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 * $DragonFly: src/sys/vfs/ufs/ffs_softdep.c,v 1.26 2005/07/26 20:53:58 dillon Exp $ 41 */ 42 43 /* 44 * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide. 45 */ 46 #ifndef DIAGNOSTIC 47 #define DIAGNOSTIC 48 #endif 49 #ifndef DEBUG 50 #define DEBUG 51 #endif 52 53 #include <sys/param.h> 54 #include <sys/kernel.h> 55 #include <sys/systm.h> 56 #include <sys/buf.h> 57 #include <sys/malloc.h> 58 #include <sys/mount.h> 59 #include <sys/proc.h> 60 #include <sys/syslog.h> 61 #include <sys/vnode.h> 62 #include <sys/conf.h> 63 #include <sys/buf2.h> 64 #include "dir.h" 65 #include "quota.h" 66 #include "inode.h" 67 #include "ufsmount.h" 68 #include "fs.h" 69 #include "softdep.h" 70 #include "ffs_extern.h" 71 #include "ufs_extern.h" 72 73 #include <sys/thread2.h> 74 75 /* 76 * These definitions need to be adapted to the system to which 77 * this file is being ported. 78 */ 79 /* 80 * malloc types defined for the softdep system. 81 */ 82 MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies"); 83 MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies"); 84 MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation"); 85 MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map"); 86 MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode"); 87 MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies"); 88 MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block"); 89 MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode"); 90 MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode"); 91 MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated"); 92 MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry"); 93 MALLOC_DEFINE(M_MKDIR, "mkdir","New directory"); 94 MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted"); 95 96 #define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE) 97 98 #define D_PAGEDEP 0 99 #define D_INODEDEP 1 100 #define D_NEWBLK 2 101 #define D_BMSAFEMAP 3 102 #define D_ALLOCDIRECT 4 103 #define D_INDIRDEP 5 104 #define D_ALLOCINDIR 6 105 #define D_FREEFRAG 7 106 #define D_FREEBLKS 8 107 #define D_FREEFILE 9 108 #define D_DIRADD 10 109 #define D_MKDIR 11 110 #define D_DIRREM 12 111 #define D_LAST D_DIRREM 112 113 /* 114 * translate from workitem type to memory type 115 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 116 */ 117 static struct malloc_type *memtype[] = { 118 M_PAGEDEP, 119 M_INODEDEP, 120 M_NEWBLK, 121 M_BMSAFEMAP, 122 M_ALLOCDIRECT, 123 M_INDIRDEP, 124 M_ALLOCINDIR, 125 M_FREEFRAG, 126 M_FREEBLKS, 127 M_FREEFILE, 128 M_DIRADD, 129 M_MKDIR, 130 M_DIRREM 131 }; 132 133 #define DtoM(type) (memtype[type]) 134 135 /* 136 * Names of malloc types. 137 */ 138 #define TYPENAME(type) \ 139 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???") 140 /* 141 * End system adaptaion definitions. 142 */ 143 144 /* 145 * Internal function prototypes. 146 */ 147 static void softdep_error(char *, int); 148 static void drain_output(struct vnode *, int); 149 static int getdirtybuf(struct buf **, int); 150 static void clear_remove(struct thread *); 151 static void clear_inodedeps(struct thread *); 152 static int flush_pagedep_deps(struct vnode *, struct mount *, 153 struct diraddhd *); 154 static int flush_inodedep_deps(struct fs *, ino_t); 155 static int handle_written_filepage(struct pagedep *, struct buf *); 156 static void diradd_inode_written(struct diradd *, struct inodedep *); 157 static int handle_written_inodeblock(struct inodedep *, struct buf *); 158 static void handle_allocdirect_partdone(struct allocdirect *); 159 static void handle_allocindir_partdone(struct allocindir *); 160 static void initiate_write_filepage(struct pagedep *, struct buf *); 161 static void handle_written_mkdir(struct mkdir *, int); 162 static void initiate_write_inodeblock(struct inodedep *, struct buf *); 163 static void handle_workitem_freefile(struct freefile *); 164 static void handle_workitem_remove(struct dirrem *); 165 static struct dirrem *newdirrem(struct buf *, struct inode *, 166 struct inode *, int, struct dirrem **); 167 static void free_diradd(struct diradd *); 168 static void free_allocindir(struct allocindir *, struct inodedep *); 169 static int indir_trunc (struct inode *, ufs_daddr_t, int, ufs_lbn_t, 170 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 void pause_timer(void *); 193 static int request_cleanup(int, int); 194 static int process_worklist_item(struct mount *, int); 195 static void add_to_worklist(struct worklist *); 196 197 /* 198 * Exported softdep operations. 199 */ 200 static void softdep_disk_io_initiation(struct buf *); 201 static void softdep_disk_write_complete(struct buf *); 202 static void softdep_deallocate_dependencies(struct buf *); 203 static int softdep_fsync(struct vnode *); 204 static int softdep_process_worklist(struct mount *); 205 static void softdep_move_dependencies(struct buf *, struct buf *); 206 static int softdep_count_dependencies(struct buf *bp, int); 207 208 static struct bio_ops softdep_bioops = { 209 softdep_disk_io_initiation, /* io_start */ 210 softdep_disk_write_complete, /* io_complete */ 211 softdep_deallocate_dependencies, /* io_deallocate */ 212 softdep_fsync, /* io_fsync */ 213 softdep_process_worklist, /* io_sync */ 214 softdep_move_dependencies, /* io_movedeps */ 215 softdep_count_dependencies, /* io_countdeps */ 216 }; 217 218 /* 219 * Locking primitives. 220 * 221 * For a uniprocessor, all we need to do is protect against disk 222 * interrupts. For a multiprocessor, this lock would have to be 223 * a mutex. A single mutex is used throughout this file, though 224 * finer grain locking could be used if contention warranted it. 225 * 226 * For a multiprocessor, the sleep call would accept a lock and 227 * release it after the sleep processing was complete. In a uniprocessor 228 * implementation there is no such interlock, so we simple mark 229 * the places where it needs to be done with the `interlocked' form 230 * of the lock calls. Since the uniprocessor sleep already interlocks 231 * the spl, there is nothing that really needs to be done. 232 */ 233 #ifndef /* NOT */ DEBUG 234 static struct lockit { 235 } lk = { 0 }; 236 #define ACQUIRE_LOCK(lk) crit_enter_id("softupdates"); 237 #define FREE_LOCK(lk) crit_exit_id("softupdates"); 238 239 #else /* DEBUG */ 240 #define NOHOLDER ((struct thread *)-1) 241 #define SPECIAL_FLAG ((struct thread *)-2) 242 static struct lockit { 243 int lkt_spl; 244 struct thread *lkt_held; 245 } lk = { 0, NOHOLDER }; 246 static int lockcnt; 247 248 static void acquire_lock(struct lockit *); 249 static void free_lock(struct lockit *); 250 void softdep_panic(char *); 251 252 #define ACQUIRE_LOCK(lk) acquire_lock(lk) 253 #define FREE_LOCK(lk) free_lock(lk) 254 255 static void 256 acquire_lock(lk) 257 struct lockit *lk; 258 { 259 thread_t holder; 260 261 if (lk->lkt_held != NOHOLDER) { 262 holder = lk->lkt_held; 263 FREE_LOCK(lk); 264 if (holder == curthread) 265 panic("softdep_lock: locking against myself"); 266 else 267 panic("softdep_lock: lock held by %p", holder); 268 } 269 crit_enter_id("softupdates"); 270 lk->lkt_held = curthread; 271 lockcnt++; 272 } 273 274 static void 275 free_lock(lk) 276 struct lockit *lk; 277 { 278 279 if (lk->lkt_held == NOHOLDER) 280 panic("softdep_unlock: lock not held"); 281 lk->lkt_held = NOHOLDER; 282 crit_exit_id("softupdates"); 283 } 284 285 /* 286 * Function to release soft updates lock and panic. 287 */ 288 void 289 softdep_panic(msg) 290 char *msg; 291 { 292 293 if (lk.lkt_held != NOHOLDER) 294 FREE_LOCK(&lk); 295 panic(msg); 296 } 297 #endif /* DEBUG */ 298 299 static int interlocked_sleep(struct lockit *, int, void *, int, 300 const char *, int); 301 302 /* 303 * When going to sleep, we must save our SPL so that it does 304 * not get lost if some other process uses the lock while we 305 * are sleeping. We restore it after we have slept. This routine 306 * wraps the interlocking with functions that sleep. The list 307 * below enumerates the available set of operations. 308 */ 309 #define UNKNOWN 0 310 #define SLEEP 1 311 #define LOCKBUF 2 312 313 static int 314 interlocked_sleep(lk, op, ident, flags, wmesg, timo) 315 struct lockit *lk; 316 int op; 317 void *ident; 318 int flags; 319 const char *wmesg; 320 int timo; 321 { 322 thread_t holder; 323 int s, retval; 324 325 s = lk->lkt_spl; 326 # ifdef DEBUG 327 if (lk->lkt_held == NOHOLDER) 328 panic("interlocked_sleep: lock not held"); 329 lk->lkt_held = NOHOLDER; 330 # endif /* DEBUG */ 331 switch (op) { 332 case SLEEP: 333 retval = tsleep(ident, flags, wmesg, timo); 334 break; 335 case LOCKBUF: 336 retval = BUF_LOCK((struct buf *)ident, flags); 337 break; 338 default: 339 panic("interlocked_sleep: unknown operation"); 340 } 341 # ifdef DEBUG 342 if (lk->lkt_held != NOHOLDER) { 343 holder = lk->lkt_held; 344 FREE_LOCK(lk); 345 if (holder == curthread) 346 panic("interlocked_sleep: locking against self"); 347 else 348 panic("interlocked_sleep: lock held by %p", holder); 349 } 350 lk->lkt_held = curthread; 351 lockcnt++; 352 # endif /* DEBUG */ 353 lk->lkt_spl = s; 354 return (retval); 355 } 356 357 /* 358 * Place holder for real semaphores. 359 */ 360 struct sema { 361 int value; 362 thread_t holder; 363 char *name; 364 int prio; 365 int timo; 366 }; 367 static void sema_init(struct sema *, char *, int, int); 368 static int sema_get(struct sema *, struct lockit *); 369 static void sema_release(struct sema *); 370 371 static void 372 sema_init(semap, name, prio, timo) 373 struct sema *semap; 374 char *name; 375 int prio, timo; 376 { 377 378 semap->holder = NOHOLDER; 379 semap->value = 0; 380 semap->name = name; 381 semap->prio = prio; 382 semap->timo = timo; 383 } 384 385 static int 386 sema_get(semap, interlock) 387 struct sema *semap; 388 struct lockit *interlock; 389 { 390 391 if (semap->value++ > 0) { 392 if (interlock != NULL) { 393 interlocked_sleep(interlock, SLEEP, (caddr_t)semap, 394 semap->prio, semap->name, semap->timo); 395 FREE_LOCK(interlock); 396 } else { 397 tsleep((caddr_t)semap, semap->prio, semap->name, 398 semap->timo); 399 } 400 return (0); 401 } 402 semap->holder = curthread; 403 if (interlock != NULL) 404 FREE_LOCK(interlock); 405 return (1); 406 } 407 408 static void 409 sema_release(semap) 410 struct sema *semap; 411 { 412 413 if (semap->value <= 0 || semap->holder != curthread) { 414 if (lk.lkt_held != NOHOLDER) 415 FREE_LOCK(&lk); 416 panic("sema_release: not held"); 417 } 418 if (--semap->value > 0) { 419 semap->value = 0; 420 wakeup(semap); 421 } 422 semap->holder = NOHOLDER; 423 } 424 425 /* 426 * Worklist queue management. 427 * These routines require that the lock be held. 428 */ 429 #ifndef /* NOT */ DEBUG 430 #define WORKLIST_INSERT(head, item) do { \ 431 (item)->wk_state |= ONWORKLIST; \ 432 LIST_INSERT_HEAD(head, item, wk_list); \ 433 } while (0) 434 #define WORKLIST_REMOVE(item) do { \ 435 (item)->wk_state &= ~ONWORKLIST; \ 436 LIST_REMOVE(item, wk_list); \ 437 } while (0) 438 #define WORKITEM_FREE(item, type) FREE(item, DtoM(type)) 439 440 #else /* DEBUG */ 441 static void worklist_insert(struct workhead *, struct worklist *); 442 static void worklist_remove(struct worklist *); 443 static void workitem_free(struct worklist *, int); 444 445 #define WORKLIST_INSERT(head, item) worklist_insert(head, item) 446 #define WORKLIST_REMOVE(item) worklist_remove(item) 447 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type) 448 449 static void 450 worklist_insert(head, item) 451 struct workhead *head; 452 struct worklist *item; 453 { 454 455 if (lk.lkt_held == NOHOLDER) 456 panic("worklist_insert: lock not held"); 457 if (item->wk_state & ONWORKLIST) { 458 FREE_LOCK(&lk); 459 panic("worklist_insert: already on list"); 460 } 461 item->wk_state |= ONWORKLIST; 462 LIST_INSERT_HEAD(head, item, wk_list); 463 } 464 465 static void 466 worklist_remove(item) 467 struct worklist *item; 468 { 469 470 if (lk.lkt_held == NOHOLDER) 471 panic("worklist_remove: lock not held"); 472 if ((item->wk_state & ONWORKLIST) == 0) { 473 FREE_LOCK(&lk); 474 panic("worklist_remove: not on list"); 475 } 476 item->wk_state &= ~ONWORKLIST; 477 LIST_REMOVE(item, wk_list); 478 } 479 480 static void 481 workitem_free(item, type) 482 struct worklist *item; 483 int type; 484 { 485 486 if (item->wk_state & ONWORKLIST) { 487 if (lk.lkt_held != NOHOLDER) 488 FREE_LOCK(&lk); 489 panic("workitem_free: still on list"); 490 } 491 if (item->wk_type != type) { 492 if (lk.lkt_held != NOHOLDER) 493 FREE_LOCK(&lk); 494 panic("workitem_free: type mismatch"); 495 } 496 FREE(item, DtoM(type)); 497 } 498 #endif /* DEBUG */ 499 500 /* 501 * Workitem queue management 502 */ 503 static struct workhead softdep_workitem_pending; 504 static int num_on_worklist; /* number of worklist items to be processed */ 505 static int softdep_worklist_busy; /* 1 => trying to do unmount */ 506 static int softdep_worklist_req; /* serialized waiters */ 507 static int max_softdeps; /* maximum number of structs before slowdown */ 508 static int tickdelay = 2; /* number of ticks to pause during slowdown */ 509 static int *stat_countp; /* statistic to count in proc_waiting timeout */ 510 static int proc_waiting; /* tracks whether we have a timeout posted */ 511 static struct callout handle; /* handle on posted proc_waiting timeout */ 512 static struct thread *filesys_syncer; /* proc of filesystem syncer process */ 513 static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 514 #define FLUSH_INODES 1 515 static int req_clear_remove; /* syncer process flush some freeblks */ 516 #define FLUSH_REMOVE 2 517 /* 518 * runtime statistics 519 */ 520 static int stat_worklist_push; /* number of worklist cleanups */ 521 static int stat_blk_limit_push; /* number of times block limit neared */ 522 static int stat_ino_limit_push; /* number of times inode limit neared */ 523 static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 524 static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 525 static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 526 static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 527 static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 528 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 529 static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 530 #ifdef DEBUG 531 #include <vm/vm.h> 532 #include <sys/sysctl.h> 533 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, ""); 534 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, ""); 535 SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,""); 536 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,""); 537 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,""); 538 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, ""); 539 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, ""); 540 SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, ""); 541 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, ""); 542 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, ""); 543 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, ""); 544 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, ""); 545 #endif /* DEBUG */ 546 547 /* 548 * Add an item to the end of the work queue. 549 * This routine requires that the lock be held. 550 * This is the only routine that adds items to the list. 551 * The following routine is the only one that removes items 552 * and does so in order from first to last. 553 */ 554 static void 555 add_to_worklist(wk) 556 struct worklist *wk; 557 { 558 static struct worklist *worklist_tail; 559 560 if (wk->wk_state & ONWORKLIST) { 561 if (lk.lkt_held != NOHOLDER) 562 FREE_LOCK(&lk); 563 panic("add_to_worklist: already on list"); 564 } 565 wk->wk_state |= ONWORKLIST; 566 if (LIST_FIRST(&softdep_workitem_pending) == NULL) 567 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list); 568 else 569 LIST_INSERT_AFTER(worklist_tail, wk, wk_list); 570 worklist_tail = wk; 571 num_on_worklist += 1; 572 } 573 574 /* 575 * Process that runs once per second to handle items in the background queue. 576 * 577 * Note that we ensure that everything is done in the order in which they 578 * appear in the queue. The code below depends on this property to ensure 579 * that blocks of a file are freed before the inode itself is freed. This 580 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 581 * until all the old ones have been purged from the dependency lists. 582 */ 583 static int 584 softdep_process_worklist(matchmnt) 585 struct mount *matchmnt; 586 { 587 thread_t td = curthread; 588 int matchcnt, loopcount; 589 long starttime; 590 591 /* 592 * Record the process identifier of our caller so that we can give 593 * this process preferential treatment in request_cleanup below. 594 */ 595 filesys_syncer = td; 596 matchcnt = 0; 597 598 /* 599 * There is no danger of having multiple processes run this 600 * code, but we have to single-thread it when softdep_flushfiles() 601 * is in operation to get an accurate count of the number of items 602 * related to its mount point that are in the list. 603 */ 604 if (matchmnt == NULL) { 605 if (softdep_worklist_busy < 0) 606 return(-1); 607 softdep_worklist_busy += 1; 608 } 609 610 /* 611 * If requested, try removing inode or removal dependencies. 612 */ 613 if (req_clear_inodedeps) { 614 clear_inodedeps(td); 615 req_clear_inodedeps -= 1; 616 wakeup_one(&proc_waiting); 617 } 618 if (req_clear_remove) { 619 clear_remove(td); 620 req_clear_remove -= 1; 621 wakeup_one(&proc_waiting); 622 } 623 loopcount = 1; 624 starttime = time_second; 625 while (num_on_worklist > 0) { 626 matchcnt += process_worklist_item(matchmnt, 0); 627 628 /* 629 * If a umount operation wants to run the worklist 630 * accurately, abort. 631 */ 632 if (softdep_worklist_req && matchmnt == NULL) { 633 matchcnt = -1; 634 break; 635 } 636 637 /* 638 * If requested, try removing inode or removal dependencies. 639 */ 640 if (req_clear_inodedeps) { 641 clear_inodedeps(td); 642 req_clear_inodedeps -= 1; 643 wakeup_one(&proc_waiting); 644 } 645 if (req_clear_remove) { 646 clear_remove(td); 647 req_clear_remove -= 1; 648 wakeup_one(&proc_waiting); 649 } 650 /* 651 * We do not generally want to stop for buffer space, but if 652 * we are really being a buffer hog, we will stop and wait. 653 */ 654 if (loopcount++ % 128 == 0) 655 bwillwrite(); 656 /* 657 * Never allow processing to run for more than one 658 * second. Otherwise the other syncer tasks may get 659 * excessively backlogged. 660 */ 661 if (starttime != time_second && matchmnt == NULL) { 662 matchcnt = -1; 663 break; 664 } 665 } 666 if (matchmnt == NULL) { 667 --softdep_worklist_busy; 668 if (softdep_worklist_req && softdep_worklist_busy == 0) 669 wakeup(&softdep_worklist_req); 670 } 671 return (matchcnt); 672 } 673 674 /* 675 * Process one item on the worklist. 676 */ 677 static int 678 process_worklist_item(matchmnt, flags) 679 struct mount *matchmnt; 680 int flags; 681 { 682 struct worklist *wk; 683 struct dirrem *dirrem; 684 struct fs *matchfs; 685 struct vnode *vp; 686 int matchcnt = 0; 687 688 matchfs = NULL; 689 if (matchmnt != NULL) 690 matchfs = VFSTOUFS(matchmnt)->um_fs; 691 ACQUIRE_LOCK(&lk); 692 /* 693 * Normally we just process each item on the worklist in order. 694 * However, if we are in a situation where we cannot lock any 695 * inodes, we have to skip over any dirrem requests whose 696 * vnodes are resident and locked. 697 */ 698 LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) { 699 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM) 700 break; 701 dirrem = WK_DIRREM(wk); 702 vp = ufs_ihashlookup(VFSTOUFS(dirrem->dm_mnt)->um_dev, 703 dirrem->dm_oldinum); 704 if (vp == NULL || !VOP_ISLOCKED(vp, curthread)) 705 break; 706 } 707 if (wk == 0) { 708 FREE_LOCK(&lk); 709 return (0); 710 } 711 WORKLIST_REMOVE(wk); 712 num_on_worklist -= 1; 713 FREE_LOCK(&lk); 714 switch (wk->wk_type) { 715 716 case D_DIRREM: 717 /* removal of a directory entry */ 718 if (WK_DIRREM(wk)->dm_mnt == matchmnt) 719 matchcnt += 1; 720 handle_workitem_remove(WK_DIRREM(wk)); 721 break; 722 723 case D_FREEBLKS: 724 /* releasing blocks and/or fragments from a file */ 725 if (WK_FREEBLKS(wk)->fb_fs == matchfs) 726 matchcnt += 1; 727 handle_workitem_freeblocks(WK_FREEBLKS(wk)); 728 break; 729 730 case D_FREEFRAG: 731 /* releasing a fragment when replaced as a file grows */ 732 if (WK_FREEFRAG(wk)->ff_fs == matchfs) 733 matchcnt += 1; 734 handle_workitem_freefrag(WK_FREEFRAG(wk)); 735 break; 736 737 case D_FREEFILE: 738 /* releasing an inode when its link count drops to 0 */ 739 if (WK_FREEFILE(wk)->fx_fs == matchfs) 740 matchcnt += 1; 741 handle_workitem_freefile(WK_FREEFILE(wk)); 742 break; 743 744 default: 745 panic("%s_process_worklist: Unknown type %s", 746 "softdep", TYPENAME(wk->wk_type)); 747 /* NOTREACHED */ 748 } 749 return (matchcnt); 750 } 751 752 /* 753 * Move dependencies from one buffer to another. 754 */ 755 static void 756 softdep_move_dependencies(oldbp, newbp) 757 struct buf *oldbp; 758 struct buf *newbp; 759 { 760 struct worklist *wk, *wktail; 761 762 if (LIST_FIRST(&newbp->b_dep) != NULL) 763 panic("softdep_move_dependencies: need merge code"); 764 wktail = 0; 765 ACQUIRE_LOCK(&lk); 766 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 767 LIST_REMOVE(wk, wk_list); 768 if (wktail == 0) 769 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 770 else 771 LIST_INSERT_AFTER(wktail, wk, wk_list); 772 wktail = wk; 773 } 774 FREE_LOCK(&lk); 775 } 776 777 /* 778 * Purge the work list of all items associated with a particular mount point. 779 */ 780 int 781 softdep_flushfiles(struct mount *oldmnt, int flags, struct thread *td) 782 { 783 struct vnode *devvp; 784 int error, loopcnt; 785 786 /* 787 * Await our turn to clear out the queue, then serialize access. 788 */ 789 while (softdep_worklist_busy != 0) { 790 softdep_worklist_req += 1; 791 tsleep(&softdep_worklist_req, 0, "softflush", 0); 792 softdep_worklist_req -= 1; 793 } 794 softdep_worklist_busy = -1; 795 796 if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0) { 797 softdep_worklist_busy = 0; 798 if (softdep_worklist_req) 799 wakeup(&softdep_worklist_req); 800 return (error); 801 } 802 /* 803 * Alternately flush the block device associated with the mount 804 * point and process any dependencies that the flushing 805 * creates. In theory, this loop can happen at most twice, 806 * but we give it a few extra just to be sure. 807 */ 808 devvp = VFSTOUFS(oldmnt)->um_devvp; 809 for (loopcnt = 10; loopcnt > 0; ) { 810 if (softdep_process_worklist(oldmnt) == 0) { 811 loopcnt--; 812 /* 813 * Do another flush in case any vnodes were brought in 814 * as part of the cleanup operations. 815 */ 816 if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0) 817 break; 818 /* 819 * If we still found nothing to do, we are really done. 820 */ 821 if (softdep_process_worklist(oldmnt) == 0) 822 break; 823 } 824 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, td); 825 error = VOP_FSYNC(devvp, MNT_WAIT, td); 826 VOP_UNLOCK(devvp, 0, td); 827 if (error) 828 break; 829 } 830 softdep_worklist_busy = 0; 831 if (softdep_worklist_req) 832 wakeup(&softdep_worklist_req); 833 834 /* 835 * If we are unmounting then it is an error to fail. If we 836 * are simply trying to downgrade to read-only, then filesystem 837 * activity can keep us busy forever, so we just fail with EBUSY. 838 */ 839 if (loopcnt == 0) { 840 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 841 panic("softdep_flushfiles: looping"); 842 error = EBUSY; 843 } 844 return (error); 845 } 846 847 /* 848 * Structure hashing. 849 * 850 * There are three types of structures that can be looked up: 851 * 1) pagedep structures identified by mount point, inode number, 852 * and logical block. 853 * 2) inodedep structures identified by mount point and inode number. 854 * 3) newblk structures identified by mount point and 855 * physical block number. 856 * 857 * The "pagedep" and "inodedep" dependency structures are hashed 858 * separately from the file blocks and inodes to which they correspond. 859 * This separation helps when the in-memory copy of an inode or 860 * file block must be replaced. It also obviates the need to access 861 * an inode or file page when simply updating (or de-allocating) 862 * dependency structures. Lookup of newblk structures is needed to 863 * find newly allocated blocks when trying to associate them with 864 * their allocdirect or allocindir structure. 865 * 866 * The lookup routines optionally create and hash a new instance when 867 * an existing entry is not found. 868 */ 869 #define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 870 #define NODELAY 0x0002 /* cannot do background work */ 871 872 /* 873 * Structures and routines associated with pagedep caching. 874 */ 875 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl; 876 u_long pagedep_hash; /* size of hash table - 1 */ 877 #define PAGEDEP_HASH(mp, inum, lbn) \ 878 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \ 879 pagedep_hash]) 880 static struct sema pagedep_in_progress; 881 882 /* 883 * Look up a pagedep. Return 1 if found, 0 if not found. 884 * If not found, allocate if DEPALLOC flag is passed. 885 * Found or allocated entry is returned in pagedeppp. 886 * This routine must be called with splbio interrupts blocked. 887 */ 888 static int 889 pagedep_lookup(ip, lbn, flags, pagedeppp) 890 struct inode *ip; 891 ufs_lbn_t lbn; 892 int flags; 893 struct pagedep **pagedeppp; 894 { 895 struct pagedep *pagedep; 896 struct pagedep_hashhead *pagedephd; 897 struct mount *mp; 898 int i; 899 900 #ifdef DEBUG 901 if (lk.lkt_held == NOHOLDER) 902 panic("pagedep_lookup: lock not held"); 903 #endif 904 mp = ITOV(ip)->v_mount; 905 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn); 906 top: 907 LIST_FOREACH(pagedep, pagedephd, pd_hash) 908 if (ip->i_number == pagedep->pd_ino && 909 lbn == pagedep->pd_lbn && 910 mp == pagedep->pd_mnt) 911 break; 912 if (pagedep) { 913 *pagedeppp = pagedep; 914 return (1); 915 } 916 if ((flags & DEPALLOC) == 0) { 917 *pagedeppp = NULL; 918 return (0); 919 } 920 if (sema_get(&pagedep_in_progress, &lk) == 0) { 921 ACQUIRE_LOCK(&lk); 922 goto top; 923 } 924 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP, 925 M_SOFTDEP_FLAGS); 926 bzero(pagedep, sizeof(struct pagedep)); 927 pagedep->pd_list.wk_type = D_PAGEDEP; 928 pagedep->pd_mnt = mp; 929 pagedep->pd_ino = ip->i_number; 930 pagedep->pd_lbn = lbn; 931 LIST_INIT(&pagedep->pd_dirremhd); 932 LIST_INIT(&pagedep->pd_pendinghd); 933 for (i = 0; i < DAHASHSZ; i++) 934 LIST_INIT(&pagedep->pd_diraddhd[i]); 935 ACQUIRE_LOCK(&lk); 936 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 937 sema_release(&pagedep_in_progress); 938 *pagedeppp = pagedep; 939 return (0); 940 } 941 942 /* 943 * Structures and routines associated with inodedep caching. 944 */ 945 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl; 946 static u_long inodedep_hash; /* size of hash table - 1 */ 947 static long num_inodedep; /* number of inodedep allocated */ 948 #define INODEDEP_HASH(fs, inum) \ 949 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash]) 950 static struct sema inodedep_in_progress; 951 952 /* 953 * Look up a inodedep. Return 1 if found, 0 if not found. 954 * If not found, allocate if DEPALLOC flag is passed. 955 * Found or allocated entry is returned in inodedeppp. 956 * This routine must be called with splbio interrupts blocked. 957 */ 958 static int 959 inodedep_lookup(fs, inum, flags, inodedeppp) 960 struct fs *fs; 961 ino_t inum; 962 int flags; 963 struct inodedep **inodedeppp; 964 { 965 struct inodedep *inodedep; 966 struct inodedep_hashhead *inodedephd; 967 int firsttry; 968 969 #ifdef DEBUG 970 if (lk.lkt_held == NOHOLDER) 971 panic("inodedep_lookup: lock not held"); 972 #endif 973 firsttry = 1; 974 inodedephd = INODEDEP_HASH(fs, inum); 975 top: 976 LIST_FOREACH(inodedep, inodedephd, id_hash) 977 if (inum == inodedep->id_ino && fs == inodedep->id_fs) 978 break; 979 if (inodedep) { 980 *inodedeppp = inodedep; 981 return (1); 982 } 983 if ((flags & DEPALLOC) == 0) { 984 *inodedeppp = NULL; 985 return (0); 986 } 987 /* 988 * If we are over our limit, try to improve the situation. 989 */ 990 if (num_inodedep > max_softdeps && firsttry && 991 speedup_syncer() == 0 && (flags & NODELAY) == 0 && 992 request_cleanup(FLUSH_INODES, 1)) { 993 firsttry = 0; 994 goto top; 995 } 996 if (sema_get(&inodedep_in_progress, &lk) == 0) { 997 ACQUIRE_LOCK(&lk); 998 goto top; 999 } 1000 num_inodedep += 1; 1001 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep), 1002 M_INODEDEP, M_SOFTDEP_FLAGS); 1003 inodedep->id_list.wk_type = D_INODEDEP; 1004 inodedep->id_fs = fs; 1005 inodedep->id_ino = inum; 1006 inodedep->id_state = ALLCOMPLETE; 1007 inodedep->id_nlinkdelta = 0; 1008 inodedep->id_savedino = NULL; 1009 inodedep->id_savedsize = -1; 1010 inodedep->id_buf = NULL; 1011 LIST_INIT(&inodedep->id_pendinghd); 1012 LIST_INIT(&inodedep->id_inowait); 1013 LIST_INIT(&inodedep->id_bufwait); 1014 TAILQ_INIT(&inodedep->id_inoupdt); 1015 TAILQ_INIT(&inodedep->id_newinoupdt); 1016 ACQUIRE_LOCK(&lk); 1017 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 1018 sema_release(&inodedep_in_progress); 1019 *inodedeppp = inodedep; 1020 return (0); 1021 } 1022 1023 /* 1024 * Structures and routines associated with newblk caching. 1025 */ 1026 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl; 1027 u_long newblk_hash; /* size of hash table - 1 */ 1028 #define NEWBLK_HASH(fs, inum) \ 1029 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash]) 1030 static struct sema newblk_in_progress; 1031 1032 /* 1033 * Look up a newblk. Return 1 if found, 0 if not found. 1034 * If not found, allocate if DEPALLOC flag is passed. 1035 * Found or allocated entry is returned in newblkpp. 1036 */ 1037 static int 1038 newblk_lookup(fs, newblkno, flags, newblkpp) 1039 struct fs *fs; 1040 ufs_daddr_t newblkno; 1041 int flags; 1042 struct newblk **newblkpp; 1043 { 1044 struct newblk *newblk; 1045 struct newblk_hashhead *newblkhd; 1046 1047 newblkhd = NEWBLK_HASH(fs, newblkno); 1048 top: 1049 LIST_FOREACH(newblk, newblkhd, nb_hash) 1050 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs) 1051 break; 1052 if (newblk) { 1053 *newblkpp = newblk; 1054 return (1); 1055 } 1056 if ((flags & DEPALLOC) == 0) { 1057 *newblkpp = NULL; 1058 return (0); 1059 } 1060 if (sema_get(&newblk_in_progress, 0) == 0) 1061 goto top; 1062 MALLOC(newblk, struct newblk *, sizeof(struct newblk), 1063 M_NEWBLK, M_SOFTDEP_FLAGS); 1064 newblk->nb_state = 0; 1065 newblk->nb_fs = fs; 1066 newblk->nb_newblkno = newblkno; 1067 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 1068 sema_release(&newblk_in_progress); 1069 *newblkpp = newblk; 1070 return (0); 1071 } 1072 1073 /* 1074 * Executed during filesystem system initialization before 1075 * mounting any filesystems. 1076 */ 1077 void 1078 softdep_initialize() 1079 { 1080 callout_init(&handle); 1081 bioops = softdep_bioops; /* XXX hack */ 1082 1083 LIST_INIT(&mkdirlisthd); 1084 LIST_INIT(&softdep_workitem_pending); 1085 max_softdeps = min(desiredvnodes * 8, 1086 M_INODEDEP->ks_limit / (2 * sizeof(struct inodedep))); 1087 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 1088 &pagedep_hash); 1089 sema_init(&pagedep_in_progress, "pagedep", 0, 0); 1090 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash); 1091 sema_init(&inodedep_in_progress, "inodedep", 0, 0); 1092 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash); 1093 sema_init(&newblk_in_progress, "newblk", 0, 0); 1094 } 1095 1096 /* 1097 * Called at mount time to notify the dependency code that a 1098 * filesystem wishes to use it. 1099 */ 1100 int 1101 softdep_mount(devvp, mp, fs) 1102 struct vnode *devvp; 1103 struct mount *mp; 1104 struct fs *fs; 1105 { 1106 struct csum cstotal; 1107 struct cg *cgp; 1108 struct buf *bp; 1109 int error, cyl; 1110 1111 mp->mnt_flag &= ~MNT_ASYNC; 1112 mp->mnt_flag |= MNT_SOFTDEP; 1113 /* 1114 * When doing soft updates, the counters in the 1115 * superblock may have gotten out of sync, so we have 1116 * to scan the cylinder groups and recalculate them. 1117 */ 1118 if (fs->fs_clean != 0) 1119 return (0); 1120 bzero(&cstotal, sizeof cstotal); 1121 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 1122 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 1123 fs->fs_cgsize, &bp)) != 0) { 1124 brelse(bp); 1125 return (error); 1126 } 1127 cgp = (struct cg *)bp->b_data; 1128 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 1129 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 1130 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 1131 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 1132 fs->fs_cs(fs, cyl) = cgp->cg_cs; 1133 brelse(bp); 1134 } 1135 #ifdef DEBUG 1136 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 1137 printf("ffs_mountfs: superblock updated for soft updates\n"); 1138 #endif 1139 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 1140 return (0); 1141 } 1142 1143 /* 1144 * Protecting the freemaps (or bitmaps). 1145 * 1146 * To eliminate the need to execute fsck before mounting a filesystem 1147 * after a power failure, one must (conservatively) guarantee that the 1148 * on-disk copy of the bitmaps never indicate that a live inode or block is 1149 * free. So, when a block or inode is allocated, the bitmap should be 1150 * updated (on disk) before any new pointers. When a block or inode is 1151 * freed, the bitmap should not be updated until all pointers have been 1152 * reset. The latter dependency is handled by the delayed de-allocation 1153 * approach described below for block and inode de-allocation. The former 1154 * dependency is handled by calling the following procedure when a block or 1155 * inode is allocated. When an inode is allocated an "inodedep" is created 1156 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 1157 * Each "inodedep" is also inserted into the hash indexing structure so 1158 * that any additional link additions can be made dependent on the inode 1159 * allocation. 1160 * 1161 * The ufs filesystem maintains a number of free block counts (e.g., per 1162 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 1163 * in addition to the bitmaps. These counts are used to improve efficiency 1164 * during allocation and therefore must be consistent with the bitmaps. 1165 * There is no convenient way to guarantee post-crash consistency of these 1166 * counts with simple update ordering, for two main reasons: (1) The counts 1167 * and bitmaps for a single cylinder group block are not in the same disk 1168 * sector. If a disk write is interrupted (e.g., by power failure), one may 1169 * be written and the other not. (2) Some of the counts are located in the 1170 * superblock rather than the cylinder group block. So, we focus our soft 1171 * updates implementation on protecting the bitmaps. When mounting a 1172 * filesystem, we recompute the auxiliary counts from the bitmaps. 1173 */ 1174 1175 /* 1176 * Called just after updating the cylinder group block to allocate an inode. 1177 */ 1178 void 1179 softdep_setup_inomapdep(bp, ip, newinum) 1180 struct buf *bp; /* buffer for cylgroup block with inode map */ 1181 struct inode *ip; /* inode related to allocation */ 1182 ino_t newinum; /* new inode number being allocated */ 1183 { 1184 struct inodedep *inodedep; 1185 struct bmsafemap *bmsafemap; 1186 1187 /* 1188 * Create a dependency for the newly allocated inode. 1189 * Panic if it already exists as something is seriously wrong. 1190 * Otherwise add it to the dependency list for the buffer holding 1191 * the cylinder group map from which it was allocated. 1192 */ 1193 ACQUIRE_LOCK(&lk); 1194 if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) { 1195 FREE_LOCK(&lk); 1196 panic("softdep_setup_inomapdep: found inode"); 1197 } 1198 inodedep->id_buf = bp; 1199 inodedep->id_state &= ~DEPCOMPLETE; 1200 bmsafemap = bmsafemap_lookup(bp); 1201 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 1202 FREE_LOCK(&lk); 1203 } 1204 1205 /* 1206 * Called just after updating the cylinder group block to 1207 * allocate block or fragment. 1208 */ 1209 void 1210 softdep_setup_blkmapdep(bp, fs, newblkno) 1211 struct buf *bp; /* buffer for cylgroup block with block map */ 1212 struct fs *fs; /* filesystem doing allocation */ 1213 ufs_daddr_t newblkno; /* number of newly allocated block */ 1214 { 1215 struct newblk *newblk; 1216 struct bmsafemap *bmsafemap; 1217 1218 /* 1219 * Create a dependency for the newly allocated block. 1220 * Add it to the dependency list for the buffer holding 1221 * the cylinder group map from which it was allocated. 1222 */ 1223 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0) 1224 panic("softdep_setup_blkmapdep: found block"); 1225 ACQUIRE_LOCK(&lk); 1226 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp); 1227 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 1228 FREE_LOCK(&lk); 1229 } 1230 1231 /* 1232 * Find the bmsafemap associated with a cylinder group buffer. 1233 * If none exists, create one. The buffer must be locked when 1234 * this routine is called and this routine must be called with 1235 * splbio interrupts blocked. 1236 */ 1237 static struct bmsafemap * 1238 bmsafemap_lookup(bp) 1239 struct buf *bp; 1240 { 1241 struct bmsafemap *bmsafemap; 1242 struct worklist *wk; 1243 1244 #ifdef DEBUG 1245 if (lk.lkt_held == NOHOLDER) 1246 panic("bmsafemap_lookup: lock not held"); 1247 #endif 1248 LIST_FOREACH(wk, &bp->b_dep, wk_list) 1249 if (wk->wk_type == D_BMSAFEMAP) 1250 return (WK_BMSAFEMAP(wk)); 1251 FREE_LOCK(&lk); 1252 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap), 1253 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 1254 bmsafemap->sm_list.wk_type = D_BMSAFEMAP; 1255 bmsafemap->sm_list.wk_state = 0; 1256 bmsafemap->sm_buf = bp; 1257 LIST_INIT(&bmsafemap->sm_allocdirecthd); 1258 LIST_INIT(&bmsafemap->sm_allocindirhd); 1259 LIST_INIT(&bmsafemap->sm_inodedephd); 1260 LIST_INIT(&bmsafemap->sm_newblkhd); 1261 ACQUIRE_LOCK(&lk); 1262 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 1263 return (bmsafemap); 1264 } 1265 1266 /* 1267 * Direct block allocation dependencies. 1268 * 1269 * When a new block is allocated, the corresponding disk locations must be 1270 * initialized (with zeros or new data) before the on-disk inode points to 1271 * them. Also, the freemap from which the block was allocated must be 1272 * updated (on disk) before the inode's pointer. These two dependencies are 1273 * independent of each other and are needed for all file blocks and indirect 1274 * blocks that are pointed to directly by the inode. Just before the 1275 * "in-core" version of the inode is updated with a newly allocated block 1276 * number, a procedure (below) is called to setup allocation dependency 1277 * structures. These structures are removed when the corresponding 1278 * dependencies are satisfied or when the block allocation becomes obsolete 1279 * (i.e., the file is deleted, the block is de-allocated, or the block is a 1280 * fragment that gets upgraded). All of these cases are handled in 1281 * procedures described later. 1282 * 1283 * When a file extension causes a fragment to be upgraded, either to a larger 1284 * fragment or to a full block, the on-disk location may change (if the 1285 * previous fragment could not simply be extended). In this case, the old 1286 * fragment must be de-allocated, but not until after the inode's pointer has 1287 * been updated. In most cases, this is handled by later procedures, which 1288 * will construct a "freefrag" structure to be added to the workitem queue 1289 * when the inode update is complete (or obsolete). The main exception to 1290 * this is when an allocation occurs while a pending allocation dependency 1291 * (for the same block pointer) remains. This case is handled in the main 1292 * allocation dependency setup procedure by immediately freeing the 1293 * unreferenced fragments. 1294 */ 1295 void 1296 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 1297 struct inode *ip; /* inode to which block is being added */ 1298 ufs_lbn_t lbn; /* block pointer within inode */ 1299 ufs_daddr_t newblkno; /* disk block number being added */ 1300 ufs_daddr_t oldblkno; /* previous block number, 0 unless frag */ 1301 long newsize; /* size of new block */ 1302 long oldsize; /* size of new block */ 1303 struct buf *bp; /* bp for allocated block */ 1304 { 1305 struct allocdirect *adp, *oldadp; 1306 struct allocdirectlst *adphead; 1307 struct bmsafemap *bmsafemap; 1308 struct inodedep *inodedep; 1309 struct pagedep *pagedep; 1310 struct newblk *newblk; 1311 1312 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect), 1313 M_ALLOCDIRECT, M_SOFTDEP_FLAGS); 1314 bzero(adp, sizeof(struct allocdirect)); 1315 adp->ad_list.wk_type = D_ALLOCDIRECT; 1316 adp->ad_lbn = lbn; 1317 adp->ad_newblkno = newblkno; 1318 adp->ad_oldblkno = oldblkno; 1319 adp->ad_newsize = newsize; 1320 adp->ad_oldsize = oldsize; 1321 adp->ad_state = ATTACHED; 1322 if (newblkno == oldblkno) 1323 adp->ad_freefrag = NULL; 1324 else 1325 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1326 1327 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1328 panic("softdep_setup_allocdirect: lost block"); 1329 1330 ACQUIRE_LOCK(&lk); 1331 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep); 1332 adp->ad_inodedep = inodedep; 1333 1334 if (newblk->nb_state == DEPCOMPLETE) { 1335 adp->ad_state |= DEPCOMPLETE; 1336 adp->ad_buf = NULL; 1337 } else { 1338 bmsafemap = newblk->nb_bmsafemap; 1339 adp->ad_buf = bmsafemap->sm_buf; 1340 LIST_REMOVE(newblk, nb_deps); 1341 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1342 } 1343 LIST_REMOVE(newblk, nb_hash); 1344 FREE(newblk, M_NEWBLK); 1345 1346 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list); 1347 if (lbn >= NDADDR) { 1348 /* allocating an indirect block */ 1349 if (oldblkno != 0) { 1350 FREE_LOCK(&lk); 1351 panic("softdep_setup_allocdirect: non-zero indir"); 1352 } 1353 } else { 1354 /* 1355 * Allocating a direct block. 1356 * 1357 * If we are allocating a directory block, then we must 1358 * allocate an associated pagedep to track additions and 1359 * deletions. 1360 */ 1361 if ((ip->i_mode & IFMT) == IFDIR && 1362 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1363 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 1364 } 1365 /* 1366 * The list of allocdirects must be kept in sorted and ascending 1367 * order so that the rollback routines can quickly determine the 1368 * first uncommitted block (the size of the file stored on disk 1369 * ends at the end of the lowest committed fragment, or if there 1370 * are no fragments, at the end of the highest committed block). 1371 * Since files generally grow, the typical case is that the new 1372 * block is to be added at the end of the list. We speed this 1373 * special case by checking against the last allocdirect in the 1374 * list before laboriously traversing the list looking for the 1375 * insertion point. 1376 */ 1377 adphead = &inodedep->id_newinoupdt; 1378 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1379 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1380 /* insert at end of list */ 1381 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1382 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1383 allocdirect_merge(adphead, adp, oldadp); 1384 FREE_LOCK(&lk); 1385 return; 1386 } 1387 TAILQ_FOREACH(oldadp, adphead, ad_next) { 1388 if (oldadp->ad_lbn >= lbn) 1389 break; 1390 } 1391 if (oldadp == NULL) { 1392 FREE_LOCK(&lk); 1393 panic("softdep_setup_allocdirect: lost entry"); 1394 } 1395 /* insert in middle of list */ 1396 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1397 if (oldadp->ad_lbn == lbn) 1398 allocdirect_merge(adphead, adp, oldadp); 1399 FREE_LOCK(&lk); 1400 } 1401 1402 /* 1403 * Replace an old allocdirect dependency with a newer one. 1404 * This routine must be called with splbio interrupts blocked. 1405 */ 1406 static void 1407 allocdirect_merge(adphead, newadp, oldadp) 1408 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 1409 struct allocdirect *newadp; /* allocdirect being added */ 1410 struct allocdirect *oldadp; /* existing allocdirect being checked */ 1411 { 1412 struct freefrag *freefrag; 1413 1414 #ifdef DEBUG 1415 if (lk.lkt_held == NOHOLDER) 1416 panic("allocdirect_merge: lock not held"); 1417 #endif 1418 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 1419 newadp->ad_oldsize != oldadp->ad_newsize || 1420 newadp->ad_lbn >= NDADDR) { 1421 FREE_LOCK(&lk); 1422 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d", 1423 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn, 1424 NDADDR); 1425 } 1426 newadp->ad_oldblkno = oldadp->ad_oldblkno; 1427 newadp->ad_oldsize = oldadp->ad_oldsize; 1428 /* 1429 * If the old dependency had a fragment to free or had never 1430 * previously had a block allocated, then the new dependency 1431 * can immediately post its freefrag and adopt the old freefrag. 1432 * This action is done by swapping the freefrag dependencies. 1433 * The new dependency gains the old one's freefrag, and the 1434 * old one gets the new one and then immediately puts it on 1435 * the worklist when it is freed by free_allocdirect. It is 1436 * not possible to do this swap when the old dependency had a 1437 * non-zero size but no previous fragment to free. This condition 1438 * arises when the new block is an extension of the old block. 1439 * Here, the first part of the fragment allocated to the new 1440 * dependency is part of the block currently claimed on disk by 1441 * the old dependency, so cannot legitimately be freed until the 1442 * conditions for the new dependency are fulfilled. 1443 */ 1444 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 1445 freefrag = newadp->ad_freefrag; 1446 newadp->ad_freefrag = oldadp->ad_freefrag; 1447 oldadp->ad_freefrag = freefrag; 1448 } 1449 free_allocdirect(adphead, oldadp, 0); 1450 } 1451 1452 /* 1453 * Allocate a new freefrag structure if needed. 1454 */ 1455 static struct freefrag * 1456 newfreefrag(ip, blkno, size) 1457 struct inode *ip; 1458 ufs_daddr_t blkno; 1459 long size; 1460 { 1461 struct freefrag *freefrag; 1462 struct fs *fs; 1463 1464 if (blkno == 0) 1465 return (NULL); 1466 fs = ip->i_fs; 1467 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 1468 panic("newfreefrag: frag size"); 1469 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag), 1470 M_FREEFRAG, M_SOFTDEP_FLAGS); 1471 freefrag->ff_list.wk_type = D_FREEFRAG; 1472 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */ 1473 freefrag->ff_inum = ip->i_number; 1474 freefrag->ff_fs = fs; 1475 freefrag->ff_devvp = ip->i_devvp; 1476 freefrag->ff_blkno = blkno; 1477 freefrag->ff_fragsize = size; 1478 return (freefrag); 1479 } 1480 1481 /* 1482 * This workitem de-allocates fragments that were replaced during 1483 * file block allocation. 1484 */ 1485 static void 1486 handle_workitem_freefrag(freefrag) 1487 struct freefrag *freefrag; 1488 { 1489 struct inode tip; 1490 1491 tip.i_fs = freefrag->ff_fs; 1492 tip.i_devvp = freefrag->ff_devvp; 1493 tip.i_dev = freefrag->ff_devvp->v_rdev; 1494 tip.i_number = freefrag->ff_inum; 1495 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */ 1496 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize); 1497 FREE(freefrag, M_FREEFRAG); 1498 } 1499 1500 /* 1501 * Indirect block allocation dependencies. 1502 * 1503 * The same dependencies that exist for a direct block also exist when 1504 * a new block is allocated and pointed to by an entry in a block of 1505 * indirect pointers. The undo/redo states described above are also 1506 * used here. Because an indirect block contains many pointers that 1507 * may have dependencies, a second copy of the entire in-memory indirect 1508 * block is kept. The buffer cache copy is always completely up-to-date. 1509 * The second copy, which is used only as a source for disk writes, 1510 * contains only the safe pointers (i.e., those that have no remaining 1511 * update dependencies). The second copy is freed when all pointers 1512 * are safe. The cache is not allowed to replace indirect blocks with 1513 * pending update dependencies. If a buffer containing an indirect 1514 * block with dependencies is written, these routines will mark it 1515 * dirty again. It can only be successfully written once all the 1516 * dependencies are removed. The ffs_fsync routine in conjunction with 1517 * softdep_sync_metadata work together to get all the dependencies 1518 * removed so that a file can be successfully written to disk. Three 1519 * procedures are used when setting up indirect block pointer 1520 * dependencies. The division is necessary because of the organization 1521 * of the "balloc" routine and because of the distinction between file 1522 * pages and file metadata blocks. 1523 */ 1524 1525 /* 1526 * Allocate a new allocindir structure. 1527 */ 1528 static struct allocindir * 1529 newallocindir(ip, ptrno, newblkno, oldblkno) 1530 struct inode *ip; /* inode for file being extended */ 1531 int ptrno; /* offset of pointer in indirect block */ 1532 ufs_daddr_t newblkno; /* disk block number being added */ 1533 ufs_daddr_t oldblkno; /* previous block number, 0 if none */ 1534 { 1535 struct allocindir *aip; 1536 1537 MALLOC(aip, struct allocindir *, sizeof(struct allocindir), 1538 M_ALLOCINDIR, M_SOFTDEP_FLAGS); 1539 bzero(aip, sizeof(struct allocindir)); 1540 aip->ai_list.wk_type = D_ALLOCINDIR; 1541 aip->ai_state = ATTACHED; 1542 aip->ai_offset = ptrno; 1543 aip->ai_newblkno = newblkno; 1544 aip->ai_oldblkno = oldblkno; 1545 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize); 1546 return (aip); 1547 } 1548 1549 /* 1550 * Called just before setting an indirect block pointer 1551 * to a newly allocated file page. 1552 */ 1553 void 1554 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 1555 struct inode *ip; /* inode for file being extended */ 1556 ufs_lbn_t lbn; /* allocated block number within file */ 1557 struct buf *bp; /* buffer with indirect blk referencing page */ 1558 int ptrno; /* offset of pointer in indirect block */ 1559 ufs_daddr_t newblkno; /* disk block number being added */ 1560 ufs_daddr_t oldblkno; /* previous block number, 0 if none */ 1561 struct buf *nbp; /* buffer holding allocated page */ 1562 { 1563 struct allocindir *aip; 1564 struct pagedep *pagedep; 1565 1566 aip = newallocindir(ip, ptrno, newblkno, oldblkno); 1567 ACQUIRE_LOCK(&lk); 1568 /* 1569 * If we are allocating a directory page, then we must 1570 * allocate an associated pagedep to track additions and 1571 * deletions. 1572 */ 1573 if ((ip->i_mode & IFMT) == IFDIR && 1574 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1575 WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list); 1576 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1577 FREE_LOCK(&lk); 1578 setup_allocindir_phase2(bp, ip, aip); 1579 } 1580 1581 /* 1582 * Called just before setting an indirect block pointer to a 1583 * newly allocated indirect block. 1584 */ 1585 void 1586 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 1587 struct buf *nbp; /* newly allocated indirect block */ 1588 struct inode *ip; /* inode for file being extended */ 1589 struct buf *bp; /* indirect block referencing allocated block */ 1590 int ptrno; /* offset of pointer in indirect block */ 1591 ufs_daddr_t newblkno; /* disk block number being added */ 1592 { 1593 struct allocindir *aip; 1594 1595 aip = newallocindir(ip, ptrno, newblkno, 0); 1596 ACQUIRE_LOCK(&lk); 1597 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1598 FREE_LOCK(&lk); 1599 setup_allocindir_phase2(bp, ip, aip); 1600 } 1601 1602 /* 1603 * Called to finish the allocation of the "aip" allocated 1604 * by one of the two routines above. 1605 */ 1606 static void 1607 setup_allocindir_phase2(bp, ip, aip) 1608 struct buf *bp; /* in-memory copy of the indirect block */ 1609 struct inode *ip; /* inode for file being extended */ 1610 struct allocindir *aip; /* allocindir allocated by the above routines */ 1611 { 1612 struct worklist *wk; 1613 struct indirdep *indirdep, *newindirdep; 1614 struct bmsafemap *bmsafemap; 1615 struct allocindir *oldaip; 1616 struct freefrag *freefrag; 1617 struct newblk *newblk; 1618 1619 if (bp->b_lblkno >= 0) 1620 panic("setup_allocindir_phase2: not indir blk"); 1621 for (indirdep = NULL, newindirdep = NULL; ; ) { 1622 ACQUIRE_LOCK(&lk); 1623 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 1624 if (wk->wk_type != D_INDIRDEP) 1625 continue; 1626 indirdep = WK_INDIRDEP(wk); 1627 break; 1628 } 1629 if (indirdep == NULL && newindirdep) { 1630 indirdep = newindirdep; 1631 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 1632 newindirdep = NULL; 1633 } 1634 FREE_LOCK(&lk); 1635 if (indirdep) { 1636 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0, 1637 &newblk) == 0) 1638 panic("setup_allocindir: lost block"); 1639 ACQUIRE_LOCK(&lk); 1640 if (newblk->nb_state == DEPCOMPLETE) { 1641 aip->ai_state |= DEPCOMPLETE; 1642 aip->ai_buf = NULL; 1643 } else { 1644 bmsafemap = newblk->nb_bmsafemap; 1645 aip->ai_buf = bmsafemap->sm_buf; 1646 LIST_REMOVE(newblk, nb_deps); 1647 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd, 1648 aip, ai_deps); 1649 } 1650 LIST_REMOVE(newblk, nb_hash); 1651 FREE(newblk, M_NEWBLK); 1652 aip->ai_indirdep = indirdep; 1653 /* 1654 * Check to see if there is an existing dependency 1655 * for this block. If there is, merge the old 1656 * dependency into the new one. 1657 */ 1658 if (aip->ai_oldblkno == 0) 1659 oldaip = NULL; 1660 else 1661 1662 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) 1663 if (oldaip->ai_offset == aip->ai_offset) 1664 break; 1665 if (oldaip != NULL) { 1666 if (oldaip->ai_newblkno != aip->ai_oldblkno) { 1667 FREE_LOCK(&lk); 1668 panic("setup_allocindir_phase2: blkno"); 1669 } 1670 aip->ai_oldblkno = oldaip->ai_oldblkno; 1671 freefrag = oldaip->ai_freefrag; 1672 oldaip->ai_freefrag = aip->ai_freefrag; 1673 aip->ai_freefrag = freefrag; 1674 free_allocindir(oldaip, NULL); 1675 } 1676 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 1677 ((ufs_daddr_t *)indirdep->ir_savebp->b_data) 1678 [aip->ai_offset] = aip->ai_oldblkno; 1679 FREE_LOCK(&lk); 1680 } 1681 if (newindirdep) { 1682 /* 1683 * Avoid any possibility of data corruption by 1684 * ensuring that our old version is thrown away. 1685 */ 1686 newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 1687 brelse(newindirdep->ir_savebp); 1688 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP); 1689 } 1690 if (indirdep) 1691 break; 1692 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep), 1693 M_INDIRDEP, M_SOFTDEP_FLAGS); 1694 newindirdep->ir_list.wk_type = D_INDIRDEP; 1695 newindirdep->ir_state = ATTACHED; 1696 LIST_INIT(&newindirdep->ir_deplisthd); 1697 LIST_INIT(&newindirdep->ir_donehd); 1698 if (bp->b_blkno == bp->b_lblkno) { 1699 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, 1700 NULL, NULL); 1701 } 1702 newindirdep->ir_savebp = 1703 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0); 1704 BUF_KERNPROC(newindirdep->ir_savebp); 1705 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 1706 } 1707 } 1708 1709 /* 1710 * Block de-allocation dependencies. 1711 * 1712 * When blocks are de-allocated, the on-disk pointers must be nullified before 1713 * the blocks are made available for use by other files. (The true 1714 * requirement is that old pointers must be nullified before new on-disk 1715 * pointers are set. We chose this slightly more stringent requirement to 1716 * reduce complexity.) Our implementation handles this dependency by updating 1717 * the inode (or indirect block) appropriately but delaying the actual block 1718 * de-allocation (i.e., freemap and free space count manipulation) until 1719 * after the updated versions reach stable storage. After the disk is 1720 * updated, the blocks can be safely de-allocated whenever it is convenient. 1721 * This implementation handles only the common case of reducing a file's 1722 * length to zero. Other cases are handled by the conventional synchronous 1723 * write approach. 1724 * 1725 * The ffs implementation with which we worked double-checks 1726 * the state of the block pointers and file size as it reduces 1727 * a file's length. Some of this code is replicated here in our 1728 * soft updates implementation. The freeblks->fb_chkcnt field is 1729 * used to transfer a part of this information to the procedure 1730 * that eventually de-allocates the blocks. 1731 * 1732 * This routine should be called from the routine that shortens 1733 * a file's length, before the inode's size or block pointers 1734 * are modified. It will save the block pointer information for 1735 * later release and zero the inode so that the calling routine 1736 * can release it. 1737 */ 1738 struct softdep_setup_freeblocks_info { 1739 struct fs *fs; 1740 struct inode *ip; 1741 }; 1742 1743 static int softdep_setup_freeblocks_bp(struct buf *bp, void *data); 1744 1745 void 1746 softdep_setup_freeblocks(ip, length) 1747 struct inode *ip; /* The inode whose length is to be reduced */ 1748 off_t length; /* The new length for the file */ 1749 { 1750 struct softdep_setup_freeblocks_info info; 1751 struct freeblks *freeblks; 1752 struct inodedep *inodedep; 1753 struct allocdirect *adp; 1754 struct vnode *vp; 1755 struct buf *bp; 1756 struct fs *fs; 1757 int i, error, delay; 1758 int count; 1759 1760 fs = ip->i_fs; 1761 if (length != 0) 1762 panic("softde_setup_freeblocks: non-zero length"); 1763 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks), 1764 M_FREEBLKS, M_SOFTDEP_FLAGS); 1765 bzero(freeblks, sizeof(struct freeblks)); 1766 freeblks->fb_list.wk_type = D_FREEBLKS; 1767 freeblks->fb_uid = ip->i_uid; 1768 freeblks->fb_previousinum = ip->i_number; 1769 freeblks->fb_devvp = ip->i_devvp; 1770 freeblks->fb_fs = fs; 1771 freeblks->fb_oldsize = ip->i_size; 1772 freeblks->fb_newsize = length; 1773 freeblks->fb_chkcnt = ip->i_blocks; 1774 for (i = 0; i < NDADDR; i++) { 1775 freeblks->fb_dblks[i] = ip->i_db[i]; 1776 ip->i_db[i] = 0; 1777 } 1778 for (i = 0; i < NIADDR; i++) { 1779 freeblks->fb_iblks[i] = ip->i_ib[i]; 1780 ip->i_ib[i] = 0; 1781 } 1782 ip->i_blocks = 0; 1783 ip->i_size = 0; 1784 /* 1785 * Push the zero'ed inode to to its disk buffer so that we are free 1786 * to delete its dependencies below. Once the dependencies are gone 1787 * the buffer can be safely released. 1788 */ 1789 if ((error = bread(ip->i_devvp, 1790 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 1791 (int)fs->fs_bsize, &bp)) != 0) 1792 softdep_error("softdep_setup_freeblocks", error); 1793 *((struct dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) = 1794 ip->i_din; 1795 /* 1796 * Find and eliminate any inode dependencies. 1797 */ 1798 ACQUIRE_LOCK(&lk); 1799 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep); 1800 if ((inodedep->id_state & IOSTARTED) != 0) { 1801 FREE_LOCK(&lk); 1802 panic("softdep_setup_freeblocks: inode busy"); 1803 } 1804 /* 1805 * Add the freeblks structure to the list of operations that 1806 * must await the zero'ed inode being written to disk. If we 1807 * still have a bitmap dependency (delay == 0), then the inode 1808 * has never been written to disk, so we can process the 1809 * freeblks below once we have deleted the dependencies. 1810 */ 1811 delay = (inodedep->id_state & DEPCOMPLETE); 1812 if (delay) 1813 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list); 1814 /* 1815 * Because the file length has been truncated to zero, any 1816 * pending block allocation dependency structures associated 1817 * with this inode are obsolete and can simply be de-allocated. 1818 * We must first merge the two dependency lists to get rid of 1819 * any duplicate freefrag structures, then purge the merged list. 1820 */ 1821 merge_inode_lists(inodedep); 1822 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 1823 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 1824 FREE_LOCK(&lk); 1825 bdwrite(bp); 1826 /* 1827 * We must wait for any I/O in progress to finish so that 1828 * all potential buffers on the dirty list will be visible. 1829 * Once they are all there, walk the list and get rid of 1830 * any dependencies. 1831 */ 1832 vp = ITOV(ip); 1833 ACQUIRE_LOCK(&lk); 1834 drain_output(vp, 1); 1835 1836 info.fs = fs; 1837 info.ip = ip; 1838 do { 1839 count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 1840 softdep_setup_freeblocks_bp, &info); 1841 } while (count > 0); 1842 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0) 1843 (void)free_inodedep(inodedep); 1844 FREE_LOCK(&lk); 1845 /* 1846 * If the inode has never been written to disk (delay == 0), 1847 * then we can process the freeblks now that we have deleted 1848 * the dependencies. 1849 */ 1850 if (!delay) 1851 handle_workitem_freeblocks(freeblks); 1852 } 1853 1854 static int 1855 softdep_setup_freeblocks_bp(struct buf *bp, void *data) 1856 { 1857 struct softdep_setup_freeblocks_info *info = data; 1858 struct inodedep *inodedep; 1859 1860 if (getdirtybuf(&bp, MNT_WAIT) == 0) 1861 return(-1); 1862 (void) inodedep_lookup(info->fs, info->ip->i_number, 0, &inodedep); 1863 deallocate_dependencies(bp, inodedep); 1864 bp->b_flags |= B_INVAL | B_NOCACHE; 1865 FREE_LOCK(&lk); 1866 brelse(bp); 1867 ACQUIRE_LOCK(&lk); 1868 return(1); 1869 } 1870 1871 /* 1872 * Reclaim any dependency structures from a buffer that is about to 1873 * be reallocated to a new vnode. The buffer must be locked, thus, 1874 * no I/O completion operations can occur while we are manipulating 1875 * its associated dependencies. The mutex is held so that other I/O's 1876 * associated with related dependencies do not occur. 1877 */ 1878 static void 1879 deallocate_dependencies(bp, inodedep) 1880 struct buf *bp; 1881 struct inodedep *inodedep; 1882 { 1883 struct worklist *wk; 1884 struct indirdep *indirdep; 1885 struct allocindir *aip; 1886 struct pagedep *pagedep; 1887 struct dirrem *dirrem; 1888 struct diradd *dap; 1889 int i; 1890 1891 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 1892 switch (wk->wk_type) { 1893 1894 case D_INDIRDEP: 1895 indirdep = WK_INDIRDEP(wk); 1896 /* 1897 * None of the indirect pointers will ever be visible, 1898 * so they can simply be tossed. GOINGAWAY ensures 1899 * that allocated pointers will be saved in the buffer 1900 * cache until they are freed. Note that they will 1901 * only be able to be found by their physical address 1902 * since the inode mapping the logical address will 1903 * be gone. The save buffer used for the safe copy 1904 * was allocated in setup_allocindir_phase2 using 1905 * the physical address so it could be used for this 1906 * purpose. Hence we swap the safe copy with the real 1907 * copy, allowing the safe copy to be freed and holding 1908 * on to the real copy for later use in indir_trunc. 1909 */ 1910 if (indirdep->ir_state & GOINGAWAY) { 1911 FREE_LOCK(&lk); 1912 panic("deallocate_dependencies: already gone"); 1913 } 1914 indirdep->ir_state |= GOINGAWAY; 1915 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 1916 free_allocindir(aip, inodedep); 1917 if (bp->b_lblkno >= 0 || 1918 bp->b_blkno != indirdep->ir_savebp->b_lblkno) { 1919 FREE_LOCK(&lk); 1920 panic("deallocate_dependencies: not indir"); 1921 } 1922 bcopy(bp->b_data, indirdep->ir_savebp->b_data, 1923 bp->b_bcount); 1924 WORKLIST_REMOVE(wk); 1925 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk); 1926 continue; 1927 1928 case D_PAGEDEP: 1929 pagedep = WK_PAGEDEP(wk); 1930 /* 1931 * None of the directory additions will ever be 1932 * visible, so they can simply be tossed. 1933 */ 1934 for (i = 0; i < DAHASHSZ; i++) 1935 while ((dap = 1936 LIST_FIRST(&pagedep->pd_diraddhd[i]))) 1937 free_diradd(dap); 1938 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0) 1939 free_diradd(dap); 1940 /* 1941 * Copy any directory remove dependencies to the list 1942 * to be processed after the zero'ed inode is written. 1943 * If the inode has already been written, then they 1944 * can be dumped directly onto the work list. 1945 */ 1946 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 1947 LIST_REMOVE(dirrem, dm_next); 1948 dirrem->dm_dirinum = pagedep->pd_ino; 1949 if (inodedep == NULL || 1950 (inodedep->id_state & ALLCOMPLETE) == 1951 ALLCOMPLETE) 1952 add_to_worklist(&dirrem->dm_list); 1953 else 1954 WORKLIST_INSERT(&inodedep->id_bufwait, 1955 &dirrem->dm_list); 1956 } 1957 WORKLIST_REMOVE(&pagedep->pd_list); 1958 LIST_REMOVE(pagedep, pd_hash); 1959 WORKITEM_FREE(pagedep, D_PAGEDEP); 1960 continue; 1961 1962 case D_ALLOCINDIR: 1963 free_allocindir(WK_ALLOCINDIR(wk), inodedep); 1964 continue; 1965 1966 case D_ALLOCDIRECT: 1967 case D_INODEDEP: 1968 FREE_LOCK(&lk); 1969 panic("deallocate_dependencies: Unexpected type %s", 1970 TYPENAME(wk->wk_type)); 1971 /* NOTREACHED */ 1972 1973 default: 1974 FREE_LOCK(&lk); 1975 panic("deallocate_dependencies: Unknown type %s", 1976 TYPENAME(wk->wk_type)); 1977 /* NOTREACHED */ 1978 } 1979 } 1980 } 1981 1982 /* 1983 * Free an allocdirect. Generate a new freefrag work request if appropriate. 1984 * This routine must be called with splbio interrupts blocked. 1985 */ 1986 static void 1987 free_allocdirect(adphead, adp, delay) 1988 struct allocdirectlst *adphead; 1989 struct allocdirect *adp; 1990 int delay; 1991 { 1992 1993 #ifdef DEBUG 1994 if (lk.lkt_held == NOHOLDER) 1995 panic("free_allocdirect: lock not held"); 1996 #endif 1997 if ((adp->ad_state & DEPCOMPLETE) == 0) 1998 LIST_REMOVE(adp, ad_deps); 1999 TAILQ_REMOVE(adphead, adp, ad_next); 2000 if ((adp->ad_state & COMPLETE) == 0) 2001 WORKLIST_REMOVE(&adp->ad_list); 2002 if (adp->ad_freefrag != NULL) { 2003 if (delay) 2004 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 2005 &adp->ad_freefrag->ff_list); 2006 else 2007 add_to_worklist(&adp->ad_freefrag->ff_list); 2008 } 2009 WORKITEM_FREE(adp, D_ALLOCDIRECT); 2010 } 2011 2012 /* 2013 * Prepare an inode to be freed. The actual free operation is not 2014 * done until the zero'ed inode has been written to disk. 2015 */ 2016 void 2017 softdep_freefile(pvp, ino, mode) 2018 struct vnode *pvp; 2019 ino_t ino; 2020 int mode; 2021 { 2022 struct inode *ip = VTOI(pvp); 2023 struct inodedep *inodedep; 2024 struct freefile *freefile; 2025 2026 /* 2027 * This sets up the inode de-allocation dependency. 2028 */ 2029 MALLOC(freefile, struct freefile *, sizeof(struct freefile), 2030 M_FREEFILE, M_SOFTDEP_FLAGS); 2031 freefile->fx_list.wk_type = D_FREEFILE; 2032 freefile->fx_list.wk_state = 0; 2033 freefile->fx_mode = mode; 2034 freefile->fx_oldinum = ino; 2035 freefile->fx_devvp = ip->i_devvp; 2036 freefile->fx_fs = ip->i_fs; 2037 2038 /* 2039 * If the inodedep does not exist, then the zero'ed inode has 2040 * been written to disk. If the allocated inode has never been 2041 * written to disk, then the on-disk inode is zero'ed. In either 2042 * case we can free the file immediately. 2043 */ 2044 ACQUIRE_LOCK(&lk); 2045 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 || 2046 check_inode_unwritten(inodedep)) { 2047 FREE_LOCK(&lk); 2048 handle_workitem_freefile(freefile); 2049 return; 2050 } 2051 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 2052 FREE_LOCK(&lk); 2053 } 2054 2055 /* 2056 * Check to see if an inode has never been written to disk. If 2057 * so free the inodedep and return success, otherwise return failure. 2058 * This routine must be called with splbio interrupts blocked. 2059 * 2060 * If we still have a bitmap dependency, then the inode has never 2061 * been written to disk. Drop the dependency as it is no longer 2062 * necessary since the inode is being deallocated. We set the 2063 * ALLCOMPLETE flags since the bitmap now properly shows that the 2064 * inode is not allocated. Even if the inode is actively being 2065 * written, it has been rolled back to its zero'ed state, so we 2066 * are ensured that a zero inode is what is on the disk. For short 2067 * lived files, this change will usually result in removing all the 2068 * dependencies from the inode so that it can be freed immediately. 2069 */ 2070 static int 2071 check_inode_unwritten(inodedep) 2072 struct inodedep *inodedep; 2073 { 2074 2075 if ((inodedep->id_state & DEPCOMPLETE) != 0 || 2076 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2077 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2078 LIST_FIRST(&inodedep->id_inowait) != NULL || 2079 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2080 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2081 inodedep->id_nlinkdelta != 0) 2082 return (0); 2083 inodedep->id_state |= ALLCOMPLETE; 2084 LIST_REMOVE(inodedep, id_deps); 2085 inodedep->id_buf = NULL; 2086 if (inodedep->id_state & ONWORKLIST) 2087 WORKLIST_REMOVE(&inodedep->id_list); 2088 if (inodedep->id_savedino != NULL) { 2089 FREE(inodedep->id_savedino, M_INODEDEP); 2090 inodedep->id_savedino = NULL; 2091 } 2092 if (free_inodedep(inodedep) == 0) { 2093 FREE_LOCK(&lk); 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(inodedep) 2104 struct inodedep *inodedep; 2105 { 2106 2107 if ((inodedep->id_state & ONWORKLIST) != 0 || 2108 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 2109 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2110 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2111 LIST_FIRST(&inodedep->id_inowait) != NULL || 2112 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2113 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2114 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL) 2115 return (0); 2116 LIST_REMOVE(inodedep, id_hash); 2117 WORKITEM_FREE(inodedep, D_INODEDEP); 2118 num_inodedep -= 1; 2119 return (1); 2120 } 2121 2122 /* 2123 * This workitem routine performs the block de-allocation. 2124 * The workitem is added to the pending list after the updated 2125 * inode block has been written to disk. As mentioned above, 2126 * checks regarding the number of blocks de-allocated (compared 2127 * to the number of blocks allocated for the file) are also 2128 * performed in this function. 2129 */ 2130 static void 2131 handle_workitem_freeblocks(freeblks) 2132 struct freeblks *freeblks; 2133 { 2134 struct inode tip; 2135 ufs_daddr_t bn; 2136 struct fs *fs; 2137 int i, level, bsize; 2138 long nblocks, blocksreleased = 0; 2139 int error, allerror = 0; 2140 ufs_lbn_t baselbns[NIADDR], tmpval; 2141 2142 tip.i_number = freeblks->fb_previousinum; 2143 tip.i_devvp = freeblks->fb_devvp; 2144 tip.i_dev = freeblks->fb_devvp->v_rdev; 2145 tip.i_fs = freeblks->fb_fs; 2146 tip.i_size = freeblks->fb_oldsize; 2147 tip.i_uid = freeblks->fb_uid; 2148 fs = freeblks->fb_fs; 2149 tmpval = 1; 2150 baselbns[0] = NDADDR; 2151 for (i = 1; i < NIADDR; i++) { 2152 tmpval *= NINDIR(fs); 2153 baselbns[i] = baselbns[i - 1] + tmpval; 2154 } 2155 nblocks = btodb(fs->fs_bsize); 2156 blocksreleased = 0; 2157 /* 2158 * Indirect blocks first. 2159 */ 2160 for (level = (NIADDR - 1); level >= 0; level--) { 2161 if ((bn = freeblks->fb_iblks[level]) == 0) 2162 continue; 2163 if ((error = indir_trunc(&tip, fsbtodb(fs, bn), level, 2164 baselbns[level], &blocksreleased)) == 0) 2165 allerror = error; 2166 ffs_blkfree(&tip, bn, fs->fs_bsize); 2167 blocksreleased += nblocks; 2168 } 2169 /* 2170 * All direct blocks or frags. 2171 */ 2172 for (i = (NDADDR - 1); i >= 0; i--) { 2173 if ((bn = freeblks->fb_dblks[i]) == 0) 2174 continue; 2175 bsize = blksize(fs, &tip, i); 2176 ffs_blkfree(&tip, bn, bsize); 2177 blocksreleased += btodb(bsize); 2178 } 2179 2180 #ifdef DIAGNOSTIC 2181 if (freeblks->fb_chkcnt != blocksreleased) 2182 printf("handle_workitem_freeblocks: block count\n"); 2183 if (allerror) 2184 softdep_error("handle_workitem_freeblks", allerror); 2185 #endif /* DIAGNOSTIC */ 2186 WORKITEM_FREE(freeblks, D_FREEBLKS); 2187 } 2188 2189 /* 2190 * Release blocks associated with the inode ip and stored in the indirect 2191 * block dbn. If level is greater than SINGLE, the block is an indirect block 2192 * and recursive calls to indirtrunc must be used to cleanse other indirect 2193 * blocks. 2194 */ 2195 static int 2196 indir_trunc(ip, dbn, level, lbn, countp) 2197 struct inode *ip; 2198 ufs_daddr_t dbn; 2199 int level; 2200 ufs_lbn_t lbn; 2201 long *countp; 2202 { 2203 struct buf *bp; 2204 ufs_daddr_t *bap; 2205 ufs_daddr_t nb; 2206 struct fs *fs; 2207 struct worklist *wk; 2208 struct indirdep *indirdep; 2209 int i, lbnadd, nblocks; 2210 int error, allerror = 0; 2211 2212 fs = ip->i_fs; 2213 lbnadd = 1; 2214 for (i = level; i > 0; i--) 2215 lbnadd *= NINDIR(fs); 2216 /* 2217 * Get buffer of block pointers to be freed. This routine is not 2218 * called until the zero'ed inode has been written, so it is safe 2219 * to free blocks as they are encountered. Because the inode has 2220 * been zero'ed, calls to bmap on these blocks will fail. So, we 2221 * have to use the on-disk address and the block device for the 2222 * filesystem to look them up. If the file was deleted before its 2223 * indirect blocks were all written to disk, the routine that set 2224 * us up (deallocate_dependencies) will have arranged to leave 2225 * a complete copy of the indirect block in memory for our use. 2226 * Otherwise we have to read the blocks in from the disk. 2227 */ 2228 ACQUIRE_LOCK(&lk); 2229 if ((bp = incore(ip->i_devvp, dbn)) != NULL && 2230 (wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2231 if (wk->wk_type != D_INDIRDEP || 2232 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp || 2233 (indirdep->ir_state & GOINGAWAY) == 0) { 2234 FREE_LOCK(&lk); 2235 panic("indir_trunc: lost indirdep"); 2236 } 2237 WORKLIST_REMOVE(wk); 2238 WORKITEM_FREE(indirdep, D_INDIRDEP); 2239 if (LIST_FIRST(&bp->b_dep) != NULL) { 2240 FREE_LOCK(&lk); 2241 panic("indir_trunc: dangling dep"); 2242 } 2243 FREE_LOCK(&lk); 2244 } else { 2245 FREE_LOCK(&lk); 2246 error = bread(ip->i_devvp, dbn, (int)fs->fs_bsize, &bp); 2247 if (error) 2248 return (error); 2249 } 2250 /* 2251 * Recursively free indirect blocks. 2252 */ 2253 bap = (ufs_daddr_t *)bp->b_data; 2254 nblocks = btodb(fs->fs_bsize); 2255 for (i = NINDIR(fs) - 1; i >= 0; i--) { 2256 if ((nb = bap[i]) == 0) 2257 continue; 2258 if (level != 0) { 2259 if ((error = indir_trunc(ip, fsbtodb(fs, nb), 2260 level - 1, lbn + (i * lbnadd), countp)) != 0) 2261 allerror = error; 2262 } 2263 ffs_blkfree(ip, nb, fs->fs_bsize); 2264 *countp += nblocks; 2265 } 2266 bp->b_flags |= B_INVAL | B_NOCACHE; 2267 brelse(bp); 2268 return (allerror); 2269 } 2270 2271 /* 2272 * Free an allocindir. 2273 * This routine must be called with splbio interrupts blocked. 2274 */ 2275 static void 2276 free_allocindir(aip, inodedep) 2277 struct allocindir *aip; 2278 struct inodedep *inodedep; 2279 { 2280 struct freefrag *freefrag; 2281 2282 #ifdef DEBUG 2283 if (lk.lkt_held == NOHOLDER) 2284 panic("free_allocindir: lock not held"); 2285 #endif 2286 if ((aip->ai_state & DEPCOMPLETE) == 0) 2287 LIST_REMOVE(aip, ai_deps); 2288 if (aip->ai_state & ONWORKLIST) 2289 WORKLIST_REMOVE(&aip->ai_list); 2290 LIST_REMOVE(aip, ai_next); 2291 if ((freefrag = aip->ai_freefrag) != NULL) { 2292 if (inodedep == NULL) 2293 add_to_worklist(&freefrag->ff_list); 2294 else 2295 WORKLIST_INSERT(&inodedep->id_bufwait, 2296 &freefrag->ff_list); 2297 } 2298 WORKITEM_FREE(aip, D_ALLOCINDIR); 2299 } 2300 2301 /* 2302 * Directory entry addition dependencies. 2303 * 2304 * When adding a new directory entry, the inode (with its incremented link 2305 * count) must be written to disk before the directory entry's pointer to it. 2306 * Also, if the inode is newly allocated, the corresponding freemap must be 2307 * updated (on disk) before the directory entry's pointer. These requirements 2308 * are met via undo/redo on the directory entry's pointer, which consists 2309 * simply of the inode number. 2310 * 2311 * As directory entries are added and deleted, the free space within a 2312 * directory block can become fragmented. The ufs filesystem will compact 2313 * a fragmented directory block to make space for a new entry. When this 2314 * occurs, the offsets of previously added entries change. Any "diradd" 2315 * dependency structures corresponding to these entries must be updated with 2316 * the new offsets. 2317 */ 2318 2319 /* 2320 * This routine is called after the in-memory inode's link 2321 * count has been incremented, but before the directory entry's 2322 * pointer to the inode has been set. 2323 */ 2324 void 2325 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp) 2326 struct buf *bp; /* buffer containing directory block */ 2327 struct inode *dp; /* inode for directory */ 2328 off_t diroffset; /* offset of new entry in directory */ 2329 ino_t newinum; /* inode referenced by new directory entry */ 2330 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 2331 { 2332 int offset; /* offset of new entry within directory block */ 2333 ufs_lbn_t lbn; /* block in directory containing new entry */ 2334 struct fs *fs; 2335 struct diradd *dap; 2336 struct pagedep *pagedep; 2337 struct inodedep *inodedep; 2338 struct mkdir *mkdir1, *mkdir2; 2339 2340 /* 2341 * Whiteouts have no dependencies. 2342 */ 2343 if (newinum == WINO) { 2344 if (newdirbp != NULL) 2345 bdwrite(newdirbp); 2346 return; 2347 } 2348 2349 fs = dp->i_fs; 2350 lbn = lblkno(fs, diroffset); 2351 offset = blkoff(fs, diroffset); 2352 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, 2353 M_SOFTDEP_FLAGS); 2354 bzero(dap, sizeof(struct diradd)); 2355 dap->da_list.wk_type = D_DIRADD; 2356 dap->da_offset = offset; 2357 dap->da_newinum = newinum; 2358 dap->da_state = ATTACHED; 2359 if (newdirbp == NULL) { 2360 dap->da_state |= DEPCOMPLETE; 2361 ACQUIRE_LOCK(&lk); 2362 } else { 2363 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 2364 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2365 M_SOFTDEP_FLAGS); 2366 mkdir1->md_list.wk_type = D_MKDIR; 2367 mkdir1->md_state = MKDIR_BODY; 2368 mkdir1->md_diradd = dap; 2369 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2370 M_SOFTDEP_FLAGS); 2371 mkdir2->md_list.wk_type = D_MKDIR; 2372 mkdir2->md_state = MKDIR_PARENT; 2373 mkdir2->md_diradd = dap; 2374 /* 2375 * Dependency on "." and ".." being written to disk. 2376 */ 2377 mkdir1->md_buf = newdirbp; 2378 ACQUIRE_LOCK(&lk); 2379 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 2380 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list); 2381 FREE_LOCK(&lk); 2382 bdwrite(newdirbp); 2383 /* 2384 * Dependency on link count increase for parent directory 2385 */ 2386 ACQUIRE_LOCK(&lk); 2387 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0 2388 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2389 dap->da_state &= ~MKDIR_PARENT; 2390 WORKITEM_FREE(mkdir2, D_MKDIR); 2391 } else { 2392 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 2393 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 2394 } 2395 } 2396 /* 2397 * Link into parent directory pagedep to await its being written. 2398 */ 2399 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2400 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2401 dap->da_pagedep = pagedep; 2402 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 2403 da_pdlist); 2404 /* 2405 * Link into its inodedep. Put it on the id_bufwait list if the inode 2406 * is not yet written. If it is written, do the post-inode write 2407 * processing to put it on the id_pendinghd list. 2408 */ 2409 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep); 2410 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 2411 diradd_inode_written(dap, inodedep); 2412 else 2413 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2414 FREE_LOCK(&lk); 2415 } 2416 2417 /* 2418 * This procedure is called to change the offset of a directory 2419 * entry when compacting a directory block which must be owned 2420 * exclusively by the caller. Note that the actual entry movement 2421 * must be done in this procedure to ensure that no I/O completions 2422 * occur while the move is in progress. 2423 */ 2424 void 2425 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize) 2426 struct inode *dp; /* inode for directory */ 2427 caddr_t base; /* address of dp->i_offset */ 2428 caddr_t oldloc; /* address of old directory location */ 2429 caddr_t newloc; /* address of new directory location */ 2430 int entrysize; /* size of directory entry */ 2431 { 2432 int offset, oldoffset, newoffset; 2433 struct pagedep *pagedep; 2434 struct diradd *dap; 2435 ufs_lbn_t lbn; 2436 2437 ACQUIRE_LOCK(&lk); 2438 lbn = lblkno(dp->i_fs, dp->i_offset); 2439 offset = blkoff(dp->i_fs, dp->i_offset); 2440 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) 2441 goto done; 2442 oldoffset = offset + (oldloc - base); 2443 newoffset = offset + (newloc - base); 2444 2445 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) { 2446 if (dap->da_offset != oldoffset) 2447 continue; 2448 dap->da_offset = newoffset; 2449 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) 2450 break; 2451 LIST_REMOVE(dap, da_pdlist); 2452 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], 2453 dap, da_pdlist); 2454 break; 2455 } 2456 if (dap == NULL) { 2457 2458 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) { 2459 if (dap->da_offset == oldoffset) { 2460 dap->da_offset = newoffset; 2461 break; 2462 } 2463 } 2464 } 2465 done: 2466 bcopy(oldloc, newloc, entrysize); 2467 FREE_LOCK(&lk); 2468 } 2469 2470 /* 2471 * Free a diradd dependency structure. This routine must be called 2472 * with splbio interrupts blocked. 2473 */ 2474 static void 2475 free_diradd(dap) 2476 struct diradd *dap; 2477 { 2478 struct dirrem *dirrem; 2479 struct pagedep *pagedep; 2480 struct inodedep *inodedep; 2481 struct mkdir *mkdir, *nextmd; 2482 2483 #ifdef DEBUG 2484 if (lk.lkt_held == NOHOLDER) 2485 panic("free_diradd: lock not held"); 2486 #endif 2487 WORKLIST_REMOVE(&dap->da_list); 2488 LIST_REMOVE(dap, da_pdlist); 2489 if ((dap->da_state & DIRCHG) == 0) { 2490 pagedep = dap->da_pagedep; 2491 } else { 2492 dirrem = dap->da_previous; 2493 pagedep = dirrem->dm_pagedep; 2494 dirrem->dm_dirinum = pagedep->pd_ino; 2495 add_to_worklist(&dirrem->dm_list); 2496 } 2497 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum, 2498 0, &inodedep) != 0) 2499 (void) free_inodedep(inodedep); 2500 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2501 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 2502 nextmd = LIST_NEXT(mkdir, md_mkdirs); 2503 if (mkdir->md_diradd != dap) 2504 continue; 2505 dap->da_state &= ~mkdir->md_state; 2506 WORKLIST_REMOVE(&mkdir->md_list); 2507 LIST_REMOVE(mkdir, md_mkdirs); 2508 WORKITEM_FREE(mkdir, D_MKDIR); 2509 } 2510 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2511 FREE_LOCK(&lk); 2512 panic("free_diradd: unfound ref"); 2513 } 2514 } 2515 WORKITEM_FREE(dap, D_DIRADD); 2516 } 2517 2518 /* 2519 * Directory entry removal dependencies. 2520 * 2521 * When removing a directory entry, the entry's inode pointer must be 2522 * zero'ed on disk before the corresponding inode's link count is decremented 2523 * (possibly freeing the inode for re-use). This dependency is handled by 2524 * updating the directory entry but delaying the inode count reduction until 2525 * after the directory block has been written to disk. After this point, the 2526 * inode count can be decremented whenever it is convenient. 2527 */ 2528 2529 /* 2530 * This routine should be called immediately after removing 2531 * a directory entry. The inode's link count should not be 2532 * decremented by the calling procedure -- the soft updates 2533 * code will do this task when it is safe. 2534 */ 2535 void 2536 softdep_setup_remove(bp, dp, ip, isrmdir) 2537 struct buf *bp; /* buffer containing directory block */ 2538 struct inode *dp; /* inode for the directory being modified */ 2539 struct inode *ip; /* inode for directory entry being removed */ 2540 int isrmdir; /* indicates if doing RMDIR */ 2541 { 2542 struct dirrem *dirrem, *prevdirrem; 2543 2544 /* 2545 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 2546 */ 2547 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2548 2549 /* 2550 * If the COMPLETE flag is clear, then there were no active 2551 * entries and we want to roll back to a zeroed entry until 2552 * the new inode is committed to disk. If the COMPLETE flag is 2553 * set then we have deleted an entry that never made it to 2554 * disk. If the entry we deleted resulted from a name change, 2555 * then the old name still resides on disk. We cannot delete 2556 * its inode (returned to us in prevdirrem) until the zeroed 2557 * directory entry gets to disk. The new inode has never been 2558 * referenced on the disk, so can be deleted immediately. 2559 */ 2560 if ((dirrem->dm_state & COMPLETE) == 0) { 2561 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 2562 dm_next); 2563 FREE_LOCK(&lk); 2564 } else { 2565 if (prevdirrem != NULL) 2566 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 2567 prevdirrem, dm_next); 2568 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 2569 FREE_LOCK(&lk); 2570 handle_workitem_remove(dirrem); 2571 } 2572 } 2573 2574 /* 2575 * Allocate a new dirrem if appropriate and return it along with 2576 * its associated pagedep. Called without a lock, returns with lock. 2577 */ 2578 static long num_dirrem; /* number of dirrem allocated */ 2579 static struct dirrem * 2580 newdirrem(bp, dp, ip, isrmdir, prevdirremp) 2581 struct buf *bp; /* buffer containing directory block */ 2582 struct inode *dp; /* inode for the directory being modified */ 2583 struct inode *ip; /* inode for directory entry being removed */ 2584 int isrmdir; /* indicates if doing RMDIR */ 2585 struct dirrem **prevdirremp; /* previously referenced inode, if any */ 2586 { 2587 int offset; 2588 ufs_lbn_t lbn; 2589 struct diradd *dap; 2590 struct dirrem *dirrem; 2591 struct pagedep *pagedep; 2592 2593 /* 2594 * Whiteouts have no deletion dependencies. 2595 */ 2596 if (ip == NULL) 2597 panic("newdirrem: whiteout"); 2598 /* 2599 * If we are over our limit, try to improve the situation. 2600 * Limiting the number of dirrem structures will also limit 2601 * the number of freefile and freeblks structures. 2602 */ 2603 if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0) 2604 (void) request_cleanup(FLUSH_REMOVE, 0); 2605 num_dirrem += 1; 2606 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem), 2607 M_DIRREM, M_SOFTDEP_FLAGS); 2608 bzero(dirrem, sizeof(struct dirrem)); 2609 dirrem->dm_list.wk_type = D_DIRREM; 2610 dirrem->dm_state = isrmdir ? RMDIR : 0; 2611 dirrem->dm_mnt = ITOV(ip)->v_mount; 2612 dirrem->dm_oldinum = ip->i_number; 2613 *prevdirremp = NULL; 2614 2615 ACQUIRE_LOCK(&lk); 2616 lbn = lblkno(dp->i_fs, dp->i_offset); 2617 offset = blkoff(dp->i_fs, dp->i_offset); 2618 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2619 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2620 dirrem->dm_pagedep = pagedep; 2621 /* 2622 * Check for a diradd dependency for the same directory entry. 2623 * If present, then both dependencies become obsolete and can 2624 * be de-allocated. Check for an entry on both the pd_dirraddhd 2625 * list and the pd_pendinghd list. 2626 */ 2627 2628 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 2629 if (dap->da_offset == offset) 2630 break; 2631 if (dap == NULL) { 2632 2633 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 2634 if (dap->da_offset == offset) 2635 break; 2636 if (dap == NULL) 2637 return (dirrem); 2638 } 2639 /* 2640 * Must be ATTACHED at this point. 2641 */ 2642 if ((dap->da_state & ATTACHED) == 0) { 2643 FREE_LOCK(&lk); 2644 panic("newdirrem: not ATTACHED"); 2645 } 2646 if (dap->da_newinum != ip->i_number) { 2647 FREE_LOCK(&lk); 2648 panic("newdirrem: inum %d should be %d", 2649 ip->i_number, dap->da_newinum); 2650 } 2651 /* 2652 * If we are deleting a changed name that never made it to disk, 2653 * then return the dirrem describing the previous inode (which 2654 * represents the inode currently referenced from this entry on disk). 2655 */ 2656 if ((dap->da_state & DIRCHG) != 0) { 2657 *prevdirremp = dap->da_previous; 2658 dap->da_state &= ~DIRCHG; 2659 dap->da_pagedep = pagedep; 2660 } 2661 /* 2662 * We are deleting an entry that never made it to disk. 2663 * Mark it COMPLETE so we can delete its inode immediately. 2664 */ 2665 dirrem->dm_state |= COMPLETE; 2666 free_diradd(dap); 2667 return (dirrem); 2668 } 2669 2670 /* 2671 * Directory entry change dependencies. 2672 * 2673 * Changing an existing directory entry requires that an add operation 2674 * be completed first followed by a deletion. The semantics for the addition 2675 * are identical to the description of adding a new entry above except 2676 * that the rollback is to the old inode number rather than zero. Once 2677 * the addition dependency is completed, the removal is done as described 2678 * in the removal routine above. 2679 */ 2680 2681 /* 2682 * This routine should be called immediately after changing 2683 * a directory entry. The inode's link count should not be 2684 * decremented by the calling procedure -- the soft updates 2685 * code will perform this task when it is safe. 2686 */ 2687 void 2688 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 2689 struct buf *bp; /* buffer containing directory block */ 2690 struct inode *dp; /* inode for the directory being modified */ 2691 struct inode *ip; /* inode for directory entry being removed */ 2692 ino_t newinum; /* new inode number for changed entry */ 2693 int isrmdir; /* indicates if doing RMDIR */ 2694 { 2695 int offset; 2696 struct diradd *dap = NULL; 2697 struct dirrem *dirrem, *prevdirrem; 2698 struct pagedep *pagedep; 2699 struct inodedep *inodedep; 2700 2701 offset = blkoff(dp->i_fs, dp->i_offset); 2702 2703 /* 2704 * Whiteouts do not need diradd dependencies. 2705 */ 2706 if (newinum != WINO) { 2707 MALLOC(dap, struct diradd *, sizeof(struct diradd), 2708 M_DIRADD, M_SOFTDEP_FLAGS); 2709 bzero(dap, sizeof(struct diradd)); 2710 dap->da_list.wk_type = D_DIRADD; 2711 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 2712 dap->da_offset = offset; 2713 dap->da_newinum = newinum; 2714 } 2715 2716 /* 2717 * Allocate a new dirrem and ACQUIRE_LOCK. 2718 */ 2719 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2720 pagedep = dirrem->dm_pagedep; 2721 /* 2722 * The possible values for isrmdir: 2723 * 0 - non-directory file rename 2724 * 1 - directory rename within same directory 2725 * inum - directory rename to new directory of given inode number 2726 * When renaming to a new directory, we are both deleting and 2727 * creating a new directory entry, so the link count on the new 2728 * directory should not change. Thus we do not need the followup 2729 * dirrem which is usually done in handle_workitem_remove. We set 2730 * the DIRCHG flag to tell handle_workitem_remove to skip the 2731 * followup dirrem. 2732 */ 2733 if (isrmdir > 1) 2734 dirrem->dm_state |= DIRCHG; 2735 2736 /* 2737 * Whiteouts have no additional dependencies, 2738 * so just put the dirrem on the correct list. 2739 */ 2740 if (newinum == WINO) { 2741 if ((dirrem->dm_state & COMPLETE) == 0) { 2742 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 2743 dm_next); 2744 } else { 2745 dirrem->dm_dirinum = pagedep->pd_ino; 2746 add_to_worklist(&dirrem->dm_list); 2747 } 2748 FREE_LOCK(&lk); 2749 return; 2750 } 2751 2752 /* 2753 * If the COMPLETE flag is clear, then there were no active 2754 * entries and we want to roll back to the previous inode until 2755 * the new inode is committed to disk. If the COMPLETE flag is 2756 * set, then we have deleted an entry that never made it to disk. 2757 * If the entry we deleted resulted from a name change, then the old 2758 * inode reference still resides on disk. Any rollback that we do 2759 * needs to be to that old inode (returned to us in prevdirrem). If 2760 * the entry we deleted resulted from a create, then there is 2761 * no entry on the disk, so we want to roll back to zero rather 2762 * than the uncommitted inode. In either of the COMPLETE cases we 2763 * want to immediately free the unwritten and unreferenced inode. 2764 */ 2765 if ((dirrem->dm_state & COMPLETE) == 0) { 2766 dap->da_previous = dirrem; 2767 } else { 2768 if (prevdirrem != NULL) { 2769 dap->da_previous = prevdirrem; 2770 } else { 2771 dap->da_state &= ~DIRCHG; 2772 dap->da_pagedep = pagedep; 2773 } 2774 dirrem->dm_dirinum = pagedep->pd_ino; 2775 add_to_worklist(&dirrem->dm_list); 2776 } 2777 /* 2778 * Link into its inodedep. Put it on the id_bufwait list if the inode 2779 * is not yet written. If it is written, do the post-inode write 2780 * processing to put it on the id_pendinghd list. 2781 */ 2782 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 || 2783 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2784 dap->da_state |= COMPLETE; 2785 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 2786 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 2787 } else { 2788 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 2789 dap, da_pdlist); 2790 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2791 } 2792 FREE_LOCK(&lk); 2793 } 2794 2795 /* 2796 * Called whenever the link count on an inode is changed. 2797 * It creates an inode dependency so that the new reference(s) 2798 * to the inode cannot be committed to disk until the updated 2799 * inode has been written. 2800 */ 2801 void 2802 softdep_change_linkcnt(ip) 2803 struct inode *ip; /* the inode with the increased link count */ 2804 { 2805 struct inodedep *inodedep; 2806 2807 ACQUIRE_LOCK(&lk); 2808 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 2809 if (ip->i_nlink < ip->i_effnlink) { 2810 FREE_LOCK(&lk); 2811 panic("softdep_change_linkcnt: bad delta"); 2812 } 2813 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2814 FREE_LOCK(&lk); 2815 } 2816 2817 /* 2818 * This workitem decrements the inode's link count. 2819 * If the link count reaches zero, the file is removed. 2820 */ 2821 static void 2822 handle_workitem_remove(dirrem) 2823 struct dirrem *dirrem; 2824 { 2825 struct thread *td = curthread; /* XXX */ 2826 struct inodedep *inodedep; 2827 struct vnode *vp; 2828 struct inode *ip; 2829 ino_t oldinum; 2830 int error; 2831 2832 if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) { 2833 softdep_error("handle_workitem_remove: vget", error); 2834 return; 2835 } 2836 ip = VTOI(vp); 2837 ACQUIRE_LOCK(&lk); 2838 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){ 2839 FREE_LOCK(&lk); 2840 panic("handle_workitem_remove: lost inodedep"); 2841 } 2842 /* 2843 * Normal file deletion. 2844 */ 2845 if ((dirrem->dm_state & RMDIR) == 0) { 2846 ip->i_nlink--; 2847 ip->i_flag |= IN_CHANGE; 2848 if (ip->i_nlink < ip->i_effnlink) { 2849 FREE_LOCK(&lk); 2850 panic("handle_workitem_remove: bad file delta"); 2851 } 2852 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2853 FREE_LOCK(&lk); 2854 vput(vp); 2855 num_dirrem -= 1; 2856 WORKITEM_FREE(dirrem, D_DIRREM); 2857 return; 2858 } 2859 /* 2860 * Directory deletion. Decrement reference count for both the 2861 * just deleted parent directory entry and the reference for ".". 2862 * Next truncate the directory to length zero. When the 2863 * truncation completes, arrange to have the reference count on 2864 * the parent decremented to account for the loss of "..". 2865 */ 2866 ip->i_nlink -= 2; 2867 ip->i_flag |= IN_CHANGE; 2868 if (ip->i_nlink < ip->i_effnlink) { 2869 FREE_LOCK(&lk); 2870 panic("handle_workitem_remove: bad dir delta"); 2871 } 2872 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2873 FREE_LOCK(&lk); 2874 if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, proc0.p_ucred, td)) != 0) 2875 softdep_error("handle_workitem_remove: truncate", error); 2876 /* 2877 * Rename a directory to a new parent. Since, we are both deleting 2878 * and creating a new directory entry, the link count on the new 2879 * directory should not change. Thus we skip the followup dirrem. 2880 */ 2881 if (dirrem->dm_state & DIRCHG) { 2882 vput(vp); 2883 num_dirrem -= 1; 2884 WORKITEM_FREE(dirrem, D_DIRREM); 2885 return; 2886 } 2887 /* 2888 * If the inodedep does not exist, then the zero'ed inode has 2889 * been written to disk. If the allocated inode has never been 2890 * written to disk, then the on-disk inode is zero'ed. In either 2891 * case we can remove the file immediately. 2892 */ 2893 ACQUIRE_LOCK(&lk); 2894 dirrem->dm_state = 0; 2895 oldinum = dirrem->dm_oldinum; 2896 dirrem->dm_oldinum = dirrem->dm_dirinum; 2897 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 || 2898 check_inode_unwritten(inodedep)) { 2899 FREE_LOCK(&lk); 2900 vput(vp); 2901 handle_workitem_remove(dirrem); 2902 return; 2903 } 2904 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 2905 FREE_LOCK(&lk); 2906 vput(vp); 2907 } 2908 2909 /* 2910 * Inode de-allocation dependencies. 2911 * 2912 * When an inode's link count is reduced to zero, it can be de-allocated. We 2913 * found it convenient to postpone de-allocation until after the inode is 2914 * written to disk with its new link count (zero). At this point, all of the 2915 * on-disk inode's block pointers are nullified and, with careful dependency 2916 * list ordering, all dependencies related to the inode will be satisfied and 2917 * the corresponding dependency structures de-allocated. So, if/when the 2918 * inode is reused, there will be no mixing of old dependencies with new 2919 * ones. This artificial dependency is set up by the block de-allocation 2920 * procedure above (softdep_setup_freeblocks) and completed by the 2921 * following procedure. 2922 */ 2923 static void 2924 handle_workitem_freefile(freefile) 2925 struct freefile *freefile; 2926 { 2927 struct vnode vp; 2928 struct inode tip; 2929 struct inodedep *idp; 2930 int error; 2931 2932 #ifdef DEBUG 2933 ACQUIRE_LOCK(&lk); 2934 error = inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp); 2935 FREE_LOCK(&lk); 2936 if (error) 2937 panic("handle_workitem_freefile: inodedep survived"); 2938 #endif 2939 tip.i_devvp = freefile->fx_devvp; 2940 tip.i_dev = freefile->fx_devvp->v_rdev; 2941 tip.i_fs = freefile->fx_fs; 2942 vp.v_data = &tip; 2943 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0) 2944 softdep_error("handle_workitem_freefile", error); 2945 WORKITEM_FREE(freefile, D_FREEFILE); 2946 } 2947 2948 /* 2949 * Disk writes. 2950 * 2951 * The dependency structures constructed above are most actively used when file 2952 * system blocks are written to disk. No constraints are placed on when a 2953 * block can be written, but unsatisfied update dependencies are made safe by 2954 * modifying (or replacing) the source memory for the duration of the disk 2955 * write. When the disk write completes, the memory block is again brought 2956 * up-to-date. 2957 * 2958 * In-core inode structure reclamation. 2959 * 2960 * Because there are a finite number of "in-core" inode structures, they are 2961 * reused regularly. By transferring all inode-related dependencies to the 2962 * in-memory inode block and indexing them separately (via "inodedep"s), we 2963 * can allow "in-core" inode structures to be reused at any time and avoid 2964 * any increase in contention. 2965 * 2966 * Called just before entering the device driver to initiate a new disk I/O. 2967 * The buffer must be locked, thus, no I/O completion operations can occur 2968 * while we are manipulating its associated dependencies. 2969 */ 2970 static void 2971 softdep_disk_io_initiation(bp) 2972 struct buf *bp; /* structure describing disk write to occur */ 2973 { 2974 struct worklist *wk, *nextwk; 2975 struct indirdep *indirdep; 2976 2977 /* 2978 * We only care about write operations. There should never 2979 * be dependencies for reads. 2980 */ 2981 if (bp->b_flags & B_READ) 2982 panic("softdep_disk_io_initiation: read"); 2983 /* 2984 * Do any necessary pre-I/O processing. 2985 */ 2986 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = nextwk) { 2987 nextwk = LIST_NEXT(wk, wk_list); 2988 switch (wk->wk_type) { 2989 2990 case D_PAGEDEP: 2991 initiate_write_filepage(WK_PAGEDEP(wk), bp); 2992 continue; 2993 2994 case D_INODEDEP: 2995 initiate_write_inodeblock(WK_INODEDEP(wk), bp); 2996 continue; 2997 2998 case D_INDIRDEP: 2999 indirdep = WK_INDIRDEP(wk); 3000 if (indirdep->ir_state & GOINGAWAY) 3001 panic("disk_io_initiation: indirdep gone"); 3002 /* 3003 * If there are no remaining dependencies, this 3004 * will be writing the real pointers, so the 3005 * dependency can be freed. 3006 */ 3007 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) { 3008 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 3009 brelse(indirdep->ir_savebp); 3010 /* inline expand WORKLIST_REMOVE(wk); */ 3011 wk->wk_state &= ~ONWORKLIST; 3012 LIST_REMOVE(wk, wk_list); 3013 WORKITEM_FREE(indirdep, D_INDIRDEP); 3014 continue; 3015 } 3016 /* 3017 * Replace up-to-date version with safe version. 3018 */ 3019 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount, 3020 M_INDIRDEP, M_SOFTDEP_FLAGS); 3021 ACQUIRE_LOCK(&lk); 3022 indirdep->ir_state &= ~ATTACHED; 3023 indirdep->ir_state |= UNDONE; 3024 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 3025 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 3026 bp->b_bcount); 3027 FREE_LOCK(&lk); 3028 continue; 3029 3030 case D_MKDIR: 3031 case D_BMSAFEMAP: 3032 case D_ALLOCDIRECT: 3033 case D_ALLOCINDIR: 3034 continue; 3035 3036 default: 3037 panic("handle_disk_io_initiation: Unexpected type %s", 3038 TYPENAME(wk->wk_type)); 3039 /* NOTREACHED */ 3040 } 3041 } 3042 } 3043 3044 /* 3045 * Called from within the procedure above to deal with unsatisfied 3046 * allocation dependencies in a directory. The buffer must be locked, 3047 * thus, no I/O completion operations can occur while we are 3048 * manipulating its associated dependencies. 3049 */ 3050 static void 3051 initiate_write_filepage(pagedep, bp) 3052 struct pagedep *pagedep; 3053 struct buf *bp; 3054 { 3055 struct diradd *dap; 3056 struct direct *ep; 3057 int i; 3058 3059 if (pagedep->pd_state & IOSTARTED) { 3060 /* 3061 * This can only happen if there is a driver that does not 3062 * understand chaining. Here biodone will reissue the call 3063 * to strategy for the incomplete buffers. 3064 */ 3065 printf("initiate_write_filepage: already started\n"); 3066 return; 3067 } 3068 pagedep->pd_state |= IOSTARTED; 3069 ACQUIRE_LOCK(&lk); 3070 for (i = 0; i < DAHASHSZ; i++) { 3071 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 3072 ep = (struct direct *) 3073 ((char *)bp->b_data + dap->da_offset); 3074 if (ep->d_ino != dap->da_newinum) { 3075 FREE_LOCK(&lk); 3076 panic("%s: dir inum %d != new %d", 3077 "initiate_write_filepage", 3078 ep->d_ino, dap->da_newinum); 3079 } 3080 if (dap->da_state & DIRCHG) 3081 ep->d_ino = dap->da_previous->dm_oldinum; 3082 else 3083 ep->d_ino = 0; 3084 dap->da_state &= ~ATTACHED; 3085 dap->da_state |= UNDONE; 3086 } 3087 } 3088 FREE_LOCK(&lk); 3089 } 3090 3091 /* 3092 * Called from within the procedure above to deal with unsatisfied 3093 * allocation dependencies in an inodeblock. The buffer must be 3094 * locked, thus, no I/O completion operations can occur while we 3095 * are manipulating its associated dependencies. 3096 */ 3097 static void 3098 initiate_write_inodeblock(inodedep, bp) 3099 struct inodedep *inodedep; 3100 struct buf *bp; /* The inode block */ 3101 { 3102 struct allocdirect *adp, *lastadp; 3103 struct dinode *dp; 3104 struct fs *fs; 3105 ufs_lbn_t prevlbn = 0; 3106 int i, deplist; 3107 3108 if (inodedep->id_state & IOSTARTED) 3109 panic("initiate_write_inodeblock: already started"); 3110 inodedep->id_state |= IOSTARTED; 3111 fs = inodedep->id_fs; 3112 dp = (struct dinode *)bp->b_data + 3113 ino_to_fsbo(fs, inodedep->id_ino); 3114 /* 3115 * If the bitmap is not yet written, then the allocated 3116 * inode cannot be written to disk. 3117 */ 3118 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 3119 if (inodedep->id_savedino != NULL) 3120 panic("initiate_write_inodeblock: already doing I/O"); 3121 MALLOC(inodedep->id_savedino, struct dinode *, 3122 sizeof(struct dinode), M_INODEDEP, M_SOFTDEP_FLAGS); 3123 *inodedep->id_savedino = *dp; 3124 bzero((caddr_t)dp, sizeof(struct dinode)); 3125 return; 3126 } 3127 /* 3128 * If no dependencies, then there is nothing to roll back. 3129 */ 3130 inodedep->id_savedsize = dp->di_size; 3131 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 3132 return; 3133 /* 3134 * Set the dependencies to busy. 3135 */ 3136 ACQUIRE_LOCK(&lk); 3137 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3138 adp = TAILQ_NEXT(adp, ad_next)) { 3139 #ifdef DIAGNOSTIC 3140 if (deplist != 0 && prevlbn >= adp->ad_lbn) { 3141 FREE_LOCK(&lk); 3142 panic("softdep_write_inodeblock: lbn order"); 3143 } 3144 prevlbn = adp->ad_lbn; 3145 if (adp->ad_lbn < NDADDR && 3146 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) { 3147 FREE_LOCK(&lk); 3148 panic("%s: direct pointer #%ld mismatch %d != %d", 3149 "softdep_write_inodeblock", adp->ad_lbn, 3150 dp->di_db[adp->ad_lbn], adp->ad_newblkno); 3151 } 3152 if (adp->ad_lbn >= NDADDR && 3153 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) { 3154 FREE_LOCK(&lk); 3155 panic("%s: indirect pointer #%ld mismatch %d != %d", 3156 "softdep_write_inodeblock", adp->ad_lbn - NDADDR, 3157 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno); 3158 } 3159 deplist |= 1 << adp->ad_lbn; 3160 if ((adp->ad_state & ATTACHED) == 0) { 3161 FREE_LOCK(&lk); 3162 panic("softdep_write_inodeblock: Unknown state 0x%x", 3163 adp->ad_state); 3164 } 3165 #endif /* DIAGNOSTIC */ 3166 adp->ad_state &= ~ATTACHED; 3167 adp->ad_state |= UNDONE; 3168 } 3169 /* 3170 * The on-disk inode cannot claim to be any larger than the last 3171 * fragment that has been written. Otherwise, the on-disk inode 3172 * might have fragments that were not the last block in the file 3173 * which would corrupt the filesystem. 3174 */ 3175 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3176 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 3177 if (adp->ad_lbn >= NDADDR) 3178 break; 3179 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 3180 /* keep going until hitting a rollback to a frag */ 3181 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 3182 continue; 3183 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 3184 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 3185 #ifdef DIAGNOSTIC 3186 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) { 3187 FREE_LOCK(&lk); 3188 panic("softdep_write_inodeblock: lost dep1"); 3189 } 3190 #endif /* DIAGNOSTIC */ 3191 dp->di_db[i] = 0; 3192 } 3193 for (i = 0; i < NIADDR; i++) { 3194 #ifdef DIAGNOSTIC 3195 if (dp->di_ib[i] != 0 && 3196 (deplist & ((1 << NDADDR) << i)) == 0) { 3197 FREE_LOCK(&lk); 3198 panic("softdep_write_inodeblock: lost dep2"); 3199 } 3200 #endif /* DIAGNOSTIC */ 3201 dp->di_ib[i] = 0; 3202 } 3203 FREE_LOCK(&lk); 3204 return; 3205 } 3206 /* 3207 * If we have zero'ed out the last allocated block of the file, 3208 * roll back the size to the last currently allocated block. 3209 * We know that this last allocated block is a full-sized as 3210 * we already checked for fragments in the loop above. 3211 */ 3212 if (lastadp != NULL && 3213 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 3214 for (i = lastadp->ad_lbn; i >= 0; i--) 3215 if (dp->di_db[i] != 0) 3216 break; 3217 dp->di_size = (i + 1) * fs->fs_bsize; 3218 } 3219 /* 3220 * The only dependencies are for indirect blocks. 3221 * 3222 * The file size for indirect block additions is not guaranteed. 3223 * Such a guarantee would be non-trivial to achieve. The conventional 3224 * synchronous write implementation also does not make this guarantee. 3225 * Fsck should catch and fix discrepancies. Arguably, the file size 3226 * can be over-estimated without destroying integrity when the file 3227 * moves into the indirect blocks (i.e., is large). If we want to 3228 * postpone fsck, we are stuck with this argument. 3229 */ 3230 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 3231 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 3232 FREE_LOCK(&lk); 3233 } 3234 3235 /* 3236 * This routine is called during the completion interrupt 3237 * service routine for a disk write (from the procedure called 3238 * by the device driver to inform the filesystem caches of 3239 * a request completion). It should be called early in this 3240 * procedure, before the block is made available to other 3241 * processes or other routines are called. 3242 */ 3243 static void 3244 softdep_disk_write_complete(bp) 3245 struct buf *bp; /* describes the completed disk write */ 3246 { 3247 struct worklist *wk; 3248 struct workhead reattach; 3249 struct newblk *newblk; 3250 struct allocindir *aip; 3251 struct allocdirect *adp; 3252 struct indirdep *indirdep; 3253 struct inodedep *inodedep; 3254 struct bmsafemap *bmsafemap; 3255 3256 #ifdef DEBUG 3257 if (lk.lkt_held != NOHOLDER) 3258 panic("softdep_disk_write_complete: lock is held"); 3259 lk.lkt_held = SPECIAL_FLAG; 3260 #endif 3261 LIST_INIT(&reattach); 3262 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 3263 WORKLIST_REMOVE(wk); 3264 switch (wk->wk_type) { 3265 3266 case D_PAGEDEP: 3267 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 3268 WORKLIST_INSERT(&reattach, wk); 3269 continue; 3270 3271 case D_INODEDEP: 3272 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 3273 WORKLIST_INSERT(&reattach, wk); 3274 continue; 3275 3276 case D_BMSAFEMAP: 3277 bmsafemap = WK_BMSAFEMAP(wk); 3278 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) { 3279 newblk->nb_state |= DEPCOMPLETE; 3280 newblk->nb_bmsafemap = NULL; 3281 LIST_REMOVE(newblk, nb_deps); 3282 } 3283 while ((adp = 3284 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) { 3285 adp->ad_state |= DEPCOMPLETE; 3286 adp->ad_buf = NULL; 3287 LIST_REMOVE(adp, ad_deps); 3288 handle_allocdirect_partdone(adp); 3289 } 3290 while ((aip = 3291 LIST_FIRST(&bmsafemap->sm_allocindirhd))) { 3292 aip->ai_state |= DEPCOMPLETE; 3293 aip->ai_buf = NULL; 3294 LIST_REMOVE(aip, ai_deps); 3295 handle_allocindir_partdone(aip); 3296 } 3297 while ((inodedep = 3298 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { 3299 inodedep->id_state |= DEPCOMPLETE; 3300 LIST_REMOVE(inodedep, id_deps); 3301 inodedep->id_buf = NULL; 3302 } 3303 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 3304 continue; 3305 3306 case D_MKDIR: 3307 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 3308 continue; 3309 3310 case D_ALLOCDIRECT: 3311 adp = WK_ALLOCDIRECT(wk); 3312 adp->ad_state |= COMPLETE; 3313 handle_allocdirect_partdone(adp); 3314 continue; 3315 3316 case D_ALLOCINDIR: 3317 aip = WK_ALLOCINDIR(wk); 3318 aip->ai_state |= COMPLETE; 3319 handle_allocindir_partdone(aip); 3320 continue; 3321 3322 case D_INDIRDEP: 3323 indirdep = WK_INDIRDEP(wk); 3324 if (indirdep->ir_state & GOINGAWAY) { 3325 lk.lkt_held = NOHOLDER; 3326 panic("disk_write_complete: indirdep gone"); 3327 } 3328 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 3329 FREE(indirdep->ir_saveddata, M_INDIRDEP); 3330 indirdep->ir_saveddata = 0; 3331 indirdep->ir_state &= ~UNDONE; 3332 indirdep->ir_state |= ATTACHED; 3333 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 3334 handle_allocindir_partdone(aip); 3335 if (aip == LIST_FIRST(&indirdep->ir_donehd)) { 3336 lk.lkt_held = NOHOLDER; 3337 panic("disk_write_complete: not gone"); 3338 } 3339 } 3340 WORKLIST_INSERT(&reattach, wk); 3341 if ((bp->b_flags & B_DELWRI) == 0) 3342 stat_indir_blk_ptrs++; 3343 bdirty(bp); 3344 continue; 3345 3346 default: 3347 lk.lkt_held = NOHOLDER; 3348 panic("handle_disk_write_complete: Unknown type %s", 3349 TYPENAME(wk->wk_type)); 3350 /* NOTREACHED */ 3351 } 3352 } 3353 /* 3354 * Reattach any requests that must be redone. 3355 */ 3356 while ((wk = LIST_FIRST(&reattach)) != NULL) { 3357 WORKLIST_REMOVE(wk); 3358 WORKLIST_INSERT(&bp->b_dep, wk); 3359 } 3360 #ifdef DEBUG 3361 if (lk.lkt_held != SPECIAL_FLAG) 3362 panic("softdep_disk_write_complete: lock lost"); 3363 lk.lkt_held = NOHOLDER; 3364 #endif 3365 } 3366 3367 /* 3368 * Called from within softdep_disk_write_complete above. Note that 3369 * this routine is always called from interrupt level with further 3370 * splbio interrupts blocked. 3371 */ 3372 static void 3373 handle_allocdirect_partdone(adp) 3374 struct allocdirect *adp; /* the completed allocdirect */ 3375 { 3376 struct allocdirect *listadp; 3377 struct inodedep *inodedep; 3378 long bsize; 3379 3380 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3381 return; 3382 if (adp->ad_buf != NULL) { 3383 lk.lkt_held = NOHOLDER; 3384 panic("handle_allocdirect_partdone: dangling dep"); 3385 } 3386 /* 3387 * The on-disk inode cannot claim to be any larger than the last 3388 * fragment that has been written. Otherwise, the on-disk inode 3389 * might have fragments that were not the last block in the file 3390 * which would corrupt the filesystem. Thus, we cannot free any 3391 * allocdirects after one whose ad_oldblkno claims a fragment as 3392 * these blocks must be rolled back to zero before writing the inode. 3393 * We check the currently active set of allocdirects in id_inoupdt. 3394 */ 3395 inodedep = adp->ad_inodedep; 3396 bsize = inodedep->id_fs->fs_bsize; 3397 TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) { 3398 /* found our block */ 3399 if (listadp == adp) 3400 break; 3401 /* continue if ad_oldlbn is not a fragment */ 3402 if (listadp->ad_oldsize == 0 || 3403 listadp->ad_oldsize == bsize) 3404 continue; 3405 /* hit a fragment */ 3406 return; 3407 } 3408 /* 3409 * If we have reached the end of the current list without 3410 * finding the just finished dependency, then it must be 3411 * on the future dependency list. Future dependencies cannot 3412 * be freed until they are moved to the current list. 3413 */ 3414 if (listadp == NULL) { 3415 #ifdef DEBUG 3416 TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next) 3417 /* found our block */ 3418 if (listadp == adp) 3419 break; 3420 if (listadp == NULL) { 3421 lk.lkt_held = NOHOLDER; 3422 panic("handle_allocdirect_partdone: lost dep"); 3423 } 3424 #endif /* DEBUG */ 3425 return; 3426 } 3427 /* 3428 * If we have found the just finished dependency, then free 3429 * it along with anything that follows it that is complete. 3430 */ 3431 for (; adp; adp = listadp) { 3432 listadp = TAILQ_NEXT(adp, ad_next); 3433 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3434 return; 3435 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 3436 } 3437 } 3438 3439 /* 3440 * Called from within softdep_disk_write_complete above. Note that 3441 * this routine is always called from interrupt level with further 3442 * splbio interrupts blocked. 3443 */ 3444 static void 3445 handle_allocindir_partdone(aip) 3446 struct allocindir *aip; /* the completed allocindir */ 3447 { 3448 struct indirdep *indirdep; 3449 3450 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 3451 return; 3452 if (aip->ai_buf != NULL) { 3453 lk.lkt_held = NOHOLDER; 3454 panic("handle_allocindir_partdone: dangling dependency"); 3455 } 3456 indirdep = aip->ai_indirdep; 3457 if (indirdep->ir_state & UNDONE) { 3458 LIST_REMOVE(aip, ai_next); 3459 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 3460 return; 3461 } 3462 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 3463 aip->ai_newblkno; 3464 LIST_REMOVE(aip, ai_next); 3465 if (aip->ai_freefrag != NULL) 3466 add_to_worklist(&aip->ai_freefrag->ff_list); 3467 WORKITEM_FREE(aip, D_ALLOCINDIR); 3468 } 3469 3470 /* 3471 * Called from within softdep_disk_write_complete above to restore 3472 * in-memory inode block contents to their most up-to-date state. Note 3473 * that this routine is always called from interrupt level with further 3474 * splbio interrupts blocked. 3475 */ 3476 static int 3477 handle_written_inodeblock(inodedep, bp) 3478 struct inodedep *inodedep; 3479 struct buf *bp; /* buffer containing the inode block */ 3480 { 3481 struct worklist *wk, *filefree; 3482 struct allocdirect *adp, *nextadp; 3483 struct dinode *dp; 3484 int hadchanges; 3485 3486 if ((inodedep->id_state & IOSTARTED) == 0) { 3487 lk.lkt_held = NOHOLDER; 3488 panic("handle_written_inodeblock: not started"); 3489 } 3490 inodedep->id_state &= ~IOSTARTED; 3491 inodedep->id_state |= COMPLETE; 3492 dp = (struct dinode *)bp->b_data + 3493 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 3494 /* 3495 * If we had to rollback the inode allocation because of 3496 * bitmaps being incomplete, then simply restore it. 3497 * Keep the block dirty so that it will not be reclaimed until 3498 * all associated dependencies have been cleared and the 3499 * corresponding updates written to disk. 3500 */ 3501 if (inodedep->id_savedino != NULL) { 3502 *dp = *inodedep->id_savedino; 3503 FREE(inodedep->id_savedino, M_INODEDEP); 3504 inodedep->id_savedino = NULL; 3505 if ((bp->b_flags & B_DELWRI) == 0) 3506 stat_inode_bitmap++; 3507 bdirty(bp); 3508 return (1); 3509 } 3510 /* 3511 * Roll forward anything that had to be rolled back before 3512 * the inode could be updated. 3513 */ 3514 hadchanges = 0; 3515 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 3516 nextadp = TAILQ_NEXT(adp, ad_next); 3517 if (adp->ad_state & ATTACHED) { 3518 lk.lkt_held = NOHOLDER; 3519 panic("handle_written_inodeblock: new entry"); 3520 } 3521 if (adp->ad_lbn < NDADDR) { 3522 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) { 3523 lk.lkt_held = NOHOLDER; 3524 panic("%s: %s #%ld mismatch %d != %d", 3525 "handle_written_inodeblock", 3526 "direct pointer", adp->ad_lbn, 3527 dp->di_db[adp->ad_lbn], adp->ad_oldblkno); 3528 } 3529 dp->di_db[adp->ad_lbn] = adp->ad_newblkno; 3530 } else { 3531 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) { 3532 lk.lkt_held = NOHOLDER; 3533 panic("%s: %s #%ld allocated as %d", 3534 "handle_written_inodeblock", 3535 "indirect pointer", adp->ad_lbn - NDADDR, 3536 dp->di_ib[adp->ad_lbn - NDADDR]); 3537 } 3538 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno; 3539 } 3540 adp->ad_state &= ~UNDONE; 3541 adp->ad_state |= ATTACHED; 3542 hadchanges = 1; 3543 } 3544 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 3545 stat_direct_blk_ptrs++; 3546 /* 3547 * Reset the file size to its most up-to-date value. 3548 */ 3549 if (inodedep->id_savedsize == -1) { 3550 lk.lkt_held = NOHOLDER; 3551 panic("handle_written_inodeblock: bad size"); 3552 } 3553 if (dp->di_size != inodedep->id_savedsize) { 3554 dp->di_size = inodedep->id_savedsize; 3555 hadchanges = 1; 3556 } 3557 inodedep->id_savedsize = -1; 3558 /* 3559 * If there were any rollbacks in the inode block, then it must be 3560 * marked dirty so that its will eventually get written back in 3561 * its correct form. 3562 */ 3563 if (hadchanges) 3564 bdirty(bp); 3565 /* 3566 * Process any allocdirects that completed during the update. 3567 */ 3568 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 3569 handle_allocdirect_partdone(adp); 3570 /* 3571 * Process deallocations that were held pending until the 3572 * inode had been written to disk. Freeing of the inode 3573 * is delayed until after all blocks have been freed to 3574 * avoid creation of new <vfsid, inum, lbn> triples 3575 * before the old ones have been deleted. 3576 */ 3577 filefree = NULL; 3578 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 3579 WORKLIST_REMOVE(wk); 3580 switch (wk->wk_type) { 3581 3582 case D_FREEFILE: 3583 /* 3584 * We defer adding filefree to the worklist until 3585 * all other additions have been made to ensure 3586 * that it will be done after all the old blocks 3587 * have been freed. 3588 */ 3589 if (filefree != NULL) { 3590 lk.lkt_held = NOHOLDER; 3591 panic("handle_written_inodeblock: filefree"); 3592 } 3593 filefree = wk; 3594 continue; 3595 3596 case D_MKDIR: 3597 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 3598 continue; 3599 3600 case D_DIRADD: 3601 diradd_inode_written(WK_DIRADD(wk), inodedep); 3602 continue; 3603 3604 case D_FREEBLKS: 3605 case D_FREEFRAG: 3606 case D_DIRREM: 3607 add_to_worklist(wk); 3608 continue; 3609 3610 default: 3611 lk.lkt_held = NOHOLDER; 3612 panic("handle_written_inodeblock: Unknown type %s", 3613 TYPENAME(wk->wk_type)); 3614 /* NOTREACHED */ 3615 } 3616 } 3617 if (filefree != NULL) { 3618 if (free_inodedep(inodedep) == 0) { 3619 lk.lkt_held = NOHOLDER; 3620 panic("handle_written_inodeblock: live inodedep"); 3621 } 3622 add_to_worklist(filefree); 3623 return (0); 3624 } 3625 3626 /* 3627 * If no outstanding dependencies, free it. 3628 */ 3629 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0) 3630 return (0); 3631 return (hadchanges); 3632 } 3633 3634 /* 3635 * Process a diradd entry after its dependent inode has been written. 3636 * This routine must be called with splbio interrupts blocked. 3637 */ 3638 static void 3639 diradd_inode_written(dap, inodedep) 3640 struct diradd *dap; 3641 struct inodedep *inodedep; 3642 { 3643 struct pagedep *pagedep; 3644 3645 dap->da_state |= COMPLETE; 3646 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3647 if (dap->da_state & DIRCHG) 3648 pagedep = dap->da_previous->dm_pagedep; 3649 else 3650 pagedep = dap->da_pagedep; 3651 LIST_REMOVE(dap, da_pdlist); 3652 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3653 } 3654 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 3655 } 3656 3657 /* 3658 * Handle the completion of a mkdir dependency. 3659 */ 3660 static void 3661 handle_written_mkdir(mkdir, type) 3662 struct mkdir *mkdir; 3663 int type; 3664 { 3665 struct diradd *dap; 3666 struct pagedep *pagedep; 3667 3668 if (mkdir->md_state != type) { 3669 lk.lkt_held = NOHOLDER; 3670 panic("handle_written_mkdir: bad type"); 3671 } 3672 dap = mkdir->md_diradd; 3673 dap->da_state &= ~type; 3674 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 3675 dap->da_state |= DEPCOMPLETE; 3676 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3677 if (dap->da_state & DIRCHG) 3678 pagedep = dap->da_previous->dm_pagedep; 3679 else 3680 pagedep = dap->da_pagedep; 3681 LIST_REMOVE(dap, da_pdlist); 3682 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3683 } 3684 LIST_REMOVE(mkdir, md_mkdirs); 3685 WORKITEM_FREE(mkdir, D_MKDIR); 3686 } 3687 3688 /* 3689 * Called from within softdep_disk_write_complete above. 3690 * A write operation was just completed. Removed inodes can 3691 * now be freed and associated block pointers may be committed. 3692 * Note that this routine is always called from interrupt level 3693 * with further splbio interrupts blocked. 3694 */ 3695 static int 3696 handle_written_filepage(pagedep, bp) 3697 struct pagedep *pagedep; 3698 struct buf *bp; /* buffer containing the written page */ 3699 { 3700 struct dirrem *dirrem; 3701 struct diradd *dap, *nextdap; 3702 struct direct *ep; 3703 int i, chgs; 3704 3705 if ((pagedep->pd_state & IOSTARTED) == 0) { 3706 lk.lkt_held = NOHOLDER; 3707 panic("handle_written_filepage: not started"); 3708 } 3709 pagedep->pd_state &= ~IOSTARTED; 3710 /* 3711 * Process any directory removals that have been committed. 3712 */ 3713 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 3714 LIST_REMOVE(dirrem, dm_next); 3715 dirrem->dm_dirinum = pagedep->pd_ino; 3716 add_to_worklist(&dirrem->dm_list); 3717 } 3718 /* 3719 * Free any directory additions that have been committed. 3720 */ 3721 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 3722 free_diradd(dap); 3723 /* 3724 * Uncommitted directory entries must be restored. 3725 */ 3726 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 3727 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 3728 dap = nextdap) { 3729 nextdap = LIST_NEXT(dap, da_pdlist); 3730 if (dap->da_state & ATTACHED) { 3731 lk.lkt_held = NOHOLDER; 3732 panic("handle_written_filepage: attached"); 3733 } 3734 ep = (struct direct *) 3735 ((char *)bp->b_data + dap->da_offset); 3736 ep->d_ino = dap->da_newinum; 3737 dap->da_state &= ~UNDONE; 3738 dap->da_state |= ATTACHED; 3739 chgs = 1; 3740 /* 3741 * If the inode referenced by the directory has 3742 * been written out, then the dependency can be 3743 * moved to the pending list. 3744 */ 3745 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3746 LIST_REMOVE(dap, da_pdlist); 3747 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 3748 da_pdlist); 3749 } 3750 } 3751 } 3752 /* 3753 * If there were any rollbacks in the directory, then it must be 3754 * marked dirty so that its will eventually get written back in 3755 * its correct form. 3756 */ 3757 if (chgs) { 3758 if ((bp->b_flags & B_DELWRI) == 0) 3759 stat_dir_entry++; 3760 bdirty(bp); 3761 } 3762 /* 3763 * If no dependencies remain, the pagedep will be freed. 3764 * Otherwise it will remain to update the page before it 3765 * is written back to disk. 3766 */ 3767 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) { 3768 for (i = 0; i < DAHASHSZ; i++) 3769 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL) 3770 break; 3771 if (i == DAHASHSZ) { 3772 LIST_REMOVE(pagedep, pd_hash); 3773 WORKITEM_FREE(pagedep, D_PAGEDEP); 3774 return (0); 3775 } 3776 } 3777 return (1); 3778 } 3779 3780 /* 3781 * Writing back in-core inode structures. 3782 * 3783 * The filesystem only accesses an inode's contents when it occupies an 3784 * "in-core" inode structure. These "in-core" structures are separate from 3785 * the page frames used to cache inode blocks. Only the latter are 3786 * transferred to/from the disk. So, when the updated contents of the 3787 * "in-core" inode structure are copied to the corresponding in-memory inode 3788 * block, the dependencies are also transferred. The following procedure is 3789 * called when copying a dirty "in-core" inode to a cached inode block. 3790 */ 3791 3792 /* 3793 * Called when an inode is loaded from disk. If the effective link count 3794 * differed from the actual link count when it was last flushed, then we 3795 * need to ensure that the correct effective link count is put back. 3796 */ 3797 void 3798 softdep_load_inodeblock(ip) 3799 struct inode *ip; /* the "in_core" copy of the inode */ 3800 { 3801 struct inodedep *inodedep; 3802 3803 /* 3804 * Check for alternate nlink count. 3805 */ 3806 ip->i_effnlink = ip->i_nlink; 3807 ACQUIRE_LOCK(&lk); 3808 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3809 FREE_LOCK(&lk); 3810 return; 3811 } 3812 ip->i_effnlink -= inodedep->id_nlinkdelta; 3813 FREE_LOCK(&lk); 3814 } 3815 3816 /* 3817 * This routine is called just before the "in-core" inode 3818 * information is to be copied to the in-memory inode block. 3819 * Recall that an inode block contains several inodes. If 3820 * the force flag is set, then the dependencies will be 3821 * cleared so that the update can always be made. Note that 3822 * the buffer is locked when this routine is called, so we 3823 * will never be in the middle of writing the inode block 3824 * to disk. 3825 */ 3826 void 3827 softdep_update_inodeblock(ip, bp, waitfor) 3828 struct inode *ip; /* the "in_core" copy of the inode */ 3829 struct buf *bp; /* the buffer containing the inode block */ 3830 int waitfor; /* nonzero => update must be allowed */ 3831 { 3832 struct inodedep *inodedep; 3833 struct worklist *wk; 3834 int error, gotit; 3835 3836 /* 3837 * If the effective link count is not equal to the actual link 3838 * count, then we must track the difference in an inodedep while 3839 * the inode is (potentially) tossed out of the cache. Otherwise, 3840 * if there is no existing inodedep, then there are no dependencies 3841 * to track. 3842 */ 3843 ACQUIRE_LOCK(&lk); 3844 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3845 FREE_LOCK(&lk); 3846 if (ip->i_effnlink != ip->i_nlink) 3847 panic("softdep_update_inodeblock: bad link count"); 3848 return; 3849 } 3850 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) { 3851 FREE_LOCK(&lk); 3852 panic("softdep_update_inodeblock: bad delta"); 3853 } 3854 /* 3855 * Changes have been initiated. Anything depending on these 3856 * changes cannot occur until this inode has been written. 3857 */ 3858 inodedep->id_state &= ~COMPLETE; 3859 if ((inodedep->id_state & ONWORKLIST) == 0) 3860 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 3861 /* 3862 * Any new dependencies associated with the incore inode must 3863 * now be moved to the list associated with the buffer holding 3864 * the in-memory copy of the inode. Once merged process any 3865 * allocdirects that are completed by the merger. 3866 */ 3867 merge_inode_lists(inodedep); 3868 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL) 3869 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); 3870 /* 3871 * Now that the inode has been pushed into the buffer, the 3872 * operations dependent on the inode being written to disk 3873 * can be moved to the id_bufwait so that they will be 3874 * processed when the buffer I/O completes. 3875 */ 3876 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 3877 WORKLIST_REMOVE(wk); 3878 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 3879 } 3880 /* 3881 * Newly allocated inodes cannot be written until the bitmap 3882 * that allocates them have been written (indicated by 3883 * DEPCOMPLETE being set in id_state). If we are doing a 3884 * forced sync (e.g., an fsync on a file), we force the bitmap 3885 * to be written so that the update can be done. 3886 */ 3887 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) { 3888 FREE_LOCK(&lk); 3889 return; 3890 } 3891 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 3892 FREE_LOCK(&lk); 3893 if (gotit && 3894 (error = VOP_BWRITE(inodedep->id_buf->b_vp, inodedep->id_buf)) != 0) 3895 softdep_error("softdep_update_inodeblock: bwrite", error); 3896 if ((inodedep->id_state & DEPCOMPLETE) == 0) 3897 panic("softdep_update_inodeblock: update failed"); 3898 } 3899 3900 /* 3901 * Merge the new inode dependency list (id_newinoupdt) into the old 3902 * inode dependency list (id_inoupdt). This routine must be called 3903 * with splbio interrupts blocked. 3904 */ 3905 static void 3906 merge_inode_lists(inodedep) 3907 struct inodedep *inodedep; 3908 { 3909 struct allocdirect *listadp, *newadp; 3910 3911 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 3912 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) { 3913 if (listadp->ad_lbn < newadp->ad_lbn) { 3914 listadp = TAILQ_NEXT(listadp, ad_next); 3915 continue; 3916 } 3917 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 3918 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 3919 if (listadp->ad_lbn == newadp->ad_lbn) { 3920 allocdirect_merge(&inodedep->id_inoupdt, newadp, 3921 listadp); 3922 listadp = newadp; 3923 } 3924 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 3925 } 3926 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) { 3927 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 3928 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next); 3929 } 3930 } 3931 3932 /* 3933 * If we are doing an fsync, then we must ensure that any directory 3934 * entries for the inode have been written after the inode gets to disk. 3935 */ 3936 static int 3937 softdep_fsync(vp) 3938 struct vnode *vp; /* the "in_core" copy of the inode */ 3939 { 3940 struct inodedep *inodedep; 3941 struct pagedep *pagedep; 3942 struct worklist *wk; 3943 struct diradd *dap; 3944 struct mount *mnt; 3945 struct vnode *pvp; 3946 struct inode *ip; 3947 struct buf *bp; 3948 struct fs *fs; 3949 struct thread *td = curthread; /* XXX */ 3950 int error, flushparent; 3951 ino_t parentino; 3952 ufs_lbn_t lbn; 3953 3954 ip = VTOI(vp); 3955 fs = ip->i_fs; 3956 ACQUIRE_LOCK(&lk); 3957 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) { 3958 FREE_LOCK(&lk); 3959 return (0); 3960 } 3961 if (LIST_FIRST(&inodedep->id_inowait) != NULL || 3962 LIST_FIRST(&inodedep->id_bufwait) != NULL || 3963 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 3964 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) { 3965 FREE_LOCK(&lk); 3966 panic("softdep_fsync: pending ops"); 3967 } 3968 for (error = 0, flushparent = 0; ; ) { 3969 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 3970 break; 3971 if (wk->wk_type != D_DIRADD) { 3972 FREE_LOCK(&lk); 3973 panic("softdep_fsync: Unexpected type %s", 3974 TYPENAME(wk->wk_type)); 3975 } 3976 dap = WK_DIRADD(wk); 3977 /* 3978 * Flush our parent if this directory entry 3979 * has a MKDIR_PARENT dependency. 3980 */ 3981 if (dap->da_state & DIRCHG) 3982 pagedep = dap->da_previous->dm_pagedep; 3983 else 3984 pagedep = dap->da_pagedep; 3985 mnt = pagedep->pd_mnt; 3986 parentino = pagedep->pd_ino; 3987 lbn = pagedep->pd_lbn; 3988 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) { 3989 FREE_LOCK(&lk); 3990 panic("softdep_fsync: dirty"); 3991 } 3992 flushparent = dap->da_state & MKDIR_PARENT; 3993 /* 3994 * If we are being fsync'ed as part of vgone'ing this vnode, 3995 * then we will not be able to release and recover the 3996 * vnode below, so we just have to give up on writing its 3997 * directory entry out. It will eventually be written, just 3998 * not now, but then the user was not asking to have it 3999 * written, so we are not breaking any promises. 4000 */ 4001 if (vp->v_flag & VRECLAIMED) 4002 break; 4003 /* 4004 * We prevent deadlock by always fetching inodes from the 4005 * root, moving down the directory tree. Thus, when fetching 4006 * our parent directory, we must unlock ourselves before 4007 * requesting the lock on our parent. See the comment in 4008 * ufs_lookup for details on possible races. 4009 */ 4010 FREE_LOCK(&lk); 4011 VOP_UNLOCK(vp, 0, td); 4012 error = VFS_VGET(mnt, parentino, &pvp); 4013 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 4014 if (error != 0) 4015 return (error); 4016 if (flushparent) { 4017 if ((error = UFS_UPDATE(pvp, 1)) != 0) { 4018 vput(pvp); 4019 return (error); 4020 } 4021 } 4022 /* 4023 * Flush directory page containing the inode's name. 4024 */ 4025 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), &bp); 4026 if (error == 0) 4027 error = VOP_BWRITE(bp->b_vp, bp); 4028 vput(pvp); 4029 if (error != 0) 4030 return (error); 4031 ACQUIRE_LOCK(&lk); 4032 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) 4033 break; 4034 } 4035 FREE_LOCK(&lk); 4036 return (0); 4037 } 4038 4039 /* 4040 * Flush all the dirty bitmaps associated with the block device 4041 * before flushing the rest of the dirty blocks so as to reduce 4042 * the number of dependencies that will have to be rolled back. 4043 */ 4044 static int softdep_fsync_mountdev_bp(struct buf *bp, void *data); 4045 4046 void 4047 softdep_fsync_mountdev(vp) 4048 struct vnode *vp; 4049 { 4050 if (!vn_isdisk(vp, NULL)) 4051 panic("softdep_fsync_mountdev: vnode not a disk"); 4052 ACQUIRE_LOCK(&lk); 4053 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 4054 softdep_fsync_mountdev_bp, NULL); 4055 drain_output(vp, 1); 4056 FREE_LOCK(&lk); 4057 } 4058 4059 static int 4060 softdep_fsync_mountdev_bp(struct buf *bp, void *data) 4061 { 4062 struct worklist *wk; 4063 4064 /* 4065 * If it is already scheduled, skip to the next buffer. 4066 */ 4067 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) 4068 return(0); 4069 if ((bp->b_flags & B_DELWRI) == 0) { 4070 FREE_LOCK(&lk); 4071 panic("softdep_fsync_mountdev: not dirty"); 4072 } 4073 /* 4074 * We are only interested in bitmaps with outstanding 4075 * dependencies. 4076 */ 4077 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 4078 wk->wk_type != D_BMSAFEMAP || 4079 (bp->b_xflags & BX_BKGRDINPROG)) { 4080 BUF_UNLOCK(bp); 4081 return(0); 4082 } 4083 bremfree(bp); 4084 FREE_LOCK(&lk); 4085 (void) bawrite(bp); 4086 ACQUIRE_LOCK(&lk); 4087 return(0); 4088 } 4089 4090 /* 4091 * This routine is called when we are trying to synchronously flush a 4092 * file. This routine must eliminate any filesystem metadata dependencies 4093 * so that the syncing routine can succeed by pushing the dirty blocks 4094 * associated with the file. If any I/O errors occur, they are returned. 4095 */ 4096 struct softdep_sync_metadata_info { 4097 struct vnode *vp; 4098 int waitfor; 4099 }; 4100 4101 static int softdep_sync_metadata_bp(struct buf *bp, void *data); 4102 4103 int 4104 softdep_sync_metadata(struct vnode *vp, struct thread *td) 4105 { 4106 struct softdep_sync_metadata_info info; 4107 int error, waitfor; 4108 4109 /* 4110 * Check whether this vnode is involved in a filesystem 4111 * that is doing soft dependency processing. 4112 */ 4113 if (!vn_isdisk(vp, NULL)) { 4114 if (!DOINGSOFTDEP(vp)) 4115 return (0); 4116 } else 4117 if (vp->v_rdev->si_mountpoint == NULL || 4118 (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0) 4119 return (0); 4120 /* 4121 * Ensure that any direct block dependencies have been cleared. 4122 */ 4123 ACQUIRE_LOCK(&lk); 4124 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) { 4125 FREE_LOCK(&lk); 4126 return (error); 4127 } 4128 /* 4129 * For most files, the only metadata dependencies are the 4130 * cylinder group maps that allocate their inode or blocks. 4131 * The block allocation dependencies can be found by traversing 4132 * the dependency lists for any buffers that remain on their 4133 * dirty buffer list. The inode allocation dependency will 4134 * be resolved when the inode is updated with MNT_WAIT. 4135 * This work is done in two passes. The first pass grabs most 4136 * of the buffers and begins asynchronously writing them. The 4137 * only way to wait for these asynchronous writes is to sleep 4138 * on the filesystem vnode which may stay busy for a long time 4139 * if the filesystem is active. So, instead, we make a second 4140 * pass over the dependencies blocking on each write. In the 4141 * usual case we will be blocking against a write that we 4142 * initiated, so when it is done the dependency will have been 4143 * resolved. Thus the second pass is expected to end quickly. 4144 */ 4145 waitfor = MNT_NOWAIT; 4146 top: 4147 /* 4148 * We must wait for any I/O in progress to finish so that 4149 * all potential buffers on the dirty list will be visible. 4150 */ 4151 drain_output(vp, 1); 4152 info.vp = vp; 4153 info.waitfor = waitfor; 4154 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 4155 softdep_sync_metadata_bp, &info); 4156 if (error < 0) { 4157 FREE_LOCK(&lk); 4158 return(-error); /* error code */ 4159 } 4160 4161 /* 4162 * The brief unlock is to allow any pent up dependency 4163 * processing to be done. Then proceed with the second pass. 4164 */ 4165 if (waitfor == MNT_NOWAIT) { 4166 waitfor = MNT_WAIT; 4167 FREE_LOCK(&lk); 4168 ACQUIRE_LOCK(&lk); 4169 goto top; 4170 } 4171 4172 /* 4173 * If we have managed to get rid of all the dirty buffers, 4174 * then we are done. For certain directories and block 4175 * devices, we may need to do further work. 4176 * 4177 * We must wait for any I/O in progress to finish so that 4178 * all potential buffers on the dirty list will be visible. 4179 */ 4180 drain_output(vp, 1); 4181 if (RB_EMPTY(&vp->v_rbdirty_tree)) { 4182 FREE_LOCK(&lk); 4183 return (0); 4184 } 4185 4186 FREE_LOCK(&lk); 4187 /* 4188 * If we are trying to sync a block device, some of its buffers may 4189 * contain metadata that cannot be written until the contents of some 4190 * partially written files have been written to disk. The only easy 4191 * way to accomplish this is to sync the entire filesystem (luckily 4192 * this happens rarely). 4193 */ 4194 if (vn_isdisk(vp, NULL) && 4195 vp->v_rdev && 4196 vp->v_rdev->si_mountpoint && !VOP_ISLOCKED(vp, NULL) && 4197 (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT, td)) != 0) 4198 return (error); 4199 return (0); 4200 } 4201 4202 static int 4203 softdep_sync_metadata_bp(struct buf *bp, void *data) 4204 { 4205 struct softdep_sync_metadata_info *info = data; 4206 struct pagedep *pagedep; 4207 struct allocdirect *adp; 4208 struct allocindir *aip; 4209 struct worklist *wk; 4210 struct buf *nbp; 4211 int error; 4212 int i; 4213 4214 if (getdirtybuf(&bp, MNT_WAIT) == 0) 4215 return (0); 4216 4217 /* 4218 * As we hold the buffer locked, none of its dependencies 4219 * will disappear. 4220 */ 4221 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 4222 switch (wk->wk_type) { 4223 4224 case D_ALLOCDIRECT: 4225 adp = WK_ALLOCDIRECT(wk); 4226 if (adp->ad_state & DEPCOMPLETE) 4227 break; 4228 nbp = adp->ad_buf; 4229 if (getdirtybuf(&nbp, info->waitfor) == 0) 4230 break; 4231 FREE_LOCK(&lk); 4232 if (info->waitfor == MNT_NOWAIT) { 4233 bawrite(nbp); 4234 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 4235 bawrite(bp); 4236 ACQUIRE_LOCK(&lk); 4237 return (-error); 4238 } 4239 ACQUIRE_LOCK(&lk); 4240 break; 4241 4242 case D_ALLOCINDIR: 4243 aip = WK_ALLOCINDIR(wk); 4244 if (aip->ai_state & DEPCOMPLETE) 4245 break; 4246 nbp = aip->ai_buf; 4247 if (getdirtybuf(&nbp, info->waitfor) == 0) 4248 break; 4249 FREE_LOCK(&lk); 4250 if (info->waitfor == MNT_NOWAIT) { 4251 bawrite(nbp); 4252 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 4253 bawrite(bp); 4254 ACQUIRE_LOCK(&lk); 4255 return (-error); 4256 } 4257 ACQUIRE_LOCK(&lk); 4258 break; 4259 4260 case D_INDIRDEP: 4261 restart: 4262 4263 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) { 4264 if (aip->ai_state & DEPCOMPLETE) 4265 continue; 4266 nbp = aip->ai_buf; 4267 if (getdirtybuf(&nbp, MNT_WAIT) == 0) 4268 goto restart; 4269 FREE_LOCK(&lk); 4270 if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 4271 bawrite(bp); 4272 ACQUIRE_LOCK(&lk); 4273 return (-error); 4274 } 4275 ACQUIRE_LOCK(&lk); 4276 goto restart; 4277 } 4278 break; 4279 4280 case D_INODEDEP: 4281 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs, 4282 WK_INODEDEP(wk)->id_ino)) != 0) { 4283 FREE_LOCK(&lk); 4284 bawrite(bp); 4285 ACQUIRE_LOCK(&lk); 4286 return (-error); 4287 } 4288 break; 4289 4290 case D_PAGEDEP: 4291 /* 4292 * We are trying to sync a directory that may 4293 * have dependencies on both its own metadata 4294 * and/or dependencies on the inodes of any 4295 * recently allocated files. We walk its diradd 4296 * lists pushing out the associated inode. 4297 */ 4298 pagedep = WK_PAGEDEP(wk); 4299 for (i = 0; i < DAHASHSZ; i++) { 4300 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 4301 continue; 4302 if ((error = 4303 flush_pagedep_deps(info->vp, 4304 pagedep->pd_mnt, 4305 &pagedep->pd_diraddhd[i]))) { 4306 FREE_LOCK(&lk); 4307 bawrite(bp); 4308 ACQUIRE_LOCK(&lk); 4309 return (-error); 4310 } 4311 } 4312 break; 4313 4314 case D_MKDIR: 4315 /* 4316 * This case should never happen if the vnode has 4317 * been properly sync'ed. However, if this function 4318 * is used at a place where the vnode has not yet 4319 * been sync'ed, this dependency can show up. So, 4320 * rather than panic, just flush it. 4321 */ 4322 nbp = WK_MKDIR(wk)->md_buf; 4323 if (getdirtybuf(&nbp, info->waitfor) == 0) 4324 break; 4325 FREE_LOCK(&lk); 4326 if (info->waitfor == MNT_NOWAIT) { 4327 bawrite(nbp); 4328 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 4329 bawrite(bp); 4330 ACQUIRE_LOCK(&lk); 4331 return (-error); 4332 } 4333 ACQUIRE_LOCK(&lk); 4334 break; 4335 4336 case D_BMSAFEMAP: 4337 /* 4338 * This case should never happen if the vnode has 4339 * been properly sync'ed. However, if this function 4340 * is used at a place where the vnode has not yet 4341 * been sync'ed, this dependency can show up. So, 4342 * rather than panic, just flush it. 4343 */ 4344 nbp = WK_BMSAFEMAP(wk)->sm_buf; 4345 if (getdirtybuf(&nbp, info->waitfor) == 0) 4346 break; 4347 FREE_LOCK(&lk); 4348 if (info->waitfor == MNT_NOWAIT) { 4349 bawrite(nbp); 4350 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 4351 bawrite(bp); 4352 ACQUIRE_LOCK(&lk); 4353 return (-error); 4354 } 4355 ACQUIRE_LOCK(&lk); 4356 break; 4357 4358 default: 4359 FREE_LOCK(&lk); 4360 panic("softdep_sync_metadata: Unknown type %s", 4361 TYPENAME(wk->wk_type)); 4362 /* NOTREACHED */ 4363 } 4364 } 4365 FREE_LOCK(&lk); 4366 bawrite(bp); 4367 ACQUIRE_LOCK(&lk); 4368 return(0); 4369 } 4370 4371 /* 4372 * Flush the dependencies associated with an inodedep. 4373 * Called with splbio blocked. 4374 */ 4375 static int 4376 flush_inodedep_deps(fs, ino) 4377 struct fs *fs; 4378 ino_t ino; 4379 { 4380 struct inodedep *inodedep; 4381 struct allocdirect *adp; 4382 int error, waitfor; 4383 struct buf *bp; 4384 4385 /* 4386 * This work is done in two passes. The first pass grabs most 4387 * of the buffers and begins asynchronously writing them. The 4388 * only way to wait for these asynchronous writes is to sleep 4389 * on the filesystem vnode which may stay busy for a long time 4390 * if the filesystem is active. So, instead, we make a second 4391 * pass over the dependencies blocking on each write. In the 4392 * usual case we will be blocking against a write that we 4393 * initiated, so when it is done the dependency will have been 4394 * resolved. Thus the second pass is expected to end quickly. 4395 * We give a brief window at the top of the loop to allow 4396 * any pending I/O to complete. 4397 */ 4398 for (waitfor = MNT_NOWAIT; ; ) { 4399 FREE_LOCK(&lk); 4400 ACQUIRE_LOCK(&lk); 4401 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4402 return (0); 4403 TAILQ_FOREACH(adp, &inodedep->id_inoupdt, ad_next) { 4404 if (adp->ad_state & DEPCOMPLETE) 4405 continue; 4406 bp = adp->ad_buf; 4407 if (getdirtybuf(&bp, waitfor) == 0) { 4408 if (waitfor == MNT_NOWAIT) 4409 continue; 4410 break; 4411 } 4412 FREE_LOCK(&lk); 4413 if (waitfor == MNT_NOWAIT) { 4414 bawrite(bp); 4415 } else if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) { 4416 ACQUIRE_LOCK(&lk); 4417 return (error); 4418 } 4419 ACQUIRE_LOCK(&lk); 4420 break; 4421 } 4422 if (adp != NULL) 4423 continue; 4424 TAILQ_FOREACH(adp, &inodedep->id_newinoupdt, ad_next) { 4425 if (adp->ad_state & DEPCOMPLETE) 4426 continue; 4427 bp = adp->ad_buf; 4428 if (getdirtybuf(&bp, waitfor) == 0) { 4429 if (waitfor == MNT_NOWAIT) 4430 continue; 4431 break; 4432 } 4433 FREE_LOCK(&lk); 4434 if (waitfor == MNT_NOWAIT) { 4435 bawrite(bp); 4436 } else if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) { 4437 ACQUIRE_LOCK(&lk); 4438 return (error); 4439 } 4440 ACQUIRE_LOCK(&lk); 4441 break; 4442 } 4443 if (adp != NULL) 4444 continue; 4445 /* 4446 * If pass2, we are done, otherwise do pass 2. 4447 */ 4448 if (waitfor == MNT_WAIT) 4449 break; 4450 waitfor = MNT_WAIT; 4451 } 4452 /* 4453 * Try freeing inodedep in case all dependencies have been removed. 4454 */ 4455 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0) 4456 (void) free_inodedep(inodedep); 4457 return (0); 4458 } 4459 4460 /* 4461 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 4462 * Called with splbio blocked. 4463 */ 4464 static int 4465 flush_pagedep_deps(pvp, mp, diraddhdp) 4466 struct vnode *pvp; 4467 struct mount *mp; 4468 struct diraddhd *diraddhdp; 4469 { 4470 struct thread *td = curthread; /* XXX */ 4471 struct inodedep *inodedep; 4472 struct ufsmount *ump; 4473 struct diradd *dap; 4474 struct vnode *vp; 4475 int gotit, error = 0; 4476 struct buf *bp; 4477 ino_t inum; 4478 4479 ump = VFSTOUFS(mp); 4480 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 4481 /* 4482 * Flush ourselves if this directory entry 4483 * has a MKDIR_PARENT dependency. 4484 */ 4485 if (dap->da_state & MKDIR_PARENT) { 4486 FREE_LOCK(&lk); 4487 if ((error = UFS_UPDATE(pvp, 1)) != 0) 4488 break; 4489 ACQUIRE_LOCK(&lk); 4490 /* 4491 * If that cleared dependencies, go on to next. 4492 */ 4493 if (dap != LIST_FIRST(diraddhdp)) 4494 continue; 4495 if (dap->da_state & MKDIR_PARENT) { 4496 FREE_LOCK(&lk); 4497 panic("flush_pagedep_deps: MKDIR_PARENT"); 4498 } 4499 } 4500 /* 4501 * A newly allocated directory must have its "." and 4502 * ".." entries written out before its name can be 4503 * committed in its parent. We do not want or need 4504 * the full semantics of a synchronous VOP_FSYNC as 4505 * that may end up here again, once for each directory 4506 * level in the filesystem. Instead, we push the blocks 4507 * and wait for them to clear. We have to fsync twice 4508 * because the first call may choose to defer blocks 4509 * that still have dependencies, but deferral will 4510 * happen at most once. 4511 */ 4512 inum = dap->da_newinum; 4513 if (dap->da_state & MKDIR_BODY) { 4514 FREE_LOCK(&lk); 4515 if ((error = VFS_VGET(mp, inum, &vp)) != 0) 4516 break; 4517 if ((error=VOP_FSYNC(vp, MNT_NOWAIT, td)) || 4518 (error=VOP_FSYNC(vp, MNT_NOWAIT, td))) { 4519 vput(vp); 4520 break; 4521 } 4522 drain_output(vp, 0); 4523 vput(vp); 4524 ACQUIRE_LOCK(&lk); 4525 /* 4526 * If that cleared dependencies, go on to next. 4527 */ 4528 if (dap != LIST_FIRST(diraddhdp)) 4529 continue; 4530 if (dap->da_state & MKDIR_BODY) { 4531 FREE_LOCK(&lk); 4532 panic("flush_pagedep_deps: MKDIR_BODY"); 4533 } 4534 } 4535 /* 4536 * Flush the inode on which the directory entry depends. 4537 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 4538 * the only remaining dependency is that the updated inode 4539 * count must get pushed to disk. The inode has already 4540 * been pushed into its inode buffer (via VOP_UPDATE) at 4541 * the time of the reference count change. So we need only 4542 * locate that buffer, ensure that there will be no rollback 4543 * caused by a bitmap dependency, then write the inode buffer. 4544 */ 4545 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) { 4546 FREE_LOCK(&lk); 4547 panic("flush_pagedep_deps: lost inode"); 4548 } 4549 /* 4550 * If the inode still has bitmap dependencies, 4551 * push them to disk. 4552 */ 4553 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4554 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 4555 FREE_LOCK(&lk); 4556 if (gotit && 4557 (error = VOP_BWRITE(inodedep->id_buf->b_vp, 4558 inodedep->id_buf)) != 0) 4559 break; 4560 ACQUIRE_LOCK(&lk); 4561 if (dap != LIST_FIRST(diraddhdp)) 4562 continue; 4563 } 4564 /* 4565 * If the inode is still sitting in a buffer waiting 4566 * to be written, push it to disk. 4567 */ 4568 FREE_LOCK(&lk); 4569 if ((error = bread(ump->um_devvp, 4570 fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), 4571 (int)ump->um_fs->fs_bsize, &bp)) != 0) 4572 break; 4573 if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) 4574 break; 4575 ACQUIRE_LOCK(&lk); 4576 /* 4577 * If we have failed to get rid of all the dependencies 4578 * then something is seriously wrong. 4579 */ 4580 if (dap == LIST_FIRST(diraddhdp)) { 4581 FREE_LOCK(&lk); 4582 panic("flush_pagedep_deps: flush failed"); 4583 } 4584 } 4585 if (error) 4586 ACQUIRE_LOCK(&lk); 4587 return (error); 4588 } 4589 4590 /* 4591 * A large burst of file addition or deletion activity can drive the 4592 * memory load excessively high. First attempt to slow things down 4593 * using the techniques below. If that fails, this routine requests 4594 * the offending operations to fall back to running synchronously 4595 * until the memory load returns to a reasonable level. 4596 */ 4597 int 4598 softdep_slowdown(vp) 4599 struct vnode *vp; 4600 { 4601 int max_softdeps_hard; 4602 4603 max_softdeps_hard = max_softdeps * 11 / 10; 4604 if (num_dirrem < max_softdeps_hard / 2 && 4605 num_inodedep < max_softdeps_hard) 4606 return (0); 4607 stat_sync_limit_hit += 1; 4608 return (1); 4609 } 4610 4611 /* 4612 * If memory utilization has gotten too high, deliberately slow things 4613 * down and speed up the I/O processing. 4614 */ 4615 static int 4616 request_cleanup(resource, islocked) 4617 int resource; 4618 int islocked; 4619 { 4620 struct thread *td = curthread; /* XXX */ 4621 4622 /* 4623 * We never hold up the filesystem syncer process. 4624 */ 4625 if (td == filesys_syncer) 4626 return (0); 4627 /* 4628 * First check to see if the work list has gotten backlogged. 4629 * If it has, co-opt this process to help clean up two entries. 4630 * Because this process may hold inodes locked, we cannot 4631 * handle any remove requests that might block on a locked 4632 * inode as that could lead to deadlock. 4633 */ 4634 if (num_on_worklist > max_softdeps / 10) { 4635 if (islocked) 4636 FREE_LOCK(&lk); 4637 process_worklist_item(NULL, LK_NOWAIT); 4638 process_worklist_item(NULL, LK_NOWAIT); 4639 stat_worklist_push += 2; 4640 if (islocked) 4641 ACQUIRE_LOCK(&lk); 4642 return(1); 4643 } 4644 4645 /* 4646 * If we are resource constrained on inode dependencies, try 4647 * flushing some dirty inodes. Otherwise, we are constrained 4648 * by file deletions, so try accelerating flushes of directories 4649 * with removal dependencies. We would like to do the cleanup 4650 * here, but we probably hold an inode locked at this point and 4651 * that might deadlock against one that we try to clean. So, 4652 * the best that we can do is request the syncer daemon to do 4653 * the cleanup for us. 4654 */ 4655 switch (resource) { 4656 4657 case FLUSH_INODES: 4658 stat_ino_limit_push += 1; 4659 req_clear_inodedeps += 1; 4660 stat_countp = &stat_ino_limit_hit; 4661 break; 4662 4663 case FLUSH_REMOVE: 4664 stat_blk_limit_push += 1; 4665 req_clear_remove += 1; 4666 stat_countp = &stat_blk_limit_hit; 4667 break; 4668 4669 default: 4670 if (islocked) 4671 FREE_LOCK(&lk); 4672 panic("request_cleanup: unknown type"); 4673 } 4674 /* 4675 * Hopefully the syncer daemon will catch up and awaken us. 4676 * We wait at most tickdelay before proceeding in any case. 4677 */ 4678 if (islocked == 0) 4679 ACQUIRE_LOCK(&lk); 4680 crit_enter(); 4681 proc_waiting += 1; 4682 if (!callout_active(&handle)) 4683 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2, 4684 pause_timer, NULL); 4685 interlocked_sleep(&lk, SLEEP, (caddr_t)&proc_waiting, 0, 4686 "softupdate", 0); 4687 proc_waiting -= 1; 4688 crit_exit(); 4689 if (islocked == 0) 4690 FREE_LOCK(&lk); 4691 return (1); 4692 } 4693 4694 /* 4695 * Awaken processes pausing in request_cleanup and clear proc_waiting 4696 * to indicate that there is no longer a timer running. 4697 */ 4698 void 4699 pause_timer(arg) 4700 void *arg; 4701 { 4702 *stat_countp += 1; 4703 wakeup_one(&proc_waiting); 4704 if (proc_waiting > 0) 4705 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2, 4706 pause_timer, NULL); 4707 else 4708 callout_deactivate(&handle); 4709 } 4710 4711 /* 4712 * Flush out a directory with at least one removal dependency in an effort to 4713 * reduce the number of dirrem, freefile, and freeblks dependency structures. 4714 */ 4715 static void 4716 clear_remove(struct thread *td) 4717 { 4718 struct pagedep_hashhead *pagedephd; 4719 struct pagedep *pagedep; 4720 static int next = 0; 4721 struct mount *mp; 4722 struct vnode *vp; 4723 int error, cnt; 4724 ino_t ino; 4725 4726 ACQUIRE_LOCK(&lk); 4727 for (cnt = 0; cnt < pagedep_hash; cnt++) { 4728 pagedephd = &pagedep_hashtbl[next++]; 4729 if (next >= pagedep_hash) 4730 next = 0; 4731 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 4732 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL) 4733 continue; 4734 mp = pagedep->pd_mnt; 4735 ino = pagedep->pd_ino; 4736 FREE_LOCK(&lk); 4737 if ((error = VFS_VGET(mp, ino, &vp)) != 0) { 4738 softdep_error("clear_remove: vget", error); 4739 return; 4740 } 4741 if ((error = VOP_FSYNC(vp, MNT_NOWAIT, td))) 4742 softdep_error("clear_remove: fsync", error); 4743 drain_output(vp, 0); 4744 vput(vp); 4745 return; 4746 } 4747 } 4748 FREE_LOCK(&lk); 4749 } 4750 4751 /* 4752 * Clear out a block of dirty inodes in an effort to reduce 4753 * the number of inodedep dependency structures. 4754 */ 4755 struct clear_inodedeps_info { 4756 struct fs *fs; 4757 struct mount *mp; 4758 }; 4759 4760 static int 4761 clear_inodedeps_mountlist_callback(struct mount *mp, void *data) 4762 { 4763 struct clear_inodedeps_info *info = data; 4764 4765 if ((mp->mnt_flag & MNT_SOFTDEP) && info->fs == VFSTOUFS(mp)->um_fs) { 4766 info->mp = mp; 4767 return(-1); 4768 } 4769 return(0); 4770 } 4771 4772 static void 4773 clear_inodedeps(struct thread *td) 4774 { 4775 struct clear_inodedeps_info info; 4776 struct inodedep_hashhead *inodedephd; 4777 struct inodedep *inodedep; 4778 static int next = 0; 4779 struct vnode *vp; 4780 struct fs *fs; 4781 int error, cnt; 4782 ino_t firstino, lastino, ino; 4783 4784 ACQUIRE_LOCK(&lk); 4785 /* 4786 * Pick a random inode dependency to be cleared. 4787 * We will then gather up all the inodes in its block 4788 * that have dependencies and flush them out. 4789 */ 4790 for (cnt = 0; cnt < inodedep_hash; cnt++) { 4791 inodedephd = &inodedep_hashtbl[next++]; 4792 if (next >= inodedep_hash) 4793 next = 0; 4794 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 4795 break; 4796 } 4797 if (inodedep == NULL) { 4798 FREE_LOCK(&lk); 4799 return; 4800 } 4801 /* 4802 * Ugly code to find mount point given pointer to superblock. 4803 */ 4804 fs = inodedep->id_fs; 4805 info.mp = NULL; 4806 info.fs = fs; 4807 mountlist_scan(clear_inodedeps_mountlist_callback, 4808 &info, MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 4809 /* 4810 * Find the last inode in the block with dependencies. 4811 */ 4812 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 4813 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 4814 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0) 4815 break; 4816 /* 4817 * Asynchronously push all but the last inode with dependencies. 4818 * Synchronously push the last inode with dependencies to ensure 4819 * that the inode block gets written to free up the inodedeps. 4820 */ 4821 for (ino = firstino; ino <= lastino; ino++) { 4822 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4823 continue; 4824 FREE_LOCK(&lk); 4825 if ((error = VFS_VGET(info.mp, ino, &vp)) != 0) { 4826 softdep_error("clear_inodedeps: vget", error); 4827 return; 4828 } 4829 if (ino == lastino) { 4830 if ((error = VOP_FSYNC(vp, MNT_WAIT, td))) 4831 softdep_error("clear_inodedeps: fsync1", error); 4832 } else { 4833 if ((error = VOP_FSYNC(vp, MNT_NOWAIT, td))) 4834 softdep_error("clear_inodedeps: fsync2", error); 4835 drain_output(vp, 0); 4836 } 4837 vput(vp); 4838 ACQUIRE_LOCK(&lk); 4839 } 4840 FREE_LOCK(&lk); 4841 } 4842 4843 /* 4844 * Function to determine if the buffer has outstanding dependencies 4845 * that will cause a roll-back if the buffer is written. If wantcount 4846 * is set, return number of dependencies, otherwise just yes or no. 4847 */ 4848 static int 4849 softdep_count_dependencies(bp, wantcount) 4850 struct buf *bp; 4851 int wantcount; 4852 { 4853 struct worklist *wk; 4854 struct inodedep *inodedep; 4855 struct indirdep *indirdep; 4856 struct allocindir *aip; 4857 struct pagedep *pagedep; 4858 struct diradd *dap; 4859 int i, retval; 4860 4861 retval = 0; 4862 ACQUIRE_LOCK(&lk); 4863 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 4864 switch (wk->wk_type) { 4865 4866 case D_INODEDEP: 4867 inodedep = WK_INODEDEP(wk); 4868 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4869 /* bitmap allocation dependency */ 4870 retval += 1; 4871 if (!wantcount) 4872 goto out; 4873 } 4874 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 4875 /* direct block pointer dependency */ 4876 retval += 1; 4877 if (!wantcount) 4878 goto out; 4879 } 4880 continue; 4881 4882 case D_INDIRDEP: 4883 indirdep = WK_INDIRDEP(wk); 4884 4885 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 4886 /* indirect block pointer dependency */ 4887 retval += 1; 4888 if (!wantcount) 4889 goto out; 4890 } 4891 continue; 4892 4893 case D_PAGEDEP: 4894 pagedep = WK_PAGEDEP(wk); 4895 for (i = 0; i < DAHASHSZ; i++) { 4896 4897 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 4898 /* directory entry dependency */ 4899 retval += 1; 4900 if (!wantcount) 4901 goto out; 4902 } 4903 } 4904 continue; 4905 4906 case D_BMSAFEMAP: 4907 case D_ALLOCDIRECT: 4908 case D_ALLOCINDIR: 4909 case D_MKDIR: 4910 /* never a dependency on these blocks */ 4911 continue; 4912 4913 default: 4914 FREE_LOCK(&lk); 4915 panic("softdep_check_for_rollback: Unexpected type %s", 4916 TYPENAME(wk->wk_type)); 4917 /* NOTREACHED */ 4918 } 4919 } 4920 out: 4921 FREE_LOCK(&lk); 4922 return retval; 4923 } 4924 4925 /* 4926 * Acquire exclusive access to a buffer. 4927 * Must be called with splbio blocked. 4928 * Return 1 if buffer was acquired. 4929 */ 4930 static int 4931 getdirtybuf(bpp, waitfor) 4932 struct buf **bpp; 4933 int waitfor; 4934 { 4935 struct buf *bp; 4936 int error; 4937 4938 for (;;) { 4939 if ((bp = *bpp) == NULL) 4940 return (0); 4941 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 4942 if ((bp->b_xflags & BX_BKGRDINPROG) == 0) 4943 break; 4944 BUF_UNLOCK(bp); 4945 if (waitfor != MNT_WAIT) 4946 return (0); 4947 bp->b_xflags |= BX_BKGRDWAIT; 4948 interlocked_sleep(&lk, SLEEP, &bp->b_xflags, 0, 4949 "getbuf", 0); 4950 continue; 4951 } 4952 if (waitfor != MNT_WAIT) 4953 return (0); 4954 error = interlocked_sleep(&lk, LOCKBUF, bp, 4955 LK_EXCLUSIVE | LK_SLEEPFAIL, 0, 0); 4956 if (error != ENOLCK) { 4957 FREE_LOCK(&lk); 4958 panic("getdirtybuf: inconsistent lock"); 4959 } 4960 } 4961 if ((bp->b_flags & B_DELWRI) == 0) { 4962 BUF_UNLOCK(bp); 4963 return (0); 4964 } 4965 bremfree(bp); 4966 return (1); 4967 } 4968 4969 /* 4970 * Wait for pending output on a vnode to complete. 4971 * Must be called with vnode locked. 4972 */ 4973 static void 4974 drain_output(vp, islocked) 4975 struct vnode *vp; 4976 int islocked; 4977 { 4978 4979 if (!islocked) 4980 ACQUIRE_LOCK(&lk); 4981 while (vp->v_numoutput) { 4982 vp->v_flag |= VBWAIT; 4983 interlocked_sleep(&lk, SLEEP, (caddr_t)&vp->v_numoutput, 4984 0, "drainvp", 0); 4985 } 4986 if (!islocked) 4987 FREE_LOCK(&lk); 4988 } 4989 4990 /* 4991 * Called whenever a buffer that is being invalidated or reallocated 4992 * contains dependencies. This should only happen if an I/O error has 4993 * occurred. The routine is called with the buffer locked. 4994 */ 4995 static void 4996 softdep_deallocate_dependencies(bp) 4997 struct buf *bp; 4998 { 4999 5000 if ((bp->b_flags & B_ERROR) == 0) 5001 panic("softdep_deallocate_dependencies: dangling deps"); 5002 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntfromname, bp->b_error); 5003 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 5004 } 5005 5006 /* 5007 * Function to handle asynchronous write errors in the filesystem. 5008 */ 5009 void 5010 softdep_error(func, error) 5011 char *func; 5012 int error; 5013 { 5014 5015 /* XXX should do something better! */ 5016 printf("%s: got error %d while accessing filesystem\n", func, error); 5017 } 5018