencies 3880 * to track. 3881 */ 3882 ACQUIRE_LOCK(&lk); 3883 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3884 FREE_LOCK(&lk); 3885 if (ip->i_effnlink != ip->i_nlink) 3886 panic("softdep_update_inodeblock: bad link count"); 3887 return; 3888 } 3889 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) { 3890 panic("softdep_update_inodeblock: bad delta"); 3891 } 3892 /* 3893 * Changes have been initiated. Anything depending on these 3894 * changes cannot occur until this inode has been written. 3895 */ 3896 inodedep->id_state &= ~COMPLETE; 3897 if ((inodedep->id_state & ONWORKLIST) == 0) 3898 WORKLIST_INSERT_BP(bp, &inodedep->id_list); 3899 /* 3900 * Any new dependencies associated with the incore inode must 3901 * now be moved to the list associated with the buffer holding 3902 * the in-memory copy of the inode. Once merged process any 3903 * allocdirects that are completed by the merger. 3904 */ 3905 merge_inode_lists(inodedep); 3906 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL) 3907 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); 3908 /* 3909 * Now that the inode has been pushed into the buffer, the 3910 * operations dependent on the inode being written to disk 3911 * can be moved to the id_bufwait so that they will be 3912 * processed when the buffer I/O completes. 3913 */ 3914 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 3915 WORKLIST_REMOVE(wk); 3916 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 3917 } 3918 /* 3919 * Newly allocated inodes cannot be written until the bitmap 3920 * that allocates them have been written (indicated by 3921 * DEPCOMPLETE being set in id_state). If we are doing a 3922 * forced sync (e.g., an fsync on a file), we force the bitmap 3923 * to be written so that the update can be done. 3924 */ 3925 if (waitfor == 0) { 3926 FREE_LOCK(&lk); 3927 return; 3928 } 3929 retry: 3930 if ((inodedep->id_state & DEPCOMPLETE) != 0) { 3931 FREE_LOCK(&lk); 3932 return; 3933 } 3934 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 3935 if (gotit == 0) { 3936 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) != 0) 3937 goto retry; 3938 FREE_LOCK(&lk); 3939 return; 3940 } 3941 ibp = inodedep->id_buf; 3942 FREE_LOCK(&lk); 3943 if ((error = bwrite(ibp)) != 0) 3944 softdep_error("softdep_update_inodeblock: bwrite", error); 3945 } 3946 3947 /* 3948 * Merge the new inode dependency list (id_newinoupdt) into the old 3949 * inode dependency list (id_inoupdt). This routine must be called 3950 * with splbio interrupts blocked. 3951 */ 3952 static void 3953 merge_inode_lists(struct inodedep *inodedep) 3954 { 3955 struct allocdirect *listadp, *newadp; 3956 3957 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 3958 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) { 3959 if (listadp->ad_lbn < newadp->ad_lbn) { 3960 listadp = TAILQ_NEXT(listadp, ad_next); 3961 continue; 3962 } 3963 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 3964 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 3965 if (listadp->ad_lbn == newadp->ad_lbn) { 3966 allocdirect_merge(&inodedep->id_inoupdt, newadp, 3967 listadp); 3968 listadp = newadp; 3969 } 3970 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 3971 } 3972 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) { 3973 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 3974 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next); 3975 } 3976 } 3977 3978 /* 3979 * If we are doing an fsync, then we must ensure that any directory 3980 * entries for the inode have been written after the inode gets to disk. 3981 * 3982 * bioops callback - hold io_token 3983 * 3984 * Parameters: 3985 * vp: the "in_core" copy of the inode 3986 */ 3987 static int 3988 softdep_fsync(struct vnode *vp) 3989 { 3990 struct inodedep *inodedep; 3991 struct pagedep *pagedep; 3992 struct worklist *wk; 3993 struct diradd *dap; 3994 struct mount *mnt; 3995 struct vnode *pvp; 3996 struct inode *ip; 3997 struct buf *bp; 3998 struct fs *fs; 3999 int error, flushparent; 4000 ino_t parentino; 4001 ufs_lbn_t lbn; 4002 4003 /* 4004 * Move check from original kernel code, possibly not needed any 4005 * more with the per-mount bioops. 4006 */ 4007 if ((vp->v_mount->mnt_flag & MNT_SOFTDEP) == 0) 4008 return (0); 4009 4010 ip = VTOI(vp); 4011 fs = ip->i_fs; 4012 ACQUIRE_LOCK(&lk); 4013 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) { 4014 FREE_LOCK(&lk); 4015 return (0); 4016 } 4017 if (LIST_FIRST(&inodedep->id_inowait) != NULL || 4018 LIST_FIRST(&inodedep->id_bufwait) != NULL || 4019 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 4020 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) { 4021 panic("softdep_fsync: pending ops"); 4022 } 4023 for (error = 0, flushparent = 0; ; ) { 4024 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 4025 break; 4026 if (wk->wk_type != D_DIRADD) { 4027 panic("softdep_fsync: Unexpected type %s", 4028 TYPENAME(wk->wk_type)); 4029 } 4030 dap = WK_DIRADD(wk); 4031 /* 4032 * Flush our parent if this directory entry 4033 * has a MKDIR_PARENT dependency. 4034 */ 4035 if (dap->da_state & DIRCHG) 4036 pagedep = dap->da_previous->dm_pagedep; 4037 else 4038 pagedep = dap->da_pagedep; 4039 mnt = pagedep->pd_mnt; 4040 parentino = pagedep->pd_ino; 4041 lbn = pagedep->pd_lbn; 4042 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) { 4043 panic("softdep_fsync: dirty"); 4044 } 4045 flushparent = dap->da_state & MKDIR_PARENT; 4046 /* 4047 * If we are being fsync'ed as part of vgone'ing this vnode, 4048 * then we will not be able to release and recover the 4049 * vnode below, so we just have to give up on writing its 4050 * directory entry out. It will eventually be written, just 4051 * not now, but then the user was not asking to have it 4052 * written, so we are not breaking any promises. 4053 */ 4054 if (vp->v_flag & VRECLAIMED) 4055 break; 4056 /* 4057 * We prevent deadlock by always fetching inodes from the 4058 * root, moving down the directory tree. Thus, when fetching 4059 * our parent directory, we must unlock ourselves before 4060 * requesting the lock on our parent. See the comment in 4061 * ufs_lookup for details on possible races. 4062 */ 4063 FREE_LOCK(&lk); 4064 vn_unlock(vp); 4065 error = VFS_VGET(mnt, NULL, parentino, &pvp); 4066 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4067 if (error != 0) { 4068 return (error); 4069 } 4070 if (flushparent) { 4071 if ((error = ffs_update(pvp, 1)) != 0) { 4072 vput(pvp); 4073 return (error); 4074 } 4075 } 4076 /* 4077 * Flush directory page containing the inode's name. 4078 */ 4079 error = bread(pvp, lblktodoff(fs, lbn), blksize(fs, VTOI(pvp), lbn), &bp); 4080 if (error == 0) 4081 error = bwrite(bp); 4082 vput(pvp); 4083 if (error != 0) { 4084 return (error); 4085 } 4086 ACQUIRE_LOCK(&lk); 4087 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) 4088 break; 4089 } 4090 FREE_LOCK(&lk); 4091 return (0); 4092 } 4093 4094 /* 4095 * Flush all the dirty bitmaps associated with the block device 4096 * before flushing the rest of the dirty blocks so as to reduce 4097 * the number of dependencies that will have to be rolled back. 4098 */ 4099 static int softdep_fsync_mountdev_bp(struct buf *bp, void *data); 4100 4101 void 4102 softdep_fsync_mountdev(struct vnode *vp) 4103 { 4104 if (!vn_isdisk(vp, NULL)) 4105 panic("softdep_fsync_mountdev: vnode not a disk"); 4106 ACQUIRE_LOCK(&lk); 4107 lwkt_gettoken(&vp->v_token); 4108 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 4109 softdep_fsync_mountdev_bp, vp); 4110 lwkt_reltoken(&vp->v_token); 4111 drain_output(vp, 1); 4112 FREE_LOCK(&lk); 4113 } 4114 4115 static int 4116 softdep_fsync_mountdev_bp(struct buf *bp, void *data) 4117 { 4118 struct worklist *wk; 4119 struct vnode *vp = data; 4120 4121 /* 4122 * If it is already scheduled, skip to the next buffer. 4123 */ 4124 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) 4125 return(0); 4126 if (bp->b_vp != vp || (bp->b_flags & B_DELWRI) == 0) { 4127 BUF_UNLOCK(bp); 4128 kprintf("softdep_fsync_mountdev_bp: warning, buffer %p ripped out from under vnode %p\n", bp, vp); 4129 return(0); 4130 } 4131 /* 4132 * We are only interested in bitmaps with outstanding 4133 * dependencies. 4134 */ 4135 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 4136 wk->wk_type != D_BMSAFEMAP) { 4137 BUF_UNLOCK(bp); 4138 return(0); 4139 } 4140 bremfree(bp); 4141 FREE_LOCK(&lk); 4142 (void) bawrite(bp); 4143 ACQUIRE_LOCK(&lk); 4144 return(0); 4145 } 4146 4147 /* 4148 * This routine is called when we are trying to synchronously flush a 4149 * file. This routine must eliminate any filesystem metadata dependencies 4150 * so that the syncing routine can succeed by pushing the dirty blocks 4151 * associated with the file. If any I/O errors occur, they are returned. 4152 */ 4153 struct softdep_sync_metadata_info { 4154 struct vnode *vp; 4155 int waitfor; 4156 }; 4157 4158 static int softdep_sync_metadata_bp(struct buf *bp, void *data); 4159 4160 int 4161 softdep_sync_metadata(struct vnode *vp, struct thread *td) 4162 { 4163 struct softdep_sync_metadata_info info; 4164 int error, waitfor; 4165 4166 /* 4167 * Check whether this vnode is involved in a filesystem 4168 * that is doing soft dependency processing. 4169 */ 4170 if (!vn_isdisk(vp, NULL)) { 4171 if (!DOINGSOFTDEP(vp)) 4172 return (0); 4173 } else 4174 if (vp->v_rdev->si_mountpoint == NULL || 4175 (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0) 4176 return (0); 4177 /* 4178 * Ensure that any direct block dependencies have been cleared. 4179 */ 4180 ACQUIRE_LOCK(&lk); 4181 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) { 4182 FREE_LOCK(&lk); 4183 return (error); 4184 } 4185 /* 4186 * For most files, the only metadata dependencies are the 4187 * cylinder group maps that allocate their inode or blocks. 4188 * The block allocation dependencies can be found by traversing 4189 * the dependency lists for any buffers that remain on their 4190 * dirty buffer list. The inode allocation dependency will 4191 * be resolved when the inode is updated with MNT_WAIT. 4192 * This work is done in two passes. The first pass grabs most 4193 * of the buffers and begins asynchronously writing them. The 4194 * only way to wait for these asynchronous writes is to sleep 4195 * on the filesystem vnode which may stay busy for a long time 4196 * if the filesystem is active. So, instead, we make a second 4197 * pass over the dependencies blocking on each write. In the 4198 * usual case we will be blocking against a write that we 4199 * initiated, so when it is done the dependency will have been 4200 * resolved. Thus the second pass is expected to end quickly. 4201 */ 4202 waitfor = MNT_NOWAIT; 4203 top: 4204 /* 4205 * We must wait for any I/O in progress to finish so that 4206 * all potential buffers on the dirty list will be visible. 4207 */ 4208 drain_output(vp, 1); 4209 4210 info.vp = vp; 4211 info.waitfor = waitfor; 4212 lwkt_gettoken(&vp->v_token); 4213 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 4214 softdep_sync_metadata_bp, &info); 4215 lwkt_reltoken(&vp->v_token); 4216 if (error < 0) { 4217 FREE_LOCK(&lk); 4218 return(-error); /* error code */ 4219 } 4220 4221 /* 4222 * The brief unlock is to allow any pent up dependency 4223 * processing to be done. Then proceed with the second pass. 4224 */ 4225 if (waitfor & MNT_NOWAIT) { 4226 waitfor = MNT_WAIT; 4227 FREE_LOCK(&lk); 4228 ACQUIRE_LOCK(&lk); 4229 goto top; 4230 } 4231 4232 /* 4233 * If we have managed to get rid of all the dirty buffers, 4234 * then we are done. For certain directories and block 4235 * devices, we may need to do further work. 4236 * 4237 * We must wait for any I/O in progress to finish so that 4238 * all potential buffers on the dirty list will be visible. 4239 */ 4240 drain_output(vp, 1); 4241 if (RB_EMPTY(&vp->v_rbdirty_tree)) { 4242 FREE_LOCK(&lk); 4243 return (0); 4244 } 4245 4246 FREE_LOCK(&lk); 4247 /* 4248 * If we are trying to sync a block device, some of its buffers may 4249 * contain metadata that cannot be written until the contents of some 4250 * partially written files have been written to disk. The only easy 4251 * way to accomplish this is to sync the entire filesystem (luckily 4252 * this happens rarely). 4253 */ 4254 if (vn_isdisk(vp, NULL) && 4255 vp->v_rdev && 4256 vp->v_rdev->si_mountpoint && !vn_islocked(vp) && 4257 (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT)) != 0) 4258 return (error); 4259 return (0); 4260 } 4261 4262 static int 4263 softdep_sync_metadata_bp(struct buf *bp, void *data) 4264 { 4265 struct softdep_sync_metadata_info *info = data; 4266 struct pagedep *pagedep; 4267 struct allocdirect *adp; 4268 struct allocindir *aip; 4269 struct worklist *wk; 4270 struct buf *nbp; 4271 int error; 4272 int i; 4273 4274 if (getdirtybuf(&bp, MNT_WAIT) == 0) { 4275 kprintf("softdep_sync_metadata_bp(1): caught buf %p going away\n", bp); 4276 return (1); 4277 } 4278 if (bp->b_vp != info->vp || (bp->b_flags & B_DELWRI) == 0) { 4279 kprintf("softdep_sync_metadata_bp(2): caught buf %p going away vp %p\n", bp, info->vp); 4280 BUF_UNLOCK(bp); 4281 return(1); 4282 } 4283 4284 /* 4285 * As we hold the buffer locked, none of its dependencies 4286 * will disappear. 4287 */ 4288 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 4289 switch (wk->wk_type) { 4290 4291 case D_ALLOCDIRECT: 4292 adp = WK_ALLOCDIRECT(wk); 4293 if (adp->ad_state & DEPCOMPLETE) 4294 break; 4295 nbp = adp->ad_buf; 4296 if (getdirtybuf(&nbp, info->waitfor) == 0) 4297 break; 4298 FREE_LOCK(&lk); 4299 if (info->waitfor & MNT_NOWAIT) { 4300 bawrite(nbp); 4301 } else if ((error = bwrite(nbp)) != 0) { 4302 bawrite(bp); 4303 ACQUIRE_LOCK(&lk); 4304 return (-error); 4305 } 4306 ACQUIRE_LOCK(&lk); 4307 break; 4308 4309 case D_ALLOCINDIR: 4310 aip = WK_ALLOCINDIR(wk); 4311 if (aip->ai_state & DEPCOMPLETE) 4312 break; 4313 nbp = aip->ai_buf; 4314 if (getdirtybuf(&nbp, info->waitfor) == 0) 4315 break; 4316 FREE_LOCK(&lk); 4317 if (info->waitfor & MNT_NOWAIT) { 4318 bawrite(nbp); 4319 } else if ((error = bwrite(nbp)) != 0) { 4320 bawrite(bp); 4321 ACQUIRE_LOCK(&lk); 4322 return (-error); 4323 } 4324 ACQUIRE_LOCK(&lk); 4325 break; 4326 4327 case D_INDIRDEP: 4328 restart: 4329 4330 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) { 4331 if (aip->ai_state & DEPCOMPLETE) 4332 continue; 4333 nbp = aip->ai_buf; 4334 if (getdirtybuf(&nbp, MNT_WAIT) == 0) 4335 goto restart; 4336 FREE_LOCK(&lk); 4337 if ((error = bwrite(nbp)) != 0) { 4338 bawrite(bp); 4339 ACQUIRE_LOCK(&lk); 4340 return (-error); 4341 } 4342 ACQUIRE_LOCK(&lk); 4343 goto restart; 4344 } 4345 break; 4346 4347 case D_INODEDEP: 4348 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs, 4349 WK_INODEDEP(wk)->id_ino)) != 0) { 4350 FREE_LOCK(&lk); 4351 bawrite(bp); 4352 ACQUIRE_LOCK(&lk); 4353 return (-error); 4354 } 4355 break; 4356 4357 case D_PAGEDEP: 4358 /* 4359 * We are trying to sync a directory that may 4360 * have dependencies on both its own metadata 4361 * and/or dependencies on the inodes of any 4362 * recently allocated files. We walk its diradd 4363 * lists pushing out the associated inode. 4364 */ 4365 pagedep = WK_PAGEDEP(wk); 4366 for (i = 0; i < DAHASHSZ; i++) { 4367 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == NULL) 4368 continue; 4369 if ((error = 4370 flush_pagedep_deps(info->vp, 4371 pagedep->pd_mnt, 4372 &pagedep->pd_diraddhd[i]))) { 4373 FREE_LOCK(&lk); 4374 bawrite(bp); 4375 ACQUIRE_LOCK(&lk); 4376 return (-error); 4377 } 4378 } 4379 break; 4380 4381 case D_MKDIR: 4382 /* 4383 * This case should never happen if the vnode has 4384 * been properly sync'ed. However, if this function 4385 * is used at a place where the vnode has not yet 4386 * been sync'ed, this dependency can show up. So, 4387 * rather than panic, just flush it. 4388 */ 4389 nbp = WK_MKDIR(wk)->md_buf; 4390 if (getdirtybuf(&nbp, info->waitfor) == 0) 4391 break; 4392 FREE_LOCK(&lk); 4393 if (info->waitfor & MNT_NOWAIT) { 4394 bawrite(nbp); 4395 } else if ((error = bwrite(nbp)) != 0) { 4396 bawrite(bp); 4397 ACQUIRE_LOCK(&lk); 4398 return (-error); 4399 } 4400 ACQUIRE_LOCK(&lk); 4401 break; 4402 4403 case D_BMSAFEMAP: 4404 /* 4405 * This case should never happen if the vnode has 4406 * been properly sync'ed. However, if this function 4407 * is used at a place where the vnode has not yet 4408 * been sync'ed, this dependency can show up. So, 4409 * rather than panic, just flush it. 4410 * 4411 * nbp can wind up == bp if a device node for the 4412 * same filesystem is being fsynced at the same time, 4413 * leading to a panic if we don't catch the case. 4414 */ 4415 nbp = WK_BMSAFEMAP(wk)->sm_buf; 4416 if (nbp == bp) 4417 break; 4418 if (getdirtybuf(&nbp, info->waitfor) == 0) 4419 break; 4420 FREE_LOCK(&lk); 4421 if (info->waitfor & MNT_NOWAIT) { 4422 bawrite(nbp); 4423 } else if ((error = bwrite(nbp)) != 0) { 4424 bawrite(bp); 4425 ACQUIRE_LOCK(&lk); 4426 return (-error); 4427 } 4428 ACQUIRE_LOCK(&lk); 4429 break; 4430 4431 default: 4432 panic("softdep_sync_metadata: Unknown type %s", 4433 TYPENAME(wk->wk_type)); 4434 /* NOTREACHED */ 4435 } 4436 } 4437 FREE_LOCK(&lk); 4438 bawrite(bp); 4439 ACQUIRE_LOCK(&lk); 4440 return(0); 4441 } 4442 4443 /* 4444 * Flush the dependencies associated with an inodedep. 4445 * Called with splbio blocked. 4446 */ 4447 static int 4448 flush_inodedep_deps(struct fs *fs, ino_t ino) 4449 { 4450 struct inodedep *inodedep; 4451 struct allocdirect *adp; 4452 int error, waitfor; 4453 struct buf *bp; 4454 4455 /* 4456 * This work is done in two passes. The first pass grabs most 4457 * of the buffers and begins asynchronously writing them. The 4458 * only way to wait for these asynchronous writes is to sleep 4459 * on the filesystem vnode which may stay busy for a long time 4460 * if the filesystem is active. So, instead, we make a second 4461 * pass over the dependencies blocking on each write. In the 4462 * usual case we will be blocking against a write that we 4463 * initiated, so when it is done the dependency will have been 4464 * resolved. Thus the second pass is expected to end quickly. 4465 * We give a brief window at the top of the loop to allow 4466 * any pending I/O to complete. 4467 */ 4468 for (waitfor = MNT_NOWAIT; ; ) { 4469 FREE_LOCK(&lk); 4470 ACQUIRE_LOCK(&lk); 4471 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4472 return (0); 4473 TAILQ_FOREACH(adp, &inodedep->id_inoupdt, ad_next) { 4474 if (adp->ad_state & DEPCOMPLETE) 4475 continue; 4476 bp = adp->ad_buf; 4477 if (getdirtybuf(&bp, waitfor) == 0) { 4478 if (waitfor & MNT_NOWAIT) 4479 continue; 4480 break; 4481 } 4482 FREE_LOCK(&lk); 4483 if (waitfor & MNT_NOWAIT) { 4484 bawrite(bp); 4485 } else if ((error = bwrite(bp)) != 0) { 4486 ACQUIRE_LOCK(&lk); 4487 return (error); 4488 } 4489 ACQUIRE_LOCK(&lk); 4490 break; 4491 } 4492 if (adp != NULL) 4493 continue; 4494 TAILQ_FOREACH(adp, &inodedep->id_newinoupdt, ad_next) { 4495 if (adp->ad_state & DEPCOMPLETE) 4496 continue; 4497 bp = adp->ad_buf; 4498 if (getdirtybuf(&bp, waitfor) == 0) { 4499 if (waitfor & MNT_NOWAIT) 4500 continue; 4501 break; 4502 } 4503 FREE_LOCK(&lk); 4504 if (waitfor & MNT_NOWAIT) { 4505 bawrite(bp); 4506 } else if ((error = bwrite(bp)) != 0) { 4507 ACQUIRE_LOCK(&lk); 4508 return (error); 4509 } 4510 ACQUIRE_LOCK(&lk); 4511 break; 4512 } 4513 if (adp != NULL) 4514 continue; 4515 /* 4516 * If pass2, we are done, otherwise do pass 2. 4517 */ 4518 if (waitfor == MNT_WAIT) 4519 break; 4520 waitfor = MNT_WAIT; 4521 } 4522 /* 4523 * Try freeing inodedep in case all dependencies have been removed. 4524 */ 4525 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0) 4526 (void) free_inodedep(inodedep); 4527 return (0); 4528 } 4529 4530 /* 4531 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 4532 * Called with splbio blocked. 4533 */ 4534 static int 4535 flush_pagedep_deps(struct vnode *pvp, struct mount *mp, 4536 struct diraddhd *diraddhdp) 4537 { 4538 struct inodedep *inodedep; 4539 struct ufsmount *ump; 4540 struct diradd *dap; 4541 struct worklist *wk; 4542 struct vnode *vp; 4543 int gotit, error = 0; 4544 struct buf *bp; 4545 ino_t inum; 4546 4547 ump = VFSTOUFS(mp); 4548 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 4549 /* 4550 * Flush ourselves if this directory entry 4551 * has a MKDIR_PARENT dependency. 4552 */ 4553 if (dap->da_state & MKDIR_PARENT) { 4554 FREE_LOCK(&lk); 4555 if ((error = ffs_update(pvp, 1)) != 0) 4556 break; 4557 ACQUIRE_LOCK(&lk); 4558 /* 4559 * If that cleared dependencies, go on to next. 4560 */ 4561 if (dap != LIST_FIRST(diraddhdp)) 4562 continue; 4563 if (dap->da_state & MKDIR_PARENT) { 4564 panic("flush_pagedep_deps: MKDIR_PARENT"); 4565 } 4566 } 4567 /* 4568 * A newly allocated directory must have its "." and 4569 * ".." entries written out before its name can be 4570 * committed in its parent. We do not want or need 4571 * the full semantics of a synchronous VOP_FSYNC as 4572 * that may end up here again, once for each directory 4573 * level in the filesystem. Instead, we push the blocks 4574 * and wait for them to clear. We have to fsync twice 4575 * because the first call may choose to defer blocks 4576 * that still have dependencies, but deferral will 4577 * happen at most once. 4578 */ 4579 inum = dap->da_newinum; 4580 if (dap->da_state & MKDIR_BODY) { 4581 FREE_LOCK(&lk); 4582 if ((error = VFS_VGET(mp, NULL, inum, &vp)) != 0) 4583 break; 4584 if ((error=VOP_FSYNC(vp, MNT_NOWAIT, 0)) || 4585 (error=VOP_FSYNC(vp, MNT_NOWAIT, 0))) { 4586 vput(vp); 4587 break; 4588 } 4589 drain_output(vp, 0); 4590 /* 4591 * If first block is still dirty with a D_MKDIR 4592 * dependency then it needs to be written now. 4593 */ 4594 error = 0; 4595 ACQUIRE_LOCK(&lk); 4596 bp = findblk(vp, 0, FINDBLK_TEST); 4597 if (bp == NULL) { 4598 FREE_LOCK(&lk); 4599 goto mkdir_body_continue; 4600 } 4601 LIST_FOREACH(wk, &bp->b_dep, wk_list) 4602 if (wk->wk_type == D_MKDIR) { 4603 gotit = getdirtybuf(&bp, MNT_WAIT); 4604 FREE_LOCK(&lk); 4605 if (gotit && (error = bwrite(bp)) != 0) 4606 goto mkdir_body_continue; 4607 break; 4608 } 4609 if (wk == NULL) 4610 FREE_LOCK(&lk); 4611 mkdir_body_continue: 4612 vput(vp); 4613 /* Flushing of first block failed. */ 4614 if (error) 4615 break; 4616 ACQUIRE_LOCK(&lk); 4617 /* 4618 * If that cleared dependencies, go on to next. 4619 */ 4620 if (dap != LIST_FIRST(diraddhdp)) 4621 continue; 4622 if (dap->da_state & MKDIR_BODY) { 4623 panic("flush_pagedep_deps: %p MKDIR_BODY", dap); 4624 } 4625 } 4626 /* 4627 * Flush the inode on which the directory entry depends. 4628 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 4629 * the only remaining dependency is that the updated inode 4630 * count must get pushed to disk. The inode has already 4631 * been pushed into its inode buffer (via VOP_UPDATE) at 4632 * the time of the reference count change. So we need only 4633 * locate that buffer, ensure that there will be no rollback 4634 * caused by a bitmap dependency, then write the inode buffer. 4635 */ 4636 retry_lookup: 4637 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) { 4638 panic("flush_pagedep_deps: lost inode"); 4639 } 4640 /* 4641 * If the inode still has bitmap dependencies, 4642 * push them to disk. 4643 */ 4644 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4645 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 4646 if (gotit == 0) 4647 goto retry_lookup; 4648 FREE_LOCK(&lk); 4649 if (gotit && (error = bwrite(inodedep->id_buf)) != 0) 4650 break; 4651 ACQUIRE_LOCK(&lk); 4652 if (dap != LIST_FIRST(diraddhdp)) 4653 continue; 4654 } 4655 /* 4656 * If the inode is still sitting in a buffer waiting 4657 * to be written, push it to disk. 4658 */ 4659 FREE_LOCK(&lk); 4660 if ((error = bread(ump->um_devvp, 4661 fsbtodoff(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), 4662 (int)ump->um_fs->fs_bsize, &bp)) != 0) 4663 break; 4664 if ((error = bwrite(bp)) != 0) 4665 break; 4666 ACQUIRE_LOCK(&lk); 4667 /* 4668 * If we have failed to get rid of all the dependencies 4669 * then something is seriously wrong. 4670 */ 4671 if (dap == LIST_FIRST(diraddhdp)) { 4672 panic("flush_pagedep_deps: flush failed"); 4673 } 4674 } 4675 if (error) 4676 ACQUIRE_LOCK(&lk); 4677 return (error); 4678 } 4679 4680 /* 4681 * A large burst of file addition or deletion activity can drive the 4682 * memory load excessively high. First attempt to slow things down 4683 * using the techniques below. If that fails, this routine requests 4684 * the offending operations to fall back to running synchronously 4685 * until the memory load returns to a reasonable level. 4686 */ 4687 int 4688 softdep_slowdown(struct vnode *vp) 4689 { 4690 int max_softdeps_hard; 4691 4692 max_softdeps_hard = max_softdeps * 11 / 10; 4693 if (num_dirrem < max_softdeps_hard / 2 && 4694 num_inodedep < max_softdeps_hard) 4695 return (0); 4696 stat_sync_limit_hit += 1; 4697 return (1); 4698 } 4699 4700 /* 4701 * If memory utilization has gotten too high, deliberately slow things 4702 * down and speed up the I/O processing. 4703 */ 4704 static int 4705 request_cleanup(int resource) 4706 { 4707 struct thread *td = curthread; /* XXX */ 4708 4709 KKASSERT(lock_held(&lk) > 0); 4710 4711 /* 4712 * We never hold up the filesystem syncer process. 4713 */ 4714 if (td == filesys_syncer) 4715 return (0); 4716 /* 4717 * First check to see if the work list has gotten backlogged. 4718 * If it has, co-opt this process to help clean up two entries. 4719 * Because this process may hold inodes locked, we cannot 4720 * handle any remove requests that might block on a locked 4721 * inode as that could lead to deadlock. 4722 */ 4723 if (num_on_worklist > max_softdeps / 10) { 4724 process_worklist_item(NULL, LK_NOWAIT); 4725 process_worklist_item(NULL, LK_NOWAIT); 4726 stat_worklist_push += 2; 4727 return(1); 4728 } 4729 4730 /* 4731 * If we are resource constrained on inode dependencies, try 4732 * flushing some dirty inodes. Otherwise, we are constrained 4733 * by file deletions, so try accelerating flushes of directories 4734 * with removal dependencies. We would like to do the cleanup 4735 * here, but we probably hold an inode locked at this point and 4736 * that might deadlock against one that we try to clean. So, 4737 * the best that we can do is request the syncer daemon to do 4738 * the cleanup for us. 4739 */ 4740 switch (resource) { 4741 4742 case FLUSH_INODES: 4743 stat_ino_limit_push += 1; 4744 req_clear_inodedeps += 1; 4745 stat_countp = &stat_ino_limit_hit; 4746 break; 4747 4748 case FLUSH_REMOVE: 4749 stat_blk_limit_push += 1; 4750 req_clear_remove += 1; 4751 stat_countp = &stat_blk_limit_hit; 4752 break; 4753 4754 default: 4755 panic("request_cleanup: unknown type"); 4756 } 4757 /* 4758 * Hopefully the syncer daemon will catch up and awaken us. 4759 * We wait at most tickdelay before proceeding in any case. 4760 */ 4761 lksleep(&proc_waiting, &lk, 0, "softupdate", 4762 tickdelay > 2 ? tickdelay : 2); 4763 return (1); 4764 } 4765 4766 /* 4767 * Flush out a directory with at least one removal dependency in an effort to 4768 * reduce the number of dirrem, freefile, and freeblks dependency structures. 4769 */ 4770 static void 4771 clear_remove(struct thread *td) 4772 { 4773 struct pagedep_hashhead *pagedephd; 4774 struct pagedep *pagedep; 4775 static int next = 0; 4776 struct mount *mp; 4777 struct vnode *vp; 4778 int error, cnt; 4779 ino_t ino; 4780 4781 ACQUIRE_LOCK(&lk); 4782 for (cnt = 0; cnt < pagedep_hash; cnt++) { 4783 pagedephd = &pagedep_hashtbl[next++]; 4784 if (next >= pagedep_hash) 4785 next = 0; 4786 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 4787 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL) 4788 continue; 4789 mp = pagedep->pd_mnt; 4790 ino = pagedep->pd_ino; 4791 FREE_LOCK(&lk); 4792 if ((error = VFS_VGET(mp, NULL, ino, &vp)) != 0) { 4793 softdep_error("clear_remove: vget", error); 4794 return; 4795 } 4796 if ((error = VOP_FSYNC(vp, MNT_NOWAIT, 0))) 4797 softdep_error("clear_remove: fsync", error); 4798 drain_output(vp, 0); 4799 vput(vp); 4800 return; 4801 } 4802 } 4803 FREE_LOCK(&lk); 4804 } 4805 4806 /* 4807 * Clear out a block of dirty inodes in an effort to reduce 4808 * the number of inodedep dependency structures. 4809 */ 4810 struct clear_inodedeps_info { 4811 struct fs *fs; 4812 struct mount *mp; 4813 }; 4814 4815 static int 4816 clear_inodedeps_mountlist_callback(struct mount *mp, void *data) 4817 { 4818 struct clear_inodedeps_info *info = data; 4819 4820 if ((mp->mnt_flag & MNT_SOFTDEP) && info->fs == VFSTOUFS(mp)->um_fs) { 4821 info->mp = mp; 4822 return(-1); 4823 } 4824 return(0); 4825 } 4826 4827 static void 4828 clear_inodedeps(struct thread *td) 4829 { 4830 struct clear_inodedeps_info info; 4831 struct inodedep_hashhead *inodedephd; 4832 struct inodedep *inodedep; 4833 static int next = 0; 4834 struct vnode *vp; 4835 struct fs *fs; 4836 int error, cnt; 4837 ino_t firstino, lastino, ino; 4838 4839 ACQUIRE_LOCK(&lk); 4840 /* 4841 * Pick a random inode dependency to be cleared. 4842 * We will then gather up all the inodes in its block 4843 * that have dependencies and flush them out. 4844 */ 4845 for (cnt = 0; cnt < inodedep_hash; cnt++) { 4846 inodedephd = &inodedep_hashtbl[next++]; 4847 if (next >= inodedep_hash) 4848 next = 0; 4849 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 4850 break; 4851 } 4852 if (inodedep == NULL) { 4853 FREE_LOCK(&lk); 4854 return; 4855 } 4856 /* 4857 * Ugly code to find mount point given pointer to superblock. 4858 */ 4859 fs = inodedep->id_fs; 4860 info.mp = NULL; 4861 info.fs = fs; 4862 mountlist_scan(clear_inodedeps_mountlist_callback, 4863 &info, MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 4864 /* 4865 * Find the last inode in the block with dependencies. 4866 */ 4867 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 4868 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 4869 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0) 4870 break; 4871 /* 4872 * Asynchronously push all but the last inode with dependencies. 4873 * Synchronously push the last inode with dependencies to ensure 4874 * that the inode block gets written to free up the inodedeps. 4875 */ 4876 for (ino = firstino; ino <= lastino; ino++) { 4877 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4878 continue; 4879 FREE_LOCK(&lk); 4880 if ((error = VFS_VGET(info.mp, NULL, ino, &vp)) != 0) { 4881 softdep_error("clear_inodedeps: vget", error); 4882 return; 4883 } 4884 if (ino == lastino) { 4885 if ((error = VOP_FSYNC(vp, MNT_WAIT, 0))) 4886 softdep_error("clear_inodedeps: fsync1", error); 4887 } else { 4888 if ((error = VOP_FSYNC(vp, MNT_NOWAIT, 0))) 4889 softdep_error("clear_inodedeps: fsync2", error); 4890 drain_output(vp, 0); 4891 } 4892 vput(vp); 4893 ACQUIRE_LOCK(&lk); 4894 } 4895 FREE_LOCK(&lk); 4896 } 4897 4898 /* 4899 * Function to determine if the buffer has outstanding dependencies 4900 * that will cause a roll-back if the buffer is written. If wantcount 4901 * is set, return number of dependencies, otherwise just yes or no. 4902 * 4903 * bioops callback - hold io_token 4904 */ 4905 static int 4906 softdep_count_dependencies(struct buf *bp, int wantcount) 4907 { 4908 struct worklist *wk; 4909 struct inodedep *inodedep; 4910 struct indirdep *indirdep; 4911 struct allocindir *aip; 4912 struct pagedep *pagedep; 4913 struct diradd *dap; 4914 int i, retval; 4915 4916 retval = 0; 4917 ACQUIRE_LOCK(&lk); 4918 4919 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 4920 switch (wk->wk_type) { 4921 4922 case D_INODEDEP: 4923 inodedep = WK_INODEDEP(wk); 4924 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4925 /* bitmap allocation dependency */ 4926 retval += 1; 4927 if (!wantcount) 4928 goto out; 4929 } 4930 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 4931 /* direct block pointer dependency */ 4932 retval += 1; 4933 if (!wantcount) 4934 goto out; 4935 } 4936 continue; 4937 4938 case D_INDIRDEP: 4939 indirdep = WK_INDIRDEP(wk); 4940 4941 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 4942 /* indirect block pointer dependency */ 4943 retval += 1; 4944 if (!wantcount) 4945 goto out; 4946 } 4947 continue; 4948 4949 case D_PAGEDEP: 4950 pagedep = WK_PAGEDEP(wk); 4951 for (i = 0; i < DAHASHSZ; i++) { 4952 4953 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 4954 /* directory entry dependency */ 4955 retval += 1; 4956 if (!wantcount) 4957 goto out; 4958 } 4959 } 4960 continue; 4961 4962 case D_BMSAFEMAP: 4963 case D_ALLOCDIRECT: 4964 case D_ALLOCINDIR: 4965 case D_MKDIR: 4966 /* never a dependency on these blocks */ 4967 continue; 4968 4969 default: 4970 panic("softdep_check_for_rollback: Unexpected type %s", 4971 TYPENAME(wk->wk_type)); 4972 /* NOTREACHED */ 4973 } 4974 } 4975 out: 4976 FREE_LOCK(&lk); 4977 4978 return retval; 4979 } 4980 4981 /* 4982 * Acquire exclusive access to a buffer. Requires softdep lock 4983 * to be held on entry. If waitfor is MNT_WAIT, may release/reacquire 4984 * softdep lock. 4985 * 4986 * Returns 1 if the buffer was locked, 0 if it was not locked or 4987 * if we had to block. 4988 * 4989 * NOTE! In order to return 1 we must acquire the buffer lock prior 4990 * to any release of &lk. Once we release &lk it's all over. 4991 * We may still have to block on the (type-stable) bp in that 4992 * case, but we must then unlock it and return 0. 4993 */ 4994 static int 4995 getdirtybuf(struct buf **bpp, int waitfor) 4996 { 4997 struct buf *bp; 4998 int error; 4999 5000 /* 5001 * If the contents of *bpp is NULL the caller presumably lost a race. 5002 */ 5003 bp = *bpp; 5004 if (bp == NULL) 5005 return (0); 5006 5007 /* 5008 * Try to obtain the buffer lock without deadlocking on &lk. 5009 */ 5010 KKASSERT(lock_held(&lk) > 0); 5011 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT); 5012 if (error == 0) { 5013 /* 5014 * If the buffer is no longer dirty the OS already wrote it 5015 * out, return failure. 5016 */ 5017 if ((bp->b_flags & B_DELWRI) == 0) { 5018 BUF_UNLOCK(bp); 5019 return (0); 5020 } 5021 5022 /* 5023 * Finish nominal buffer locking sequence return success. 5024 */ 5025 bremfree(bp); 5026 return (1); 5027 } 5028 5029 /* 5030 * Failure case. 5031 * 5032 * If we are not being asked to wait, return 0 immediately. 5033 */ 5034 if (waitfor != MNT_WAIT) 5035 return (0); 5036 5037 /* 5038 * Once we release the softdep lock we can never return success, 5039 * but we still have to block on the type-stable buf for the caller 5040 * to be able to retry without livelocking the system. 5041 * 5042 * The caller will normally retry in this case. 5043 */ 5044 FREE_LOCK(&lk); 5045 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL); 5046 ACQUIRE_LOCK(&lk); 5047 if (error == 0) 5048 BUF_UNLOCK(bp); 5049 return (0); 5050 } 5051 5052 /* 5053 * Wait for pending output on a vnode to complete. 5054 * Must be called with vnode locked. 5055 */ 5056 static void 5057 drain_output(struct vnode *vp, int islocked) 5058 { 5059 5060 if (!islocked) 5061 ACQUIRE_LOCK(&lk); 5062 while (bio_track_active(&vp->v_track_write)) { 5063 FREE_LOCK(&lk); 5064 bio_track_wait(&vp->v_track_write, 0, 0); 5065 ACQUIRE_LOCK(&lk); 5066 } 5067 if (!islocked) 5068 FREE_LOCK(&lk); 5069 } 5070 5071 /* 5072 * Called whenever a buffer that is being invalidated or reallocated 5073 * contains dependencies. This should only happen if an I/O error has 5074 * occurred. The routine is called with the buffer locked. 5075 * 5076 * bioops callback - hold io_token 5077 */ 5078 static void 5079 softdep_deallocate_dependencies(struct buf *bp) 5080 { 5081 /* nothing to do, mp lock not needed */ 5082 if ((bp->b_flags & B_ERROR) == 0) 5083 panic("softdep_deallocate_dependencies: dangling deps"); 5084 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntfromname, bp->b_error); 5085 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 5086 } 5087 5088 /* 5089 * Function to handle asynchronous write errors in the filesystem. 5090 */ 5091 void 5092 softdep_error(char *func, int error) 5093 { 5094 /* XXX should do something better! */ 5095 kprintf("%s: got error %d while accessing filesystem\n", func, error); 5096 } 5097