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 1.20 2004/11/09 04:41:47 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 /* 74 * These definitions need to be adapted to the system to which 75 * this file is being ported. 76 */ 77 /* 78 * malloc types defined for the softdep system. 79 */ 80 MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies"); 81 MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies"); 82 MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation"); 83 MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map"); 84 MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode"); 85 MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies"); 86 MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block"); 87 MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode"); 88 MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode"); 89 MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated"); 90 MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry"); 91 MALLOC_DEFINE(M_MKDIR, "mkdir","New directory"); 92 MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted"); 93 94 #define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE) 95 96 #define D_PAGEDEP 0 97 #define D_INODEDEP 1 98 #define D_NEWBLK 2 99 #define D_BMSAFEMAP 3 100 #define D_ALLOCDIRECT 4 101 #define D_INDIRDEP 5 102 #define D_ALLOCINDIR 6 103 #define D_FREEFRAG 7 104 #define D_FREEBLKS 8 105 #define D_FREEFILE 9 106 #define D_DIRADD 10 107 #define D_MKDIR 11 108 #define D_DIRREM 12 109 #define D_LAST D_DIRREM 110 111 /* 112 * translate from workitem type to memory type 113 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 114 */ 115 static struct malloc_type *memtype[] = { 116 M_PAGEDEP, 117 M_INODEDEP, 118 M_NEWBLK, 119 M_BMSAFEMAP, 120 M_ALLOCDIRECT, 121 M_INDIRDEP, 122 M_ALLOCINDIR, 123 M_FREEFRAG, 124 M_FREEBLKS, 125 M_FREEFILE, 126 M_DIRADD, 127 M_MKDIR, 128 M_DIRREM 129 }; 130 131 #define DtoM(type) (memtype[type]) 132 133 /* 134 * Names of malloc types. 135 */ 136 #define TYPENAME(type) \ 137 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???") 138 /* 139 * End system adaptaion definitions. 140 */ 141 142 /* 143 * Internal function prototypes. 144 */ 145 static void softdep_error(char *, int); 146 static void drain_output(struct vnode *, int); 147 static int getdirtybuf(struct buf **, int); 148 static void clear_remove(struct thread *); 149 static void clear_inodedeps(struct thread *); 150 static int flush_pagedep_deps(struct vnode *, struct mount *, 151 struct diraddhd *); 152 static int flush_inodedep_deps(struct fs *, ino_t); 153 static int handle_written_filepage(struct pagedep *, struct buf *); 154 static void diradd_inode_written(struct diradd *, struct inodedep *); 155 static int handle_written_inodeblock(struct inodedep *, struct buf *); 156 static void handle_allocdirect_partdone(struct allocdirect *); 157 static void handle_allocindir_partdone(struct allocindir *); 158 static void initiate_write_filepage(struct pagedep *, struct buf *); 159 static void handle_written_mkdir(struct mkdir *, int); 160 static void initiate_write_inodeblock(struct inodedep *, struct buf *); 161 static void handle_workitem_freefile(struct freefile *); 162 static void handle_workitem_remove(struct dirrem *); 163 static struct dirrem *newdirrem(struct buf *, struct inode *, 164 struct inode *, int, struct dirrem **); 165 static void free_diradd(struct diradd *); 166 static void free_allocindir(struct allocindir *, struct inodedep *); 167 static int indir_trunc (struct inode *, ufs_daddr_t, int, ufs_lbn_t, 168 long *); 169 static void deallocate_dependencies(struct buf *, struct inodedep *); 170 static void free_allocdirect(struct allocdirectlst *, 171 struct allocdirect *, int); 172 static int check_inode_unwritten(struct inodedep *); 173 static int free_inodedep(struct inodedep *); 174 static void handle_workitem_freeblocks(struct freeblks *); 175 static void merge_inode_lists(struct inodedep *); 176 static void setup_allocindir_phase2(struct buf *, struct inode *, 177 struct allocindir *); 178 static struct allocindir *newallocindir(struct inode *, int, ufs_daddr_t, 179 ufs_daddr_t); 180 static void handle_workitem_freefrag(struct freefrag *); 181 static struct freefrag *newfreefrag(struct inode *, ufs_daddr_t, long); 182 static void allocdirect_merge(struct allocdirectlst *, 183 struct allocdirect *, struct allocdirect *); 184 static struct bmsafemap *bmsafemap_lookup(struct buf *); 185 static int newblk_lookup(struct fs *, ufs_daddr_t, int, 186 struct newblk **); 187 static int inodedep_lookup(struct fs *, ino_t, int, struct inodedep **); 188 static int pagedep_lookup(struct inode *, ufs_lbn_t, int, 189 struct pagedep **); 190 static void pause_timer(void *); 191 static int request_cleanup(int, int); 192 static int process_worklist_item(struct mount *, int); 193 static void add_to_worklist(struct worklist *); 194 195 /* 196 * Exported softdep operations. 197 */ 198 static void softdep_disk_io_initiation(struct buf *); 199 static void softdep_disk_write_complete(struct buf *); 200 static void softdep_deallocate_dependencies(struct buf *); 201 static int softdep_fsync(struct vnode *); 202 static int softdep_process_worklist(struct mount *); 203 static void softdep_move_dependencies(struct buf *, struct buf *); 204 static int softdep_count_dependencies(struct buf *bp, int); 205 206 static struct bio_ops softdep_bioops = { 207 softdep_disk_io_initiation, /* io_start */ 208 softdep_disk_write_complete, /* io_complete */ 209 softdep_deallocate_dependencies, /* io_deallocate */ 210 softdep_fsync, /* io_fsync */ 211 softdep_process_worklist, /* io_sync */ 212 softdep_move_dependencies, /* io_movedeps */ 213 softdep_count_dependencies, /* io_countdeps */ 214 }; 215 216 /* 217 * Locking primitives. 218 * 219 * For a uniprocessor, all we need to do is protect against disk 220 * interrupts. For a multiprocessor, this lock would have to be 221 * a mutex. A single mutex is used throughout this file, though 222 * finer grain locking could be used if contention warranted it. 223 * 224 * For a multiprocessor, the sleep call would accept a lock and 225 * release it after the sleep processing was complete. In a uniprocessor 226 * implementation there is no such interlock, so we simple mark 227 * the places where it needs to be done with the `interlocked' form 228 * of the lock calls. Since the uniprocessor sleep already interlocks 229 * the spl, there is nothing that really needs to be done. 230 */ 231 #ifndef /* NOT */ DEBUG 232 static struct lockit { 233 int lkt_spl; 234 } lk = { 0 }; 235 #define ACQUIRE_LOCK(lk) (lk)->lkt_spl = splbio() 236 #define FREE_LOCK(lk) splx((lk)->lkt_spl) 237 238 #else /* DEBUG */ 239 #define NOHOLDER ((struct thread *)-1) 240 #define SPECIAL_FLAG ((struct thread *)-2) 241 static struct lockit { 242 int lkt_spl; 243 struct thread *lkt_held; 244 } lk = { 0, NOHOLDER }; 245 static int lockcnt; 246 247 static void acquire_lock(struct lockit *); 248 static void free_lock(struct lockit *); 249 void softdep_panic(char *); 250 251 #define ACQUIRE_LOCK(lk) acquire_lock(lk) 252 #define FREE_LOCK(lk) free_lock(lk) 253 254 static void 255 acquire_lock(lk) 256 struct lockit *lk; 257 { 258 thread_t holder; 259 260 if (lk->lkt_held != NOHOLDER) { 261 holder = lk->lkt_held; 262 FREE_LOCK(lk); 263 if (holder == curthread) 264 panic("softdep_lock: locking against myself"); 265 else 266 panic("softdep_lock: lock held by %p", holder); 267 } 268 lk->lkt_spl = splbio(); 269 lk->lkt_held = curthread; 270 lockcnt++; 271 } 272 273 static void 274 free_lock(lk) 275 struct lockit *lk; 276 { 277 278 if (lk->lkt_held == NOHOLDER) 279 panic("softdep_unlock: lock not held"); 280 lk->lkt_held = NOHOLDER; 281 splx(lk->lkt_spl); 282 } 283 284 /* 285 * Function to release soft updates lock and panic. 286 */ 287 void 288 softdep_panic(msg) 289 char *msg; 290 { 291 292 if (lk.lkt_held != NOHOLDER) 293 FREE_LOCK(&lk); 294 panic(msg); 295 } 296 #endif /* DEBUG */ 297 298 static int interlocked_sleep(struct lockit *, int, void *, int, 299 const char *, int); 300 301 /* 302 * When going to sleep, we must save our SPL so that it does 303 * not get lost if some other process uses the lock while we 304 * are sleeping. We restore it after we have slept. This routine 305 * wraps the interlocking with functions that sleep. The list 306 * below enumerates the available set of operations. 307 */ 308 #define UNKNOWN 0 309 #define SLEEP 1 310 #define LOCKBUF 2 311 312 static int 313 interlocked_sleep(lk, op, ident, flags, wmesg, timo) 314 struct lockit *lk; 315 int op; 316 void *ident; 317 int flags; 318 const char *wmesg; 319 int timo; 320 { 321 thread_t holder; 322 int s, retval; 323 324 s = lk->lkt_spl; 325 # ifdef DEBUG 326 if (lk->lkt_held == NOHOLDER) 327 panic("interlocked_sleep: lock not held"); 328 lk->lkt_held = NOHOLDER; 329 # endif /* DEBUG */ 330 switch (op) { 331 case SLEEP: 332 retval = tsleep(ident, flags, wmesg, timo); 333 break; 334 case LOCKBUF: 335 retval = BUF_LOCK((struct buf *)ident, flags); 336 break; 337 default: 338 panic("interlocked_sleep: unknown operation"); 339 } 340 # ifdef DEBUG 341 if (lk->lkt_held != NOHOLDER) { 342 holder = lk->lkt_held; 343 FREE_LOCK(lk); 344 if (holder == curthread) 345 panic("interlocked_sleep: locking against self"); 346 else 347 panic("interlocked_sleep: lock held by %p", holder); 348 } 349 lk->lkt_held = curthread; 350 lockcnt++; 351 # endif /* DEBUG */ 352 lk->lkt_spl = s; 353 return (retval); 354 } 355 356 /* 357 * Place holder for real semaphores. 358 */ 359 struct sema { 360 int value; 361 thread_t holder; 362 char *name; 363 int prio; 364 int timo; 365 }; 366 static void sema_init(struct sema *, char *, int, int); 367 static int sema_get(struct sema *, struct lockit *); 368 static void sema_release(struct sema *); 369 370 static void 371 sema_init(semap, name, prio, timo) 372 struct sema *semap; 373 char *name; 374 int prio, timo; 375 { 376 377 semap->holder = NOHOLDER; 378 semap->value = 0; 379 semap->name = name; 380 semap->prio = prio; 381 semap->timo = timo; 382 } 383 384 static int 385 sema_get(semap, interlock) 386 struct sema *semap; 387 struct lockit *interlock; 388 { 389 390 if (semap->value++ > 0) { 391 if (interlock != NULL) { 392 interlocked_sleep(interlock, SLEEP, (caddr_t)semap, 393 semap->prio, semap->name, semap->timo); 394 FREE_LOCK(interlock); 395 } else { 396 tsleep((caddr_t)semap, semap->prio, semap->name, 397 semap->timo); 398 } 399 return (0); 400 } 401 semap->holder = curthread; 402 if (interlock != NULL) 403 FREE_LOCK(interlock); 404 return (1); 405 } 406 407 static void 408 sema_release(semap) 409 struct sema *semap; 410 { 411 412 if (semap->value <= 0 || semap->holder != curthread) { 413 if (lk.lkt_held != NOHOLDER) 414 FREE_LOCK(&lk); 415 panic("sema_release: not held"); 416 } 417 if (--semap->value > 0) { 418 semap->value = 0; 419 wakeup(semap); 420 } 421 semap->holder = NOHOLDER; 422 } 423 424 /* 425 * Worklist queue management. 426 * These routines require that the lock be held. 427 */ 428 #ifndef /* NOT */ DEBUG 429 #define WORKLIST_INSERT(head, item) do { \ 430 (item)->wk_state |= ONWORKLIST; \ 431 LIST_INSERT_HEAD(head, item, wk_list); \ 432 } while (0) 433 #define WORKLIST_REMOVE(item) do { \ 434 (item)->wk_state &= ~ONWORKLIST; \ 435 LIST_REMOVE(item, wk_list); \ 436 } while (0) 437 #define WORKITEM_FREE(item, type) FREE(item, DtoM(type)) 438 439 #else /* DEBUG */ 440 static void worklist_insert(struct workhead *, struct worklist *); 441 static void worklist_remove(struct worklist *); 442 static void workitem_free(struct worklist *, int); 443 444 #define WORKLIST_INSERT(head, item) worklist_insert(head, item) 445 #define WORKLIST_REMOVE(item) worklist_remove(item) 446 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type) 447 448 static void 449 worklist_insert(head, item) 450 struct workhead *head; 451 struct worklist *item; 452 { 453 454 if (lk.lkt_held == NOHOLDER) 455 panic("worklist_insert: lock not held"); 456 if (item->wk_state & ONWORKLIST) { 457 FREE_LOCK(&lk); 458 panic("worklist_insert: already on list"); 459 } 460 item->wk_state |= ONWORKLIST; 461 LIST_INSERT_HEAD(head, item, wk_list); 462 } 463 464 static void 465 worklist_remove(item) 466 struct worklist *item; 467 { 468 469 if (lk.lkt_held == NOHOLDER) 470 panic("worklist_remove: lock not held"); 471 if ((item->wk_state & ONWORKLIST) == 0) { 472 FREE_LOCK(&lk); 473 panic("worklist_remove: not on list"); 474 } 475 item->wk_state &= ~ONWORKLIST; 476 LIST_REMOVE(item, wk_list); 477 } 478 479 static void 480 workitem_free(item, type) 481 struct worklist *item; 482 int type; 483 { 484 485 if (item->wk_state & ONWORKLIST) { 486 if (lk.lkt_held != NOHOLDER) 487 FREE_LOCK(&lk); 488 panic("workitem_free: still on list"); 489 } 490 if (item->wk_type != type) { 491 if (lk.lkt_held != NOHOLDER) 492 FREE_LOCK(&lk); 493 panic("workitem_free: type mismatch"); 494 } 495 FREE(item, DtoM(type)); 496 } 497 #endif /* DEBUG */ 498 499 /* 500 * Workitem queue management 501 */ 502 static struct workhead softdep_workitem_pending; 503 static int num_on_worklist; /* number of worklist items to be processed */ 504 static int softdep_worklist_busy; /* 1 => trying to do unmount */ 505 static int softdep_worklist_req; /* serialized waiters */ 506 static int max_softdeps; /* maximum number of structs before slowdown */ 507 static int tickdelay = 2; /* number of ticks to pause during slowdown */ 508 static int *stat_countp; /* statistic to count in proc_waiting timeout */ 509 static int proc_waiting; /* tracks whether we have a timeout posted */ 510 static struct callout handle; /* handle on posted proc_waiting timeout */ 511 static struct thread *filesys_syncer; /* proc of filesystem syncer process */ 512 static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 513 #define FLUSH_INODES 1 514 static int req_clear_remove; /* syncer process flush some freeblks */ 515 #define FLUSH_REMOVE 2 516 /* 517 * runtime statistics 518 */ 519 static int stat_worklist_push; /* number of worklist cleanups */ 520 static int stat_blk_limit_push; /* number of times block limit neared */ 521 static int stat_ino_limit_push; /* number of times inode limit neared */ 522 static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 523 static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 524 static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 525 static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 526 static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 527 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 528 static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 529 #ifdef DEBUG 530 #include <vm/vm.h> 531 #include <sys/sysctl.h> 532 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, ""); 533 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, ""); 534 SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,""); 535 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,""); 536 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,""); 537 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, ""); 538 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, ""); 539 SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, ""); 540 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, ""); 541 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, ""); 542 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, ""); 543 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, ""); 544 #endif /* DEBUG */ 545 546 /* 547 * Add an item to the end of the work queue. 548 * This routine requires that the lock be held. 549 * This is the only routine that adds items to the list. 550 * The following routine is the only one that removes items 551 * and does so in order from first to last. 552 */ 553 static void 554 add_to_worklist(wk) 555 struct worklist *wk; 556 { 557 static struct worklist *worklist_tail; 558 559 if (wk->wk_state & ONWORKLIST) { 560 if (lk.lkt_held != NOHOLDER) 561 FREE_LOCK(&lk); 562 panic("add_to_worklist: already on list"); 563 } 564 wk->wk_state |= ONWORKLIST; 565 if (LIST_FIRST(&softdep_workitem_pending) == NULL) 566 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list); 567 else 568 LIST_INSERT_AFTER(worklist_tail, wk, wk_list); 569 worklist_tail = wk; 570 num_on_worklist += 1; 571 } 572 573 /* 574 * Process that runs once per second to handle items in the background queue. 575 * 576 * Note that we ensure that everything is done in the order in which they 577 * appear in the queue. The code below depends on this property to ensure 578 * that blocks of a file are freed before the inode itself is freed. This 579 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 580 * until all the old ones have been purged from the dependency lists. 581 */ 582 static int 583 softdep_process_worklist(matchmnt) 584 struct mount *matchmnt; 585 { 586 thread_t td = curthread; 587 int matchcnt, loopcount; 588 long starttime; 589 590 /* 591 * Record the process identifier of our caller so that we can give 592 * this process preferential treatment in request_cleanup below. 593 */ 594 filesys_syncer = td; 595 matchcnt = 0; 596 597 /* 598 * There is no danger of having multiple processes run this 599 * code, but we have to single-thread it when softdep_flushfiles() 600 * is in operation to get an accurate count of the number of items 601 * related to its mount point that are in the list. 602 */ 603 if (matchmnt == NULL) { 604 if (softdep_worklist_busy < 0) 605 return(-1); 606 softdep_worklist_busy += 1; 607 } 608 609 /* 610 * If requested, try removing inode or removal dependencies. 611 */ 612 if (req_clear_inodedeps) { 613 clear_inodedeps(td); 614 req_clear_inodedeps -= 1; 615 wakeup_one(&proc_waiting); 616 } 617 if (req_clear_remove) { 618 clear_remove(td); 619 req_clear_remove -= 1; 620 wakeup_one(&proc_waiting); 621 } 622 loopcount = 1; 623 starttime = time_second; 624 while (num_on_worklist > 0) { 625 matchcnt += process_worklist_item(matchmnt, 0); 626 627 /* 628 * If a umount operation wants to run the worklist 629 * accurately, abort. 630 */ 631 if (softdep_worklist_req && matchmnt == NULL) { 632 matchcnt = -1; 633 break; 634 } 635 636 /* 637 * If requested, try removing inode or removal dependencies. 638 */ 639 if (req_clear_inodedeps) { 640 clear_inodedeps(td); 641 req_clear_inodedeps -= 1; 642 wakeup_one(&proc_waiting); 643 } 644 if (req_clear_remove) { 645 clear_remove(td); 646 req_clear_remove -= 1; 647 wakeup_one(&proc_waiting); 648 } 649 /* 650 * We do not generally want to stop for buffer space, but if 651 * we are really being a buffer hog, we will stop and wait. 652 */ 653 if (loopcount++ % 128 == 0) 654 bwillwrite(); 655 /* 656 * Never allow processing to run for more than one 657 * second. Otherwise the other syncer tasks may get 658 * excessively backlogged. 659 */ 660 if (starttime != time_second && matchmnt == NULL) { 661 matchcnt = -1; 662 break; 663 } 664 } 665 if (matchmnt == NULL) { 666 --softdep_worklist_busy; 667 if (softdep_worklist_req && softdep_worklist_busy == 0) 668 wakeup(&softdep_worklist_req); 669 } 670 return (matchcnt); 671 } 672 673 /* 674 * Process one item on the worklist. 675 */ 676 static int 677 process_worklist_item(matchmnt, flags) 678 struct mount *matchmnt; 679 int flags; 680 { 681 struct worklist *wk; 682 struct dirrem *dirrem; 683 struct fs *matchfs; 684 struct vnode *vp; 685 int matchcnt = 0; 686 687 matchfs = NULL; 688 if (matchmnt != NULL) 689 matchfs = VFSTOUFS(matchmnt)->um_fs; 690 ACQUIRE_LOCK(&lk); 691 /* 692 * Normally we just process each item on the worklist in order. 693 * However, if we are in a situation where we cannot lock any 694 * inodes, we have to skip over any dirrem requests whose 695 * vnodes are resident and locked. 696 */ 697 LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) { 698 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM) 699 break; 700 dirrem = WK_DIRREM(wk); 701 vp = ufs_ihashlookup(VFSTOUFS(dirrem->dm_mnt)->um_dev, 702 dirrem->dm_oldinum); 703 if (vp == NULL || !VOP_ISLOCKED(vp, curthread)) 704 break; 705 } 706 if (wk == 0) { 707 FREE_LOCK(&lk); 708 return (0); 709 } 710 WORKLIST_REMOVE(wk); 711 num_on_worklist -= 1; 712 FREE_LOCK(&lk); 713 switch (wk->wk_type) { 714 715 case D_DIRREM: 716 /* removal of a directory entry */ 717 if (WK_DIRREM(wk)->dm_mnt == matchmnt) 718 matchcnt += 1; 719 handle_workitem_remove(WK_DIRREM(wk)); 720 break; 721 722 case D_FREEBLKS: 723 /* releasing blocks and/or fragments from a file */ 724 if (WK_FREEBLKS(wk)->fb_fs == matchfs) 725 matchcnt += 1; 726 handle_workitem_freeblocks(WK_FREEBLKS(wk)); 727 break; 728 729 case D_FREEFRAG: 730 /* releasing a fragment when replaced as a file grows */ 731 if (WK_FREEFRAG(wk)->ff_fs == matchfs) 732 matchcnt += 1; 733 handle_workitem_freefrag(WK_FREEFRAG(wk)); 734 break; 735 736 case D_FREEFILE: 737 /* releasing an inode when its link count drops to 0 */ 738 if (WK_FREEFILE(wk)->fx_fs == matchfs) 739 matchcnt += 1; 740 handle_workitem_freefile(WK_FREEFILE(wk)); 741 break; 742 743 default: 744 panic("%s_process_worklist: Unknown type %s", 745 "softdep", TYPENAME(wk->wk_type)); 746 /* NOTREACHED */ 747 } 748 return (matchcnt); 749 } 750 751 /* 752 * Move dependencies from one buffer to another. 753 */ 754 static void 755 softdep_move_dependencies(oldbp, newbp) 756 struct buf *oldbp; 757 struct buf *newbp; 758 { 759 struct worklist *wk, *wktail; 760 761 if (LIST_FIRST(&newbp->b_dep) != NULL) 762 panic("softdep_move_dependencies: need merge code"); 763 wktail = 0; 764 ACQUIRE_LOCK(&lk); 765 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 766 LIST_REMOVE(wk, wk_list); 767 if (wktail == 0) 768 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 769 else 770 LIST_INSERT_AFTER(wktail, wk, wk_list); 771 wktail = wk; 772 } 773 FREE_LOCK(&lk); 774 } 775 776 /* 777 * Purge the work list of all items associated with a particular mount point. 778 */ 779 int 780 softdep_flushfiles(struct mount *oldmnt, int flags, struct thread *td) 781 { 782 struct vnode *devvp; 783 int error, loopcnt; 784 785 /* 786 * Await our turn to clear out the queue, then serialize access. 787 */ 788 while (softdep_worklist_busy != 0) { 789 softdep_worklist_req += 1; 790 tsleep(&softdep_worklist_req, 0, "softflush", 0); 791 softdep_worklist_req -= 1; 792 } 793 softdep_worklist_busy = -1; 794 795 if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0) { 796 softdep_worklist_busy = 0; 797 if (softdep_worklist_req) 798 wakeup(&softdep_worklist_req); 799 return (error); 800 } 801 /* 802 * Alternately flush the block device associated with the mount 803 * point and process any dependencies that the flushing 804 * creates. In theory, this loop can happen at most twice, 805 * but we give it a few extra just to be sure. 806 */ 807 devvp = VFSTOUFS(oldmnt)->um_devvp; 808 for (loopcnt = 10; loopcnt > 0; ) { 809 if (softdep_process_worklist(oldmnt) == 0) { 810 loopcnt--; 811 /* 812 * Do another flush in case any vnodes were brought in 813 * as part of the cleanup operations. 814 */ 815 if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0) 816 break; 817 /* 818 * If we still found nothing to do, we are really done. 819 */ 820 if (softdep_process_worklist(oldmnt) == 0) 821 break; 822 } 823 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, td); 824 error = VOP_FSYNC(devvp, MNT_WAIT, td); 825 VOP_UNLOCK(devvp, 0, td); 826 if (error) 827 break; 828 } 829 softdep_worklist_busy = 0; 830 if (softdep_worklist_req) 831 wakeup(&softdep_worklist_req); 832 833 /* 834 * If we are unmounting then it is an error to fail. If we 835 * are simply trying to downgrade to read-only, then filesystem 836 * activity can keep us busy forever, so we just fail with EBUSY. 837 */ 838 if (loopcnt == 0) { 839 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 840 panic("softdep_flushfiles: looping"); 841 error = EBUSY; 842 } 843 return (error); 844 } 845 846 /* 847 * Structure hashing. 848 * 849 * There are three types of structures that can be looked up: 850 * 1) pagedep structures identified by mount point, inode number, 851 * and logical block. 852 * 2) inodedep structures identified by mount point and inode number. 853 * 3) newblk structures identified by mount point and 854 * physical block number. 855 * 856 * The "pagedep" and "inodedep" dependency structures are hashed 857 * separately from the file blocks and inodes to which they correspond. 858 * This separation helps when the in-memory copy of an inode or 859 * file block must be replaced. It also obviates the need to access 860 * an inode or file page when simply updating (or de-allocating) 861 * dependency structures. Lookup of newblk structures is needed to 862 * find newly allocated blocks when trying to associate them with 863 * their allocdirect or allocindir structure. 864 * 865 * The lookup routines optionally create and hash a new instance when 866 * an existing entry is not found. 867 */ 868 #define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 869 #define NODELAY 0x0002 /* cannot do background work */ 870 871 /* 872 * Structures and routines associated with pagedep caching. 873 */ 874 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl; 875 u_long pagedep_hash; /* size of hash table - 1 */ 876 #define PAGEDEP_HASH(mp, inum, lbn) \ 877 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \ 878 pagedep_hash]) 879 static struct sema pagedep_in_progress; 880 881 /* 882 * Look up a pagedep. Return 1 if found, 0 if not found. 883 * If not found, allocate if DEPALLOC flag is passed. 884 * Found or allocated entry is returned in pagedeppp. 885 * This routine must be called with splbio interrupts blocked. 886 */ 887 static int 888 pagedep_lookup(ip, lbn, flags, pagedeppp) 889 struct inode *ip; 890 ufs_lbn_t lbn; 891 int flags; 892 struct pagedep **pagedeppp; 893 { 894 struct pagedep *pagedep; 895 struct pagedep_hashhead *pagedephd; 896 struct mount *mp; 897 int i; 898 899 #ifdef DEBUG 900 if (lk.lkt_held == NOHOLDER) 901 panic("pagedep_lookup: lock not held"); 902 #endif 903 mp = ITOV(ip)->v_mount; 904 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn); 905 top: 906 LIST_FOREACH(pagedep, pagedephd, pd_hash) 907 if (ip->i_number == pagedep->pd_ino && 908 lbn == pagedep->pd_lbn && 909 mp == pagedep->pd_mnt) 910 break; 911 if (pagedep) { 912 *pagedeppp = pagedep; 913 return (1); 914 } 915 if ((flags & DEPALLOC) == 0) { 916 *pagedeppp = NULL; 917 return (0); 918 } 919 if (sema_get(&pagedep_in_progress, &lk) == 0) { 920 ACQUIRE_LOCK(&lk); 921 goto top; 922 } 923 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP, 924 M_SOFTDEP_FLAGS); 925 bzero(pagedep, sizeof(struct pagedep)); 926 pagedep->pd_list.wk_type = D_PAGEDEP; 927 pagedep->pd_mnt = mp; 928 pagedep->pd_ino = ip->i_number; 929 pagedep->pd_lbn = lbn; 930 LIST_INIT(&pagedep->pd_dirremhd); 931 LIST_INIT(&pagedep->pd_pendinghd); 932 for (i = 0; i < DAHASHSZ; i++) 933 LIST_INIT(&pagedep->pd_diraddhd[i]); 934 ACQUIRE_LOCK(&lk); 935 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 936 sema_release(&pagedep_in_progress); 937 *pagedeppp = pagedep; 938 return (0); 939 } 940 941 /* 942 * Structures and routines associated with inodedep caching. 943 */ 944 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl; 945 static u_long inodedep_hash; /* size of hash table - 1 */ 946 static long num_inodedep; /* number of inodedep allocated */ 947 #define INODEDEP_HASH(fs, inum) \ 948 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash]) 949 static struct sema inodedep_in_progress; 950 951 /* 952 * Look up a inodedep. Return 1 if found, 0 if not found. 953 * If not found, allocate if DEPALLOC flag is passed. 954 * Found or allocated entry is returned in inodedeppp. 955 * This routine must be called with splbio interrupts blocked. 956 */ 957 static int 958 inodedep_lookup(fs, inum, flags, inodedeppp) 959 struct fs *fs; 960 ino_t inum; 961 int flags; 962 struct inodedep **inodedeppp; 963 { 964 struct inodedep *inodedep; 965 struct inodedep_hashhead *inodedephd; 966 int firsttry; 967 968 #ifdef DEBUG 969 if (lk.lkt_held == NOHOLDER) 970 panic("inodedep_lookup: lock not held"); 971 #endif 972 firsttry = 1; 973 inodedephd = INODEDEP_HASH(fs, inum); 974 top: 975 LIST_FOREACH(inodedep, inodedephd, id_hash) 976 if (inum == inodedep->id_ino && fs == inodedep->id_fs) 977 break; 978 if (inodedep) { 979 *inodedeppp = inodedep; 980 return (1); 981 } 982 if ((flags & DEPALLOC) == 0) { 983 *inodedeppp = NULL; 984 return (0); 985 } 986 /* 987 * If we are over our limit, try to improve the situation. 988 */ 989 if (num_inodedep > max_softdeps && firsttry && 990 speedup_syncer() == 0 && (flags & NODELAY) == 0 && 991 request_cleanup(FLUSH_INODES, 1)) { 992 firsttry = 0; 993 goto top; 994 } 995 if (sema_get(&inodedep_in_progress, &lk) == 0) { 996 ACQUIRE_LOCK(&lk); 997 goto top; 998 } 999 num_inodedep += 1; 1000 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep), 1001 M_INODEDEP, M_SOFTDEP_FLAGS); 1002 inodedep->id_list.wk_type = D_INODEDEP; 1003 inodedep->id_fs = fs; 1004 inodedep->id_ino = inum; 1005 inodedep->id_state = ALLCOMPLETE; 1006 inodedep->id_nlinkdelta = 0; 1007 inodedep->id_savedino = NULL; 1008 inodedep->id_savedsize = -1; 1009 inodedep->id_buf = NULL; 1010 LIST_INIT(&inodedep->id_pendinghd); 1011 LIST_INIT(&inodedep->id_inowait); 1012 LIST_INIT(&inodedep->id_bufwait); 1013 TAILQ_INIT(&inodedep->id_inoupdt); 1014 TAILQ_INIT(&inodedep->id_newinoupdt); 1015 ACQUIRE_LOCK(&lk); 1016 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 1017 sema_release(&inodedep_in_progress); 1018 *inodedeppp = inodedep; 1019 return (0); 1020 } 1021 1022 /* 1023 * Structures and routines associated with newblk caching. 1024 */ 1025 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl; 1026 u_long newblk_hash; /* size of hash table - 1 */ 1027 #define NEWBLK_HASH(fs, inum) \ 1028 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash]) 1029 static struct sema newblk_in_progress; 1030 1031 /* 1032 * Look up a newblk. Return 1 if found, 0 if not found. 1033 * If not found, allocate if DEPALLOC flag is passed. 1034 * Found or allocated entry is returned in newblkpp. 1035 */ 1036 static int 1037 newblk_lookup(fs, newblkno, flags, newblkpp) 1038 struct fs *fs; 1039 ufs_daddr_t newblkno; 1040 int flags; 1041 struct newblk **newblkpp; 1042 { 1043 struct newblk *newblk; 1044 struct newblk_hashhead *newblkhd; 1045 1046 newblkhd = NEWBLK_HASH(fs, newblkno); 1047 top: 1048 LIST_FOREACH(newblk, newblkhd, nb_hash) 1049 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs) 1050 break; 1051 if (newblk) { 1052 *newblkpp = newblk; 1053 return (1); 1054 } 1055 if ((flags & DEPALLOC) == 0) { 1056 *newblkpp = NULL; 1057 return (0); 1058 } 1059 if (sema_get(&newblk_in_progress, 0) == 0) 1060 goto top; 1061 MALLOC(newblk, struct newblk *, sizeof(struct newblk), 1062 M_NEWBLK, M_SOFTDEP_FLAGS); 1063 newblk->nb_state = 0; 1064 newblk->nb_fs = fs; 1065 newblk->nb_newblkno = newblkno; 1066 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 1067 sema_release(&newblk_in_progress); 1068 *newblkpp = newblk; 1069 return (0); 1070 } 1071 1072 /* 1073 * Executed during filesystem system initialization before 1074 * mounting any filesystems. 1075 */ 1076 void 1077 softdep_initialize() 1078 { 1079 callout_init(&handle); 1080 bioops = softdep_bioops; /* XXX hack */ 1081 1082 LIST_INIT(&mkdirlisthd); 1083 LIST_INIT(&softdep_workitem_pending); 1084 max_softdeps = min(desiredvnodes * 8, 1085 M_INODEDEP->ks_limit / (2 * sizeof(struct inodedep))); 1086 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 1087 &pagedep_hash); 1088 sema_init(&pagedep_in_progress, "pagedep", 0, 0); 1089 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash); 1090 sema_init(&inodedep_in_progress, "inodedep", 0, 0); 1091 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash); 1092 sema_init(&newblk_in_progress, "newblk", 0, 0); 1093 } 1094 1095 /* 1096 * Called at mount time to notify the dependency code that a 1097 * filesystem wishes to use it. 1098 */ 1099 int 1100 softdep_mount(devvp, mp, fs) 1101 struct vnode *devvp; 1102 struct mount *mp; 1103 struct fs *fs; 1104 { 1105 struct csum cstotal; 1106 struct cg *cgp; 1107 struct buf *bp; 1108 int error, cyl; 1109 1110 mp->mnt_flag &= ~MNT_ASYNC; 1111 mp->mnt_flag |= MNT_SOFTDEP; 1112 /* 1113 * When doing soft updates, the counters in the 1114 * superblock may have gotten out of sync, so we have 1115 * to scan the cylinder groups and recalculate them. 1116 */ 1117 if (fs->fs_clean != 0) 1118 return (0); 1119 bzero(&cstotal, sizeof cstotal); 1120 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 1121 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 1122 fs->fs_cgsize, &bp)) != 0) { 1123 brelse(bp); 1124 return (error); 1125 } 1126 cgp = (struct cg *)bp->b_data; 1127 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 1128 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 1129 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 1130 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 1131 fs->fs_cs(fs, cyl) = cgp->cg_cs; 1132 brelse(bp); 1133 } 1134 #ifdef DEBUG 1135 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 1136 printf("ffs_mountfs: superblock updated for soft updates\n"); 1137 #endif 1138 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 1139 return (0); 1140 } 1141 1142 /* 1143 * Protecting the freemaps (or bitmaps). 1144 * 1145 * To eliminate the need to execute fsck before mounting a filesystem 1146 * after a power failure, one must (conservatively) guarantee that the 1147 * on-disk copy of the bitmaps never indicate that a live inode or block is 1148 * free. So, when a block or inode is allocated, the bitmap should be 1149 * updated (on disk) before any new pointers. When a block or inode is 1150 * freed, the bitmap should not be updated until all pointers have been 1151 * reset. The latter dependency is handled by the delayed de-allocation 1152 * approach described below for block and inode de-allocation. The former 1153 * dependency is handled by calling the following procedure when a block or 1154 * inode is allocated. When an inode is allocated an "inodedep" is created 1155 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 1156 * Each "inodedep" is also inserted into the hash indexing structure so 1157 * that any additional link additions can be made dependent on the inode 1158 * allocation. 1159 * 1160 * The ufs filesystem maintains a number of free block counts (e.g., per 1161 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 1162 * in addition to the bitmaps. These counts are used to improve efficiency 1163 * during allocation and therefore must be consistent with the bitmaps. 1164 * There is no convenient way to guarantee post-crash consistency of these 1165 * counts with simple update ordering, for two main reasons: (1) The counts 1166 * and bitmaps for a single cylinder group block are not in the same disk 1167 * sector. If a disk write is interrupted (e.g., by power failure), one may 1168 * be written and the other not. (2) Some of the counts are located in the 1169 * superblock rather than the cylinder group block. So, we focus our soft 1170 * updates implementation on protecting the bitmaps. When mounting a 1171 * filesystem, we recompute the auxiliary counts from the bitmaps. 1172 */ 1173 1174 /* 1175 * Called just after updating the cylinder group block to allocate an inode. 1176 */ 1177 void 1178 softdep_setup_inomapdep(bp, ip, newinum) 1179 struct buf *bp; /* buffer for cylgroup block with inode map */ 1180 struct inode *ip; /* inode related to allocation */ 1181 ino_t newinum; /* new inode number being allocated */ 1182 { 1183 struct inodedep *inodedep; 1184 struct bmsafemap *bmsafemap; 1185 1186 /* 1187 * Create a dependency for the newly allocated inode. 1188 * Panic if it already exists as something is seriously wrong. 1189 * Otherwise add it to the dependency list for the buffer holding 1190 * the cylinder group map from which it was allocated. 1191 */ 1192 ACQUIRE_LOCK(&lk); 1193 if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) { 1194 FREE_LOCK(&lk); 1195 panic("softdep_setup_inomapdep: found inode"); 1196 } 1197 inodedep->id_buf = bp; 1198 inodedep->id_state &= ~DEPCOMPLETE; 1199 bmsafemap = bmsafemap_lookup(bp); 1200 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 1201 FREE_LOCK(&lk); 1202 } 1203 1204 /* 1205 * Called just after updating the cylinder group block to 1206 * allocate block or fragment. 1207 */ 1208 void 1209 softdep_setup_blkmapdep(bp, fs, newblkno) 1210 struct buf *bp; /* buffer for cylgroup block with block map */ 1211 struct fs *fs; /* filesystem doing allocation */ 1212 ufs_daddr_t newblkno; /* number of newly allocated block */ 1213 { 1214 struct newblk *newblk; 1215 struct bmsafemap *bmsafemap; 1216 1217 /* 1218 * Create a dependency for the newly allocated block. 1219 * Add it to the dependency list for the buffer holding 1220 * the cylinder group map from which it was allocated. 1221 */ 1222 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0) 1223 panic("softdep_setup_blkmapdep: found block"); 1224 ACQUIRE_LOCK(&lk); 1225 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp); 1226 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 1227 FREE_LOCK(&lk); 1228 } 1229 1230 /* 1231 * Find the bmsafemap associated with a cylinder group buffer. 1232 * If none exists, create one. The buffer must be locked when 1233 * this routine is called and this routine must be called with 1234 * splbio interrupts blocked. 1235 */ 1236 static struct bmsafemap * 1237 bmsafemap_lookup(bp) 1238 struct buf *bp; 1239 { 1240 struct bmsafemap *bmsafemap; 1241 struct worklist *wk; 1242 1243 #ifdef DEBUG 1244 if (lk.lkt_held == NOHOLDER) 1245 panic("bmsafemap_lookup: lock not held"); 1246 #endif 1247 LIST_FOREACH(wk, &bp->b_dep, wk_list) 1248 if (wk->wk_type == D_BMSAFEMAP) 1249 return (WK_BMSAFEMAP(wk)); 1250 FREE_LOCK(&lk); 1251 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap), 1252 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 1253 bmsafemap->sm_list.wk_type = D_BMSAFEMAP; 1254 bmsafemap->sm_list.wk_state = 0; 1255 bmsafemap->sm_buf = bp; 1256 LIST_INIT(&bmsafemap->sm_allocdirecthd); 1257 LIST_INIT(&bmsafemap->sm_allocindirhd); 1258 LIST_INIT(&bmsafemap->sm_inodedephd); 1259 LIST_INIT(&bmsafemap->sm_newblkhd); 1260 ACQUIRE_LOCK(&lk); 1261 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 1262 return (bmsafemap); 1263 } 1264 1265 /* 1266 * Direct block allocation dependencies. 1267 * 1268 * When a new block is allocated, the corresponding disk locations must be 1269 * initialized (with zeros or new data) before the on-disk inode points to 1270 * them. Also, the freemap from which the block was allocated must be 1271 * updated (on disk) before the inode's pointer. These two dependencies are 1272 * independent of each other and are needed for all file blocks and indirect 1273 * blocks that are pointed to directly by the inode. Just before the 1274 * "in-core" version of the inode is updated with a newly allocated block 1275 * number, a procedure (below) is called to setup allocation dependency 1276 * structures. These structures are removed when the corresponding 1277 * dependencies are satisfied or when the block allocation becomes obsolete 1278 * (i.e., the file is deleted, the block is de-allocated, or the block is a 1279 * fragment that gets upgraded). All of these cases are handled in 1280 * procedures described later. 1281 * 1282 * When a file extension causes a fragment to be upgraded, either to a larger 1283 * fragment or to a full block, the on-disk location may change (if the 1284 * previous fragment could not simply be extended). In this case, the old 1285 * fragment must be de-allocated, but not until after the inode's pointer has 1286 * been updated. In most cases, this is handled by later procedures, which 1287 * will construct a "freefrag" structure to be added to the workitem queue 1288 * when the inode update is complete (or obsolete). The main exception to 1289 * this is when an allocation occurs while a pending allocation dependency 1290 * (for the same block pointer) remains. This case is handled in the main 1291 * allocation dependency setup procedure by immediately freeing the 1292 * unreferenced fragments. 1293 */ 1294 void 1295 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 1296 struct inode *ip; /* inode to which block is being added */ 1297 ufs_lbn_t lbn; /* block pointer within inode */ 1298 ufs_daddr_t newblkno; /* disk block number being added */ 1299 ufs_daddr_t oldblkno; /* previous block number, 0 unless frag */ 1300 long newsize; /* size of new block */ 1301 long oldsize; /* size of new block */ 1302 struct buf *bp; /* bp for allocated block */ 1303 { 1304 struct allocdirect *adp, *oldadp; 1305 struct allocdirectlst *adphead; 1306 struct bmsafemap *bmsafemap; 1307 struct inodedep *inodedep; 1308 struct pagedep *pagedep; 1309 struct newblk *newblk; 1310 1311 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect), 1312 M_ALLOCDIRECT, M_SOFTDEP_FLAGS); 1313 bzero(adp, sizeof(struct allocdirect)); 1314 adp->ad_list.wk_type = D_ALLOCDIRECT; 1315 adp->ad_lbn = lbn; 1316 adp->ad_newblkno = newblkno; 1317 adp->ad_oldblkno = oldblkno; 1318 adp->ad_newsize = newsize; 1319 adp->ad_oldsize = oldsize; 1320 adp->ad_state = ATTACHED; 1321 if (newblkno == oldblkno) 1322 adp->ad_freefrag = NULL; 1323 else 1324 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1325 1326 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1327 panic("softdep_setup_allocdirect: lost block"); 1328 1329 ACQUIRE_LOCK(&lk); 1330 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep); 1331 adp->ad_inodedep = inodedep; 1332 1333 if (newblk->nb_state == DEPCOMPLETE) { 1334 adp->ad_state |= DEPCOMPLETE; 1335 adp->ad_buf = NULL; 1336 } else { 1337 bmsafemap = newblk->nb_bmsafemap; 1338 adp->ad_buf = bmsafemap->sm_buf; 1339 LIST_REMOVE(newblk, nb_deps); 1340 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1341 } 1342 LIST_REMOVE(newblk, nb_hash); 1343 FREE(newblk, M_NEWBLK); 1344 1345 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list); 1346 if (lbn >= NDADDR) { 1347 /* allocating an indirect block */ 1348 if (oldblkno != 0) { 1349 FREE_LOCK(&lk); 1350 panic("softdep_setup_allocdirect: non-zero indir"); 1351 } 1352 } else { 1353 /* 1354 * Allocating a direct block. 1355 * 1356 * If we are allocating a directory block, then we must 1357 * allocate an associated pagedep to track additions and 1358 * deletions. 1359 */ 1360 if ((ip->i_mode & IFMT) == IFDIR && 1361 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1362 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 1363 } 1364 /* 1365 * The list of allocdirects must be kept in sorted and ascending 1366 * order so that the rollback routines can quickly determine the 1367 * first uncommitted block (the size of the file stored on disk 1368 * ends at the end of the lowest committed fragment, or if there 1369 * are no fragments, at the end of the highest committed block). 1370 * Since files generally grow, the typical case is that the new 1371 * block is to be added at the end of the list. We speed this 1372 * special case by checking against the last allocdirect in the 1373 * list before laboriously traversing the list looking for the 1374 * insertion point. 1375 */ 1376 adphead = &inodedep->id_newinoupdt; 1377 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1378 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1379 /* insert at end of list */ 1380 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1381 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1382 allocdirect_merge(adphead, adp, oldadp); 1383 FREE_LOCK(&lk); 1384 return; 1385 } 1386 TAILQ_FOREACH(oldadp, adphead, ad_next) { 1387 if (oldadp->ad_lbn >= lbn) 1388 break; 1389 } 1390 if (oldadp == NULL) { 1391 FREE_LOCK(&lk); 1392 panic("softdep_setup_allocdirect: lost entry"); 1393 } 1394 /* insert in middle of list */ 1395 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1396 if (oldadp->ad_lbn == lbn) 1397 allocdirect_merge(adphead, adp, oldadp); 1398 FREE_LOCK(&lk); 1399 } 1400 1401 /* 1402 * Replace an old allocdirect dependency with a newer one. 1403 * This routine must be called with splbio interrupts blocked. 1404 */ 1405 static void 1406 allocdirect_merge(adphead, newadp, oldadp) 1407 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 1408 struct allocdirect *newadp; /* allocdirect being added */ 1409 struct allocdirect *oldadp; /* existing allocdirect being checked */ 1410 { 1411 struct freefrag *freefrag; 1412 1413 #ifdef DEBUG 1414 if (lk.lkt_held == NOHOLDER) 1415 panic("allocdirect_merge: lock not held"); 1416 #endif 1417 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 1418 newadp->ad_oldsize != oldadp->ad_newsize || 1419 newadp->ad_lbn >= NDADDR) { 1420 FREE_LOCK(&lk); 1421 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d", 1422 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn, 1423 NDADDR); 1424 } 1425 newadp->ad_oldblkno = oldadp->ad_oldblkno; 1426 newadp->ad_oldsize = oldadp->ad_oldsize; 1427 /* 1428 * If the old dependency had a fragment to free or had never 1429 * previously had a block allocated, then the new dependency 1430 * can immediately post its freefrag and adopt the old freefrag. 1431 * This action is done by swapping the freefrag dependencies. 1432 * The new dependency gains the old one's freefrag, and the 1433 * old one gets the new one and then immediately puts it on 1434 * the worklist when it is freed by free_allocdirect. It is 1435 * not possible to do this swap when the old dependency had a 1436 * non-zero size but no previous fragment to free. This condition 1437 * arises when the new block is an extension of the old block. 1438 * Here, the first part of the fragment allocated to the new 1439 * dependency is part of the block currently claimed on disk by 1440 * the old dependency, so cannot legitimately be freed until the 1441 * conditions for the new dependency are fulfilled. 1442 */ 1443 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 1444 freefrag = newadp->ad_freefrag; 1445 newadp->ad_freefrag = oldadp->ad_freefrag; 1446 oldadp->ad_freefrag = freefrag; 1447 } 1448 free_allocdirect(adphead, oldadp, 0); 1449 } 1450 1451 /* 1452 * Allocate a new freefrag structure if needed. 1453 */ 1454 static struct freefrag * 1455 newfreefrag(ip, blkno, size) 1456 struct inode *ip; 1457 ufs_daddr_t blkno; 1458 long size; 1459 { 1460 struct freefrag *freefrag; 1461 struct fs *fs; 1462 1463 if (blkno == 0) 1464 return (NULL); 1465 fs = ip->i_fs; 1466 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 1467 panic("newfreefrag: frag size"); 1468 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag), 1469 M_FREEFRAG, M_SOFTDEP_FLAGS); 1470 freefrag->ff_list.wk_type = D_FREEFRAG; 1471 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */ 1472 freefrag->ff_inum = ip->i_number; 1473 freefrag->ff_fs = fs; 1474 freefrag->ff_devvp = ip->i_devvp; 1475 freefrag->ff_blkno = blkno; 1476 freefrag->ff_fragsize = size; 1477 return (freefrag); 1478 } 1479 1480 /* 1481 * This workitem de-allocates fragments that were replaced during 1482 * file block allocation. 1483 */ 1484 static void 1485 handle_workitem_freefrag(freefrag) 1486 struct freefrag *freefrag; 1487 { 1488 struct inode tip; 1489 1490 tip.i_fs = freefrag->ff_fs; 1491 tip.i_devvp = freefrag->ff_devvp; 1492 tip.i_dev = freefrag->ff_devvp->v_rdev; 1493 tip.i_number = freefrag->ff_inum; 1494 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */ 1495 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize); 1496 FREE(freefrag, M_FREEFRAG); 1497 } 1498 1499 /* 1500 * Indirect block allocation dependencies. 1501 * 1502 * The same dependencies that exist for a direct block also exist when 1503 * a new block is allocated and pointed to by an entry in a block of 1504 * indirect pointers. The undo/redo states described above are also 1505 * used here. Because an indirect block contains many pointers that 1506 * may have dependencies, a second copy of the entire in-memory indirect 1507 * block is kept. The buffer cache copy is always completely up-to-date. 1508 * The second copy, which is used only as a source for disk writes, 1509 * contains only the safe pointers (i.e., those that have no remaining 1510 * update dependencies). The second copy is freed when all pointers 1511 * are safe. The cache is not allowed to replace indirect blocks with 1512 * pending update dependencies. If a buffer containing an indirect 1513 * block with dependencies is written, these routines will mark it 1514 * dirty again. It can only be successfully written once all the 1515 * dependencies are removed. The ffs_fsync routine in conjunction with 1516 * softdep_sync_metadata work together to get all the dependencies 1517 * removed so that a file can be successfully written to disk. Three 1518 * procedures are used when setting up indirect block pointer 1519 * dependencies. The division is necessary because of the organization 1520 * of the "balloc" routine and because of the distinction between file 1521 * pages and file metadata blocks. 1522 */ 1523 1524 /* 1525 * Allocate a new allocindir structure. 1526 */ 1527 static struct allocindir * 1528 newallocindir(ip, ptrno, newblkno, oldblkno) 1529 struct inode *ip; /* inode for file being extended */ 1530 int ptrno; /* offset of pointer in indirect block */ 1531 ufs_daddr_t newblkno; /* disk block number being added */ 1532 ufs_daddr_t oldblkno; /* previous block number, 0 if none */ 1533 { 1534 struct allocindir *aip; 1535 1536 MALLOC(aip, struct allocindir *, sizeof(struct allocindir), 1537 M_ALLOCINDIR, M_SOFTDEP_FLAGS); 1538 bzero(aip, sizeof(struct allocindir)); 1539 aip->ai_list.wk_type = D_ALLOCINDIR; 1540 aip->ai_state = ATTACHED; 1541 aip->ai_offset = ptrno; 1542 aip->ai_newblkno = newblkno; 1543 aip->ai_oldblkno = oldblkno; 1544 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize); 1545 return (aip); 1546 } 1547 1548 /* 1549 * Called just before setting an indirect block pointer 1550 * to a newly allocated file page. 1551 */ 1552 void 1553 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 1554 struct inode *ip; /* inode for file being extended */ 1555 ufs_lbn_t lbn; /* allocated block number within file */ 1556 struct buf *bp; /* buffer with indirect blk referencing page */ 1557 int ptrno; /* offset of pointer in indirect block */ 1558 ufs_daddr_t newblkno; /* disk block number being added */ 1559 ufs_daddr_t oldblkno; /* previous block number, 0 if none */ 1560 struct buf *nbp; /* buffer holding allocated page */ 1561 { 1562 struct allocindir *aip; 1563 struct pagedep *pagedep; 1564 1565 aip = newallocindir(ip, ptrno, newblkno, oldblkno); 1566 ACQUIRE_LOCK(&lk); 1567 /* 1568 * If we are allocating a directory page, then we must 1569 * allocate an associated pagedep to track additions and 1570 * deletions. 1571 */ 1572 if ((ip->i_mode & IFMT) == IFDIR && 1573 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1574 WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list); 1575 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1576 FREE_LOCK(&lk); 1577 setup_allocindir_phase2(bp, ip, aip); 1578 } 1579 1580 /* 1581 * Called just before setting an indirect block pointer to a 1582 * newly allocated indirect block. 1583 */ 1584 void 1585 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 1586 struct buf *nbp; /* newly allocated indirect block */ 1587 struct inode *ip; /* inode for file being extended */ 1588 struct buf *bp; /* indirect block referencing allocated block */ 1589 int ptrno; /* offset of pointer in indirect block */ 1590 ufs_daddr_t newblkno; /* disk block number being added */ 1591 { 1592 struct allocindir *aip; 1593 1594 aip = newallocindir(ip, ptrno, newblkno, 0); 1595 ACQUIRE_LOCK(&lk); 1596 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1597 FREE_LOCK(&lk); 1598 setup_allocindir_phase2(bp, ip, aip); 1599 } 1600 1601 /* 1602 * Called to finish the allocation of the "aip" allocated 1603 * by one of the two routines above. 1604 */ 1605 static void 1606 setup_allocindir_phase2(bp, ip, aip) 1607 struct buf *bp; /* in-memory copy of the indirect block */ 1608 struct inode *ip; /* inode for file being extended */ 1609 struct allocindir *aip; /* allocindir allocated by the above routines */ 1610 { 1611 struct worklist *wk; 1612 struct indirdep *indirdep, *newindirdep; 1613 struct bmsafemap *bmsafemap; 1614 struct allocindir *oldaip; 1615 struct freefrag *freefrag; 1616 struct newblk *newblk; 1617 1618 if (bp->b_lblkno >= 0) 1619 panic("setup_allocindir_phase2: not indir blk"); 1620 for (indirdep = NULL, newindirdep = NULL; ; ) { 1621 ACQUIRE_LOCK(&lk); 1622 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 1623 if (wk->wk_type != D_INDIRDEP) 1624 continue; 1625 indirdep = WK_INDIRDEP(wk); 1626 break; 1627 } 1628 if (indirdep == NULL && newindirdep) { 1629 indirdep = newindirdep; 1630 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 1631 newindirdep = NULL; 1632 } 1633 FREE_LOCK(&lk); 1634 if (indirdep) { 1635 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0, 1636 &newblk) == 0) 1637 panic("setup_allocindir: lost block"); 1638 ACQUIRE_LOCK(&lk); 1639 if (newblk->nb_state == DEPCOMPLETE) { 1640 aip->ai_state |= DEPCOMPLETE; 1641 aip->ai_buf = NULL; 1642 } else { 1643 bmsafemap = newblk->nb_bmsafemap; 1644 aip->ai_buf = bmsafemap->sm_buf; 1645 LIST_REMOVE(newblk, nb_deps); 1646 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd, 1647 aip, ai_deps); 1648 } 1649 LIST_REMOVE(newblk, nb_hash); 1650 FREE(newblk, M_NEWBLK); 1651 aip->ai_indirdep = indirdep; 1652 /* 1653 * Check to see if there is an existing dependency 1654 * for this block. If there is, merge the old 1655 * dependency into the new one. 1656 */ 1657 if (aip->ai_oldblkno == 0) 1658 oldaip = NULL; 1659 else 1660 1661 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) 1662 if (oldaip->ai_offset == aip->ai_offset) 1663 break; 1664 if (oldaip != NULL) { 1665 if (oldaip->ai_newblkno != aip->ai_oldblkno) { 1666 FREE_LOCK(&lk); 1667 panic("setup_allocindir_phase2: blkno"); 1668 } 1669 aip->ai_oldblkno = oldaip->ai_oldblkno; 1670 freefrag = oldaip->ai_freefrag; 1671 oldaip->ai_freefrag = aip->ai_freefrag; 1672 aip->ai_freefrag = freefrag; 1673 free_allocindir(oldaip, NULL); 1674 } 1675 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 1676 ((ufs_daddr_t *)indirdep->ir_savebp->b_data) 1677 [aip->ai_offset] = aip->ai_oldblkno; 1678 FREE_LOCK(&lk); 1679 } 1680 if (newindirdep) { 1681 /* 1682 * Avoid any possibility of data corruption by 1683 * ensuring that our old version is thrown away. 1684 */ 1685 newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 1686 brelse(newindirdep->ir_savebp); 1687 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP); 1688 } 1689 if (indirdep) 1690 break; 1691 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep), 1692 M_INDIRDEP, M_SOFTDEP_FLAGS); 1693 newindirdep->ir_list.wk_type = D_INDIRDEP; 1694 newindirdep->ir_state = ATTACHED; 1695 LIST_INIT(&newindirdep->ir_deplisthd); 1696 LIST_INIT(&newindirdep->ir_donehd); 1697 if (bp->b_blkno == bp->b_lblkno) { 1698 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, 1699 NULL, NULL); 1700 } 1701 newindirdep->ir_savebp = 1702 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0); 1703 BUF_KERNPROC(newindirdep->ir_savebp); 1704 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 1705 } 1706 } 1707 1708 /* 1709 * Block de-allocation dependencies. 1710 * 1711 * When blocks are de-allocated, the on-disk pointers must be nullified before 1712 * the blocks are made available for use by other files. (The true 1713 * requirement is that old pointers must be nullified before new on-disk 1714 * pointers are set. We chose this slightly more stringent requirement to 1715 * reduce complexity.) Our implementation handles this dependency by updating 1716 * the inode (or indirect block) appropriately but delaying the actual block 1717 * de-allocation (i.e., freemap and free space count manipulation) until 1718 * after the updated versions reach stable storage. After the disk is 1719 * updated, the blocks can be safely de-allocated whenever it is convenient. 1720 * This implementation handles only the common case of reducing a file's 1721 * length to zero. Other cases are handled by the conventional synchronous 1722 * write approach. 1723 * 1724 * The ffs implementation with which we worked double-checks 1725 * the state of the block pointers and file size as it reduces 1726 * a file's length. Some of this code is replicated here in our 1727 * soft updates implementation. The freeblks->fb_chkcnt field is 1728 * used to transfer a part of this information to the procedure 1729 * that eventually de-allocates the blocks. 1730 * 1731 * This routine should be called from the routine that shortens 1732 * a file's length, before the inode's size or block pointers 1733 * are modified. It will save the block pointer information for 1734 * later release and zero the inode so that the calling routine 1735 * can release it. 1736 */ 1737 void 1738 softdep_setup_freeblocks(ip, length) 1739 struct inode *ip; /* The inode whose length is to be reduced */ 1740 off_t length; /* The new length for the file */ 1741 { 1742 struct freeblks *freeblks; 1743 struct inodedep *inodedep; 1744 struct allocdirect *adp; 1745 struct vnode *vp; 1746 struct buf *bp; 1747 struct fs *fs; 1748 int i, error, delay; 1749 1750 fs = ip->i_fs; 1751 if (length != 0) 1752 panic("softde_setup_freeblocks: non-zero length"); 1753 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks), 1754 M_FREEBLKS, M_SOFTDEP_FLAGS); 1755 bzero(freeblks, sizeof(struct freeblks)); 1756 freeblks->fb_list.wk_type = D_FREEBLKS; 1757 freeblks->fb_uid = ip->i_uid; 1758 freeblks->fb_previousinum = ip->i_number; 1759 freeblks->fb_devvp = ip->i_devvp; 1760 freeblks->fb_fs = fs; 1761 freeblks->fb_oldsize = ip->i_size; 1762 freeblks->fb_newsize = length; 1763 freeblks->fb_chkcnt = ip->i_blocks; 1764 for (i = 0; i < NDADDR; i++) { 1765 freeblks->fb_dblks[i] = ip->i_db[i]; 1766 ip->i_db[i] = 0; 1767 } 1768 for (i = 0; i < NIADDR; i++) { 1769 freeblks->fb_iblks[i] = ip->i_ib[i]; 1770 ip->i_ib[i] = 0; 1771 } 1772 ip->i_blocks = 0; 1773 ip->i_size = 0; 1774 /* 1775 * Push the zero'ed inode to to its disk buffer so that we are free 1776 * to delete its dependencies below. Once the dependencies are gone 1777 * the buffer can be safely released. 1778 */ 1779 if ((error = bread(ip->i_devvp, 1780 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 1781 (int)fs->fs_bsize, &bp)) != 0) 1782 softdep_error("softdep_setup_freeblocks", error); 1783 *((struct dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) = 1784 ip->i_din; 1785 /* 1786 * Find and eliminate any inode dependencies. 1787 */ 1788 ACQUIRE_LOCK(&lk); 1789 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep); 1790 if ((inodedep->id_state & IOSTARTED) != 0) { 1791 FREE_LOCK(&lk); 1792 panic("softdep_setup_freeblocks: inode busy"); 1793 } 1794 /* 1795 * Add the freeblks structure to the list of operations that 1796 * must await the zero'ed inode being written to disk. If we 1797 * still have a bitmap dependency (delay == 0), then the inode 1798 * has never been written to disk, so we can process the 1799 * freeblks below once we have deleted the dependencies. 1800 */ 1801 delay = (inodedep->id_state & DEPCOMPLETE); 1802 if (delay) 1803 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list); 1804 /* 1805 * Because the file length has been truncated to zero, any 1806 * pending block allocation dependency structures associated 1807 * with this inode are obsolete and can simply be de-allocated. 1808 * We must first merge the two dependency lists to get rid of 1809 * any duplicate freefrag structures, then purge the merged list. 1810 */ 1811 merge_inode_lists(inodedep); 1812 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 1813 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 1814 FREE_LOCK(&lk); 1815 bdwrite(bp); 1816 /* 1817 * We must wait for any I/O in progress to finish so that 1818 * all potential buffers on the dirty list will be visible. 1819 * Once they are all there, walk the list and get rid of 1820 * any dependencies. 1821 */ 1822 vp = ITOV(ip); 1823 ACQUIRE_LOCK(&lk); 1824 drain_output(vp, 1); 1825 while (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT)) { 1826 bp = TAILQ_FIRST(&vp->v_dirtyblkhd); 1827 (void) inodedep_lookup(fs, ip->i_number, 0, &inodedep); 1828 deallocate_dependencies(bp, inodedep); 1829 bp->b_flags |= B_INVAL | B_NOCACHE; 1830 FREE_LOCK(&lk); 1831 brelse(bp); 1832 ACQUIRE_LOCK(&lk); 1833 } 1834 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0) 1835 (void)free_inodedep(inodedep); 1836 FREE_LOCK(&lk); 1837 /* 1838 * If the inode has never been written to disk (delay == 0), 1839 * then we can process the freeblks now that we have deleted 1840 * the dependencies. 1841 */ 1842 if (!delay) 1843 handle_workitem_freeblocks(freeblks); 1844 } 1845 1846 /* 1847 * Reclaim any dependency structures from a buffer that is about to 1848 * be reallocated to a new vnode. The buffer must be locked, thus, 1849 * no I/O completion operations can occur while we are manipulating 1850 * its associated dependencies. The mutex is held so that other I/O's 1851 * associated with related dependencies do not occur. 1852 */ 1853 static void 1854 deallocate_dependencies(bp, inodedep) 1855 struct buf *bp; 1856 struct inodedep *inodedep; 1857 { 1858 struct worklist *wk; 1859 struct indirdep *indirdep; 1860 struct allocindir *aip; 1861 struct pagedep *pagedep; 1862 struct dirrem *dirrem; 1863 struct diradd *dap; 1864 int i; 1865 1866 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 1867 switch (wk->wk_type) { 1868 1869 case D_INDIRDEP: 1870 indirdep = WK_INDIRDEP(wk); 1871 /* 1872 * None of the indirect pointers will ever be visible, 1873 * so they can simply be tossed. GOINGAWAY ensures 1874 * that allocated pointers will be saved in the buffer 1875 * cache until they are freed. Note that they will 1876 * only be able to be found by their physical address 1877 * since the inode mapping the logical address will 1878 * be gone. The save buffer used for the safe copy 1879 * was allocated in setup_allocindir_phase2 using 1880 * the physical address so it could be used for this 1881 * purpose. Hence we swap the safe copy with the real 1882 * copy, allowing the safe copy to be freed and holding 1883 * on to the real copy for later use in indir_trunc. 1884 */ 1885 if (indirdep->ir_state & GOINGAWAY) { 1886 FREE_LOCK(&lk); 1887 panic("deallocate_dependencies: already gone"); 1888 } 1889 indirdep->ir_state |= GOINGAWAY; 1890 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 1891 free_allocindir(aip, inodedep); 1892 if (bp->b_lblkno >= 0 || 1893 bp->b_blkno != indirdep->ir_savebp->b_lblkno) { 1894 FREE_LOCK(&lk); 1895 panic("deallocate_dependencies: not indir"); 1896 } 1897 bcopy(bp->b_data, indirdep->ir_savebp->b_data, 1898 bp->b_bcount); 1899 WORKLIST_REMOVE(wk); 1900 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk); 1901 continue; 1902 1903 case D_PAGEDEP: 1904 pagedep = WK_PAGEDEP(wk); 1905 /* 1906 * None of the directory additions will ever be 1907 * visible, so they can simply be tossed. 1908 */ 1909 for (i = 0; i < DAHASHSZ; i++) 1910 while ((dap = 1911 LIST_FIRST(&pagedep->pd_diraddhd[i]))) 1912 free_diradd(dap); 1913 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0) 1914 free_diradd(dap); 1915 /* 1916 * Copy any directory remove dependencies to the list 1917 * to be processed after the zero'ed inode is written. 1918 * If the inode has already been written, then they 1919 * can be dumped directly onto the work list. 1920 */ 1921 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 1922 LIST_REMOVE(dirrem, dm_next); 1923 dirrem->dm_dirinum = pagedep->pd_ino; 1924 if (inodedep == NULL || 1925 (inodedep->id_state & ALLCOMPLETE) == 1926 ALLCOMPLETE) 1927 add_to_worklist(&dirrem->dm_list); 1928 else 1929 WORKLIST_INSERT(&inodedep->id_bufwait, 1930 &dirrem->dm_list); 1931 } 1932 WORKLIST_REMOVE(&pagedep->pd_list); 1933 LIST_REMOVE(pagedep, pd_hash); 1934 WORKITEM_FREE(pagedep, D_PAGEDEP); 1935 continue; 1936 1937 case D_ALLOCINDIR: 1938 free_allocindir(WK_ALLOCINDIR(wk), inodedep); 1939 continue; 1940 1941 case D_ALLOCDIRECT: 1942 case D_INODEDEP: 1943 FREE_LOCK(&lk); 1944 panic("deallocate_dependencies: Unexpected type %s", 1945 TYPENAME(wk->wk_type)); 1946 /* NOTREACHED */ 1947 1948 default: 1949 FREE_LOCK(&lk); 1950 panic("deallocate_dependencies: Unknown type %s", 1951 TYPENAME(wk->wk_type)); 1952 /* NOTREACHED */ 1953 } 1954 } 1955 } 1956 1957 /* 1958 * Free an allocdirect. Generate a new freefrag work request if appropriate. 1959 * This routine must be called with splbio interrupts blocked. 1960 */ 1961 static void 1962 free_allocdirect(adphead, adp, delay) 1963 struct allocdirectlst *adphead; 1964 struct allocdirect *adp; 1965 int delay; 1966 { 1967 1968 #ifdef DEBUG 1969 if (lk.lkt_held == NOHOLDER) 1970 panic("free_allocdirect: lock not held"); 1971 #endif 1972 if ((adp->ad_state & DEPCOMPLETE) == 0) 1973 LIST_REMOVE(adp, ad_deps); 1974 TAILQ_REMOVE(adphead, adp, ad_next); 1975 if ((adp->ad_state & COMPLETE) == 0) 1976 WORKLIST_REMOVE(&adp->ad_list); 1977 if (adp->ad_freefrag != NULL) { 1978 if (delay) 1979 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 1980 &adp->ad_freefrag->ff_list); 1981 else 1982 add_to_worklist(&adp->ad_freefrag->ff_list); 1983 } 1984 WORKITEM_FREE(adp, D_ALLOCDIRECT); 1985 } 1986 1987 /* 1988 * Prepare an inode to be freed. The actual free operation is not 1989 * done until the zero'ed inode has been written to disk. 1990 */ 1991 void 1992 softdep_freefile(pvp, ino, mode) 1993 struct vnode *pvp; 1994 ino_t ino; 1995 int mode; 1996 { 1997 struct inode *ip = VTOI(pvp); 1998 struct inodedep *inodedep; 1999 struct freefile *freefile; 2000 2001 /* 2002 * This sets up the inode de-allocation dependency. 2003 */ 2004 MALLOC(freefile, struct freefile *, sizeof(struct freefile), 2005 M_FREEFILE, M_SOFTDEP_FLAGS); 2006 freefile->fx_list.wk_type = D_FREEFILE; 2007 freefile->fx_list.wk_state = 0; 2008 freefile->fx_mode = mode; 2009 freefile->fx_oldinum = ino; 2010 freefile->fx_devvp = ip->i_devvp; 2011 freefile->fx_fs = ip->i_fs; 2012 2013 /* 2014 * If the inodedep does not exist, then the zero'ed inode has 2015 * been written to disk. If the allocated inode has never been 2016 * written to disk, then the on-disk inode is zero'ed. In either 2017 * case we can free the file immediately. 2018 */ 2019 ACQUIRE_LOCK(&lk); 2020 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 || 2021 check_inode_unwritten(inodedep)) { 2022 FREE_LOCK(&lk); 2023 handle_workitem_freefile(freefile); 2024 return; 2025 } 2026 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 2027 FREE_LOCK(&lk); 2028 } 2029 2030 /* 2031 * Check to see if an inode has never been written to disk. If 2032 * so free the inodedep and return success, otherwise return failure. 2033 * This routine must be called with splbio interrupts blocked. 2034 * 2035 * If we still have a bitmap dependency, then the inode has never 2036 * been written to disk. Drop the dependency as it is no longer 2037 * necessary since the inode is being deallocated. We set the 2038 * ALLCOMPLETE flags since the bitmap now properly shows that the 2039 * inode is not allocated. Even if the inode is actively being 2040 * written, it has been rolled back to its zero'ed state, so we 2041 * are ensured that a zero inode is what is on the disk. For short 2042 * lived files, this change will usually result in removing all the 2043 * dependencies from the inode so that it can be freed immediately. 2044 */ 2045 static int 2046 check_inode_unwritten(inodedep) 2047 struct inodedep *inodedep; 2048 { 2049 2050 if ((inodedep->id_state & DEPCOMPLETE) != 0 || 2051 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2052 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2053 LIST_FIRST(&inodedep->id_inowait) != NULL || 2054 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2055 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2056 inodedep->id_nlinkdelta != 0) 2057 return (0); 2058 inodedep->id_state |= ALLCOMPLETE; 2059 LIST_REMOVE(inodedep, id_deps); 2060 inodedep->id_buf = NULL; 2061 if (inodedep->id_state & ONWORKLIST) 2062 WORKLIST_REMOVE(&inodedep->id_list); 2063 if (inodedep->id_savedino != NULL) { 2064 FREE(inodedep->id_savedino, M_INODEDEP); 2065 inodedep->id_savedino = NULL; 2066 } 2067 if (free_inodedep(inodedep) == 0) { 2068 FREE_LOCK(&lk); 2069 panic("check_inode_unwritten: busy inode"); 2070 } 2071 return (1); 2072 } 2073 2074 /* 2075 * Try to free an inodedep structure. Return 1 if it could be freed. 2076 */ 2077 static int 2078 free_inodedep(inodedep) 2079 struct inodedep *inodedep; 2080 { 2081 2082 if ((inodedep->id_state & ONWORKLIST) != 0 || 2083 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 2084 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2085 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2086 LIST_FIRST(&inodedep->id_inowait) != NULL || 2087 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2088 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2089 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL) 2090 return (0); 2091 LIST_REMOVE(inodedep, id_hash); 2092 WORKITEM_FREE(inodedep, D_INODEDEP); 2093 num_inodedep -= 1; 2094 return (1); 2095 } 2096 2097 /* 2098 * This workitem routine performs the block de-allocation. 2099 * The workitem is added to the pending list after the updated 2100 * inode block has been written to disk. As mentioned above, 2101 * checks regarding the number of blocks de-allocated (compared 2102 * to the number of blocks allocated for the file) are also 2103 * performed in this function. 2104 */ 2105 static void 2106 handle_workitem_freeblocks(freeblks) 2107 struct freeblks *freeblks; 2108 { 2109 struct inode tip; 2110 ufs_daddr_t bn; 2111 struct fs *fs; 2112 int i, level, bsize; 2113 long nblocks, blocksreleased = 0; 2114 int error, allerror = 0; 2115 ufs_lbn_t baselbns[NIADDR], tmpval; 2116 2117 tip.i_number = freeblks->fb_previousinum; 2118 tip.i_devvp = freeblks->fb_devvp; 2119 tip.i_dev = freeblks->fb_devvp->v_rdev; 2120 tip.i_fs = freeblks->fb_fs; 2121 tip.i_size = freeblks->fb_oldsize; 2122 tip.i_uid = freeblks->fb_uid; 2123 fs = freeblks->fb_fs; 2124 tmpval = 1; 2125 baselbns[0] = NDADDR; 2126 for (i = 1; i < NIADDR; i++) { 2127 tmpval *= NINDIR(fs); 2128 baselbns[i] = baselbns[i - 1] + tmpval; 2129 } 2130 nblocks = btodb(fs->fs_bsize); 2131 blocksreleased = 0; 2132 /* 2133 * Indirect blocks first. 2134 */ 2135 for (level = (NIADDR - 1); level >= 0; level--) { 2136 if ((bn = freeblks->fb_iblks[level]) == 0) 2137 continue; 2138 if ((error = indir_trunc(&tip, fsbtodb(fs, bn), level, 2139 baselbns[level], &blocksreleased)) == 0) 2140 allerror = error; 2141 ffs_blkfree(&tip, bn, fs->fs_bsize); 2142 blocksreleased += nblocks; 2143 } 2144 /* 2145 * All direct blocks or frags. 2146 */ 2147 for (i = (NDADDR - 1); i >= 0; i--) { 2148 if ((bn = freeblks->fb_dblks[i]) == 0) 2149 continue; 2150 bsize = blksize(fs, &tip, i); 2151 ffs_blkfree(&tip, bn, bsize); 2152 blocksreleased += btodb(bsize); 2153 } 2154 2155 #ifdef DIAGNOSTIC 2156 if (freeblks->fb_chkcnt != blocksreleased) 2157 printf("handle_workitem_freeblocks: block count\n"); 2158 if (allerror) 2159 softdep_error("handle_workitem_freeblks", allerror); 2160 #endif /* DIAGNOSTIC */ 2161 WORKITEM_FREE(freeblks, D_FREEBLKS); 2162 } 2163 2164 /* 2165 * Release blocks associated with the inode ip and stored in the indirect 2166 * block dbn. If level is greater than SINGLE, the block is an indirect block 2167 * and recursive calls to indirtrunc must be used to cleanse other indirect 2168 * blocks. 2169 */ 2170 static int 2171 indir_trunc(ip, dbn, level, lbn, countp) 2172 struct inode *ip; 2173 ufs_daddr_t dbn; 2174 int level; 2175 ufs_lbn_t lbn; 2176 long *countp; 2177 { 2178 struct buf *bp; 2179 ufs_daddr_t *bap; 2180 ufs_daddr_t nb; 2181 struct fs *fs; 2182 struct worklist *wk; 2183 struct indirdep *indirdep; 2184 int i, lbnadd, nblocks; 2185 int error, allerror = 0; 2186 2187 fs = ip->i_fs; 2188 lbnadd = 1; 2189 for (i = level; i > 0; i--) 2190 lbnadd *= NINDIR(fs); 2191 /* 2192 * Get buffer of block pointers to be freed. This routine is not 2193 * called until the zero'ed inode has been written, so it is safe 2194 * to free blocks as they are encountered. Because the inode has 2195 * been zero'ed, calls to bmap on these blocks will fail. So, we 2196 * have to use the on-disk address and the block device for the 2197 * filesystem to look them up. If the file was deleted before its 2198 * indirect blocks were all written to disk, the routine that set 2199 * us up (deallocate_dependencies) will have arranged to leave 2200 * a complete copy of the indirect block in memory for our use. 2201 * Otherwise we have to read the blocks in from the disk. 2202 */ 2203 ACQUIRE_LOCK(&lk); 2204 if ((bp = incore(ip->i_devvp, dbn)) != NULL && 2205 (wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2206 if (wk->wk_type != D_INDIRDEP || 2207 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp || 2208 (indirdep->ir_state & GOINGAWAY) == 0) { 2209 FREE_LOCK(&lk); 2210 panic("indir_trunc: lost indirdep"); 2211 } 2212 WORKLIST_REMOVE(wk); 2213 WORKITEM_FREE(indirdep, D_INDIRDEP); 2214 if (LIST_FIRST(&bp->b_dep) != NULL) { 2215 FREE_LOCK(&lk); 2216 panic("indir_trunc: dangling dep"); 2217 } 2218 FREE_LOCK(&lk); 2219 } else { 2220 FREE_LOCK(&lk); 2221 error = bread(ip->i_devvp, dbn, (int)fs->fs_bsize, &bp); 2222 if (error) 2223 return (error); 2224 } 2225 /* 2226 * Recursively free indirect blocks. 2227 */ 2228 bap = (ufs_daddr_t *)bp->b_data; 2229 nblocks = btodb(fs->fs_bsize); 2230 for (i = NINDIR(fs) - 1; i >= 0; i--) { 2231 if ((nb = bap[i]) == 0) 2232 continue; 2233 if (level != 0) { 2234 if ((error = indir_trunc(ip, fsbtodb(fs, nb), 2235 level - 1, lbn + (i * lbnadd), countp)) != 0) 2236 allerror = error; 2237 } 2238 ffs_blkfree(ip, nb, fs->fs_bsize); 2239 *countp += nblocks; 2240 } 2241 bp->b_flags |= B_INVAL | B_NOCACHE; 2242 brelse(bp); 2243 return (allerror); 2244 } 2245 2246 /* 2247 * Free an allocindir. 2248 * This routine must be called with splbio interrupts blocked. 2249 */ 2250 static void 2251 free_allocindir(aip, inodedep) 2252 struct allocindir *aip; 2253 struct inodedep *inodedep; 2254 { 2255 struct freefrag *freefrag; 2256 2257 #ifdef DEBUG 2258 if (lk.lkt_held == NOHOLDER) 2259 panic("free_allocindir: lock not held"); 2260 #endif 2261 if ((aip->ai_state & DEPCOMPLETE) == 0) 2262 LIST_REMOVE(aip, ai_deps); 2263 if (aip->ai_state & ONWORKLIST) 2264 WORKLIST_REMOVE(&aip->ai_list); 2265 LIST_REMOVE(aip, ai_next); 2266 if ((freefrag = aip->ai_freefrag) != NULL) { 2267 if (inodedep == NULL) 2268 add_to_worklist(&freefrag->ff_list); 2269 else 2270 WORKLIST_INSERT(&inodedep->id_bufwait, 2271 &freefrag->ff_list); 2272 } 2273 WORKITEM_FREE(aip, D_ALLOCINDIR); 2274 } 2275 2276 /* 2277 * Directory entry addition dependencies. 2278 * 2279 * When adding a new directory entry, the inode (with its incremented link 2280 * count) must be written to disk before the directory entry's pointer to it. 2281 * Also, if the inode is newly allocated, the corresponding freemap must be 2282 * updated (on disk) before the directory entry's pointer. These requirements 2283 * are met via undo/redo on the directory entry's pointer, which consists 2284 * simply of the inode number. 2285 * 2286 * As directory entries are added and deleted, the free space within a 2287 * directory block can become fragmented. The ufs filesystem will compact 2288 * a fragmented directory block to make space for a new entry. When this 2289 * occurs, the offsets of previously added entries change. Any "diradd" 2290 * dependency structures corresponding to these entries must be updated with 2291 * the new offsets. 2292 */ 2293 2294 /* 2295 * This routine is called after the in-memory inode's link 2296 * count has been incremented, but before the directory entry's 2297 * pointer to the inode has been set. 2298 */ 2299 void 2300 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp) 2301 struct buf *bp; /* buffer containing directory block */ 2302 struct inode *dp; /* inode for directory */ 2303 off_t diroffset; /* offset of new entry in directory */ 2304 ino_t newinum; /* inode referenced by new directory entry */ 2305 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 2306 { 2307 int offset; /* offset of new entry within directory block */ 2308 ufs_lbn_t lbn; /* block in directory containing new entry */ 2309 struct fs *fs; 2310 struct diradd *dap; 2311 struct pagedep *pagedep; 2312 struct inodedep *inodedep; 2313 struct mkdir *mkdir1, *mkdir2; 2314 2315 /* 2316 * Whiteouts have no dependencies. 2317 */ 2318 if (newinum == WINO) { 2319 if (newdirbp != NULL) 2320 bdwrite(newdirbp); 2321 return; 2322 } 2323 2324 fs = dp->i_fs; 2325 lbn = lblkno(fs, diroffset); 2326 offset = blkoff(fs, diroffset); 2327 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, 2328 M_SOFTDEP_FLAGS); 2329 bzero(dap, sizeof(struct diradd)); 2330 dap->da_list.wk_type = D_DIRADD; 2331 dap->da_offset = offset; 2332 dap->da_newinum = newinum; 2333 dap->da_state = ATTACHED; 2334 if (newdirbp == NULL) { 2335 dap->da_state |= DEPCOMPLETE; 2336 ACQUIRE_LOCK(&lk); 2337 } else { 2338 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 2339 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2340 M_SOFTDEP_FLAGS); 2341 mkdir1->md_list.wk_type = D_MKDIR; 2342 mkdir1->md_state = MKDIR_BODY; 2343 mkdir1->md_diradd = dap; 2344 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2345 M_SOFTDEP_FLAGS); 2346 mkdir2->md_list.wk_type = D_MKDIR; 2347 mkdir2->md_state = MKDIR_PARENT; 2348 mkdir2->md_diradd = dap; 2349 /* 2350 * Dependency on "." and ".." being written to disk. 2351 */ 2352 mkdir1->md_buf = newdirbp; 2353 ACQUIRE_LOCK(&lk); 2354 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 2355 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list); 2356 FREE_LOCK(&lk); 2357 bdwrite(newdirbp); 2358 /* 2359 * Dependency on link count increase for parent directory 2360 */ 2361 ACQUIRE_LOCK(&lk); 2362 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0 2363 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2364 dap->da_state &= ~MKDIR_PARENT; 2365 WORKITEM_FREE(mkdir2, D_MKDIR); 2366 } else { 2367 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 2368 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 2369 } 2370 } 2371 /* 2372 * Link into parent directory pagedep to await its being written. 2373 */ 2374 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2375 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2376 dap->da_pagedep = pagedep; 2377 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 2378 da_pdlist); 2379 /* 2380 * Link into its inodedep. Put it on the id_bufwait list if the inode 2381 * is not yet written. If it is written, do the post-inode write 2382 * processing to put it on the id_pendinghd list. 2383 */ 2384 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep); 2385 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 2386 diradd_inode_written(dap, inodedep); 2387 else 2388 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2389 FREE_LOCK(&lk); 2390 } 2391 2392 /* 2393 * This procedure is called to change the offset of a directory 2394 * entry when compacting a directory block which must be owned 2395 * exclusively by the caller. Note that the actual entry movement 2396 * must be done in this procedure to ensure that no I/O completions 2397 * occur while the move is in progress. 2398 */ 2399 void 2400 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize) 2401 struct inode *dp; /* inode for directory */ 2402 caddr_t base; /* address of dp->i_offset */ 2403 caddr_t oldloc; /* address of old directory location */ 2404 caddr_t newloc; /* address of new directory location */ 2405 int entrysize; /* size of directory entry */ 2406 { 2407 int offset, oldoffset, newoffset; 2408 struct pagedep *pagedep; 2409 struct diradd *dap; 2410 ufs_lbn_t lbn; 2411 2412 ACQUIRE_LOCK(&lk); 2413 lbn = lblkno(dp->i_fs, dp->i_offset); 2414 offset = blkoff(dp->i_fs, dp->i_offset); 2415 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) 2416 goto done; 2417 oldoffset = offset + (oldloc - base); 2418 newoffset = offset + (newloc - base); 2419 2420 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) { 2421 if (dap->da_offset != oldoffset) 2422 continue; 2423 dap->da_offset = newoffset; 2424 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) 2425 break; 2426 LIST_REMOVE(dap, da_pdlist); 2427 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], 2428 dap, da_pdlist); 2429 break; 2430 } 2431 if (dap == NULL) { 2432 2433 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) { 2434 if (dap->da_offset == oldoffset) { 2435 dap->da_offset = newoffset; 2436 break; 2437 } 2438 } 2439 } 2440 done: 2441 bcopy(oldloc, newloc, entrysize); 2442 FREE_LOCK(&lk); 2443 } 2444 2445 /* 2446 * Free a diradd dependency structure. This routine must be called 2447 * with splbio interrupts blocked. 2448 */ 2449 static void 2450 free_diradd(dap) 2451 struct diradd *dap; 2452 { 2453 struct dirrem *dirrem; 2454 struct pagedep *pagedep; 2455 struct inodedep *inodedep; 2456 struct mkdir *mkdir, *nextmd; 2457 2458 #ifdef DEBUG 2459 if (lk.lkt_held == NOHOLDER) 2460 panic("free_diradd: lock not held"); 2461 #endif 2462 WORKLIST_REMOVE(&dap->da_list); 2463 LIST_REMOVE(dap, da_pdlist); 2464 if ((dap->da_state & DIRCHG) == 0) { 2465 pagedep = dap->da_pagedep; 2466 } else { 2467 dirrem = dap->da_previous; 2468 pagedep = dirrem->dm_pagedep; 2469 dirrem->dm_dirinum = pagedep->pd_ino; 2470 add_to_worklist(&dirrem->dm_list); 2471 } 2472 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum, 2473 0, &inodedep) != 0) 2474 (void) free_inodedep(inodedep); 2475 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2476 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 2477 nextmd = LIST_NEXT(mkdir, md_mkdirs); 2478 if (mkdir->md_diradd != dap) 2479 continue; 2480 dap->da_state &= ~mkdir->md_state; 2481 WORKLIST_REMOVE(&mkdir->md_list); 2482 LIST_REMOVE(mkdir, md_mkdirs); 2483 WORKITEM_FREE(mkdir, D_MKDIR); 2484 } 2485 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2486 FREE_LOCK(&lk); 2487 panic("free_diradd: unfound ref"); 2488 } 2489 } 2490 WORKITEM_FREE(dap, D_DIRADD); 2491 } 2492 2493 /* 2494 * Directory entry removal dependencies. 2495 * 2496 * When removing a directory entry, the entry's inode pointer must be 2497 * zero'ed on disk before the corresponding inode's link count is decremented 2498 * (possibly freeing the inode for re-use). This dependency is handled by 2499 * updating the directory entry but delaying the inode count reduction until 2500 * after the directory block has been written to disk. After this point, the 2501 * inode count can be decremented whenever it is convenient. 2502 */ 2503 2504 /* 2505 * This routine should be called immediately after removing 2506 * a directory entry. The inode's link count should not be 2507 * decremented by the calling procedure -- the soft updates 2508 * code will do this task when it is safe. 2509 */ 2510 void 2511 softdep_setup_remove(bp, dp, ip, isrmdir) 2512 struct buf *bp; /* buffer containing directory block */ 2513 struct inode *dp; /* inode for the directory being modified */ 2514 struct inode *ip; /* inode for directory entry being removed */ 2515 int isrmdir; /* indicates if doing RMDIR */ 2516 { 2517 struct dirrem *dirrem, *prevdirrem; 2518 2519 /* 2520 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 2521 */ 2522 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2523 2524 /* 2525 * If the COMPLETE flag is clear, then there were no active 2526 * entries and we want to roll back to a zeroed entry until 2527 * the new inode is committed to disk. If the COMPLETE flag is 2528 * set then we have deleted an entry that never made it to 2529 * disk. If the entry we deleted resulted from a name change, 2530 * then the old name still resides on disk. We cannot delete 2531 * its inode (returned to us in prevdirrem) until the zeroed 2532 * directory entry gets to disk. The new inode has never been 2533 * referenced on the disk, so can be deleted immediately. 2534 */ 2535 if ((dirrem->dm_state & COMPLETE) == 0) { 2536 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 2537 dm_next); 2538 FREE_LOCK(&lk); 2539 } else { 2540 if (prevdirrem != NULL) 2541 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 2542 prevdirrem, dm_next); 2543 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 2544 FREE_LOCK(&lk); 2545 handle_workitem_remove(dirrem); 2546 } 2547 } 2548 2549 /* 2550 * Allocate a new dirrem if appropriate and return it along with 2551 * its associated pagedep. Called without a lock, returns with lock. 2552 */ 2553 static long num_dirrem; /* number of dirrem allocated */ 2554 static struct dirrem * 2555 newdirrem(bp, dp, ip, isrmdir, prevdirremp) 2556 struct buf *bp; /* buffer containing directory block */ 2557 struct inode *dp; /* inode for the directory being modified */ 2558 struct inode *ip; /* inode for directory entry being removed */ 2559 int isrmdir; /* indicates if doing RMDIR */ 2560 struct dirrem **prevdirremp; /* previously referenced inode, if any */ 2561 { 2562 int offset; 2563 ufs_lbn_t lbn; 2564 struct diradd *dap; 2565 struct dirrem *dirrem; 2566 struct pagedep *pagedep; 2567 2568 /* 2569 * Whiteouts have no deletion dependencies. 2570 */ 2571 if (ip == NULL) 2572 panic("newdirrem: whiteout"); 2573 /* 2574 * If we are over our limit, try to improve the situation. 2575 * Limiting the number of dirrem structures will also limit 2576 * the number of freefile and freeblks structures. 2577 */ 2578 if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0) 2579 (void) request_cleanup(FLUSH_REMOVE, 0); 2580 num_dirrem += 1; 2581 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem), 2582 M_DIRREM, M_SOFTDEP_FLAGS); 2583 bzero(dirrem, sizeof(struct dirrem)); 2584 dirrem->dm_list.wk_type = D_DIRREM; 2585 dirrem->dm_state = isrmdir ? RMDIR : 0; 2586 dirrem->dm_mnt = ITOV(ip)->v_mount; 2587 dirrem->dm_oldinum = ip->i_number; 2588 *prevdirremp = NULL; 2589 2590 ACQUIRE_LOCK(&lk); 2591 lbn = lblkno(dp->i_fs, dp->i_offset); 2592 offset = blkoff(dp->i_fs, dp->i_offset); 2593 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2594 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2595 dirrem->dm_pagedep = pagedep; 2596 /* 2597 * Check for a diradd dependency for the same directory entry. 2598 * If present, then both dependencies become obsolete and can 2599 * be de-allocated. Check for an entry on both the pd_dirraddhd 2600 * list and the pd_pendinghd list. 2601 */ 2602 2603 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 2604 if (dap->da_offset == offset) 2605 break; 2606 if (dap == NULL) { 2607 2608 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 2609 if (dap->da_offset == offset) 2610 break; 2611 if (dap == NULL) 2612 return (dirrem); 2613 } 2614 /* 2615 * Must be ATTACHED at this point. 2616 */ 2617 if ((dap->da_state & ATTACHED) == 0) { 2618 FREE_LOCK(&lk); 2619 panic("newdirrem: not ATTACHED"); 2620 } 2621 if (dap->da_newinum != ip->i_number) { 2622 FREE_LOCK(&lk); 2623 panic("newdirrem: inum %d should be %d", 2624 ip->i_number, dap->da_newinum); 2625 } 2626 /* 2627 * If we are deleting a changed name that never made it to disk, 2628 * then return the dirrem describing the previous inode (which 2629 * represents the inode currently referenced from this entry on disk). 2630 */ 2631 if ((dap->da_state & DIRCHG) != 0) { 2632 *prevdirremp = dap->da_previous; 2633 dap->da_state &= ~DIRCHG; 2634 dap->da_pagedep = pagedep; 2635 } 2636 /* 2637 * We are deleting an entry that never made it to disk. 2638 * Mark it COMPLETE so we can delete its inode immediately. 2639 */ 2640 dirrem->dm_state |= COMPLETE; 2641 free_diradd(dap); 2642 return (dirrem); 2643 } 2644 2645 /* 2646 * Directory entry change dependencies. 2647 * 2648 * Changing an existing directory entry requires that an add operation 2649 * be completed first followed by a deletion. The semantics for the addition 2650 * are identical to the description of adding a new entry above except 2651 * that the rollback is to the old inode number rather than zero. Once 2652 * the addition dependency is completed, the removal is done as described 2653 * in the removal routine above. 2654 */ 2655 2656 /* 2657 * This routine should be called immediately after changing 2658 * a directory entry. The inode's link count should not be 2659 * decremented by the calling procedure -- the soft updates 2660 * code will perform this task when it is safe. 2661 */ 2662 void 2663 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 2664 struct buf *bp; /* buffer containing directory block */ 2665 struct inode *dp; /* inode for the directory being modified */ 2666 struct inode *ip; /* inode for directory entry being removed */ 2667 ino_t newinum; /* new inode number for changed entry */ 2668 int isrmdir; /* indicates if doing RMDIR */ 2669 { 2670 int offset; 2671 struct diradd *dap = NULL; 2672 struct dirrem *dirrem, *prevdirrem; 2673 struct pagedep *pagedep; 2674 struct inodedep *inodedep; 2675 2676 offset = blkoff(dp->i_fs, dp->i_offset); 2677 2678 /* 2679 * Whiteouts do not need diradd dependencies. 2680 */ 2681 if (newinum != WINO) { 2682 MALLOC(dap, struct diradd *, sizeof(struct diradd), 2683 M_DIRADD, M_SOFTDEP_FLAGS); 2684 bzero(dap, sizeof(struct diradd)); 2685 dap->da_list.wk_type = D_DIRADD; 2686 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 2687 dap->da_offset = offset; 2688 dap->da_newinum = newinum; 2689 } 2690 2691 /* 2692 * Allocate a new dirrem and ACQUIRE_LOCK. 2693 */ 2694 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2695 pagedep = dirrem->dm_pagedep; 2696 /* 2697 * The possible values for isrmdir: 2698 * 0 - non-directory file rename 2699 * 1 - directory rename within same directory 2700 * inum - directory rename to new directory of given inode number 2701 * When renaming to a new directory, we are both deleting and 2702 * creating a new directory entry, so the link count on the new 2703 * directory should not change. Thus we do not need the followup 2704 * dirrem which is usually done in handle_workitem_remove. We set 2705 * the DIRCHG flag to tell handle_workitem_remove to skip the 2706 * followup dirrem. 2707 */ 2708 if (isrmdir > 1) 2709 dirrem->dm_state |= DIRCHG; 2710 2711 /* 2712 * Whiteouts have no additional dependencies, 2713 * so just put the dirrem on the correct list. 2714 */ 2715 if (newinum == WINO) { 2716 if ((dirrem->dm_state & COMPLETE) == 0) { 2717 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 2718 dm_next); 2719 } else { 2720 dirrem->dm_dirinum = pagedep->pd_ino; 2721 add_to_worklist(&dirrem->dm_list); 2722 } 2723 FREE_LOCK(&lk); 2724 return; 2725 } 2726 2727 /* 2728 * If the COMPLETE flag is clear, then there were no active 2729 * entries and we want to roll back to the previous inode until 2730 * the new inode is committed to disk. If the COMPLETE flag is 2731 * set, then we have deleted an entry that never made it to disk. 2732 * If the entry we deleted resulted from a name change, then the old 2733 * inode reference still resides on disk. Any rollback that we do 2734 * needs to be to that old inode (returned to us in prevdirrem). If 2735 * the entry we deleted resulted from a create, then there is 2736 * no entry on the disk, so we want to roll back to zero rather 2737 * than the uncommitted inode. In either of the COMPLETE cases we 2738 * want to immediately free the unwritten and unreferenced inode. 2739 */ 2740 if ((dirrem->dm_state & COMPLETE) == 0) { 2741 dap->da_previous = dirrem; 2742 } else { 2743 if (prevdirrem != NULL) { 2744 dap->da_previous = prevdirrem; 2745 } else { 2746 dap->da_state &= ~DIRCHG; 2747 dap->da_pagedep = pagedep; 2748 } 2749 dirrem->dm_dirinum = pagedep->pd_ino; 2750 add_to_worklist(&dirrem->dm_list); 2751 } 2752 /* 2753 * Link into its inodedep. Put it on the id_bufwait list if the inode 2754 * is not yet written. If it is written, do the post-inode write 2755 * processing to put it on the id_pendinghd list. 2756 */ 2757 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 || 2758 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2759 dap->da_state |= COMPLETE; 2760 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 2761 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 2762 } else { 2763 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 2764 dap, da_pdlist); 2765 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2766 } 2767 FREE_LOCK(&lk); 2768 } 2769 2770 /* 2771 * Called whenever the link count on an inode is changed. 2772 * It creates an inode dependency so that the new reference(s) 2773 * to the inode cannot be committed to disk until the updated 2774 * inode has been written. 2775 */ 2776 void 2777 softdep_change_linkcnt(ip) 2778 struct inode *ip; /* the inode with the increased link count */ 2779 { 2780 struct inodedep *inodedep; 2781 2782 ACQUIRE_LOCK(&lk); 2783 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 2784 if (ip->i_nlink < ip->i_effnlink) { 2785 FREE_LOCK(&lk); 2786 panic("softdep_change_linkcnt: bad delta"); 2787 } 2788 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2789 FREE_LOCK(&lk); 2790 } 2791 2792 /* 2793 * This workitem decrements the inode's link count. 2794 * If the link count reaches zero, the file is removed. 2795 */ 2796 static void 2797 handle_workitem_remove(dirrem) 2798 struct dirrem *dirrem; 2799 { 2800 struct thread *td = curthread; /* XXX */ 2801 struct inodedep *inodedep; 2802 struct vnode *vp; 2803 struct inode *ip; 2804 ino_t oldinum; 2805 int error; 2806 2807 if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) { 2808 softdep_error("handle_workitem_remove: vget", error); 2809 return; 2810 } 2811 ip = VTOI(vp); 2812 ACQUIRE_LOCK(&lk); 2813 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){ 2814 FREE_LOCK(&lk); 2815 panic("handle_workitem_remove: lost inodedep"); 2816 } 2817 /* 2818 * Normal file deletion. 2819 */ 2820 if ((dirrem->dm_state & RMDIR) == 0) { 2821 ip->i_nlink--; 2822 ip->i_flag |= IN_CHANGE; 2823 if (ip->i_nlink < ip->i_effnlink) { 2824 FREE_LOCK(&lk); 2825 panic("handle_workitem_remove: bad file delta"); 2826 } 2827 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2828 FREE_LOCK(&lk); 2829 vput(vp); 2830 num_dirrem -= 1; 2831 WORKITEM_FREE(dirrem, D_DIRREM); 2832 return; 2833 } 2834 /* 2835 * Directory deletion. Decrement reference count for both the 2836 * just deleted parent directory entry and the reference for ".". 2837 * Next truncate the directory to length zero. When the 2838 * truncation completes, arrange to have the reference count on 2839 * the parent decremented to account for the loss of "..". 2840 */ 2841 ip->i_nlink -= 2; 2842 ip->i_flag |= IN_CHANGE; 2843 if (ip->i_nlink < ip->i_effnlink) { 2844 FREE_LOCK(&lk); 2845 panic("handle_workitem_remove: bad dir delta"); 2846 } 2847 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2848 FREE_LOCK(&lk); 2849 if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, proc0.p_ucred, td)) != 0) 2850 softdep_error("handle_workitem_remove: truncate", error); 2851 /* 2852 * Rename a directory to a new parent. Since, we are both deleting 2853 * and creating a new directory entry, the link count on the new 2854 * directory should not change. Thus we skip the followup dirrem. 2855 */ 2856 if (dirrem->dm_state & DIRCHG) { 2857 vput(vp); 2858 num_dirrem -= 1; 2859 WORKITEM_FREE(dirrem, D_DIRREM); 2860 return; 2861 } 2862 /* 2863 * If the inodedep does not exist, then the zero'ed inode has 2864 * been written to disk. If the allocated inode has never been 2865 * written to disk, then the on-disk inode is zero'ed. In either 2866 * case we can remove the file immediately. 2867 */ 2868 ACQUIRE_LOCK(&lk); 2869 dirrem->dm_state = 0; 2870 oldinum = dirrem->dm_oldinum; 2871 dirrem->dm_oldinum = dirrem->dm_dirinum; 2872 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 || 2873 check_inode_unwritten(inodedep)) { 2874 FREE_LOCK(&lk); 2875 vput(vp); 2876 handle_workitem_remove(dirrem); 2877 return; 2878 } 2879 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 2880 FREE_LOCK(&lk); 2881 vput(vp); 2882 } 2883 2884 /* 2885 * Inode de-allocation dependencies. 2886 * 2887 * When an inode's link count is reduced to zero, it can be de-allocated. We 2888 * found it convenient to postpone de-allocation until after the inode is 2889 * written to disk with its new link count (zero). At this point, all of the 2890 * on-disk inode's block pointers are nullified and, with careful dependency 2891 * list ordering, all dependencies related to the inode will be satisfied and 2892 * the corresponding dependency structures de-allocated. So, if/when the 2893 * inode is reused, there will be no mixing of old dependencies with new 2894 * ones. This artificial dependency is set up by the block de-allocation 2895 * procedure above (softdep_setup_freeblocks) and completed by the 2896 * following procedure. 2897 */ 2898 static void 2899 handle_workitem_freefile(freefile) 2900 struct freefile *freefile; 2901 { 2902 struct vnode vp; 2903 struct inode tip; 2904 struct inodedep *idp; 2905 int error; 2906 2907 #ifdef DEBUG 2908 ACQUIRE_LOCK(&lk); 2909 error = inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp); 2910 FREE_LOCK(&lk); 2911 if (error) 2912 panic("handle_workitem_freefile: inodedep survived"); 2913 #endif 2914 tip.i_devvp = freefile->fx_devvp; 2915 tip.i_dev = freefile->fx_devvp->v_rdev; 2916 tip.i_fs = freefile->fx_fs; 2917 vp.v_data = &tip; 2918 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0) 2919 softdep_error("handle_workitem_freefile", error); 2920 WORKITEM_FREE(freefile, D_FREEFILE); 2921 } 2922 2923 /* 2924 * Disk writes. 2925 * 2926 * The dependency structures constructed above are most actively used when file 2927 * system blocks are written to disk. No constraints are placed on when a 2928 * block can be written, but unsatisfied update dependencies are made safe by 2929 * modifying (or replacing) the source memory for the duration of the disk 2930 * write. When the disk write completes, the memory block is again brought 2931 * up-to-date. 2932 * 2933 * In-core inode structure reclamation. 2934 * 2935 * Because there are a finite number of "in-core" inode structures, they are 2936 * reused regularly. By transferring all inode-related dependencies to the 2937 * in-memory inode block and indexing them separately (via "inodedep"s), we 2938 * can allow "in-core" inode structures to be reused at any time and avoid 2939 * any increase in contention. 2940 * 2941 * Called just before entering the device driver to initiate a new disk I/O. 2942 * The buffer must be locked, thus, no I/O completion operations can occur 2943 * while we are manipulating its associated dependencies. 2944 */ 2945 static void 2946 softdep_disk_io_initiation(bp) 2947 struct buf *bp; /* structure describing disk write to occur */ 2948 { 2949 struct worklist *wk, *nextwk; 2950 struct indirdep *indirdep; 2951 2952 /* 2953 * We only care about write operations. There should never 2954 * be dependencies for reads. 2955 */ 2956 if (bp->b_flags & B_READ) 2957 panic("softdep_disk_io_initiation: read"); 2958 /* 2959 * Do any necessary pre-I/O processing. 2960 */ 2961 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = nextwk) { 2962 nextwk = LIST_NEXT(wk, wk_list); 2963 switch (wk->wk_type) { 2964 2965 case D_PAGEDEP: 2966 initiate_write_filepage(WK_PAGEDEP(wk), bp); 2967 continue; 2968 2969 case D_INODEDEP: 2970 initiate_write_inodeblock(WK_INODEDEP(wk), bp); 2971 continue; 2972 2973 case D_INDIRDEP: 2974 indirdep = WK_INDIRDEP(wk); 2975 if (indirdep->ir_state & GOINGAWAY) 2976 panic("disk_io_initiation: indirdep gone"); 2977 /* 2978 * If there are no remaining dependencies, this 2979 * will be writing the real pointers, so the 2980 * dependency can be freed. 2981 */ 2982 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) { 2983 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 2984 brelse(indirdep->ir_savebp); 2985 /* inline expand WORKLIST_REMOVE(wk); */ 2986 wk->wk_state &= ~ONWORKLIST; 2987 LIST_REMOVE(wk, wk_list); 2988 WORKITEM_FREE(indirdep, D_INDIRDEP); 2989 continue; 2990 } 2991 /* 2992 * Replace up-to-date version with safe version. 2993 */ 2994 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount, 2995 M_INDIRDEP, M_SOFTDEP_FLAGS); 2996 ACQUIRE_LOCK(&lk); 2997 indirdep->ir_state &= ~ATTACHED; 2998 indirdep->ir_state |= UNDONE; 2999 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 3000 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 3001 bp->b_bcount); 3002 FREE_LOCK(&lk); 3003 continue; 3004 3005 case D_MKDIR: 3006 case D_BMSAFEMAP: 3007 case D_ALLOCDIRECT: 3008 case D_ALLOCINDIR: 3009 continue; 3010 3011 default: 3012 panic("handle_disk_io_initiation: Unexpected type %s", 3013 TYPENAME(wk->wk_type)); 3014 /* NOTREACHED */ 3015 } 3016 } 3017 } 3018 3019 /* 3020 * Called from within the procedure above to deal with unsatisfied 3021 * allocation dependencies in a directory. The buffer must be locked, 3022 * thus, no I/O completion operations can occur while we are 3023 * manipulating its associated dependencies. 3024 */ 3025 static void 3026 initiate_write_filepage(pagedep, bp) 3027 struct pagedep *pagedep; 3028 struct buf *bp; 3029 { 3030 struct diradd *dap; 3031 struct direct *ep; 3032 int i; 3033 3034 if (pagedep->pd_state & IOSTARTED) { 3035 /* 3036 * This can only happen if there is a driver that does not 3037 * understand chaining. Here biodone will reissue the call 3038 * to strategy for the incomplete buffers. 3039 */ 3040 printf("initiate_write_filepage: already started\n"); 3041 return; 3042 } 3043 pagedep->pd_state |= IOSTARTED; 3044 ACQUIRE_LOCK(&lk); 3045 for (i = 0; i < DAHASHSZ; i++) { 3046 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 3047 ep = (struct direct *) 3048 ((char *)bp->b_data + dap->da_offset); 3049 if (ep->d_ino != dap->da_newinum) { 3050 FREE_LOCK(&lk); 3051 panic("%s: dir inum %d != new %d", 3052 "initiate_write_filepage", 3053 ep->d_ino, dap->da_newinum); 3054 } 3055 if (dap->da_state & DIRCHG) 3056 ep->d_ino = dap->da_previous->dm_oldinum; 3057 else 3058 ep->d_ino = 0; 3059 dap->da_state &= ~ATTACHED; 3060 dap->da_state |= UNDONE; 3061 } 3062 } 3063 FREE_LOCK(&lk); 3064 } 3065 3066 /* 3067 * Called from within the procedure above to deal with unsatisfied 3068 * allocation dependencies in an inodeblock. The buffer must be 3069 * locked, thus, no I/O completion operations can occur while we 3070 * are manipulating its associated dependencies. 3071 */ 3072 static void 3073 initiate_write_inodeblock(inodedep, bp) 3074 struct inodedep *inodedep; 3075 struct buf *bp; /* The inode block */ 3076 { 3077 struct allocdirect *adp, *lastadp; 3078 struct dinode *dp; 3079 struct fs *fs; 3080 ufs_lbn_t prevlbn = 0; 3081 int i, deplist; 3082 3083 if (inodedep->id_state & IOSTARTED) 3084 panic("initiate_write_inodeblock: already started"); 3085 inodedep->id_state |= IOSTARTED; 3086 fs = inodedep->id_fs; 3087 dp = (struct dinode *)bp->b_data + 3088 ino_to_fsbo(fs, inodedep->id_ino); 3089 /* 3090 * If the bitmap is not yet written, then the allocated 3091 * inode cannot be written to disk. 3092 */ 3093 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 3094 if (inodedep->id_savedino != NULL) 3095 panic("initiate_write_inodeblock: already doing I/O"); 3096 MALLOC(inodedep->id_savedino, struct dinode *, 3097 sizeof(struct dinode), M_INODEDEP, M_SOFTDEP_FLAGS); 3098 *inodedep->id_savedino = *dp; 3099 bzero((caddr_t)dp, sizeof(struct dinode)); 3100 return; 3101 } 3102 /* 3103 * If no dependencies, then there is nothing to roll back. 3104 */ 3105 inodedep->id_savedsize = dp->di_size; 3106 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 3107 return; 3108 /* 3109 * Set the dependencies to busy. 3110 */ 3111 ACQUIRE_LOCK(&lk); 3112 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3113 adp = TAILQ_NEXT(adp, ad_next)) { 3114 #ifdef DIAGNOSTIC 3115 if (deplist != 0 && prevlbn >= adp->ad_lbn) { 3116 FREE_LOCK(&lk); 3117 panic("softdep_write_inodeblock: lbn order"); 3118 } 3119 prevlbn = adp->ad_lbn; 3120 if (adp->ad_lbn < NDADDR && 3121 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) { 3122 FREE_LOCK(&lk); 3123 panic("%s: direct pointer #%ld mismatch %d != %d", 3124 "softdep_write_inodeblock", adp->ad_lbn, 3125 dp->di_db[adp->ad_lbn], adp->ad_newblkno); 3126 } 3127 if (adp->ad_lbn >= NDADDR && 3128 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) { 3129 FREE_LOCK(&lk); 3130 panic("%s: indirect pointer #%ld mismatch %d != %d", 3131 "softdep_write_inodeblock", adp->ad_lbn - NDADDR, 3132 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno); 3133 } 3134 deplist |= 1 << adp->ad_lbn; 3135 if ((adp->ad_state & ATTACHED) == 0) { 3136 FREE_LOCK(&lk); 3137 panic("softdep_write_inodeblock: Unknown state 0x%x", 3138 adp->ad_state); 3139 } 3140 #endif /* DIAGNOSTIC */ 3141 adp->ad_state &= ~ATTACHED; 3142 adp->ad_state |= UNDONE; 3143 } 3144 /* 3145 * The on-disk inode cannot claim to be any larger than the last 3146 * fragment that has been written. Otherwise, the on-disk inode 3147 * might have fragments that were not the last block in the file 3148 * which would corrupt the filesystem. 3149 */ 3150 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3151 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 3152 if (adp->ad_lbn >= NDADDR) 3153 break; 3154 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 3155 /* keep going until hitting a rollback to a frag */ 3156 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 3157 continue; 3158 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 3159 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 3160 #ifdef DIAGNOSTIC 3161 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) { 3162 FREE_LOCK(&lk); 3163 panic("softdep_write_inodeblock: lost dep1"); 3164 } 3165 #endif /* DIAGNOSTIC */ 3166 dp->di_db[i] = 0; 3167 } 3168 for (i = 0; i < NIADDR; i++) { 3169 #ifdef DIAGNOSTIC 3170 if (dp->di_ib[i] != 0 && 3171 (deplist & ((1 << NDADDR) << i)) == 0) { 3172 FREE_LOCK(&lk); 3173 panic("softdep_write_inodeblock: lost dep2"); 3174 } 3175 #endif /* DIAGNOSTIC */ 3176 dp->di_ib[i] = 0; 3177 } 3178 FREE_LOCK(&lk); 3179 return; 3180 } 3181 /* 3182 * If we have zero'ed out the last allocated block of the file, 3183 * roll back the size to the last currently allocated block. 3184 * We know that this last allocated block is a full-sized as 3185 * we already checked for fragments in the loop above. 3186 */ 3187 if (lastadp != NULL && 3188 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 3189 for (i = lastadp->ad_lbn; i >= 0; i--) 3190 if (dp->di_db[i] != 0) 3191 break; 3192 dp->di_size = (i + 1) * fs->fs_bsize; 3193 } 3194 /* 3195 * The only dependencies are for indirect blocks. 3196 * 3197 * The file size for indirect block additions is not guaranteed. 3198 * Such a guarantee would be non-trivial to achieve. The conventional 3199 * synchronous write implementation also does not make this guarantee. 3200 * Fsck should catch and fix discrepancies. Arguably, the file size 3201 * can be over-estimated without destroying integrity when the file 3202 * moves into the indirect blocks (i.e., is large). If we want to 3203 * postpone fsck, we are stuck with this argument. 3204 */ 3205 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 3206 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 3207 FREE_LOCK(&lk); 3208 } 3209 3210 /* 3211 * This routine is called during the completion interrupt 3212 * service routine for a disk write (from the procedure called 3213 * by the device driver to inform the filesystem caches of 3214 * a request completion). It should be called early in this 3215 * procedure, before the block is made available to other 3216 * processes or other routines are called. 3217 */ 3218 static void 3219 softdep_disk_write_complete(bp) 3220 struct buf *bp; /* describes the completed disk write */ 3221 { 3222 struct worklist *wk; 3223 struct workhead reattach; 3224 struct newblk *newblk; 3225 struct allocindir *aip; 3226 struct allocdirect *adp; 3227 struct indirdep *indirdep; 3228 struct inodedep *inodedep; 3229 struct bmsafemap *bmsafemap; 3230 3231 #ifdef DEBUG 3232 if (lk.lkt_held != NOHOLDER) 3233 panic("softdep_disk_write_complete: lock is held"); 3234 lk.lkt_held = SPECIAL_FLAG; 3235 #endif 3236 LIST_INIT(&reattach); 3237 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 3238 WORKLIST_REMOVE(wk); 3239 switch (wk->wk_type) { 3240 3241 case D_PAGEDEP: 3242 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 3243 WORKLIST_INSERT(&reattach, wk); 3244 continue; 3245 3246 case D_INODEDEP: 3247 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 3248 WORKLIST_INSERT(&reattach, wk); 3249 continue; 3250 3251 case D_BMSAFEMAP: 3252 bmsafemap = WK_BMSAFEMAP(wk); 3253 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) { 3254 newblk->nb_state |= DEPCOMPLETE; 3255 newblk->nb_bmsafemap = NULL; 3256 LIST_REMOVE(newblk, nb_deps); 3257 } 3258 while ((adp = 3259 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) { 3260 adp->ad_state |= DEPCOMPLETE; 3261 adp->ad_buf = NULL; 3262 LIST_REMOVE(adp, ad_deps); 3263 handle_allocdirect_partdone(adp); 3264 } 3265 while ((aip = 3266 LIST_FIRST(&bmsafemap->sm_allocindirhd))) { 3267 aip->ai_state |= DEPCOMPLETE; 3268 aip->ai_buf = NULL; 3269 LIST_REMOVE(aip, ai_deps); 3270 handle_allocindir_partdone(aip); 3271 } 3272 while ((inodedep = 3273 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { 3274 inodedep->id_state |= DEPCOMPLETE; 3275 LIST_REMOVE(inodedep, id_deps); 3276 inodedep->id_buf = NULL; 3277 } 3278 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 3279 continue; 3280 3281 case D_MKDIR: 3282 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 3283 continue; 3284 3285 case D_ALLOCDIRECT: 3286 adp = WK_ALLOCDIRECT(wk); 3287 adp->ad_state |= COMPLETE; 3288 handle_allocdirect_partdone(adp); 3289 continue; 3290 3291 case D_ALLOCINDIR: 3292 aip = WK_ALLOCINDIR(wk); 3293 aip->ai_state |= COMPLETE; 3294 handle_allocindir_partdone(aip); 3295 continue; 3296 3297 case D_INDIRDEP: 3298 indirdep = WK_INDIRDEP(wk); 3299 if (indirdep->ir_state & GOINGAWAY) { 3300 lk.lkt_held = NOHOLDER; 3301 panic("disk_write_complete: indirdep gone"); 3302 } 3303 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 3304 FREE(indirdep->ir_saveddata, M_INDIRDEP); 3305 indirdep->ir_saveddata = 0; 3306 indirdep->ir_state &= ~UNDONE; 3307 indirdep->ir_state |= ATTACHED; 3308 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 3309 handle_allocindir_partdone(aip); 3310 if (aip == LIST_FIRST(&indirdep->ir_donehd)) { 3311 lk.lkt_held = NOHOLDER; 3312 panic("disk_write_complete: not gone"); 3313 } 3314 } 3315 WORKLIST_INSERT(&reattach, wk); 3316 if ((bp->b_flags & B_DELWRI) == 0) 3317 stat_indir_blk_ptrs++; 3318 bdirty(bp); 3319 continue; 3320 3321 default: 3322 lk.lkt_held = NOHOLDER; 3323 panic("handle_disk_write_complete: Unknown type %s", 3324 TYPENAME(wk->wk_type)); 3325 /* NOTREACHED */ 3326 } 3327 } 3328 /* 3329 * Reattach any requests that must be redone. 3330 */ 3331 while ((wk = LIST_FIRST(&reattach)) != NULL) { 3332 WORKLIST_REMOVE(wk); 3333 WORKLIST_INSERT(&bp->b_dep, wk); 3334 } 3335 #ifdef DEBUG 3336 if (lk.lkt_held != SPECIAL_FLAG) 3337 panic("softdep_disk_write_complete: lock lost"); 3338 lk.lkt_held = NOHOLDER; 3339 #endif 3340 } 3341 3342 /* 3343 * Called from within softdep_disk_write_complete above. Note that 3344 * this routine is always called from interrupt level with further 3345 * splbio interrupts blocked. 3346 */ 3347 static void 3348 handle_allocdirect_partdone(adp) 3349 struct allocdirect *adp; /* the completed allocdirect */ 3350 { 3351 struct allocdirect *listadp; 3352 struct inodedep *inodedep; 3353 long bsize; 3354 3355 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3356 return; 3357 if (adp->ad_buf != NULL) { 3358 lk.lkt_held = NOHOLDER; 3359 panic("handle_allocdirect_partdone: dangling dep"); 3360 } 3361 /* 3362 * The on-disk inode cannot claim to be any larger than the last 3363 * fragment that has been written. Otherwise, the on-disk inode 3364 * might have fragments that were not the last block in the file 3365 * which would corrupt the filesystem. Thus, we cannot free any 3366 * allocdirects after one whose ad_oldblkno claims a fragment as 3367 * these blocks must be rolled back to zero before writing the inode. 3368 * We check the currently active set of allocdirects in id_inoupdt. 3369 */ 3370 inodedep = adp->ad_inodedep; 3371 bsize = inodedep->id_fs->fs_bsize; 3372 TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) { 3373 /* found our block */ 3374 if (listadp == adp) 3375 break; 3376 /* continue if ad_oldlbn is not a fragment */ 3377 if (listadp->ad_oldsize == 0 || 3378 listadp->ad_oldsize == bsize) 3379 continue; 3380 /* hit a fragment */ 3381 return; 3382 } 3383 /* 3384 * If we have reached the end of the current list without 3385 * finding the just finished dependency, then it must be 3386 * on the future dependency list. Future dependencies cannot 3387 * be freed until they are moved to the current list. 3388 */ 3389 if (listadp == NULL) { 3390 #ifdef DEBUG 3391 TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next) 3392 /* found our block */ 3393 if (listadp == adp) 3394 break; 3395 if (listadp == NULL) { 3396 lk.lkt_held = NOHOLDER; 3397 panic("handle_allocdirect_partdone: lost dep"); 3398 } 3399 #endif /* DEBUG */ 3400 return; 3401 } 3402 /* 3403 * If we have found the just finished dependency, then free 3404 * it along with anything that follows it that is complete. 3405 */ 3406 for (; adp; adp = listadp) { 3407 listadp = TAILQ_NEXT(adp, ad_next); 3408 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3409 return; 3410 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 3411 } 3412 } 3413 3414 /* 3415 * Called from within softdep_disk_write_complete above. Note that 3416 * this routine is always called from interrupt level with further 3417 * splbio interrupts blocked. 3418 */ 3419 static void 3420 handle_allocindir_partdone(aip) 3421 struct allocindir *aip; /* the completed allocindir */ 3422 { 3423 struct indirdep *indirdep; 3424 3425 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 3426 return; 3427 if (aip->ai_buf != NULL) { 3428 lk.lkt_held = NOHOLDER; 3429 panic("handle_allocindir_partdone: dangling dependency"); 3430 } 3431 indirdep = aip->ai_indirdep; 3432 if (indirdep->ir_state & UNDONE) { 3433 LIST_REMOVE(aip, ai_next); 3434 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 3435 return; 3436 } 3437 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 3438 aip->ai_newblkno; 3439 LIST_REMOVE(aip, ai_next); 3440 if (aip->ai_freefrag != NULL) 3441 add_to_worklist(&aip->ai_freefrag->ff_list); 3442 WORKITEM_FREE(aip, D_ALLOCINDIR); 3443 } 3444 3445 /* 3446 * Called from within softdep_disk_write_complete above to restore 3447 * in-memory inode block contents to their most up-to-date state. Note 3448 * that this routine is always called from interrupt level with further 3449 * splbio interrupts blocked. 3450 */ 3451 static int 3452 handle_written_inodeblock(inodedep, bp) 3453 struct inodedep *inodedep; 3454 struct buf *bp; /* buffer containing the inode block */ 3455 { 3456 struct worklist *wk, *filefree; 3457 struct allocdirect *adp, *nextadp; 3458 struct dinode *dp; 3459 int hadchanges; 3460 3461 if ((inodedep->id_state & IOSTARTED) == 0) { 3462 lk.lkt_held = NOHOLDER; 3463 panic("handle_written_inodeblock: not started"); 3464 } 3465 inodedep->id_state &= ~IOSTARTED; 3466 inodedep->id_state |= COMPLETE; 3467 dp = (struct dinode *)bp->b_data + 3468 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 3469 /* 3470 * If we had to rollback the inode allocation because of 3471 * bitmaps being incomplete, then simply restore it. 3472 * Keep the block dirty so that it will not be reclaimed until 3473 * all associated dependencies have been cleared and the 3474 * corresponding updates written to disk. 3475 */ 3476 if (inodedep->id_savedino != NULL) { 3477 *dp = *inodedep->id_savedino; 3478 FREE(inodedep->id_savedino, M_INODEDEP); 3479 inodedep->id_savedino = NULL; 3480 if ((bp->b_flags & B_DELWRI) == 0) 3481 stat_inode_bitmap++; 3482 bdirty(bp); 3483 return (1); 3484 } 3485 /* 3486 * Roll forward anything that had to be rolled back before 3487 * the inode could be updated. 3488 */ 3489 hadchanges = 0; 3490 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 3491 nextadp = TAILQ_NEXT(adp, ad_next); 3492 if (adp->ad_state & ATTACHED) { 3493 lk.lkt_held = NOHOLDER; 3494 panic("handle_written_inodeblock: new entry"); 3495 } 3496 if (adp->ad_lbn < NDADDR) { 3497 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) { 3498 lk.lkt_held = NOHOLDER; 3499 panic("%s: %s #%ld mismatch %d != %d", 3500 "handle_written_inodeblock", 3501 "direct pointer", adp->ad_lbn, 3502 dp->di_db[adp->ad_lbn], adp->ad_oldblkno); 3503 } 3504 dp->di_db[adp->ad_lbn] = adp->ad_newblkno; 3505 } else { 3506 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) { 3507 lk.lkt_held = NOHOLDER; 3508 panic("%s: %s #%ld allocated as %d", 3509 "handle_written_inodeblock", 3510 "indirect pointer", adp->ad_lbn - NDADDR, 3511 dp->di_ib[adp->ad_lbn - NDADDR]); 3512 } 3513 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno; 3514 } 3515 adp->ad_state &= ~UNDONE; 3516 adp->ad_state |= ATTACHED; 3517 hadchanges = 1; 3518 } 3519 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 3520 stat_direct_blk_ptrs++; 3521 /* 3522 * Reset the file size to its most up-to-date value. 3523 */ 3524 if (inodedep->id_savedsize == -1) { 3525 lk.lkt_held = NOHOLDER; 3526 panic("handle_written_inodeblock: bad size"); 3527 } 3528 if (dp->di_size != inodedep->id_savedsize) { 3529 dp->di_size = inodedep->id_savedsize; 3530 hadchanges = 1; 3531 } 3532 inodedep->id_savedsize = -1; 3533 /* 3534 * If there were any rollbacks in the inode block, then it must be 3535 * marked dirty so that its will eventually get written back in 3536 * its correct form. 3537 */ 3538 if (hadchanges) 3539 bdirty(bp); 3540 /* 3541 * Process any allocdirects that completed during the update. 3542 */ 3543 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 3544 handle_allocdirect_partdone(adp); 3545 /* 3546 * Process deallocations that were held pending until the 3547 * inode had been written to disk. Freeing of the inode 3548 * is delayed until after all blocks have been freed to 3549 * avoid creation of new <vfsid, inum, lbn> triples 3550 * before the old ones have been deleted. 3551 */ 3552 filefree = NULL; 3553 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 3554 WORKLIST_REMOVE(wk); 3555 switch (wk->wk_type) { 3556 3557 case D_FREEFILE: 3558 /* 3559 * We defer adding filefree to the worklist until 3560 * all other additions have been made to ensure 3561 * that it will be done after all the old blocks 3562 * have been freed. 3563 */ 3564 if (filefree != NULL) { 3565 lk.lkt_held = NOHOLDER; 3566 panic("handle_written_inodeblock: filefree"); 3567 } 3568 filefree = wk; 3569 continue; 3570 3571 case D_MKDIR: 3572 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 3573 continue; 3574 3575 case D_DIRADD: 3576 diradd_inode_written(WK_DIRADD(wk), inodedep); 3577 continue; 3578 3579 case D_FREEBLKS: 3580 case D_FREEFRAG: 3581 case D_DIRREM: 3582 add_to_worklist(wk); 3583 continue; 3584 3585 default: 3586 lk.lkt_held = NOHOLDER; 3587 panic("handle_written_inodeblock: Unknown type %s", 3588 TYPENAME(wk->wk_type)); 3589 /* NOTREACHED */ 3590 } 3591 } 3592 if (filefree != NULL) { 3593 if (free_inodedep(inodedep) == 0) { 3594 lk.lkt_held = NOHOLDER; 3595 panic("handle_written_inodeblock: live inodedep"); 3596 } 3597 add_to_worklist(filefree); 3598 return (0); 3599 } 3600 3601 /* 3602 * If no outstanding dependencies, free it. 3603 */ 3604 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0) 3605 return (0); 3606 return (hadchanges); 3607 } 3608 3609 /* 3610 * Process a diradd entry after its dependent inode has been written. 3611 * This routine must be called with splbio interrupts blocked. 3612 */ 3613 static void 3614 diradd_inode_written(dap, inodedep) 3615 struct diradd *dap; 3616 struct inodedep *inodedep; 3617 { 3618 struct pagedep *pagedep; 3619 3620 dap->da_state |= COMPLETE; 3621 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3622 if (dap->da_state & DIRCHG) 3623 pagedep = dap->da_previous->dm_pagedep; 3624 else 3625 pagedep = dap->da_pagedep; 3626 LIST_REMOVE(dap, da_pdlist); 3627 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3628 } 3629 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 3630 } 3631 3632 /* 3633 * Handle the completion of a mkdir dependency. 3634 */ 3635 static void 3636 handle_written_mkdir(mkdir, type) 3637 struct mkdir *mkdir; 3638 int type; 3639 { 3640 struct diradd *dap; 3641 struct pagedep *pagedep; 3642 3643 if (mkdir->md_state != type) { 3644 lk.lkt_held = NOHOLDER; 3645 panic("handle_written_mkdir: bad type"); 3646 } 3647 dap = mkdir->md_diradd; 3648 dap->da_state &= ~type; 3649 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 3650 dap->da_state |= DEPCOMPLETE; 3651 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3652 if (dap->da_state & DIRCHG) 3653 pagedep = dap->da_previous->dm_pagedep; 3654 else 3655 pagedep = dap->da_pagedep; 3656 LIST_REMOVE(dap, da_pdlist); 3657 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3658 } 3659 LIST_REMOVE(mkdir, md_mkdirs); 3660 WORKITEM_FREE(mkdir, D_MKDIR); 3661 } 3662 3663 /* 3664 * Called from within softdep_disk_write_complete above. 3665 * A write operation was just completed. Removed inodes can 3666 * now be freed and associated block pointers may be committed. 3667 * Note that this routine is always called from interrupt level 3668 * with further splbio interrupts blocked. 3669 */ 3670 static int 3671 handle_written_filepage(pagedep, bp) 3672 struct pagedep *pagedep; 3673 struct buf *bp; /* buffer containing the written page */ 3674 { 3675 struct dirrem *dirrem; 3676 struct diradd *dap, *nextdap; 3677 struct direct *ep; 3678 int i, chgs; 3679 3680 if ((pagedep->pd_state & IOSTARTED) == 0) { 3681 lk.lkt_held = NOHOLDER; 3682 panic("handle_written_filepage: not started"); 3683 } 3684 pagedep->pd_state &= ~IOSTARTED; 3685 /* 3686 * Process any directory removals that have been committed. 3687 */ 3688 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 3689 LIST_REMOVE(dirrem, dm_next); 3690 dirrem->dm_dirinum = pagedep->pd_ino; 3691 add_to_worklist(&dirrem->dm_list); 3692 } 3693 /* 3694 * Free any directory additions that have been committed. 3695 */ 3696 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 3697 free_diradd(dap); 3698 /* 3699 * Uncommitted directory entries must be restored. 3700 */ 3701 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 3702 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 3703 dap = nextdap) { 3704 nextdap = LIST_NEXT(dap, da_pdlist); 3705 if (dap->da_state & ATTACHED) { 3706 lk.lkt_held = NOHOLDER; 3707 panic("handle_written_filepage: attached"); 3708 } 3709 ep = (struct direct *) 3710 ((char *)bp->b_data + dap->da_offset); 3711 ep->d_ino = dap->da_newinum; 3712 dap->da_state &= ~UNDONE; 3713 dap->da_state |= ATTACHED; 3714 chgs = 1; 3715 /* 3716 * If the inode referenced by the directory has 3717 * been written out, then the dependency can be 3718 * moved to the pending list. 3719 */ 3720 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3721 LIST_REMOVE(dap, da_pdlist); 3722 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 3723 da_pdlist); 3724 } 3725 } 3726 } 3727 /* 3728 * If there were any rollbacks in the directory, then it must be 3729 * marked dirty so that its will eventually get written back in 3730 * its correct form. 3731 */ 3732 if (chgs) { 3733 if ((bp->b_flags & B_DELWRI) == 0) 3734 stat_dir_entry++; 3735 bdirty(bp); 3736 } 3737 /* 3738 * If no dependencies remain, the pagedep will be freed. 3739 * Otherwise it will remain to update the page before it 3740 * is written back to disk. 3741 */ 3742 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) { 3743 for (i = 0; i < DAHASHSZ; i++) 3744 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL) 3745 break; 3746 if (i == DAHASHSZ) { 3747 LIST_REMOVE(pagedep, pd_hash); 3748 WORKITEM_FREE(pagedep, D_PAGEDEP); 3749 return (0); 3750 } 3751 } 3752 return (1); 3753 } 3754 3755 /* 3756 * Writing back in-core inode structures. 3757 * 3758 * The filesystem only accesses an inode's contents when it occupies an 3759 * "in-core" inode structure. These "in-core" structures are separate from 3760 * the page frames used to cache inode blocks. Only the latter are 3761 * transferred to/from the disk. So, when the updated contents of the 3762 * "in-core" inode structure are copied to the corresponding in-memory inode 3763 * block, the dependencies are also transferred. The following procedure is 3764 * called when copying a dirty "in-core" inode to a cached inode block. 3765 */ 3766 3767 /* 3768 * Called when an inode is loaded from disk. If the effective link count 3769 * differed from the actual link count when it was last flushed, then we 3770 * need to ensure that the correct effective link count is put back. 3771 */ 3772 void 3773 softdep_load_inodeblock(ip) 3774 struct inode *ip; /* the "in_core" copy of the inode */ 3775 { 3776 struct inodedep *inodedep; 3777 3778 /* 3779 * Check for alternate nlink count. 3780 */ 3781 ip->i_effnlink = ip->i_nlink; 3782 ACQUIRE_LOCK(&lk); 3783 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3784 FREE_LOCK(&lk); 3785 return; 3786 } 3787 ip->i_effnlink -= inodedep->id_nlinkdelta; 3788 FREE_LOCK(&lk); 3789 } 3790 3791 /* 3792 * This routine is called just before the "in-core" inode 3793 * information is to be copied to the in-memory inode block. 3794 * Recall that an inode block contains several inodes. If 3795 * the force flag is set, then the dependencies will be 3796 * cleared so that the update can always be made. Note that 3797 * the buffer is locked when this routine is called, so we 3798 * will never be in the middle of writing the inode block 3799 * to disk. 3800 */ 3801 void 3802 softdep_update_inodeblock(ip, bp, waitfor) 3803 struct inode *ip; /* the "in_core" copy of the inode */ 3804 struct buf *bp; /* the buffer containing the inode block */ 3805 int waitfor; /* nonzero => update must be allowed */ 3806 { 3807 struct inodedep *inodedep; 3808 struct worklist *wk; 3809 int error, gotit; 3810 3811 /* 3812 * If the effective link count is not equal to the actual link 3813 * count, then we must track the difference in an inodedep while 3814 * the inode is (potentially) tossed out of the cache. Otherwise, 3815 * if there is no existing inodedep, then there are no dependencies 3816 * to track. 3817 */ 3818 ACQUIRE_LOCK(&lk); 3819 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3820 FREE_LOCK(&lk); 3821 if (ip->i_effnlink != ip->i_nlink) 3822 panic("softdep_update_inodeblock: bad link count"); 3823 return; 3824 } 3825 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) { 3826 FREE_LOCK(&lk); 3827 panic("softdep_update_inodeblock: bad delta"); 3828 } 3829 /* 3830 * Changes have been initiated. Anything depending on these 3831 * changes cannot occur until this inode has been written. 3832 */ 3833 inodedep->id_state &= ~COMPLETE; 3834 if ((inodedep->id_state & ONWORKLIST) == 0) 3835 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 3836 /* 3837 * Any new dependencies associated with the incore inode must 3838 * now be moved to the list associated with the buffer holding 3839 * the in-memory copy of the inode. Once merged process any 3840 * allocdirects that are completed by the merger. 3841 */ 3842 merge_inode_lists(inodedep); 3843 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL) 3844 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); 3845 /* 3846 * Now that the inode has been pushed into the buffer, the 3847 * operations dependent on the inode being written to disk 3848 * can be moved to the id_bufwait so that they will be 3849 * processed when the buffer I/O completes. 3850 */ 3851 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 3852 WORKLIST_REMOVE(wk); 3853 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 3854 } 3855 /* 3856 * Newly allocated inodes cannot be written until the bitmap 3857 * that allocates them have been written (indicated by 3858 * DEPCOMPLETE being set in id_state). If we are doing a 3859 * forced sync (e.g., an fsync on a file), we force the bitmap 3860 * to be written so that the update can be done. 3861 */ 3862 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) { 3863 FREE_LOCK(&lk); 3864 return; 3865 } 3866 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 3867 FREE_LOCK(&lk); 3868 if (gotit && 3869 (error = VOP_BWRITE(inodedep->id_buf->b_vp, inodedep->id_buf)) != 0) 3870 softdep_error("softdep_update_inodeblock: bwrite", error); 3871 if ((inodedep->id_state & DEPCOMPLETE) == 0) 3872 panic("softdep_update_inodeblock: update failed"); 3873 } 3874 3875 /* 3876 * Merge the new inode dependency list (id_newinoupdt) into the old 3877 * inode dependency list (id_inoupdt). This routine must be called 3878 * with splbio interrupts blocked. 3879 */ 3880 static void 3881 merge_inode_lists(inodedep) 3882 struct inodedep *inodedep; 3883 { 3884 struct allocdirect *listadp, *newadp; 3885 3886 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 3887 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) { 3888 if (listadp->ad_lbn < newadp->ad_lbn) { 3889 listadp = TAILQ_NEXT(listadp, ad_next); 3890 continue; 3891 } 3892 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 3893 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 3894 if (listadp->ad_lbn == newadp->ad_lbn) { 3895 allocdirect_merge(&inodedep->id_inoupdt, newadp, 3896 listadp); 3897 listadp = newadp; 3898 } 3899 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 3900 } 3901 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) { 3902 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 3903 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next); 3904 } 3905 } 3906 3907 /* 3908 * If we are doing an fsync, then we must ensure that any directory 3909 * entries for the inode have been written after the inode gets to disk. 3910 */ 3911 static int 3912 softdep_fsync(vp) 3913 struct vnode *vp; /* the "in_core" copy of the inode */ 3914 { 3915 struct inodedep *inodedep; 3916 struct pagedep *pagedep; 3917 struct worklist *wk; 3918 struct diradd *dap; 3919 struct mount *mnt; 3920 struct vnode *pvp; 3921 struct inode *ip; 3922 struct buf *bp; 3923 struct fs *fs; 3924 struct thread *td = curthread; /* XXX */ 3925 int error, flushparent; 3926 ino_t parentino; 3927 ufs_lbn_t lbn; 3928 3929 ip = VTOI(vp); 3930 fs = ip->i_fs; 3931 ACQUIRE_LOCK(&lk); 3932 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) { 3933 FREE_LOCK(&lk); 3934 return (0); 3935 } 3936 if (LIST_FIRST(&inodedep->id_inowait) != NULL || 3937 LIST_FIRST(&inodedep->id_bufwait) != NULL || 3938 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 3939 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) { 3940 FREE_LOCK(&lk); 3941 panic("softdep_fsync: pending ops"); 3942 } 3943 for (error = 0, flushparent = 0; ; ) { 3944 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 3945 break; 3946 if (wk->wk_type != D_DIRADD) { 3947 FREE_LOCK(&lk); 3948 panic("softdep_fsync: Unexpected type %s", 3949 TYPENAME(wk->wk_type)); 3950 } 3951 dap = WK_DIRADD(wk); 3952 /* 3953 * Flush our parent if this directory entry 3954 * has a MKDIR_PARENT dependency. 3955 */ 3956 if (dap->da_state & DIRCHG) 3957 pagedep = dap->da_previous->dm_pagedep; 3958 else 3959 pagedep = dap->da_pagedep; 3960 mnt = pagedep->pd_mnt; 3961 parentino = pagedep->pd_ino; 3962 lbn = pagedep->pd_lbn; 3963 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) { 3964 FREE_LOCK(&lk); 3965 panic("softdep_fsync: dirty"); 3966 } 3967 flushparent = dap->da_state & MKDIR_PARENT; 3968 /* 3969 * If we are being fsync'ed as part of vgone'ing this vnode, 3970 * then we will not be able to release and recover the 3971 * vnode below, so we just have to give up on writing its 3972 * directory entry out. It will eventually be written, just 3973 * not now, but then the user was not asking to have it 3974 * written, so we are not breaking any promises. 3975 */ 3976 if (vp->v_flag & VRECLAIMED) 3977 break; 3978 /* 3979 * We prevent deadlock by always fetching inodes from the 3980 * root, moving down the directory tree. Thus, when fetching 3981 * our parent directory, we must unlock ourselves before 3982 * requesting the lock on our parent. See the comment in 3983 * ufs_lookup for details on possible races. 3984 */ 3985 FREE_LOCK(&lk); 3986 VOP_UNLOCK(vp, 0, td); 3987 error = VFS_VGET(mnt, parentino, &pvp); 3988 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 3989 if (error != 0) 3990 return (error); 3991 if (flushparent) { 3992 if ((error = UFS_UPDATE(pvp, 1)) != 0) { 3993 vput(pvp); 3994 return (error); 3995 } 3996 } 3997 /* 3998 * Flush directory page containing the inode's name. 3999 */ 4000 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), &bp); 4001 if (error == 0) 4002 error = VOP_BWRITE(bp->b_vp, bp); 4003 vput(pvp); 4004 if (error != 0) 4005 return (error); 4006 ACQUIRE_LOCK(&lk); 4007 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) 4008 break; 4009 } 4010 FREE_LOCK(&lk); 4011 return (0); 4012 } 4013 4014 /* 4015 * Flush all the dirty bitmaps associated with the block device 4016 * before flushing the rest of the dirty blocks so as to reduce 4017 * the number of dependencies that will have to be rolled back. 4018 */ 4019 void 4020 softdep_fsync_mountdev(vp) 4021 struct vnode *vp; 4022 { 4023 struct buf *bp, *nbp; 4024 struct worklist *wk; 4025 4026 if (!vn_isdisk(vp, NULL)) 4027 panic("softdep_fsync_mountdev: vnode not a disk"); 4028 ACQUIRE_LOCK(&lk); 4029 for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 4030 nbp = TAILQ_NEXT(bp, b_vnbufs); 4031 /* 4032 * If it is already scheduled, skip to the next buffer. 4033 */ 4034 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) 4035 continue; 4036 if ((bp->b_flags & B_DELWRI) == 0) { 4037 FREE_LOCK(&lk); 4038 panic("softdep_fsync_mountdev: not dirty"); 4039 } 4040 /* 4041 * We are only interested in bitmaps with outstanding 4042 * dependencies. 4043 */ 4044 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 4045 wk->wk_type != D_BMSAFEMAP || 4046 (bp->b_xflags & BX_BKGRDINPROG)) { 4047 BUF_UNLOCK(bp); 4048 continue; 4049 } 4050 bremfree(bp); 4051 FREE_LOCK(&lk); 4052 (void) bawrite(bp); 4053 ACQUIRE_LOCK(&lk); 4054 /* 4055 * Since we may have slept during the I/O, we need 4056 * to start from a known point. 4057 */ 4058 nbp = TAILQ_FIRST(&vp->v_dirtyblkhd); 4059 } 4060 drain_output(vp, 1); 4061 FREE_LOCK(&lk); 4062 } 4063 4064 /* 4065 * This routine is called when we are trying to synchronously flush a 4066 * file. This routine must eliminate any filesystem metadata dependencies 4067 * so that the syncing routine can succeed by pushing the dirty blocks 4068 * associated with the file. If any I/O errors occur, they are returned. 4069 */ 4070 int 4071 softdep_sync_metadata(ap) 4072 struct vop_fsync_args /* { 4073 struct vnode *a_vp; 4074 struct ucred *a_cred; 4075 int a_waitfor; 4076 struct proc *a_p; 4077 } */ *ap; 4078 { 4079 struct vnode *vp = ap->a_vp; 4080 struct pagedep *pagedep; 4081 struct allocdirect *adp; 4082 struct allocindir *aip; 4083 struct buf *bp, *nbp; 4084 struct worklist *wk; 4085 int i, error, waitfor; 4086 4087 /* 4088 * Check whether this vnode is involved in a filesystem 4089 * that is doing soft dependency processing. 4090 */ 4091 if (!vn_isdisk(vp, NULL)) { 4092 if (!DOINGSOFTDEP(vp)) 4093 return (0); 4094 } else 4095 if (vp->v_rdev->si_mountpoint == NULL || 4096 (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0) 4097 return (0); 4098 /* 4099 * Ensure that any direct block dependencies have been cleared. 4100 */ 4101 ACQUIRE_LOCK(&lk); 4102 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) { 4103 FREE_LOCK(&lk); 4104 return (error); 4105 } 4106 /* 4107 * For most files, the only metadata dependencies are the 4108 * cylinder group maps that allocate their inode or blocks. 4109 * The block allocation dependencies can be found by traversing 4110 * the dependency lists for any buffers that remain on their 4111 * dirty buffer list. The inode allocation dependency will 4112 * be resolved when the inode is updated with MNT_WAIT. 4113 * This work is done in two passes. The first pass grabs most 4114 * of the buffers and begins asynchronously writing them. The 4115 * only way to wait for these asynchronous writes is to sleep 4116 * on the filesystem vnode which may stay busy for a long time 4117 * if the filesystem is active. So, instead, we make a second 4118 * pass over the dependencies blocking on each write. In the 4119 * usual case we will be blocking against a write that we 4120 * initiated, so when it is done the dependency will have been 4121 * resolved. Thus the second pass is expected to end quickly. 4122 */ 4123 waitfor = MNT_NOWAIT; 4124 top: 4125 /* 4126 * We must wait for any I/O in progress to finish so that 4127 * all potential buffers on the dirty list will be visible. 4128 */ 4129 drain_output(vp, 1); 4130 if (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT) == 0) { 4131 FREE_LOCK(&lk); 4132 return (0); 4133 } 4134 bp = TAILQ_FIRST(&vp->v_dirtyblkhd); 4135 loop: 4136 /* 4137 * As we hold the buffer locked, none of its dependencies 4138 * will disappear. 4139 */ 4140 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 4141 switch (wk->wk_type) { 4142 4143 case D_ALLOCDIRECT: 4144 adp = WK_ALLOCDIRECT(wk); 4145 if (adp->ad_state & DEPCOMPLETE) 4146 break; 4147 nbp = adp->ad_buf; 4148 if (getdirtybuf(&nbp, waitfor) == 0) 4149 break; 4150 FREE_LOCK(&lk); 4151 if (waitfor == MNT_NOWAIT) { 4152 bawrite(nbp); 4153 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 4154 bawrite(bp); 4155 return (error); 4156 } 4157 ACQUIRE_LOCK(&lk); 4158 break; 4159 4160 case D_ALLOCINDIR: 4161 aip = WK_ALLOCINDIR(wk); 4162 if (aip->ai_state & DEPCOMPLETE) 4163 break; 4164 nbp = aip->ai_buf; 4165 if (getdirtybuf(&nbp, waitfor) == 0) 4166 break; 4167 FREE_LOCK(&lk); 4168 if (waitfor == MNT_NOWAIT) { 4169 bawrite(nbp); 4170 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 4171 bawrite(bp); 4172 return (error); 4173 } 4174 ACQUIRE_LOCK(&lk); 4175 break; 4176 4177 case D_INDIRDEP: 4178 restart: 4179 4180 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) { 4181 if (aip->ai_state & DEPCOMPLETE) 4182 continue; 4183 nbp = aip->ai_buf; 4184 if (getdirtybuf(&nbp, MNT_WAIT) == 0) 4185 goto restart; 4186 FREE_LOCK(&lk); 4187 if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 4188 bawrite(bp); 4189 return (error); 4190 } 4191 ACQUIRE_LOCK(&lk); 4192 goto restart; 4193 } 4194 break; 4195 4196 case D_INODEDEP: 4197 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs, 4198 WK_INODEDEP(wk)->id_ino)) != 0) { 4199 FREE_LOCK(&lk); 4200 bawrite(bp); 4201 return (error); 4202 } 4203 break; 4204 4205 case D_PAGEDEP: 4206 /* 4207 * We are trying to sync a directory that may 4208 * have dependencies on both its own metadata 4209 * and/or dependencies on the inodes of any 4210 * recently allocated files. We walk its diradd 4211 * lists pushing out the associated inode. 4212 */ 4213 pagedep = WK_PAGEDEP(wk); 4214 for (i = 0; i < DAHASHSZ; i++) { 4215 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 4216 continue; 4217 if ((error = 4218 flush_pagedep_deps(vp, pagedep->pd_mnt, 4219 &pagedep->pd_diraddhd[i]))) { 4220 FREE_LOCK(&lk); 4221 bawrite(bp); 4222 return (error); 4223 } 4224 } 4225 break; 4226 4227 case D_MKDIR: 4228 /* 4229 * This case should never happen if the vnode has 4230 * been properly sync'ed. However, if this function 4231 * is used at a place where the vnode has not yet 4232 * been sync'ed, this dependency can show up. So, 4233 * rather than panic, just flush it. 4234 */ 4235 nbp = WK_MKDIR(wk)->md_buf; 4236 if (getdirtybuf(&nbp, waitfor) == 0) 4237 break; 4238 FREE_LOCK(&lk); 4239 if (waitfor == MNT_NOWAIT) { 4240 bawrite(nbp); 4241 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 4242 bawrite(bp); 4243 return (error); 4244 } 4245 ACQUIRE_LOCK(&lk); 4246 break; 4247 4248 case D_BMSAFEMAP: 4249 /* 4250 * This case should never happen if the vnode has 4251 * been properly sync'ed. However, if this function 4252 * is used at a place where the vnode has not yet 4253 * been sync'ed, this dependency can show up. So, 4254 * rather than panic, just flush it. 4255 */ 4256 nbp = WK_BMSAFEMAP(wk)->sm_buf; 4257 if (getdirtybuf(&nbp, waitfor) == 0) 4258 break; 4259 FREE_LOCK(&lk); 4260 if (waitfor == MNT_NOWAIT) { 4261 bawrite(nbp); 4262 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 4263 bawrite(bp); 4264 return (error); 4265 } 4266 ACQUIRE_LOCK(&lk); 4267 break; 4268 4269 default: 4270 FREE_LOCK(&lk); 4271 panic("softdep_sync_metadata: Unknown type %s", 4272 TYPENAME(wk->wk_type)); 4273 /* NOTREACHED */ 4274 } 4275 } 4276 (void) getdirtybuf(&TAILQ_NEXT(bp, b_vnbufs), MNT_WAIT); 4277 nbp = TAILQ_NEXT(bp, b_vnbufs); 4278 FREE_LOCK(&lk); 4279 bawrite(bp); 4280 ACQUIRE_LOCK(&lk); 4281 if (nbp != NULL) { 4282 bp = nbp; 4283 goto loop; 4284 } 4285 /* 4286 * The brief unlock is to allow any pent up dependency 4287 * processing to be done. Then proceed with the second pass. 4288 */ 4289 if (waitfor == MNT_NOWAIT) { 4290 waitfor = MNT_WAIT; 4291 FREE_LOCK(&lk); 4292 ACQUIRE_LOCK(&lk); 4293 goto top; 4294 } 4295 4296 /* 4297 * If we have managed to get rid of all the dirty buffers, 4298 * then we are done. For certain directories and block 4299 * devices, we may need to do further work. 4300 * 4301 * We must wait for any I/O in progress to finish so that 4302 * all potential buffers on the dirty list will be visible. 4303 */ 4304 drain_output(vp, 1); 4305 if (TAILQ_FIRST(&vp->v_dirtyblkhd) == NULL) { 4306 FREE_LOCK(&lk); 4307 return (0); 4308 } 4309 4310 FREE_LOCK(&lk); 4311 /* 4312 * If we are trying to sync a block device, some of its buffers may 4313 * contain metadata that cannot be written until the contents of some 4314 * partially written files have been written to disk. The only easy 4315 * way to accomplish this is to sync the entire filesystem (luckily 4316 * this happens rarely). 4317 */ 4318 if (vn_isdisk(vp, NULL) && 4319 vp->v_rdev && 4320 vp->v_rdev->si_mountpoint && !VOP_ISLOCKED(vp, NULL) && 4321 (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT, ap->a_td)) != 0) 4322 return (error); 4323 return (0); 4324 } 4325 4326 /* 4327 * Flush the dependencies associated with an inodedep. 4328 * Called with splbio blocked. 4329 */ 4330 static int 4331 flush_inodedep_deps(fs, ino) 4332 struct fs *fs; 4333 ino_t ino; 4334 { 4335 struct inodedep *inodedep; 4336 struct allocdirect *adp; 4337 int error, waitfor; 4338 struct buf *bp; 4339 4340 /* 4341 * This work is done in two passes. The first pass grabs most 4342 * of the buffers and begins asynchronously writing them. The 4343 * only way to wait for these asynchronous writes is to sleep 4344 * on the filesystem vnode which may stay busy for a long time 4345 * if the filesystem is active. So, instead, we make a second 4346 * pass over the dependencies blocking on each write. In the 4347 * usual case we will be blocking against a write that we 4348 * initiated, so when it is done the dependency will have been 4349 * resolved. Thus the second pass is expected to end quickly. 4350 * We give a brief window at the top of the loop to allow 4351 * any pending I/O to complete. 4352 */ 4353 for (waitfor = MNT_NOWAIT; ; ) { 4354 FREE_LOCK(&lk); 4355 ACQUIRE_LOCK(&lk); 4356 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4357 return (0); 4358 TAILQ_FOREACH(adp, &inodedep->id_inoupdt, ad_next) { 4359 if (adp->ad_state & DEPCOMPLETE) 4360 continue; 4361 bp = adp->ad_buf; 4362 if (getdirtybuf(&bp, waitfor) == 0) { 4363 if (waitfor == MNT_NOWAIT) 4364 continue; 4365 break; 4366 } 4367 FREE_LOCK(&lk); 4368 if (waitfor == MNT_NOWAIT) { 4369 bawrite(bp); 4370 } else if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) { 4371 ACQUIRE_LOCK(&lk); 4372 return (error); 4373 } 4374 ACQUIRE_LOCK(&lk); 4375 break; 4376 } 4377 if (adp != NULL) 4378 continue; 4379 TAILQ_FOREACH(adp, &inodedep->id_newinoupdt, ad_next) { 4380 if (adp->ad_state & DEPCOMPLETE) 4381 continue; 4382 bp = adp->ad_buf; 4383 if (getdirtybuf(&bp, waitfor) == 0) { 4384 if (waitfor == MNT_NOWAIT) 4385 continue; 4386 break; 4387 } 4388 FREE_LOCK(&lk); 4389 if (waitfor == MNT_NOWAIT) { 4390 bawrite(bp); 4391 } else if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) { 4392 ACQUIRE_LOCK(&lk); 4393 return (error); 4394 } 4395 ACQUIRE_LOCK(&lk); 4396 break; 4397 } 4398 if (adp != NULL) 4399 continue; 4400 /* 4401 * If pass2, we are done, otherwise do pass 2. 4402 */ 4403 if (waitfor == MNT_WAIT) 4404 break; 4405 waitfor = MNT_WAIT; 4406 } 4407 /* 4408 * Try freeing inodedep in case all dependencies have been removed. 4409 */ 4410 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0) 4411 (void) free_inodedep(inodedep); 4412 return (0); 4413 } 4414 4415 /* 4416 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 4417 * Called with splbio blocked. 4418 */ 4419 static int 4420 flush_pagedep_deps(pvp, mp, diraddhdp) 4421 struct vnode *pvp; 4422 struct mount *mp; 4423 struct diraddhd *diraddhdp; 4424 { 4425 struct thread *td = curthread; /* XXX */ 4426 struct inodedep *inodedep; 4427 struct ufsmount *ump; 4428 struct diradd *dap; 4429 struct vnode *vp; 4430 int gotit, error = 0; 4431 struct buf *bp; 4432 ino_t inum; 4433 4434 ump = VFSTOUFS(mp); 4435 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 4436 /* 4437 * Flush ourselves if this directory entry 4438 * has a MKDIR_PARENT dependency. 4439 */ 4440 if (dap->da_state & MKDIR_PARENT) { 4441 FREE_LOCK(&lk); 4442 if ((error = UFS_UPDATE(pvp, 1)) != 0) 4443 break; 4444 ACQUIRE_LOCK(&lk); 4445 /* 4446 * If that cleared dependencies, go on to next. 4447 */ 4448 if (dap != LIST_FIRST(diraddhdp)) 4449 continue; 4450 if (dap->da_state & MKDIR_PARENT) { 4451 FREE_LOCK(&lk); 4452 panic("flush_pagedep_deps: MKDIR_PARENT"); 4453 } 4454 } 4455 /* 4456 * A newly allocated directory must have its "." and 4457 * ".." entries written out before its name can be 4458 * committed in its parent. We do not want or need 4459 * the full semantics of a synchronous VOP_FSYNC as 4460 * that may end up here again, once for each directory 4461 * level in the filesystem. Instead, we push the blocks 4462 * and wait for them to clear. We have to fsync twice 4463 * because the first call may choose to defer blocks 4464 * that still have dependencies, but deferral will 4465 * happen at most once. 4466 */ 4467 inum = dap->da_newinum; 4468 if (dap->da_state & MKDIR_BODY) { 4469 FREE_LOCK(&lk); 4470 if ((error = VFS_VGET(mp, inum, &vp)) != 0) 4471 break; 4472 if ((error=VOP_FSYNC(vp, MNT_NOWAIT, td)) || 4473 (error=VOP_FSYNC(vp, MNT_NOWAIT, td))) { 4474 vput(vp); 4475 break; 4476 } 4477 drain_output(vp, 0); 4478 vput(vp); 4479 ACQUIRE_LOCK(&lk); 4480 /* 4481 * If that cleared dependencies, go on to next. 4482 */ 4483 if (dap != LIST_FIRST(diraddhdp)) 4484 continue; 4485 if (dap->da_state & MKDIR_BODY) { 4486 FREE_LOCK(&lk); 4487 panic("flush_pagedep_deps: MKDIR_BODY"); 4488 } 4489 } 4490 /* 4491 * Flush the inode on which the directory entry depends. 4492 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 4493 * the only remaining dependency is that the updated inode 4494 * count must get pushed to disk. The inode has already 4495 * been pushed into its inode buffer (via VOP_UPDATE) at 4496 * the time of the reference count change. So we need only 4497 * locate that buffer, ensure that there will be no rollback 4498 * caused by a bitmap dependency, then write the inode buffer. 4499 */ 4500 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) { 4501 FREE_LOCK(&lk); 4502 panic("flush_pagedep_deps: lost inode"); 4503 } 4504 /* 4505 * If the inode still has bitmap dependencies, 4506 * push them to disk. 4507 */ 4508 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4509 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 4510 FREE_LOCK(&lk); 4511 if (gotit && 4512 (error = VOP_BWRITE(inodedep->id_buf->b_vp, 4513 inodedep->id_buf)) != 0) 4514 break; 4515 ACQUIRE_LOCK(&lk); 4516 if (dap != LIST_FIRST(diraddhdp)) 4517 continue; 4518 } 4519 /* 4520 * If the inode is still sitting in a buffer waiting 4521 * to be written, push it to disk. 4522 */ 4523 FREE_LOCK(&lk); 4524 if ((error = bread(ump->um_devvp, 4525 fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), 4526 (int)ump->um_fs->fs_bsize, &bp)) != 0) 4527 break; 4528 if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) 4529 break; 4530 ACQUIRE_LOCK(&lk); 4531 /* 4532 * If we have failed to get rid of all the dependencies 4533 * then something is seriously wrong. 4534 */ 4535 if (dap == LIST_FIRST(diraddhdp)) { 4536 FREE_LOCK(&lk); 4537 panic("flush_pagedep_deps: flush failed"); 4538 } 4539 } 4540 if (error) 4541 ACQUIRE_LOCK(&lk); 4542 return (error); 4543 } 4544 4545 /* 4546 * A large burst of file addition or deletion activity can drive the 4547 * memory load excessively high. First attempt to slow things down 4548 * using the techniques below. If that fails, this routine requests 4549 * the offending operations to fall back to running synchronously 4550 * until the memory load returns to a reasonable level. 4551 */ 4552 int 4553 softdep_slowdown(vp) 4554 struct vnode *vp; 4555 { 4556 int max_softdeps_hard; 4557 4558 max_softdeps_hard = max_softdeps * 11 / 10; 4559 if (num_dirrem < max_softdeps_hard / 2 && 4560 num_inodedep < max_softdeps_hard) 4561 return (0); 4562 stat_sync_limit_hit += 1; 4563 return (1); 4564 } 4565 4566 /* 4567 * If memory utilization has gotten too high, deliberately slow things 4568 * down and speed up the I/O processing. 4569 */ 4570 static int 4571 request_cleanup(resource, islocked) 4572 int resource; 4573 int islocked; 4574 { 4575 struct thread *td = curthread; /* XXX */ 4576 int s; 4577 4578 /* 4579 * We never hold up the filesystem syncer process. 4580 */ 4581 if (td == filesys_syncer) 4582 return (0); 4583 /* 4584 * First check to see if the work list has gotten backlogged. 4585 * If it has, co-opt this process to help clean up two entries. 4586 * Because this process may hold inodes locked, we cannot 4587 * handle any remove requests that might block on a locked 4588 * inode as that could lead to deadlock. 4589 */ 4590 if (num_on_worklist > max_softdeps / 10) { 4591 if (islocked) 4592 FREE_LOCK(&lk); 4593 process_worklist_item(NULL, LK_NOWAIT); 4594 process_worklist_item(NULL, LK_NOWAIT); 4595 stat_worklist_push += 2; 4596 if (islocked) 4597 ACQUIRE_LOCK(&lk); 4598 return(1); 4599 } 4600 4601 /* 4602 * If we are resource constrained on inode dependencies, try 4603 * flushing some dirty inodes. Otherwise, we are constrained 4604 * by file deletions, so try accelerating flushes of directories 4605 * with removal dependencies. We would like to do the cleanup 4606 * here, but we probably hold an inode locked at this point and 4607 * that might deadlock against one that we try to clean. So, 4608 * the best that we can do is request the syncer daemon to do 4609 * the cleanup for us. 4610 */ 4611 switch (resource) { 4612 4613 case FLUSH_INODES: 4614 stat_ino_limit_push += 1; 4615 req_clear_inodedeps += 1; 4616 stat_countp = &stat_ino_limit_hit; 4617 break; 4618 4619 case FLUSH_REMOVE: 4620 stat_blk_limit_push += 1; 4621 req_clear_remove += 1; 4622 stat_countp = &stat_blk_limit_hit; 4623 break; 4624 4625 default: 4626 if (islocked) 4627 FREE_LOCK(&lk); 4628 panic("request_cleanup: unknown type"); 4629 } 4630 /* 4631 * Hopefully the syncer daemon will catch up and awaken us. 4632 * We wait at most tickdelay before proceeding in any case. 4633 */ 4634 if (islocked == 0) 4635 ACQUIRE_LOCK(&lk); 4636 s = splsoftclock(); 4637 proc_waiting += 1; 4638 if (!callout_active(&handle)) 4639 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2, 4640 pause_timer, NULL); 4641 interlocked_sleep(&lk, SLEEP, (caddr_t)&proc_waiting, 0, 4642 "softupdate", 0); 4643 proc_waiting -= 1; 4644 splx(s); 4645 if (islocked == 0) 4646 FREE_LOCK(&lk); 4647 return (1); 4648 } 4649 4650 /* 4651 * Awaken processes pausing in request_cleanup and clear proc_waiting 4652 * to indicate that there is no longer a timer running. 4653 */ 4654 void 4655 pause_timer(arg) 4656 void *arg; 4657 { 4658 *stat_countp += 1; 4659 wakeup_one(&proc_waiting); 4660 if (proc_waiting > 0) 4661 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2, 4662 pause_timer, NULL); 4663 else 4664 callout_deactivate(&handle); 4665 } 4666 4667 /* 4668 * Flush out a directory with at least one removal dependency in an effort to 4669 * reduce the number of dirrem, freefile, and freeblks dependency structures. 4670 */ 4671 static void 4672 clear_remove(struct thread *td) 4673 { 4674 struct pagedep_hashhead *pagedephd; 4675 struct pagedep *pagedep; 4676 static int next = 0; 4677 struct mount *mp; 4678 struct vnode *vp; 4679 int error, cnt; 4680 ino_t ino; 4681 4682 ACQUIRE_LOCK(&lk); 4683 for (cnt = 0; cnt < pagedep_hash; cnt++) { 4684 pagedephd = &pagedep_hashtbl[next++]; 4685 if (next >= pagedep_hash) 4686 next = 0; 4687 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 4688 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL) 4689 continue; 4690 mp = pagedep->pd_mnt; 4691 ino = pagedep->pd_ino; 4692 FREE_LOCK(&lk); 4693 if ((error = VFS_VGET(mp, ino, &vp)) != 0) { 4694 softdep_error("clear_remove: vget", error); 4695 return; 4696 } 4697 if ((error = VOP_FSYNC(vp, MNT_NOWAIT, td))) 4698 softdep_error("clear_remove: fsync", error); 4699 drain_output(vp, 0); 4700 vput(vp); 4701 return; 4702 } 4703 } 4704 FREE_LOCK(&lk); 4705 } 4706 4707 /* 4708 * Clear out a block of dirty inodes in an effort to reduce 4709 * the number of inodedep dependency structures. 4710 */ 4711 static void 4712 clear_inodedeps(struct thread *td) 4713 { 4714 struct inodedep_hashhead *inodedephd; 4715 struct inodedep *inodedep; 4716 static int next = 0; 4717 struct mount *mp; 4718 struct vnode *vp; 4719 struct fs *fs; 4720 int error, cnt; 4721 ino_t firstino, lastino, ino; 4722 4723 ACQUIRE_LOCK(&lk); 4724 /* 4725 * Pick a random inode dependency to be cleared. 4726 * We will then gather up all the inodes in its block 4727 * that have dependencies and flush them out. 4728 */ 4729 for (cnt = 0; cnt < inodedep_hash; cnt++) { 4730 inodedephd = &inodedep_hashtbl[next++]; 4731 if (next >= inodedep_hash) 4732 next = 0; 4733 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 4734 break; 4735 } 4736 if (inodedep == NULL) 4737 return; 4738 /* 4739 * Ugly code to find mount point given pointer to superblock. 4740 */ 4741 fs = inodedep->id_fs; 4742 TAILQ_FOREACH(mp, &mountlist, mnt_list) 4743 if ((mp->mnt_flag & MNT_SOFTDEP) && fs == VFSTOUFS(mp)->um_fs) 4744 break; 4745 /* 4746 * Find the last inode in the block with dependencies. 4747 */ 4748 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 4749 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 4750 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0) 4751 break; 4752 /* 4753 * Asynchronously push all but the last inode with dependencies. 4754 * Synchronously push the last inode with dependencies to ensure 4755 * that the inode block gets written to free up the inodedeps. 4756 */ 4757 for (ino = firstino; ino <= lastino; ino++) { 4758 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4759 continue; 4760 FREE_LOCK(&lk); 4761 if ((error = VFS_VGET(mp, ino, &vp)) != 0) { 4762 softdep_error("clear_inodedeps: vget", error); 4763 return; 4764 } 4765 if (ino == lastino) { 4766 if ((error = VOP_FSYNC(vp, MNT_WAIT, td))) 4767 softdep_error("clear_inodedeps: fsync1", error); 4768 } else { 4769 if ((error = VOP_FSYNC(vp, MNT_NOWAIT, td))) 4770 softdep_error("clear_inodedeps: fsync2", error); 4771 drain_output(vp, 0); 4772 } 4773 vput(vp); 4774 ACQUIRE_LOCK(&lk); 4775 } 4776 FREE_LOCK(&lk); 4777 } 4778 4779 /* 4780 * Function to determine if the buffer has outstanding dependencies 4781 * that will cause a roll-back if the buffer is written. If wantcount 4782 * is set, return number of dependencies, otherwise just yes or no. 4783 */ 4784 static int 4785 softdep_count_dependencies(bp, wantcount) 4786 struct buf *bp; 4787 int wantcount; 4788 { 4789 struct worklist *wk; 4790 struct inodedep *inodedep; 4791 struct indirdep *indirdep; 4792 struct allocindir *aip; 4793 struct pagedep *pagedep; 4794 struct diradd *dap; 4795 int i, retval; 4796 4797 retval = 0; 4798 ACQUIRE_LOCK(&lk); 4799 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 4800 switch (wk->wk_type) { 4801 4802 case D_INODEDEP: 4803 inodedep = WK_INODEDEP(wk); 4804 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4805 /* bitmap allocation dependency */ 4806 retval += 1; 4807 if (!wantcount) 4808 goto out; 4809 } 4810 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 4811 /* direct block pointer dependency */ 4812 retval += 1; 4813 if (!wantcount) 4814 goto out; 4815 } 4816 continue; 4817 4818 case D_INDIRDEP: 4819 indirdep = WK_INDIRDEP(wk); 4820 4821 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 4822 /* indirect block pointer dependency */ 4823 retval += 1; 4824 if (!wantcount) 4825 goto out; 4826 } 4827 continue; 4828 4829 case D_PAGEDEP: 4830 pagedep = WK_PAGEDEP(wk); 4831 for (i = 0; i < DAHASHSZ; i++) { 4832 4833 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 4834 /* directory entry dependency */ 4835 retval += 1; 4836 if (!wantcount) 4837 goto out; 4838 } 4839 } 4840 continue; 4841 4842 case D_BMSAFEMAP: 4843 case D_ALLOCDIRECT: 4844 case D_ALLOCINDIR: 4845 case D_MKDIR: 4846 /* never a dependency on these blocks */ 4847 continue; 4848 4849 default: 4850 FREE_LOCK(&lk); 4851 panic("softdep_check_for_rollback: Unexpected type %s", 4852 TYPENAME(wk->wk_type)); 4853 /* NOTREACHED */ 4854 } 4855 } 4856 out: 4857 FREE_LOCK(&lk); 4858 return retval; 4859 } 4860 4861 /* 4862 * Acquire exclusive access to a buffer. 4863 * Must be called with splbio blocked. 4864 * Return 1 if buffer was acquired. 4865 */ 4866 static int 4867 getdirtybuf(bpp, waitfor) 4868 struct buf **bpp; 4869 int waitfor; 4870 { 4871 struct buf *bp; 4872 int error; 4873 4874 for (;;) { 4875 if ((bp = *bpp) == NULL) 4876 return (0); 4877 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 4878 if ((bp->b_xflags & BX_BKGRDINPROG) == 0) 4879 break; 4880 BUF_UNLOCK(bp); 4881 if (waitfor != MNT_WAIT) 4882 return (0); 4883 bp->b_xflags |= BX_BKGRDWAIT; 4884 interlocked_sleep(&lk, SLEEP, &bp->b_xflags, 0, 4885 "getbuf", 0); 4886 continue; 4887 } 4888 if (waitfor != MNT_WAIT) 4889 return (0); 4890 error = interlocked_sleep(&lk, LOCKBUF, bp, 4891 LK_EXCLUSIVE | LK_SLEEPFAIL, 0, 0); 4892 if (error != ENOLCK) { 4893 FREE_LOCK(&lk); 4894 panic("getdirtybuf: inconsistent lock"); 4895 } 4896 } 4897 if ((bp->b_flags & B_DELWRI) == 0) { 4898 BUF_UNLOCK(bp); 4899 return (0); 4900 } 4901 bremfree(bp); 4902 return (1); 4903 } 4904 4905 /* 4906 * Wait for pending output on a vnode to complete. 4907 * Must be called with vnode locked. 4908 */ 4909 static void 4910 drain_output(vp, islocked) 4911 struct vnode *vp; 4912 int islocked; 4913 { 4914 4915 if (!islocked) 4916 ACQUIRE_LOCK(&lk); 4917 while (vp->v_numoutput) { 4918 vp->v_flag |= VBWAIT; 4919 interlocked_sleep(&lk, SLEEP, (caddr_t)&vp->v_numoutput, 4920 0, "drainvp", 0); 4921 } 4922 if (!islocked) 4923 FREE_LOCK(&lk); 4924 } 4925 4926 /* 4927 * Called whenever a buffer that is being invalidated or reallocated 4928 * contains dependencies. This should only happen if an I/O error has 4929 * occurred. The routine is called with the buffer locked. 4930 */ 4931 static void 4932 softdep_deallocate_dependencies(bp) 4933 struct buf *bp; 4934 { 4935 4936 if ((bp->b_flags & B_ERROR) == 0) 4937 panic("softdep_deallocate_dependencies: dangling deps"); 4938 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); 4939 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 4940 } 4941 4942 /* 4943 * Function to handle asynchronous write errors in the filesystem. 4944 */ 4945 void 4946 softdep_error(func, error) 4947 char *func; 4948 int error; 4949 { 4950 4951 /* XXX should do something better! */ 4952 printf("%s: got error %d while accessing filesystem\n", func, error); 4953 } 4954