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 1.57 2008/06/28 17:59:51 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 <machine/inttypes.h> 64 #include "dir.h" 65 #include "quota.h" 66 #include "inode.h" 67 #include "ufsmount.h" 68 #include "fs.h" 69 #include "softdep.h" 70 #include "ffs_extern.h" 71 #include "ufs_extern.h" 72 73 #include <sys/buf2.h> 74 #include <sys/mplock2.h> 75 #include <sys/thread2.h> 76 77 /* 78 * These definitions need to be adapted to the system to which 79 * this file is being ported. 80 */ 81 /* 82 * malloc types defined for the softdep system. 83 */ 84 MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies"); 85 MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies"); 86 MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation"); 87 MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map"); 88 MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode"); 89 MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies"); 90 MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block"); 91 MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode"); 92 MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode"); 93 MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated"); 94 MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry"); 95 MALLOC_DEFINE(M_MKDIR, "mkdir","New directory"); 96 MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted"); 97 98 #define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE) 99 100 #define D_PAGEDEP 0 101 #define D_INODEDEP 1 102 #define D_NEWBLK 2 103 #define D_BMSAFEMAP 3 104 #define D_ALLOCDIRECT 4 105 #define D_INDIRDEP 5 106 #define D_ALLOCINDIR 6 107 #define D_FREEFRAG 7 108 #define D_FREEBLKS 8 109 #define D_FREEFILE 9 110 #define D_DIRADD 10 111 #define D_MKDIR 11 112 #define D_DIRREM 12 113 #define D_LAST D_DIRREM 114 115 /* 116 * translate from workitem type to memory type 117 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 118 */ 119 static struct malloc_type *memtype[] = { 120 M_PAGEDEP, 121 M_INODEDEP, 122 M_NEWBLK, 123 M_BMSAFEMAP, 124 M_ALLOCDIRECT, 125 M_INDIRDEP, 126 M_ALLOCINDIR, 127 M_FREEFRAG, 128 M_FREEBLKS, 129 M_FREEFILE, 130 M_DIRADD, 131 M_MKDIR, 132 M_DIRREM 133 }; 134 135 #define DtoM(type) (memtype[type]) 136 137 /* 138 * Names of malloc types. 139 */ 140 #define TYPENAME(type) \ 141 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???") 142 /* 143 * End system adaptaion definitions. 144 */ 145 146 /* 147 * Internal function prototypes. 148 */ 149 static void softdep_error(char *, int); 150 static void drain_output(struct vnode *, int); 151 static int getdirtybuf(struct buf **, int); 152 static void clear_remove(struct thread *); 153 static void clear_inodedeps(struct thread *); 154 static int flush_pagedep_deps(struct vnode *, struct mount *, 155 struct diraddhd *); 156 static int flush_inodedep_deps(struct fs *, ino_t); 157 static int handle_written_filepage(struct pagedep *, struct buf *); 158 static void diradd_inode_written(struct diradd *, struct inodedep *); 159 static int handle_written_inodeblock(struct inodedep *, struct buf *); 160 static void handle_allocdirect_partdone(struct allocdirect *); 161 static void handle_allocindir_partdone(struct allocindir *); 162 static void initiate_write_filepage(struct pagedep *, struct buf *); 163 static void handle_written_mkdir(struct mkdir *, int); 164 static void initiate_write_inodeblock(struct inodedep *, struct buf *); 165 static void handle_workitem_freefile(struct freefile *); 166 static void handle_workitem_remove(struct dirrem *); 167 static struct dirrem *newdirrem(struct buf *, struct inode *, 168 struct inode *, int, struct dirrem **); 169 static void free_diradd(struct diradd *); 170 static void free_allocindir(struct allocindir *, struct inodedep *); 171 static int indir_trunc (struct inode *, off_t, int, ufs_lbn_t, long *); 172 static void deallocate_dependencies(struct buf *, struct inodedep *); 173 static void free_allocdirect(struct allocdirectlst *, 174 struct allocdirect *, int); 175 static int check_inode_unwritten(struct inodedep *); 176 static int free_inodedep(struct inodedep *); 177 static void handle_workitem_freeblocks(struct freeblks *); 178 static void merge_inode_lists(struct inodedep *); 179 static void setup_allocindir_phase2(struct buf *, struct inode *, 180 struct allocindir *); 181 static struct allocindir *newallocindir(struct inode *, int, ufs_daddr_t, 182 ufs_daddr_t); 183 static void handle_workitem_freefrag(struct freefrag *); 184 static struct freefrag *newfreefrag(struct inode *, ufs_daddr_t, long); 185 static void allocdirect_merge(struct allocdirectlst *, 186 struct allocdirect *, struct allocdirect *); 187 static struct bmsafemap *bmsafemap_lookup(struct buf *); 188 static int newblk_lookup(struct fs *, ufs_daddr_t, int, 189 struct newblk **); 190 static int inodedep_lookup(struct fs *, ino_t, int, struct inodedep **); 191 static int pagedep_lookup(struct inode *, ufs_lbn_t, int, 192 struct pagedep **); 193 static void pause_timer(void *); 194 static int request_cleanup(int, int); 195 static int process_worklist_item(struct mount *, int); 196 static void add_to_worklist(struct worklist *); 197 198 /* 199 * Exported softdep operations. 200 */ 201 static void softdep_disk_io_initiation(struct buf *); 202 static void softdep_disk_write_complete(struct buf *); 203 static void softdep_deallocate_dependencies(struct buf *); 204 static int softdep_fsync(struct vnode *); 205 static int softdep_process_worklist(struct mount *); 206 static void softdep_move_dependencies(struct buf *, struct buf *); 207 static int softdep_count_dependencies(struct buf *bp, int); 208 static int softdep_checkread(struct buf *bp); 209 static int softdep_checkwrite(struct buf *bp); 210 211 static struct bio_ops softdep_bioops = { 212 .io_start = softdep_disk_io_initiation, 213 .io_complete = softdep_disk_write_complete, 214 .io_deallocate = softdep_deallocate_dependencies, 215 .io_fsync = softdep_fsync, 216 .io_sync = softdep_process_worklist, 217 .io_movedeps = softdep_move_dependencies, 218 .io_countdeps = softdep_count_dependencies, 219 .io_checkread = softdep_checkread, 220 .io_checkwrite = softdep_checkwrite 221 }; 222 223 /* 224 * Locking primitives. 225 * 226 * For a uniprocessor, all we need to do is protect against disk 227 * interrupts. For a multiprocessor, this lock would have to be 228 * a mutex. A single mutex is used throughout this file, though 229 * finer grain locking could be used if contention warranted it. 230 * 231 * For a multiprocessor, the sleep call would accept a lock and 232 * release it after the sleep processing was complete. In a uniprocessor 233 * implementation there is no such interlock, so we simple mark 234 * the places where it needs to be done with the `interlocked' form 235 * of the lock calls. Since the uniprocessor sleep already interlocks 236 * the spl, there is nothing that really needs to be done. 237 */ 238 #ifndef /* NOT */ DEBUG 239 static struct lockit { 240 } lk = { 0 }; 241 #define ACQUIRE_LOCK(lk) crit_enter_id("softupdates"); 242 #define FREE_LOCK(lk) crit_exit_id("softupdates"); 243 244 #else /* DEBUG */ 245 #define NOHOLDER ((struct thread *)-1) 246 #define SPECIAL_FLAG ((struct thread *)-2) 247 static struct lockit { 248 int lkt_spl; 249 struct thread *lkt_held; 250 } lk = { 0, NOHOLDER }; 251 static int lockcnt; 252 253 static void acquire_lock(struct lockit *); 254 static void free_lock(struct lockit *); 255 void softdep_panic(char *); 256 257 #define ACQUIRE_LOCK(lk) acquire_lock(lk) 258 #define FREE_LOCK(lk) free_lock(lk) 259 260 static void 261 acquire_lock(struct lockit *lk) 262 { 263 thread_t holder; 264 265 if (lk->lkt_held != NOHOLDER) { 266 holder = lk->lkt_held; 267 FREE_LOCK(lk); 268 if (holder == curthread) 269 panic("softdep_lock: locking against myself"); 270 else 271 panic("softdep_lock: lock held by %p", holder); 272 } 273 crit_enter_id("softupdates"); 274 lk->lkt_held = curthread; 275 lockcnt++; 276 } 277 278 static void 279 free_lock(struct lockit *lk) 280 { 281 282 if (lk->lkt_held == NOHOLDER) 283 panic("softdep_unlock: lock not held"); 284 lk->lkt_held = NOHOLDER; 285 crit_exit_id("softupdates"); 286 } 287 288 /* 289 * Function to release soft updates lock and panic. 290 */ 291 void 292 softdep_panic(char *msg) 293 { 294 295 if (lk.lkt_held != NOHOLDER) 296 FREE_LOCK(&lk); 297 panic(msg); 298 } 299 #endif /* DEBUG */ 300 301 static int interlocked_sleep(struct lockit *, int, void *, int, 302 const char *, int); 303 304 /* 305 * When going to sleep, we must save our SPL so that it does 306 * not get lost if some other process uses the lock while we 307 * are sleeping. We restore it after we have slept. This routine 308 * wraps the interlocking with functions that sleep. The list 309 * below enumerates the available set of operations. 310 */ 311 #define UNKNOWN 0 312 #define SLEEP 1 313 #define LOCKBUF 2 314 315 static int 316 interlocked_sleep(struct lockit *lk, int op, void *ident, int flags, 317 const char *wmesg, int timo) 318 { 319 thread_t holder; 320 int s, retval; 321 322 s = lk->lkt_spl; 323 # ifdef DEBUG 324 if (lk->lkt_held == NOHOLDER) 325 panic("interlocked_sleep: lock not held"); 326 lk->lkt_held = NOHOLDER; 327 # endif /* DEBUG */ 328 switch (op) { 329 case SLEEP: 330 retval = tsleep(ident, flags, wmesg, timo); 331 break; 332 case LOCKBUF: 333 retval = BUF_LOCK((struct buf *)ident, flags); 334 break; 335 default: 336 panic("interlocked_sleep: unknown operation"); 337 } 338 # ifdef DEBUG 339 if (lk->lkt_held != NOHOLDER) { 340 holder = lk->lkt_held; 341 FREE_LOCK(lk); 342 if (holder == curthread) 343 panic("interlocked_sleep: locking against self"); 344 else 345 panic("interlocked_sleep: lock held by %p", holder); 346 } 347 lk->lkt_held = curthread; 348 lockcnt++; 349 # endif /* DEBUG */ 350 lk->lkt_spl = s; 351 return (retval); 352 } 353 354 /* 355 * Place holder for real semaphores. 356 */ 357 struct sema { 358 int value; 359 thread_t holder; 360 char *name; 361 int prio; 362 int timo; 363 }; 364 static void sema_init(struct sema *, char *, int, int); 365 static int sema_get(struct sema *, struct lockit *); 366 static void sema_release(struct sema *); 367 368 static void 369 sema_init(struct sema *semap, char *name, int prio, int timo) 370 { 371 372 semap->holder = NOHOLDER; 373 semap->value = 0; 374 semap->name = name; 375 semap->prio = prio; 376 semap->timo = timo; 377 } 378 379 static int 380 sema_get(struct sema *semap, struct lockit *interlock) 381 { 382 383 if (semap->value++ > 0) { 384 if (interlock != NULL) { 385 interlocked_sleep(interlock, SLEEP, (caddr_t)semap, 386 semap->prio, semap->name, semap->timo); 387 FREE_LOCK(interlock); 388 } else { 389 tsleep((caddr_t)semap, semap->prio, semap->name, 390 semap->timo); 391 } 392 return (0); 393 } 394 semap->holder = curthread; 395 if (interlock != NULL) 396 FREE_LOCK(interlock); 397 return (1); 398 } 399 400 static void 401 sema_release(struct sema *semap) 402 { 403 404 if (semap->value <= 0 || semap->holder != curthread) { 405 if (lk.lkt_held != NOHOLDER) 406 FREE_LOCK(&lk); 407 panic("sema_release: not held"); 408 } 409 if (--semap->value > 0) { 410 semap->value = 0; 411 wakeup(semap); 412 } 413 semap->holder = NOHOLDER; 414 } 415 416 /* 417 * Worklist queue management. 418 * These routines require that the lock be held. 419 */ 420 #ifndef /* NOT */ DEBUG 421 #define WORKLIST_INSERT(head, item) do { \ 422 (item)->wk_state |= ONWORKLIST; \ 423 LIST_INSERT_HEAD(head, item, wk_list); \ 424 } while (0) 425 426 #define WORKLIST_INSERT_BP(bp, item) do { \ 427 (item)->wk_state |= ONWORKLIST; \ 428 (bp)->b_ops = &softdep_bioops; \ 429 LIST_INSERT_HEAD(&(bp)->b_dep, item, wk_list); \ 430 } while (0) 431 432 #define WORKLIST_REMOVE(item) do { \ 433 (item)->wk_state &= ~ONWORKLIST; \ 434 LIST_REMOVE(item, wk_list); \ 435 } while (0) 436 437 #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_BP(bp, item) do { \ 445 (bp)->b_ops = &softdep_bioops; \ 446 worklist_insert(&(bp)->b_dep, item); \ 447 } while (0) 448 449 #define WORKLIST_INSERT(head, item) worklist_insert(head, item) 450 #define WORKLIST_REMOVE(item) worklist_remove(item) 451 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type) 452 453 static void 454 worklist_insert(struct workhead *head, struct worklist *item) 455 { 456 457 if (lk.lkt_held == NOHOLDER) 458 panic("worklist_insert: lock not held"); 459 if (item->wk_state & ONWORKLIST) { 460 FREE_LOCK(&lk); 461 panic("worklist_insert: already on list"); 462 } 463 item->wk_state |= ONWORKLIST; 464 LIST_INSERT_HEAD(head, item, wk_list); 465 } 466 467 static void 468 worklist_remove(struct worklist *item) 469 { 470 471 if (lk.lkt_held == NOHOLDER) 472 panic("worklist_remove: lock not held"); 473 if ((item->wk_state & ONWORKLIST) == 0) { 474 FREE_LOCK(&lk); 475 panic("worklist_remove: not on list"); 476 } 477 item->wk_state &= ~ONWORKLIST; 478 LIST_REMOVE(item, wk_list); 479 } 480 481 static void 482 workitem_free(struct worklist *item, 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 "Maximum soft dependencies before slowdown occurs"); 534 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, 535 "Ticks to delay before allocating during slowdown"); 536 SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0, 537 "Number of worklist cleanups"); 538 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0, 539 "Number of times block limit neared"); 540 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0, 541 "Number of times inode limit neared"); 542 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, 543 "Number of times block slowdown imposed"); 544 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, 545 "Number of times inode slowdown imposed "); 546 SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, 547 "Number of synchronous slowdowns imposed"); 548 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, 549 "Bufs redirtied as indir ptrs not written"); 550 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, 551 "Bufs redirtied as inode bitmap not written"); 552 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, 553 "Bufs redirtied as direct ptrs not written"); 554 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, 555 "Bufs redirtied as dir entry cannot write"); 556 #endif /* DEBUG */ 557 558 /* 559 * Add an item to the end of the work queue. 560 * This routine requires that the lock be held. 561 * This is the only routine that adds items to the list. 562 * The following routine is the only one that removes items 563 * and does so in order from first to last. 564 */ 565 static void 566 add_to_worklist(struct worklist *wk) 567 { 568 static struct worklist *worklist_tail; 569 570 if (wk->wk_state & ONWORKLIST) { 571 if (lk.lkt_held != NOHOLDER) 572 FREE_LOCK(&lk); 573 panic("add_to_worklist: already on list"); 574 } 575 wk->wk_state |= ONWORKLIST; 576 if (LIST_FIRST(&softdep_workitem_pending) == NULL) 577 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list); 578 else 579 LIST_INSERT_AFTER(worklist_tail, wk, wk_list); 580 worklist_tail = wk; 581 num_on_worklist += 1; 582 } 583 584 /* 585 * Process that runs once per second to handle items in the background queue. 586 * 587 * Note that we ensure that everything is done in the order in which they 588 * appear in the queue. The code below depends on this property to ensure 589 * that blocks of a file are freed before the inode itself is freed. This 590 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 591 * until all the old ones have been purged from the dependency lists. 592 * 593 * bioops callback - hold io_token 594 */ 595 static int 596 softdep_process_worklist(struct mount *matchmnt) 597 { 598 thread_t td = curthread; 599 int matchcnt, loopcount; 600 long starttime; 601 602 get_mplock(); 603 604 /* 605 * Record the process identifier of our caller so that we can give 606 * this process preferential treatment in request_cleanup below. 607 */ 608 filesys_syncer = td; 609 matchcnt = 0; 610 611 /* 612 * There is no danger of having multiple processes run this 613 * code, but we have to single-thread it when softdep_flushfiles() 614 * is in operation to get an accurate count of the number of items 615 * related to its mount point that are in the list. 616 */ 617 if (matchmnt == NULL) { 618 if (softdep_worklist_busy < 0) { 619 matchcnt = -1; 620 goto done; 621 } 622 softdep_worklist_busy += 1; 623 } 624 625 /* 626 * If requested, try removing inode or removal dependencies. 627 */ 628 if (req_clear_inodedeps) { 629 clear_inodedeps(td); 630 req_clear_inodedeps -= 1; 631 wakeup_one(&proc_waiting); 632 } 633 if (req_clear_remove) { 634 clear_remove(td); 635 req_clear_remove -= 1; 636 wakeup_one(&proc_waiting); 637 } 638 loopcount = 1; 639 starttime = time_second; 640 while (num_on_worklist > 0) { 641 matchcnt += process_worklist_item(matchmnt, 0); 642 643 /* 644 * If a umount operation wants to run the worklist 645 * accurately, abort. 646 */ 647 if (softdep_worklist_req && matchmnt == NULL) { 648 matchcnt = -1; 649 break; 650 } 651 652 /* 653 * If requested, try removing inode or removal dependencies. 654 */ 655 if (req_clear_inodedeps) { 656 clear_inodedeps(td); 657 req_clear_inodedeps -= 1; 658 wakeup_one(&proc_waiting); 659 } 660 if (req_clear_remove) { 661 clear_remove(td); 662 req_clear_remove -= 1; 663 wakeup_one(&proc_waiting); 664 } 665 /* 666 * We do not generally want to stop for buffer space, but if 667 * we are really being a buffer hog, we will stop and wait. 668 */ 669 if (loopcount++ % 128 == 0) 670 bwillinode(1); 671 /* 672 * Never allow processing to run for more than one 673 * second. Otherwise the other syncer tasks may get 674 * excessively backlogged. 675 */ 676 if (starttime != time_second && matchmnt == NULL) { 677 matchcnt = -1; 678 break; 679 } 680 } 681 if (matchmnt == NULL) { 682 --softdep_worklist_busy; 683 if (softdep_worklist_req && softdep_worklist_busy == 0) 684 wakeup(&softdep_worklist_req); 685 } 686 done: 687 rel_mplock(); 688 return (matchcnt); 689 } 690 691 /* 692 * Process one item on the worklist. 693 */ 694 static int 695 process_worklist_item(struct mount *matchmnt, int flags) 696 { 697 struct worklist *wk; 698 struct dirrem *dirrem; 699 struct fs *matchfs; 700 struct vnode *vp; 701 int matchcnt = 0; 702 703 matchfs = NULL; 704 if (matchmnt != NULL) 705 matchfs = VFSTOUFS(matchmnt)->um_fs; 706 ACQUIRE_LOCK(&lk); 707 /* 708 * Normally we just process each item on the worklist in order. 709 * However, if we are in a situation where we cannot lock any 710 * inodes, we have to skip over any dirrem requests whose 711 * vnodes are resident and locked. 712 */ 713 LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) { 714 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM) 715 break; 716 dirrem = WK_DIRREM(wk); 717 vp = ufs_ihashlookup(VFSTOUFS(dirrem->dm_mnt)->um_dev, 718 dirrem->dm_oldinum); 719 if (vp == NULL || !vn_islocked(vp)) 720 break; 721 } 722 if (wk == 0) { 723 FREE_LOCK(&lk); 724 return (0); 725 } 726 WORKLIST_REMOVE(wk); 727 num_on_worklist -= 1; 728 FREE_LOCK(&lk); 729 switch (wk->wk_type) { 730 731 case D_DIRREM: 732 /* removal of a directory entry */ 733 if (WK_DIRREM(wk)->dm_mnt == matchmnt) 734 matchcnt += 1; 735 handle_workitem_remove(WK_DIRREM(wk)); 736 break; 737 738 case D_FREEBLKS: 739 /* releasing blocks and/or fragments from a file */ 740 if (WK_FREEBLKS(wk)->fb_fs == matchfs) 741 matchcnt += 1; 742 handle_workitem_freeblocks(WK_FREEBLKS(wk)); 743 break; 744 745 case D_FREEFRAG: 746 /* releasing a fragment when replaced as a file grows */ 747 if (WK_FREEFRAG(wk)->ff_fs == matchfs) 748 matchcnt += 1; 749 handle_workitem_freefrag(WK_FREEFRAG(wk)); 750 break; 751 752 case D_FREEFILE: 753 /* releasing an inode when its link count drops to 0 */ 754 if (WK_FREEFILE(wk)->fx_fs == matchfs) 755 matchcnt += 1; 756 handle_workitem_freefile(WK_FREEFILE(wk)); 757 break; 758 759 default: 760 panic("%s_process_worklist: Unknown type %s", 761 "softdep", TYPENAME(wk->wk_type)); 762 /* NOTREACHED */ 763 } 764 return (matchcnt); 765 } 766 767 /* 768 * Move dependencies from one buffer to another. 769 * 770 * bioops callback - hold io_token 771 */ 772 static void 773 softdep_move_dependencies(struct buf *oldbp, struct buf *newbp) 774 { 775 struct worklist *wk, *wktail; 776 777 get_mplock(); 778 if (LIST_FIRST(&newbp->b_dep) != NULL) 779 panic("softdep_move_dependencies: need merge code"); 780 wktail = NULL; 781 ACQUIRE_LOCK(&lk); 782 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 783 LIST_REMOVE(wk, wk_list); 784 if (wktail == NULL) 785 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 786 else 787 LIST_INSERT_AFTER(wktail, wk, wk_list); 788 wktail = wk; 789 newbp->b_ops = &softdep_bioops; 790 } 791 FREE_LOCK(&lk); 792 rel_mplock(); 793 } 794 795 /* 796 * Purge the work list of all items associated with a particular mount point. 797 */ 798 int 799 softdep_flushfiles(struct mount *oldmnt, int flags) 800 { 801 struct vnode *devvp; 802 int error, loopcnt; 803 804 /* 805 * Await our turn to clear out the queue, then serialize access. 806 */ 807 while (softdep_worklist_busy != 0) { 808 softdep_worklist_req += 1; 809 tsleep(&softdep_worklist_req, 0, "softflush", 0); 810 softdep_worklist_req -= 1; 811 } 812 softdep_worklist_busy = -1; 813 814 if ((error = ffs_flushfiles(oldmnt, flags)) != 0) { 815 softdep_worklist_busy = 0; 816 if (softdep_worklist_req) 817 wakeup(&softdep_worklist_req); 818 return (error); 819 } 820 /* 821 * Alternately flush the block device associated with the mount 822 * point and process any dependencies that the flushing 823 * creates. In theory, this loop can happen at most twice, 824 * but we give it a few extra just to be sure. 825 */ 826 devvp = VFSTOUFS(oldmnt)->um_devvp; 827 for (loopcnt = 10; loopcnt > 0; ) { 828 if (softdep_process_worklist(oldmnt) == 0) { 829 loopcnt--; 830 /* 831 * Do another flush in case any vnodes were brought in 832 * as part of the cleanup operations. 833 */ 834 if ((error = ffs_flushfiles(oldmnt, flags)) != 0) 835 break; 836 /* 837 * If we still found nothing to do, we are really done. 838 */ 839 if (softdep_process_worklist(oldmnt) == 0) 840 break; 841 } 842 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 843 error = VOP_FSYNC(devvp, MNT_WAIT, 0); 844 vn_unlock(devvp); 845 if (error) 846 break; 847 } 848 softdep_worklist_busy = 0; 849 if (softdep_worklist_req) 850 wakeup(&softdep_worklist_req); 851 852 /* 853 * If we are unmounting then it is an error to fail. If we 854 * are simply trying to downgrade to read-only, then filesystem 855 * activity can keep us busy forever, so we just fail with EBUSY. 856 */ 857 if (loopcnt == 0) { 858 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 859 panic("softdep_flushfiles: looping"); 860 error = EBUSY; 861 } 862 return (error); 863 } 864 865 /* 866 * Structure hashing. 867 * 868 * There are three types of structures that can be looked up: 869 * 1) pagedep structures identified by mount point, inode number, 870 * and logical block. 871 * 2) inodedep structures identified by mount point and inode number. 872 * 3) newblk structures identified by mount point and 873 * physical block number. 874 * 875 * The "pagedep" and "inodedep" dependency structures are hashed 876 * separately from the file blocks and inodes to which they correspond. 877 * This separation helps when the in-memory copy of an inode or 878 * file block must be replaced. It also obviates the need to access 879 * an inode or file page when simply updating (or de-allocating) 880 * dependency structures. Lookup of newblk structures is needed to 881 * find newly allocated blocks when trying to associate them with 882 * their allocdirect or allocindir structure. 883 * 884 * The lookup routines optionally create and hash a new instance when 885 * an existing entry is not found. 886 */ 887 #define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 888 #define NODELAY 0x0002 /* cannot do background work */ 889 890 /* 891 * Structures and routines associated with pagedep caching. 892 */ 893 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl; 894 u_long pagedep_hash; /* size of hash table - 1 */ 895 #define PAGEDEP_HASH(mp, inum, lbn) \ 896 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \ 897 pagedep_hash]) 898 static struct sema pagedep_in_progress; 899 900 /* 901 * Helper routine for pagedep_lookup() 902 */ 903 static __inline 904 struct pagedep * 905 pagedep_find(struct pagedep_hashhead *pagedephd, ino_t ino, ufs_lbn_t lbn, 906 struct mount *mp) 907 { 908 struct pagedep *pagedep; 909 910 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 911 if (ino == pagedep->pd_ino && 912 lbn == pagedep->pd_lbn && 913 mp == pagedep->pd_mnt) { 914 return (pagedep); 915 } 916 } 917 return(NULL); 918 } 919 920 /* 921 * Look up a pagedep. Return 1 if found, 0 if not found. 922 * If not found, allocate if DEPALLOC flag is passed. 923 * Found or allocated entry is returned in pagedeppp. 924 * This routine must be called with splbio interrupts blocked. 925 */ 926 static int 927 pagedep_lookup(struct inode *ip, ufs_lbn_t lbn, int flags, 928 struct pagedep **pagedeppp) 929 { 930 struct pagedep *pagedep; 931 struct pagedep_hashhead *pagedephd; 932 struct mount *mp; 933 int i; 934 935 #ifdef DEBUG 936 if (lk.lkt_held == NOHOLDER) 937 panic("pagedep_lookup: lock not held"); 938 #endif 939 mp = ITOV(ip)->v_mount; 940 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn); 941 top: 942 *pagedeppp = pagedep_find(pagedephd, ip->i_number, lbn, mp); 943 if (*pagedeppp) 944 return(1); 945 if ((flags & DEPALLOC) == 0) 946 return (0); 947 if (sema_get(&pagedep_in_progress, &lk) == 0) { 948 ACQUIRE_LOCK(&lk); 949 goto top; 950 } 951 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP, 952 M_SOFTDEP_FLAGS | M_ZERO); 953 954 if (pagedep_find(pagedephd, ip->i_number, lbn, mp)) { 955 kprintf("pagedep_lookup: blocking race avoided\n"); 956 ACQUIRE_LOCK(&lk); 957 sema_release(&pagedep_in_progress); 958 kfree(pagedep, M_PAGEDEP); 959 goto top; 960 } 961 962 pagedep->pd_list.wk_type = D_PAGEDEP; 963 pagedep->pd_mnt = mp; 964 pagedep->pd_ino = ip->i_number; 965 pagedep->pd_lbn = lbn; 966 LIST_INIT(&pagedep->pd_dirremhd); 967 LIST_INIT(&pagedep->pd_pendinghd); 968 for (i = 0; i < DAHASHSZ; i++) 969 LIST_INIT(&pagedep->pd_diraddhd[i]); 970 ACQUIRE_LOCK(&lk); 971 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 972 sema_release(&pagedep_in_progress); 973 *pagedeppp = pagedep; 974 return (0); 975 } 976 977 /* 978 * Structures and routines associated with inodedep caching. 979 */ 980 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl; 981 static u_long inodedep_hash; /* size of hash table - 1 */ 982 static long num_inodedep; /* number of inodedep allocated */ 983 #define INODEDEP_HASH(fs, inum) \ 984 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash]) 985 static struct sema inodedep_in_progress; 986 987 /* 988 * Helper routine for inodedep_lookup() 989 */ 990 static __inline 991 struct inodedep * 992 inodedep_find(struct inodedep_hashhead *inodedephd, struct fs *fs, ino_t inum) 993 { 994 struct inodedep *inodedep; 995 996 LIST_FOREACH(inodedep, inodedephd, id_hash) { 997 if (inum == inodedep->id_ino && fs == inodedep->id_fs) 998 return(inodedep); 999 } 1000 return (NULL); 1001 } 1002 1003 /* 1004 * Look up a inodedep. Return 1 if found, 0 if not found. 1005 * If not found, allocate if DEPALLOC flag is passed. 1006 * Found or allocated entry is returned in inodedeppp. 1007 * This routine must be called with splbio interrupts blocked. 1008 */ 1009 static int 1010 inodedep_lookup(struct fs *fs, ino_t inum, int flags, 1011 struct inodedep **inodedeppp) 1012 { 1013 struct inodedep *inodedep; 1014 struct inodedep_hashhead *inodedephd; 1015 int firsttry; 1016 1017 #ifdef DEBUG 1018 if (lk.lkt_held == NOHOLDER) 1019 panic("inodedep_lookup: lock not held"); 1020 #endif 1021 firsttry = 1; 1022 inodedephd = INODEDEP_HASH(fs, inum); 1023 top: 1024 *inodedeppp = inodedep_find(inodedephd, fs, inum); 1025 if (*inodedeppp) 1026 return (1); 1027 if ((flags & DEPALLOC) == 0) 1028 return (0); 1029 /* 1030 * If we are over our limit, try to improve the situation. 1031 */ 1032 if (num_inodedep > max_softdeps && firsttry && 1033 speedup_syncer() == 0 && (flags & NODELAY) == 0 && 1034 request_cleanup(FLUSH_INODES, 1)) { 1035 firsttry = 0; 1036 goto top; 1037 } 1038 if (sema_get(&inodedep_in_progress, &lk) == 0) { 1039 ACQUIRE_LOCK(&lk); 1040 goto top; 1041 } 1042 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep), 1043 M_INODEDEP, M_SOFTDEP_FLAGS | M_ZERO); 1044 if (inodedep_find(inodedephd, fs, inum)) { 1045 kprintf("inodedep_lookup: blocking race avoided\n"); 1046 ACQUIRE_LOCK(&lk); 1047 sema_release(&inodedep_in_progress); 1048 kfree(inodedep, M_INODEDEP); 1049 goto top; 1050 } 1051 inodedep->id_list.wk_type = D_INODEDEP; 1052 inodedep->id_fs = fs; 1053 inodedep->id_ino = inum; 1054 inodedep->id_state = ALLCOMPLETE; 1055 inodedep->id_nlinkdelta = 0; 1056 inodedep->id_savedino = NULL; 1057 inodedep->id_savedsize = -1; 1058 inodedep->id_buf = NULL; 1059 LIST_INIT(&inodedep->id_pendinghd); 1060 LIST_INIT(&inodedep->id_inowait); 1061 LIST_INIT(&inodedep->id_bufwait); 1062 TAILQ_INIT(&inodedep->id_inoupdt); 1063 TAILQ_INIT(&inodedep->id_newinoupdt); 1064 ACQUIRE_LOCK(&lk); 1065 num_inodedep += 1; 1066 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 1067 sema_release(&inodedep_in_progress); 1068 *inodedeppp = inodedep; 1069 return (0); 1070 } 1071 1072 /* 1073 * Structures and routines associated with newblk caching. 1074 */ 1075 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl; 1076 u_long newblk_hash; /* size of hash table - 1 */ 1077 #define NEWBLK_HASH(fs, inum) \ 1078 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash]) 1079 static struct sema newblk_in_progress; 1080 1081 /* 1082 * Helper routine for newblk_lookup() 1083 */ 1084 static __inline 1085 struct newblk * 1086 newblk_find(struct newblk_hashhead *newblkhd, struct fs *fs, 1087 ufs_daddr_t newblkno) 1088 { 1089 struct newblk *newblk; 1090 1091 LIST_FOREACH(newblk, newblkhd, nb_hash) { 1092 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs) 1093 return (newblk); 1094 } 1095 return(NULL); 1096 } 1097 1098 /* 1099 * Look up a newblk. Return 1 if found, 0 if not found. 1100 * If not found, allocate if DEPALLOC flag is passed. 1101 * Found or allocated entry is returned in newblkpp. 1102 */ 1103 static int 1104 newblk_lookup(struct fs *fs, ufs_daddr_t newblkno, int flags, 1105 struct newblk **newblkpp) 1106 { 1107 struct newblk *newblk; 1108 struct newblk_hashhead *newblkhd; 1109 1110 newblkhd = NEWBLK_HASH(fs, newblkno); 1111 top: 1112 *newblkpp = newblk_find(newblkhd, fs, newblkno); 1113 if (*newblkpp) 1114 return(1); 1115 if ((flags & DEPALLOC) == 0) 1116 return (0); 1117 if (sema_get(&newblk_in_progress, 0) == 0) 1118 goto top; 1119 MALLOC(newblk, struct newblk *, sizeof(struct newblk), 1120 M_NEWBLK, M_SOFTDEP_FLAGS | M_ZERO); 1121 1122 if (newblk_find(newblkhd, fs, newblkno)) { 1123 kprintf("newblk_lookup: blocking race avoided\n"); 1124 sema_release(&pagedep_in_progress); 1125 kfree(newblk, M_NEWBLK); 1126 goto top; 1127 } 1128 newblk->nb_state = 0; 1129 newblk->nb_fs = fs; 1130 newblk->nb_newblkno = newblkno; 1131 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 1132 sema_release(&newblk_in_progress); 1133 *newblkpp = newblk; 1134 return (0); 1135 } 1136 1137 /* 1138 * Executed during filesystem system initialization before 1139 * mounting any filesystems. 1140 */ 1141 void 1142 softdep_initialize(void) 1143 { 1144 callout_init(&handle); 1145 1146 LIST_INIT(&mkdirlisthd); 1147 LIST_INIT(&softdep_workitem_pending); 1148 max_softdeps = min(desiredvnodes * 8, 1149 M_INODEDEP->ks_limit / (2 * sizeof(struct inodedep))); 1150 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 1151 &pagedep_hash); 1152 sema_init(&pagedep_in_progress, "pagedep", 0, 0); 1153 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash); 1154 sema_init(&inodedep_in_progress, "inodedep", 0, 0); 1155 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash); 1156 sema_init(&newblk_in_progress, "newblk", 0, 0); 1157 add_bio_ops(&softdep_bioops); 1158 } 1159 1160 /* 1161 * Called at mount time to notify the dependency code that a 1162 * filesystem wishes to use it. 1163 */ 1164 int 1165 softdep_mount(struct vnode *devvp, struct mount *mp, struct fs *fs) 1166 { 1167 struct csum cstotal; 1168 struct cg *cgp; 1169 struct buf *bp; 1170 int error, cyl; 1171 1172 mp->mnt_flag &= ~MNT_ASYNC; 1173 mp->mnt_flag |= MNT_SOFTDEP; 1174 mp->mnt_bioops = &softdep_bioops; 1175 /* 1176 * When doing soft updates, the counters in the 1177 * superblock may have gotten out of sync, so we have 1178 * to scan the cylinder groups and recalculate them. 1179 */ 1180 if (fs->fs_clean != 0) 1181 return (0); 1182 bzero(&cstotal, sizeof cstotal); 1183 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 1184 if ((error = bread(devvp, fsbtodoff(fs, cgtod(fs, cyl)), 1185 fs->fs_cgsize, &bp)) != 0) { 1186 brelse(bp); 1187 return (error); 1188 } 1189 cgp = (struct cg *)bp->b_data; 1190 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 1191 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 1192 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 1193 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 1194 fs->fs_cs(fs, cyl) = cgp->cg_cs; 1195 brelse(bp); 1196 } 1197 #ifdef DEBUG 1198 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 1199 kprintf("ffs_mountfs: superblock updated for soft updates\n"); 1200 #endif 1201 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 1202 return (0); 1203 } 1204 1205 /* 1206 * Protecting the freemaps (or bitmaps). 1207 * 1208 * To eliminate the need to execute fsck before mounting a filesystem 1209 * after a power failure, one must (conservatively) guarantee that the 1210 * on-disk copy of the bitmaps never indicate that a live inode or block is 1211 * free. So, when a block or inode is allocated, the bitmap should be 1212 * updated (on disk) before any new pointers. When a block or inode is 1213 * freed, the bitmap should not be updated until all pointers have been 1214 * reset. The latter dependency is handled by the delayed de-allocation 1215 * approach described below for block and inode de-allocation. The former 1216 * dependency is handled by calling the following procedure when a block or 1217 * inode is allocated. When an inode is allocated an "inodedep" is created 1218 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 1219 * Each "inodedep" is also inserted into the hash indexing structure so 1220 * that any additional link additions can be made dependent on the inode 1221 * allocation. 1222 * 1223 * The ufs filesystem maintains a number of free block counts (e.g., per 1224 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 1225 * in addition to the bitmaps. These counts are used to improve efficiency 1226 * during allocation and therefore must be consistent with the bitmaps. 1227 * There is no convenient way to guarantee post-crash consistency of these 1228 * counts with simple update ordering, for two main reasons: (1) The counts 1229 * and bitmaps for a single cylinder group block are not in the same disk 1230 * sector. If a disk write is interrupted (e.g., by power failure), one may 1231 * be written and the other not. (2) Some of the counts are located in the 1232 * superblock rather than the cylinder group block. So, we focus our soft 1233 * updates implementation on protecting the bitmaps. When mounting a 1234 * filesystem, we recompute the auxiliary counts from the bitmaps. 1235 */ 1236 1237 /* 1238 * Called just after updating the cylinder group block to allocate an inode. 1239 * 1240 * Parameters: 1241 * bp: buffer for cylgroup block with inode map 1242 * ip: inode related to allocation 1243 * newinum: new inode number being allocated 1244 */ 1245 void 1246 softdep_setup_inomapdep(struct buf *bp, struct inode *ip, ino_t newinum) 1247 { 1248 struct inodedep *inodedep; 1249 struct bmsafemap *bmsafemap; 1250 1251 /* 1252 * Create a dependency for the newly allocated inode. 1253 * Panic if it already exists as something is seriously wrong. 1254 * Otherwise add it to the dependency list for the buffer holding 1255 * the cylinder group map from which it was allocated. 1256 */ 1257 ACQUIRE_LOCK(&lk); 1258 if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) { 1259 FREE_LOCK(&lk); 1260 panic("softdep_setup_inomapdep: found inode"); 1261 } 1262 inodedep->id_buf = bp; 1263 inodedep->id_state &= ~DEPCOMPLETE; 1264 bmsafemap = bmsafemap_lookup(bp); 1265 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 1266 FREE_LOCK(&lk); 1267 } 1268 1269 /* 1270 * Called just after updating the cylinder group block to 1271 * allocate block or fragment. 1272 * 1273 * Parameters: 1274 * bp: buffer for cylgroup block with block map 1275 * fs: filesystem doing allocation 1276 * newblkno: number of newly allocated block 1277 */ 1278 void 1279 softdep_setup_blkmapdep(struct buf *bp, struct fs *fs, 1280 ufs_daddr_t newblkno) 1281 { 1282 struct newblk *newblk; 1283 struct bmsafemap *bmsafemap; 1284 1285 /* 1286 * Create a dependency for the newly allocated block. 1287 * Add it to the dependency list for the buffer holding 1288 * the cylinder group map from which it was allocated. 1289 */ 1290 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0) 1291 panic("softdep_setup_blkmapdep: found block"); 1292 ACQUIRE_LOCK(&lk); 1293 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp); 1294 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 1295 FREE_LOCK(&lk); 1296 } 1297 1298 /* 1299 * Find the bmsafemap associated with a cylinder group buffer. 1300 * If none exists, create one. The buffer must be locked when 1301 * this routine is called and this routine must be called with 1302 * splbio interrupts blocked. 1303 */ 1304 static struct bmsafemap * 1305 bmsafemap_lookup(struct buf *bp) 1306 { 1307 struct bmsafemap *bmsafemap; 1308 struct worklist *wk; 1309 1310 #ifdef DEBUG 1311 if (lk.lkt_held == NOHOLDER) 1312 panic("bmsafemap_lookup: lock not held"); 1313 #endif 1314 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 1315 if (wk->wk_type == D_BMSAFEMAP) 1316 return (WK_BMSAFEMAP(wk)); 1317 } 1318 FREE_LOCK(&lk); 1319 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap), 1320 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 1321 bmsafemap->sm_list.wk_type = D_BMSAFEMAP; 1322 bmsafemap->sm_list.wk_state = 0; 1323 bmsafemap->sm_buf = bp; 1324 LIST_INIT(&bmsafemap->sm_allocdirecthd); 1325 LIST_INIT(&bmsafemap->sm_allocindirhd); 1326 LIST_INIT(&bmsafemap->sm_inodedephd); 1327 LIST_INIT(&bmsafemap->sm_newblkhd); 1328 ACQUIRE_LOCK(&lk); 1329 WORKLIST_INSERT_BP(bp, &bmsafemap->sm_list); 1330 return (bmsafemap); 1331 } 1332 1333 /* 1334 * Direct block allocation dependencies. 1335 * 1336 * When a new block is allocated, the corresponding disk locations must be 1337 * initialized (with zeros or new data) before the on-disk inode points to 1338 * them. Also, the freemap from which the block was allocated must be 1339 * updated (on disk) before the inode's pointer. These two dependencies are 1340 * independent of each other and are needed for all file blocks and indirect 1341 * blocks that are pointed to directly by the inode. Just before the 1342 * "in-core" version of the inode is updated with a newly allocated block 1343 * number, a procedure (below) is called to setup allocation dependency 1344 * structures. These structures are removed when the corresponding 1345 * dependencies are satisfied or when the block allocation becomes obsolete 1346 * (i.e., the file is deleted, the block is de-allocated, or the block is a 1347 * fragment that gets upgraded). All of these cases are handled in 1348 * procedures described later. 1349 * 1350 * When a file extension causes a fragment to be upgraded, either to a larger 1351 * fragment or to a full block, the on-disk location may change (if the 1352 * previous fragment could not simply be extended). In this case, the old 1353 * fragment must be de-allocated, but not until after the inode's pointer has 1354 * been updated. In most cases, this is handled by later procedures, which 1355 * will construct a "freefrag" structure to be added to the workitem queue 1356 * when the inode update is complete (or obsolete). The main exception to 1357 * this is when an allocation occurs while a pending allocation dependency 1358 * (for the same block pointer) remains. This case is handled in the main 1359 * allocation dependency setup procedure by immediately freeing the 1360 * unreferenced fragments. 1361 * 1362 * Parameters: 1363 * ip: inode to which block is being added 1364 * lbn: block pointer within inode 1365 * newblkno: disk block number being added 1366 * oldblkno: previous block number, 0 unless frag 1367 * newsize: size of new block 1368 * oldsize: size of new block 1369 * bp: bp for allocated block 1370 */ 1371 void 1372 softdep_setup_allocdirect(struct inode *ip, ufs_lbn_t lbn, ufs_daddr_t newblkno, 1373 ufs_daddr_t oldblkno, long newsize, long oldsize, 1374 struct buf *bp) 1375 { 1376 struct allocdirect *adp, *oldadp; 1377 struct allocdirectlst *adphead; 1378 struct bmsafemap *bmsafemap; 1379 struct inodedep *inodedep; 1380 struct pagedep *pagedep; 1381 struct newblk *newblk; 1382 1383 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect), 1384 M_ALLOCDIRECT, M_SOFTDEP_FLAGS | M_ZERO); 1385 adp->ad_list.wk_type = D_ALLOCDIRECT; 1386 adp->ad_lbn = lbn; 1387 adp->ad_newblkno = newblkno; 1388 adp->ad_oldblkno = oldblkno; 1389 adp->ad_newsize = newsize; 1390 adp->ad_oldsize = oldsize; 1391 adp->ad_state = ATTACHED; 1392 if (newblkno == oldblkno) 1393 adp->ad_freefrag = NULL; 1394 else 1395 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1396 1397 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1398 panic("softdep_setup_allocdirect: lost block"); 1399 1400 ACQUIRE_LOCK(&lk); 1401 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep); 1402 adp->ad_inodedep = inodedep; 1403 1404 if (newblk->nb_state == DEPCOMPLETE) { 1405 adp->ad_state |= DEPCOMPLETE; 1406 adp->ad_buf = NULL; 1407 } else { 1408 bmsafemap = newblk->nb_bmsafemap; 1409 adp->ad_buf = bmsafemap->sm_buf; 1410 LIST_REMOVE(newblk, nb_deps); 1411 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1412 } 1413 LIST_REMOVE(newblk, nb_hash); 1414 FREE(newblk, M_NEWBLK); 1415 1416 WORKLIST_INSERT_BP(bp, &adp->ad_list); 1417 if (lbn >= NDADDR) { 1418 /* allocating an indirect block */ 1419 if (oldblkno != 0) { 1420 FREE_LOCK(&lk); 1421 panic("softdep_setup_allocdirect: non-zero indir"); 1422 } 1423 } else { 1424 /* 1425 * Allocating a direct block. 1426 * 1427 * If we are allocating a directory block, then we must 1428 * allocate an associated pagedep to track additions and 1429 * deletions. 1430 */ 1431 if ((ip->i_mode & IFMT) == IFDIR && 1432 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) { 1433 WORKLIST_INSERT_BP(bp, &pagedep->pd_list); 1434 } 1435 } 1436 /* 1437 * The list of allocdirects must be kept in sorted and ascending 1438 * order so that the rollback routines can quickly determine the 1439 * first uncommitted block (the size of the file stored on disk 1440 * ends at the end of the lowest committed fragment, or if there 1441 * are no fragments, at the end of the highest committed block). 1442 * Since files generally grow, the typical case is that the new 1443 * block is to be added at the end of the list. We speed this 1444 * special case by checking against the last allocdirect in the 1445 * list before laboriously traversing the list looking for the 1446 * insertion point. 1447 */ 1448 adphead = &inodedep->id_newinoupdt; 1449 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1450 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1451 /* insert at end of list */ 1452 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1453 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1454 allocdirect_merge(adphead, adp, oldadp); 1455 FREE_LOCK(&lk); 1456 return; 1457 } 1458 TAILQ_FOREACH(oldadp, adphead, ad_next) { 1459 if (oldadp->ad_lbn >= lbn) 1460 break; 1461 } 1462 if (oldadp == NULL) { 1463 FREE_LOCK(&lk); 1464 panic("softdep_setup_allocdirect: lost entry"); 1465 } 1466 /* insert in middle of list */ 1467 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1468 if (oldadp->ad_lbn == lbn) 1469 allocdirect_merge(adphead, adp, oldadp); 1470 FREE_LOCK(&lk); 1471 } 1472 1473 /* 1474 * Replace an old allocdirect dependency with a newer one. 1475 * This routine must be called with splbio interrupts blocked. 1476 * 1477 * Parameters: 1478 * adphead: head of list holding allocdirects 1479 * newadp: allocdirect being added 1480 * oldadp: existing allocdirect being checked 1481 */ 1482 static void 1483 allocdirect_merge(struct allocdirectlst *adphead, 1484 struct allocdirect *newadp, 1485 struct allocdirect *oldadp) 1486 { 1487 struct freefrag *freefrag; 1488 1489 #ifdef DEBUG 1490 if (lk.lkt_held == NOHOLDER) 1491 panic("allocdirect_merge: lock not held"); 1492 #endif 1493 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 1494 newadp->ad_oldsize != oldadp->ad_newsize || 1495 newadp->ad_lbn >= NDADDR) { 1496 FREE_LOCK(&lk); 1497 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d", 1498 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn, 1499 NDADDR); 1500 } 1501 newadp->ad_oldblkno = oldadp->ad_oldblkno; 1502 newadp->ad_oldsize = oldadp->ad_oldsize; 1503 /* 1504 * If the old dependency had a fragment to free or had never 1505 * previously had a block allocated, then the new dependency 1506 * can immediately post its freefrag and adopt the old freefrag. 1507 * This action is done by swapping the freefrag dependencies. 1508 * The new dependency gains the old one's freefrag, and the 1509 * old one gets the new one and then immediately puts it on 1510 * the worklist when it is freed by free_allocdirect. It is 1511 * not possible to do this swap when the old dependency had a 1512 * non-zero size but no previous fragment to free. This condition 1513 * arises when the new block is an extension of the old block. 1514 * Here, the first part of the fragment allocated to the new 1515 * dependency is part of the block currently claimed on disk by 1516 * the old dependency, so cannot legitimately be freed until the 1517 * conditions for the new dependency are fulfilled. 1518 */ 1519 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 1520 freefrag = newadp->ad_freefrag; 1521 newadp->ad_freefrag = oldadp->ad_freefrag; 1522 oldadp->ad_freefrag = freefrag; 1523 } 1524 free_allocdirect(adphead, oldadp, 0); 1525 } 1526 1527 /* 1528 * Allocate a new freefrag structure if needed. 1529 */ 1530 static struct freefrag * 1531 newfreefrag(struct inode *ip, ufs_daddr_t blkno, long size) 1532 { 1533 struct freefrag *freefrag; 1534 struct fs *fs; 1535 1536 if (blkno == 0) 1537 return (NULL); 1538 fs = ip->i_fs; 1539 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 1540 panic("newfreefrag: frag size"); 1541 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag), 1542 M_FREEFRAG, M_SOFTDEP_FLAGS); 1543 freefrag->ff_list.wk_type = D_FREEFRAG; 1544 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */ 1545 freefrag->ff_inum = ip->i_number; 1546 freefrag->ff_fs = fs; 1547 freefrag->ff_devvp = ip->i_devvp; 1548 freefrag->ff_blkno = blkno; 1549 freefrag->ff_fragsize = size; 1550 return (freefrag); 1551 } 1552 1553 /* 1554 * This workitem de-allocates fragments that were replaced during 1555 * file block allocation. 1556 */ 1557 static void 1558 handle_workitem_freefrag(struct freefrag *freefrag) 1559 { 1560 struct inode tip; 1561 1562 tip.i_fs = freefrag->ff_fs; 1563 tip.i_devvp = freefrag->ff_devvp; 1564 tip.i_dev = freefrag->ff_devvp->v_rdev; 1565 tip.i_number = freefrag->ff_inum; 1566 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */ 1567 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize); 1568 FREE(freefrag, M_FREEFRAG); 1569 } 1570 1571 /* 1572 * Indirect block allocation dependencies. 1573 * 1574 * The same dependencies that exist for a direct block also exist when 1575 * a new block is allocated and pointed to by an entry in a block of 1576 * indirect pointers. The undo/redo states described above are also 1577 * used here. Because an indirect block contains many pointers that 1578 * may have dependencies, a second copy of the entire in-memory indirect 1579 * block is kept. The buffer cache copy is always completely up-to-date. 1580 * The second copy, which is used only as a source for disk writes, 1581 * contains only the safe pointers (i.e., those that have no remaining 1582 * update dependencies). The second copy is freed when all pointers 1583 * are safe. The cache is not allowed to replace indirect blocks with 1584 * pending update dependencies. If a buffer containing an indirect 1585 * block with dependencies is written, these routines will mark it 1586 * dirty again. It can only be successfully written once all the 1587 * dependencies are removed. The ffs_fsync routine in conjunction with 1588 * softdep_sync_metadata work together to get all the dependencies 1589 * removed so that a file can be successfully written to disk. Three 1590 * procedures are used when setting up indirect block pointer 1591 * dependencies. The division is necessary because of the organization 1592 * of the "balloc" routine and because of the distinction between file 1593 * pages and file metadata blocks. 1594 */ 1595 1596 /* 1597 * Allocate a new allocindir structure. 1598 * 1599 * Parameters: 1600 * ip: inode for file being extended 1601 * ptrno: offset of pointer in indirect block 1602 * newblkno: disk block number being added 1603 * oldblkno: previous block number, 0 if none 1604 */ 1605 static struct allocindir * 1606 newallocindir(struct inode *ip, int ptrno, ufs_daddr_t newblkno, 1607 ufs_daddr_t oldblkno) 1608 { 1609 struct allocindir *aip; 1610 1611 MALLOC(aip, struct allocindir *, sizeof(struct allocindir), 1612 M_ALLOCINDIR, M_SOFTDEP_FLAGS | M_ZERO); 1613 aip->ai_list.wk_type = D_ALLOCINDIR; 1614 aip->ai_state = ATTACHED; 1615 aip->ai_offset = ptrno; 1616 aip->ai_newblkno = newblkno; 1617 aip->ai_oldblkno = oldblkno; 1618 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize); 1619 return (aip); 1620 } 1621 1622 /* 1623 * Called just before setting an indirect block pointer 1624 * to a newly allocated file page. 1625 * 1626 * Parameters: 1627 * ip: inode for file being extended 1628 * lbn: allocated block number within file 1629 * bp: buffer with indirect blk referencing page 1630 * ptrno: offset of pointer in indirect block 1631 * newblkno: disk block number being added 1632 * oldblkno: previous block number, 0 if none 1633 * nbp: buffer holding allocated page 1634 */ 1635 void 1636 softdep_setup_allocindir_page(struct inode *ip, ufs_lbn_t lbn, 1637 struct buf *bp, int ptrno, 1638 ufs_daddr_t newblkno, ufs_daddr_t oldblkno, 1639 struct buf *nbp) 1640 { 1641 struct allocindir *aip; 1642 struct pagedep *pagedep; 1643 1644 aip = newallocindir(ip, ptrno, newblkno, oldblkno); 1645 ACQUIRE_LOCK(&lk); 1646 /* 1647 * If we are allocating a directory page, then we must 1648 * allocate an associated pagedep to track additions and 1649 * deletions. 1650 */ 1651 if ((ip->i_mode & IFMT) == IFDIR && 1652 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1653 WORKLIST_INSERT_BP(nbp, &pagedep->pd_list); 1654 WORKLIST_INSERT_BP(nbp, &aip->ai_list); 1655 FREE_LOCK(&lk); 1656 setup_allocindir_phase2(bp, ip, aip); 1657 } 1658 1659 /* 1660 * Called just before setting an indirect block pointer to a 1661 * newly allocated indirect block. 1662 * Parameters: 1663 * nbp: newly allocated indirect block 1664 * ip: inode for file being extended 1665 * bp: indirect block referencing allocated block 1666 * ptrno: offset of pointer in indirect block 1667 * newblkno: disk block number being added 1668 */ 1669 void 1670 softdep_setup_allocindir_meta(struct buf *nbp, struct inode *ip, 1671 struct buf *bp, int ptrno, 1672 ufs_daddr_t newblkno) 1673 { 1674 struct allocindir *aip; 1675 1676 aip = newallocindir(ip, ptrno, newblkno, 0); 1677 ACQUIRE_LOCK(&lk); 1678 WORKLIST_INSERT_BP(nbp, &aip->ai_list); 1679 FREE_LOCK(&lk); 1680 setup_allocindir_phase2(bp, ip, aip); 1681 } 1682 1683 /* 1684 * Called to finish the allocation of the "aip" allocated 1685 * by one of the two routines above. 1686 * 1687 * Parameters: 1688 * bp: in-memory copy of the indirect block 1689 * ip: inode for file being extended 1690 * aip: allocindir allocated by the above routines 1691 */ 1692 static void 1693 setup_allocindir_phase2(struct buf *bp, struct inode *ip, 1694 struct allocindir *aip) 1695 { 1696 struct worklist *wk; 1697 struct indirdep *indirdep, *newindirdep; 1698 struct bmsafemap *bmsafemap; 1699 struct allocindir *oldaip; 1700 struct freefrag *freefrag; 1701 struct newblk *newblk; 1702 1703 if (bp->b_loffset >= 0) 1704 panic("setup_allocindir_phase2: not indir blk"); 1705 for (indirdep = NULL, newindirdep = NULL; ; ) { 1706 ACQUIRE_LOCK(&lk); 1707 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 1708 if (wk->wk_type != D_INDIRDEP) 1709 continue; 1710 indirdep = WK_INDIRDEP(wk); 1711 break; 1712 } 1713 if (indirdep == NULL && newindirdep) { 1714 indirdep = newindirdep; 1715 WORKLIST_INSERT_BP(bp, &indirdep->ir_list); 1716 newindirdep = NULL; 1717 } 1718 FREE_LOCK(&lk); 1719 if (indirdep) { 1720 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0, 1721 &newblk) == 0) 1722 panic("setup_allocindir: lost block"); 1723 ACQUIRE_LOCK(&lk); 1724 if (newblk->nb_state == DEPCOMPLETE) { 1725 aip->ai_state |= DEPCOMPLETE; 1726 aip->ai_buf = NULL; 1727 } else { 1728 bmsafemap = newblk->nb_bmsafemap; 1729 aip->ai_buf = bmsafemap->sm_buf; 1730 LIST_REMOVE(newblk, nb_deps); 1731 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd, 1732 aip, ai_deps); 1733 } 1734 LIST_REMOVE(newblk, nb_hash); 1735 FREE(newblk, M_NEWBLK); 1736 aip->ai_indirdep = indirdep; 1737 /* 1738 * Check to see if there is an existing dependency 1739 * for this block. If there is, merge the old 1740 * dependency into the new one. 1741 */ 1742 if (aip->ai_oldblkno == 0) 1743 oldaip = NULL; 1744 else 1745 1746 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) 1747 if (oldaip->ai_offset == aip->ai_offset) 1748 break; 1749 if (oldaip != NULL) { 1750 if (oldaip->ai_newblkno != aip->ai_oldblkno) { 1751 FREE_LOCK(&lk); 1752 panic("setup_allocindir_phase2: blkno"); 1753 } 1754 aip->ai_oldblkno = oldaip->ai_oldblkno; 1755 freefrag = oldaip->ai_freefrag; 1756 oldaip->ai_freefrag = aip->ai_freefrag; 1757 aip->ai_freefrag = freefrag; 1758 free_allocindir(oldaip, NULL); 1759 } 1760 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 1761 ((ufs_daddr_t *)indirdep->ir_savebp->b_data) 1762 [aip->ai_offset] = aip->ai_oldblkno; 1763 FREE_LOCK(&lk); 1764 } 1765 if (newindirdep) { 1766 /* 1767 * Avoid any possibility of data corruption by 1768 * ensuring that our old version is thrown away. 1769 */ 1770 newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 1771 brelse(newindirdep->ir_savebp); 1772 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP); 1773 } 1774 if (indirdep) 1775 break; 1776 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep), 1777 M_INDIRDEP, M_SOFTDEP_FLAGS); 1778 newindirdep->ir_list.wk_type = D_INDIRDEP; 1779 newindirdep->ir_state = ATTACHED; 1780 LIST_INIT(&newindirdep->ir_deplisthd); 1781 LIST_INIT(&newindirdep->ir_donehd); 1782 if (bp->b_bio2.bio_offset == NOOFFSET) { 1783 VOP_BMAP(bp->b_vp, bp->b_bio1.bio_offset, 1784 &bp->b_bio2.bio_offset, NULL, NULL, 1785 BUF_CMD_WRITE); 1786 } 1787 KKASSERT(bp->b_bio2.bio_offset != NOOFFSET); 1788 newindirdep->ir_savebp = getblk(ip->i_devvp, 1789 bp->b_bio2.bio_offset, 1790 bp->b_bcount, 0, 0); 1791 BUF_KERNPROC(newindirdep->ir_savebp); 1792 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 1793 } 1794 } 1795 1796 /* 1797 * Block de-allocation dependencies. 1798 * 1799 * When blocks are de-allocated, the on-disk pointers must be nullified before 1800 * the blocks are made available for use by other files. (The true 1801 * requirement is that old pointers must be nullified before new on-disk 1802 * pointers are set. We chose this slightly more stringent requirement to 1803 * reduce complexity.) Our implementation handles this dependency by updating 1804 * the inode (or indirect block) appropriately but delaying the actual block 1805 * de-allocation (i.e., freemap and free space count manipulation) until 1806 * after the updated versions reach stable storage. After the disk is 1807 * updated, the blocks can be safely de-allocated whenever it is convenient. 1808 * This implementation handles only the common case of reducing a file's 1809 * length to zero. Other cases are handled by the conventional synchronous 1810 * write approach. 1811 * 1812 * The ffs implementation with which we worked double-checks 1813 * the state of the block pointers and file size as it reduces 1814 * a file's length. Some of this code is replicated here in our 1815 * soft updates implementation. The freeblks->fb_chkcnt field is 1816 * used to transfer a part of this information to the procedure 1817 * that eventually de-allocates the blocks. 1818 * 1819 * This routine should be called from the routine that shortens 1820 * a file's length, before the inode's size or block pointers 1821 * are modified. It will save the block pointer information for 1822 * later release and zero the inode so that the calling routine 1823 * can release it. 1824 */ 1825 struct softdep_setup_freeblocks_info { 1826 struct fs *fs; 1827 struct inode *ip; 1828 }; 1829 1830 static int softdep_setup_freeblocks_bp(struct buf *bp, void *data); 1831 1832 /* 1833 * Parameters: 1834 * ip: The inode whose length is to be reduced 1835 * length: The new length for the file 1836 */ 1837 void 1838 softdep_setup_freeblocks(struct inode *ip, off_t length) 1839 { 1840 struct softdep_setup_freeblocks_info info; 1841 struct freeblks *freeblks; 1842 struct inodedep *inodedep; 1843 struct allocdirect *adp; 1844 struct vnode *vp; 1845 struct buf *bp; 1846 struct fs *fs; 1847 int i, error, delay; 1848 int count; 1849 1850 fs = ip->i_fs; 1851 if (length != 0) 1852 panic("softde_setup_freeblocks: non-zero length"); 1853 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks), 1854 M_FREEBLKS, M_SOFTDEP_FLAGS | M_ZERO); 1855 freeblks->fb_list.wk_type = D_FREEBLKS; 1856 freeblks->fb_state = ATTACHED; 1857 freeblks->fb_uid = ip->i_uid; 1858 freeblks->fb_previousinum = ip->i_number; 1859 freeblks->fb_devvp = ip->i_devvp; 1860 freeblks->fb_fs = fs; 1861 freeblks->fb_oldsize = ip->i_size; 1862 freeblks->fb_newsize = length; 1863 freeblks->fb_chkcnt = ip->i_blocks; 1864 for (i = 0; i < NDADDR; i++) { 1865 freeblks->fb_dblks[i] = ip->i_db[i]; 1866 ip->i_db[i] = 0; 1867 } 1868 for (i = 0; i < NIADDR; i++) { 1869 freeblks->fb_iblks[i] = ip->i_ib[i]; 1870 ip->i_ib[i] = 0; 1871 } 1872 ip->i_blocks = 0; 1873 ip->i_size = 0; 1874 /* 1875 * Push the zero'ed inode to to its disk buffer so that we are free 1876 * to delete its dependencies below. Once the dependencies are gone 1877 * the buffer can be safely released. 1878 */ 1879 if ((error = bread(ip->i_devvp, 1880 fsbtodoff(fs, ino_to_fsba(fs, ip->i_number)), 1881 (int)fs->fs_bsize, &bp)) != 0) 1882 softdep_error("softdep_setup_freeblocks", error); 1883 *((struct ufs1_dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) = 1884 ip->i_din; 1885 /* 1886 * Find and eliminate any inode dependencies. 1887 */ 1888 ACQUIRE_LOCK(&lk); 1889 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep); 1890 if ((inodedep->id_state & IOSTARTED) != 0) { 1891 FREE_LOCK(&lk); 1892 panic("softdep_setup_freeblocks: inode busy"); 1893 } 1894 /* 1895 * Add the freeblks structure to the list of operations that 1896 * must await the zero'ed inode being written to disk. If we 1897 * still have a bitmap dependency (delay == 0), then the inode 1898 * has never been written to disk, so we can process the 1899 * freeblks below once we have deleted the dependencies. 1900 */ 1901 delay = (inodedep->id_state & DEPCOMPLETE); 1902 if (delay) 1903 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list); 1904 /* 1905 * Because the file length has been truncated to zero, any 1906 * pending block allocation dependency structures associated 1907 * with this inode are obsolete and can simply be de-allocated. 1908 * We must first merge the two dependency lists to get rid of 1909 * any duplicate freefrag structures, then purge the merged list. 1910 */ 1911 merge_inode_lists(inodedep); 1912 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 1913 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 1914 FREE_LOCK(&lk); 1915 bdwrite(bp); 1916 /* 1917 * We must wait for any I/O in progress to finish so that 1918 * all potential buffers on the dirty list will be visible. 1919 * Once they are all there, walk the list and get rid of 1920 * any dependencies. 1921 */ 1922 vp = ITOV(ip); 1923 ACQUIRE_LOCK(&lk); 1924 drain_output(vp, 1); 1925 1926 info.fs = fs; 1927 info.ip = ip; 1928 lwkt_gettoken(&vp->v_token); 1929 do { 1930 count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 1931 softdep_setup_freeblocks_bp, &info); 1932 } while (count != 0); 1933 lwkt_reltoken(&vp->v_token); 1934 1935 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0) 1936 (void)free_inodedep(inodedep); 1937 1938 if (delay) { 1939 freeblks->fb_state |= DEPCOMPLETE; 1940 /* 1941 * If the inode with zeroed block pointers is now on disk 1942 * we can start freeing blocks. Add freeblks to the worklist 1943 * instead of calling handle_workitem_freeblocks directly as 1944 * it is more likely that additional IO is needed to complete 1945 * the request here than in the !delay case. 1946 */ 1947 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 1948 add_to_worklist(&freeblks->fb_list); 1949 } 1950 1951 FREE_LOCK(&lk); 1952 /* 1953 * If the inode has never been written to disk (delay == 0), 1954 * then we can process the freeblks now that we have deleted 1955 * the dependencies. 1956 */ 1957 if (!delay) 1958 handle_workitem_freeblocks(freeblks); 1959 } 1960 1961 static int 1962 softdep_setup_freeblocks_bp(struct buf *bp, void *data) 1963 { 1964 struct softdep_setup_freeblocks_info *info = data; 1965 struct inodedep *inodedep; 1966 1967 if (getdirtybuf(&bp, MNT_WAIT) == 0) { 1968 kprintf("softdep_setup_freeblocks_bp(1): caught bp %p going away\n", bp); 1969 return(-1); 1970 } 1971 if (bp->b_vp != ITOV(info->ip) || (bp->b_flags & B_DELWRI) == 0) { 1972 kprintf("softdep_setup_freeblocks_bp(2): caught bp %p going away\n", bp); 1973 BUF_UNLOCK(bp); 1974 return(-1); 1975 } 1976 (void) inodedep_lookup(info->fs, info->ip->i_number, 0, &inodedep); 1977 deallocate_dependencies(bp, inodedep); 1978 bp->b_flags |= B_INVAL | B_NOCACHE; 1979 FREE_LOCK(&lk); 1980 brelse(bp); 1981 ACQUIRE_LOCK(&lk); 1982 return(1); 1983 } 1984 1985 /* 1986 * Reclaim any dependency structures from a buffer that is about to 1987 * be reallocated to a new vnode. The buffer must be locked, thus, 1988 * no I/O completion operations can occur while we are manipulating 1989 * its associated dependencies. The mutex is held so that other I/O's 1990 * associated with related dependencies do not occur. 1991 */ 1992 static void 1993 deallocate_dependencies(struct buf *bp, struct inodedep *inodedep) 1994 { 1995 struct worklist *wk; 1996 struct indirdep *indirdep; 1997 struct allocindir *aip; 1998 struct pagedep *pagedep; 1999 struct dirrem *dirrem; 2000 struct diradd *dap; 2001 int i; 2002 2003 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2004 switch (wk->wk_type) { 2005 2006 case D_INDIRDEP: 2007 indirdep = WK_INDIRDEP(wk); 2008 /* 2009 * None of the indirect pointers will ever be visible, 2010 * so they can simply be tossed. GOINGAWAY ensures 2011 * that allocated pointers will be saved in the buffer 2012 * cache until they are freed. Note that they will 2013 * only be able to be found by their physical address 2014 * since the inode mapping the logical address will 2015 * be gone. The save buffer used for the safe copy 2016 * was allocated in setup_allocindir_phase2 using 2017 * the physical address so it could be used for this 2018 * purpose. Hence we swap the safe copy with the real 2019 * copy, allowing the safe copy to be freed and holding 2020 * on to the real copy for later use in indir_trunc. 2021 * 2022 * NOTE: ir_savebp is relative to the block device 2023 * so b_bio1 contains the device block number. 2024 */ 2025 if (indirdep->ir_state & GOINGAWAY) { 2026 FREE_LOCK(&lk); 2027 panic("deallocate_dependencies: already gone"); 2028 } 2029 indirdep->ir_state |= GOINGAWAY; 2030 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 2031 free_allocindir(aip, inodedep); 2032 if (bp->b_bio1.bio_offset >= 0 || 2033 bp->b_bio2.bio_offset != indirdep->ir_savebp->b_bio1.bio_offset) { 2034 FREE_LOCK(&lk); 2035 panic("deallocate_dependencies: not indir"); 2036 } 2037 bcopy(bp->b_data, indirdep->ir_savebp->b_data, 2038 bp->b_bcount); 2039 WORKLIST_REMOVE(wk); 2040 WORKLIST_INSERT_BP(indirdep->ir_savebp, wk); 2041 continue; 2042 2043 case D_PAGEDEP: 2044 pagedep = WK_PAGEDEP(wk); 2045 /* 2046 * None of the directory additions will ever be 2047 * visible, so they can simply be tossed. 2048 */ 2049 for (i = 0; i < DAHASHSZ; i++) 2050 while ((dap = 2051 LIST_FIRST(&pagedep->pd_diraddhd[i]))) 2052 free_diradd(dap); 2053 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0) 2054 free_diradd(dap); 2055 /* 2056 * Copy any directory remove dependencies to the list 2057 * to be processed after the zero'ed inode is written. 2058 * If the inode has already been written, then they 2059 * can be dumped directly onto the work list. 2060 */ 2061 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 2062 LIST_REMOVE(dirrem, dm_next); 2063 dirrem->dm_dirinum = pagedep->pd_ino; 2064 if (inodedep == NULL || 2065 (inodedep->id_state & ALLCOMPLETE) == 2066 ALLCOMPLETE) 2067 add_to_worklist(&dirrem->dm_list); 2068 else 2069 WORKLIST_INSERT(&inodedep->id_bufwait, 2070 &dirrem->dm_list); 2071 } 2072 WORKLIST_REMOVE(&pagedep->pd_list); 2073 LIST_REMOVE(pagedep, pd_hash); 2074 WORKITEM_FREE(pagedep, D_PAGEDEP); 2075 continue; 2076 2077 case D_ALLOCINDIR: 2078 free_allocindir(WK_ALLOCINDIR(wk), inodedep); 2079 continue; 2080 2081 case D_ALLOCDIRECT: 2082 case D_INODEDEP: 2083 FREE_LOCK(&lk); 2084 panic("deallocate_dependencies: Unexpected type %s", 2085 TYPENAME(wk->wk_type)); 2086 /* NOTREACHED */ 2087 2088 default: 2089 FREE_LOCK(&lk); 2090 panic("deallocate_dependencies: Unknown type %s", 2091 TYPENAME(wk->wk_type)); 2092 /* NOTREACHED */ 2093 } 2094 } 2095 } 2096 2097 /* 2098 * Free an allocdirect. Generate a new freefrag work request if appropriate. 2099 * This routine must be called with splbio interrupts blocked. 2100 */ 2101 static void 2102 free_allocdirect(struct allocdirectlst *adphead, 2103 struct allocdirect *adp, int delay) 2104 { 2105 2106 #ifdef DEBUG 2107 if (lk.lkt_held == NOHOLDER) 2108 panic("free_allocdirect: lock not held"); 2109 #endif 2110 if ((adp->ad_state & DEPCOMPLETE) == 0) 2111 LIST_REMOVE(adp, ad_deps); 2112 TAILQ_REMOVE(adphead, adp, ad_next); 2113 if ((adp->ad_state & COMPLETE) == 0) 2114 WORKLIST_REMOVE(&adp->ad_list); 2115 if (adp->ad_freefrag != NULL) { 2116 if (delay) 2117 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 2118 &adp->ad_freefrag->ff_list); 2119 else 2120 add_to_worklist(&adp->ad_freefrag->ff_list); 2121 } 2122 WORKITEM_FREE(adp, D_ALLOCDIRECT); 2123 } 2124 2125 /* 2126 * Prepare an inode to be freed. The actual free operation is not 2127 * done until the zero'ed inode has been written to disk. 2128 */ 2129 void 2130 softdep_freefile(struct vnode *pvp, ino_t ino, int mode) 2131 { 2132 struct inode *ip = VTOI(pvp); 2133 struct inodedep *inodedep; 2134 struct freefile *freefile; 2135 2136 /* 2137 * This sets up the inode de-allocation dependency. 2138 */ 2139 MALLOC(freefile, struct freefile *, sizeof(struct freefile), 2140 M_FREEFILE, M_SOFTDEP_FLAGS); 2141 freefile->fx_list.wk_type = D_FREEFILE; 2142 freefile->fx_list.wk_state = 0; 2143 freefile->fx_mode = mode; 2144 freefile->fx_oldinum = ino; 2145 freefile->fx_devvp = ip->i_devvp; 2146 freefile->fx_fs = ip->i_fs; 2147 2148 /* 2149 * If the inodedep does not exist, then the zero'ed inode has 2150 * been written to disk. If the allocated inode has never been 2151 * written to disk, then the on-disk inode is zero'ed. In either 2152 * case we can free the file immediately. 2153 */ 2154 ACQUIRE_LOCK(&lk); 2155 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 || 2156 check_inode_unwritten(inodedep)) { 2157 FREE_LOCK(&lk); 2158 handle_workitem_freefile(freefile); 2159 return; 2160 } 2161 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 2162 FREE_LOCK(&lk); 2163 } 2164 2165 /* 2166 * Check to see if an inode has never been written to disk. If 2167 * so free the inodedep and return success, otherwise return failure. 2168 * This routine must be called with splbio interrupts blocked. 2169 * 2170 * If we still have a bitmap dependency, then the inode has never 2171 * been written to disk. Drop the dependency as it is no longer 2172 * necessary since the inode is being deallocated. We set the 2173 * ALLCOMPLETE flags since the bitmap now properly shows that the 2174 * inode is not allocated. Even if the inode is actively being 2175 * written, it has been rolled back to its zero'ed state, so we 2176 * are ensured that a zero inode is what is on the disk. For short 2177 * lived files, this change will usually result in removing all the 2178 * dependencies from the inode so that it can be freed immediately. 2179 */ 2180 static int 2181 check_inode_unwritten(struct inodedep *inodedep) 2182 { 2183 2184 if ((inodedep->id_state & DEPCOMPLETE) != 0 || 2185 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2186 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2187 LIST_FIRST(&inodedep->id_inowait) != NULL || 2188 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2189 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2190 inodedep->id_nlinkdelta != 0) 2191 return (0); 2192 2193 /* 2194 * Another process might be in initiate_write_inodeblock 2195 * trying to allocate memory without holding "Softdep Lock". 2196 */ 2197 if ((inodedep->id_state & IOSTARTED) != 0 && 2198 inodedep->id_savedino == NULL) 2199 return(0); 2200 2201 inodedep->id_state |= ALLCOMPLETE; 2202 LIST_REMOVE(inodedep, id_deps); 2203 inodedep->id_buf = NULL; 2204 if (inodedep->id_state & ONWORKLIST) 2205 WORKLIST_REMOVE(&inodedep->id_list); 2206 if (inodedep->id_savedino != NULL) { 2207 FREE(inodedep->id_savedino, M_INODEDEP); 2208 inodedep->id_savedino = NULL; 2209 } 2210 if (free_inodedep(inodedep) == 0) { 2211 FREE_LOCK(&lk); 2212 panic("check_inode_unwritten: busy inode"); 2213 } 2214 return (1); 2215 } 2216 2217 /* 2218 * Try to free an inodedep structure. Return 1 if it could be freed. 2219 */ 2220 static int 2221 free_inodedep(struct inodedep *inodedep) 2222 { 2223 2224 if ((inodedep->id_state & ONWORKLIST) != 0 || 2225 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 2226 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2227 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2228 LIST_FIRST(&inodedep->id_inowait) != NULL || 2229 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2230 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2231 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL) 2232 return (0); 2233 LIST_REMOVE(inodedep, id_hash); 2234 WORKITEM_FREE(inodedep, D_INODEDEP); 2235 num_inodedep -= 1; 2236 return (1); 2237 } 2238 2239 /* 2240 * This workitem routine performs the block de-allocation. 2241 * The workitem is added to the pending list after the updated 2242 * inode block has been written to disk. As mentioned above, 2243 * checks regarding the number of blocks de-allocated (compared 2244 * to the number of blocks allocated for the file) are also 2245 * performed in this function. 2246 */ 2247 static void 2248 handle_workitem_freeblocks(struct freeblks *freeblks) 2249 { 2250 struct inode tip; 2251 ufs_daddr_t bn; 2252 struct fs *fs; 2253 int i, level, bsize; 2254 long nblocks, blocksreleased = 0; 2255 int error, allerror = 0; 2256 ufs_lbn_t baselbns[NIADDR], tmpval; 2257 2258 tip.i_number = freeblks->fb_previousinum; 2259 tip.i_devvp = freeblks->fb_devvp; 2260 tip.i_dev = freeblks->fb_devvp->v_rdev; 2261 tip.i_fs = freeblks->fb_fs; 2262 tip.i_size = freeblks->fb_oldsize; 2263 tip.i_uid = freeblks->fb_uid; 2264 fs = freeblks->fb_fs; 2265 tmpval = 1; 2266 baselbns[0] = NDADDR; 2267 for (i = 1; i < NIADDR; i++) { 2268 tmpval *= NINDIR(fs); 2269 baselbns[i] = baselbns[i - 1] + tmpval; 2270 } 2271 nblocks = btodb(fs->fs_bsize); 2272 blocksreleased = 0; 2273 /* 2274 * Indirect blocks first. 2275 */ 2276 for (level = (NIADDR - 1); level >= 0; level--) { 2277 if ((bn = freeblks->fb_iblks[level]) == 0) 2278 continue; 2279 if ((error = indir_trunc(&tip, fsbtodoff(fs, bn), level, 2280 baselbns[level], &blocksreleased)) == 0) 2281 allerror = error; 2282 ffs_blkfree(&tip, bn, fs->fs_bsize); 2283 blocksreleased += nblocks; 2284 } 2285 /* 2286 * All direct blocks or frags. 2287 */ 2288 for (i = (NDADDR - 1); i >= 0; i--) { 2289 if ((bn = freeblks->fb_dblks[i]) == 0) 2290 continue; 2291 bsize = blksize(fs, &tip, i); 2292 ffs_blkfree(&tip, bn, bsize); 2293 blocksreleased += btodb(bsize); 2294 } 2295 2296 #ifdef DIAGNOSTIC 2297 if (freeblks->fb_chkcnt != blocksreleased) 2298 kprintf("handle_workitem_freeblocks: block count\n"); 2299 if (allerror) 2300 softdep_error("handle_workitem_freeblks", allerror); 2301 #endif /* DIAGNOSTIC */ 2302 WORKITEM_FREE(freeblks, D_FREEBLKS); 2303 } 2304 2305 /* 2306 * Release blocks associated with the inode ip and stored in the indirect 2307 * block at doffset. If level is greater than SINGLE, the block is an 2308 * indirect block and recursive calls to indirtrunc must be used to 2309 * cleanse other indirect blocks. 2310 */ 2311 static int 2312 indir_trunc(struct inode *ip, off_t doffset, int level, ufs_lbn_t lbn, 2313 long *countp) 2314 { 2315 struct buf *bp; 2316 ufs_daddr_t *bap; 2317 ufs_daddr_t nb; 2318 struct fs *fs; 2319 struct worklist *wk; 2320 struct indirdep *indirdep; 2321 int i, lbnadd, nblocks; 2322 int error, allerror = 0; 2323 2324 fs = ip->i_fs; 2325 lbnadd = 1; 2326 for (i = level; i > 0; i--) 2327 lbnadd *= NINDIR(fs); 2328 /* 2329 * Get buffer of block pointers to be freed. This routine is not 2330 * called until the zero'ed inode has been written, so it is safe 2331 * to free blocks as they are encountered. Because the inode has 2332 * been zero'ed, calls to bmap on these blocks will fail. So, we 2333 * have to use the on-disk address and the block device for the 2334 * filesystem to look them up. If the file was deleted before its 2335 * indirect blocks were all written to disk, the routine that set 2336 * us up (deallocate_dependencies) will have arranged to leave 2337 * a complete copy of the indirect block in memory for our use. 2338 * Otherwise we have to read the blocks in from the disk. 2339 */ 2340 ACQUIRE_LOCK(&lk); 2341 if ((bp = findblk(ip->i_devvp, doffset, FINDBLK_TEST)) != NULL && 2342 (wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2343 /* 2344 * bp must be ir_savebp, which is held locked for our use. 2345 */ 2346 if (wk->wk_type != D_INDIRDEP || 2347 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp || 2348 (indirdep->ir_state & GOINGAWAY) == 0) { 2349 FREE_LOCK(&lk); 2350 panic("indir_trunc: lost indirdep"); 2351 } 2352 WORKLIST_REMOVE(wk); 2353 WORKITEM_FREE(indirdep, D_INDIRDEP); 2354 if (LIST_FIRST(&bp->b_dep) != NULL) { 2355 FREE_LOCK(&lk); 2356 panic("indir_trunc: dangling dep"); 2357 } 2358 FREE_LOCK(&lk); 2359 } else { 2360 FREE_LOCK(&lk); 2361 error = bread(ip->i_devvp, doffset, (int)fs->fs_bsize, &bp); 2362 if (error) 2363 return (error); 2364 } 2365 /* 2366 * Recursively free indirect blocks. 2367 */ 2368 bap = (ufs_daddr_t *)bp->b_data; 2369 nblocks = btodb(fs->fs_bsize); 2370 for (i = NINDIR(fs) - 1; i >= 0; i--) { 2371 if ((nb = bap[i]) == 0) 2372 continue; 2373 if (level != 0) { 2374 if ((error = indir_trunc(ip, fsbtodoff(fs, nb), 2375 level - 1, lbn + (i * lbnadd), countp)) != 0) 2376 allerror = error; 2377 } 2378 ffs_blkfree(ip, nb, fs->fs_bsize); 2379 *countp += nblocks; 2380 } 2381 bp->b_flags |= B_INVAL | B_NOCACHE; 2382 brelse(bp); 2383 return (allerror); 2384 } 2385 2386 /* 2387 * Free an allocindir. 2388 * This routine must be called with splbio interrupts blocked. 2389 */ 2390 static void 2391 free_allocindir(struct allocindir *aip, struct inodedep *inodedep) 2392 { 2393 struct freefrag *freefrag; 2394 2395 #ifdef DEBUG 2396 if (lk.lkt_held == NOHOLDER) 2397 panic("free_allocindir: lock not held"); 2398 #endif 2399 if ((aip->ai_state & DEPCOMPLETE) == 0) 2400 LIST_REMOVE(aip, ai_deps); 2401 if (aip->ai_state & ONWORKLIST) 2402 WORKLIST_REMOVE(&aip->ai_list); 2403 LIST_REMOVE(aip, ai_next); 2404 if ((freefrag = aip->ai_freefrag) != NULL) { 2405 if (inodedep == NULL) 2406 add_to_worklist(&freefrag->ff_list); 2407 else 2408 WORKLIST_INSERT(&inodedep->id_bufwait, 2409 &freefrag->ff_list); 2410 } 2411 WORKITEM_FREE(aip, D_ALLOCINDIR); 2412 } 2413 2414 /* 2415 * Directory entry addition dependencies. 2416 * 2417 * When adding a new directory entry, the inode (with its incremented link 2418 * count) must be written to disk before the directory entry's pointer to it. 2419 * Also, if the inode is newly allocated, the corresponding freemap must be 2420 * updated (on disk) before the directory entry's pointer. These requirements 2421 * are met via undo/redo on the directory entry's pointer, which consists 2422 * simply of the inode number. 2423 * 2424 * As directory entries are added and deleted, the free space within a 2425 * directory block can become fragmented. The ufs filesystem will compact 2426 * a fragmented directory block to make space for a new entry. When this 2427 * occurs, the offsets of previously added entries change. Any "diradd" 2428 * dependency structures corresponding to these entries must be updated with 2429 * the new offsets. 2430 */ 2431 2432 /* 2433 * This routine is called after the in-memory inode's link 2434 * count has been incremented, but before the directory entry's 2435 * pointer to the inode has been set. 2436 * 2437 * Parameters: 2438 * bp: buffer containing directory block 2439 * dp: inode for directory 2440 * diroffset: offset of new entry in directory 2441 * newinum: inode referenced by new directory entry 2442 * newdirbp: non-NULL => contents of new mkdir 2443 */ 2444 void 2445 softdep_setup_directory_add(struct buf *bp, struct inode *dp, off_t diroffset, 2446 ino_t newinum, struct buf *newdirbp) 2447 { 2448 int offset; /* offset of new entry within directory block */ 2449 ufs_lbn_t lbn; /* block in directory containing new entry */ 2450 struct fs *fs; 2451 struct diradd *dap; 2452 struct pagedep *pagedep; 2453 struct inodedep *inodedep; 2454 struct mkdir *mkdir1, *mkdir2; 2455 2456 /* 2457 * Whiteouts have no dependencies. 2458 */ 2459 if (newinum == WINO) { 2460 if (newdirbp != NULL) 2461 bdwrite(newdirbp); 2462 return; 2463 } 2464 2465 fs = dp->i_fs; 2466 lbn = lblkno(fs, diroffset); 2467 offset = blkoff(fs, diroffset); 2468 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, 2469 M_SOFTDEP_FLAGS | M_ZERO); 2470 dap->da_list.wk_type = D_DIRADD; 2471 dap->da_offset = offset; 2472 dap->da_newinum = newinum; 2473 dap->da_state = ATTACHED; 2474 if (newdirbp == NULL) { 2475 dap->da_state |= DEPCOMPLETE; 2476 ACQUIRE_LOCK(&lk); 2477 } else { 2478 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 2479 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2480 M_SOFTDEP_FLAGS); 2481 mkdir1->md_list.wk_type = D_MKDIR; 2482 mkdir1->md_state = MKDIR_BODY; 2483 mkdir1->md_diradd = dap; 2484 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2485 M_SOFTDEP_FLAGS); 2486 mkdir2->md_list.wk_type = D_MKDIR; 2487 mkdir2->md_state = MKDIR_PARENT; 2488 mkdir2->md_diradd = dap; 2489 /* 2490 * Dependency on "." and ".." being written to disk. 2491 */ 2492 mkdir1->md_buf = newdirbp; 2493 ACQUIRE_LOCK(&lk); 2494 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 2495 WORKLIST_INSERT_BP(newdirbp, &mkdir1->md_list); 2496 FREE_LOCK(&lk); 2497 bdwrite(newdirbp); 2498 /* 2499 * Dependency on link count increase for parent directory 2500 */ 2501 ACQUIRE_LOCK(&lk); 2502 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0 2503 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2504 dap->da_state &= ~MKDIR_PARENT; 2505 WORKITEM_FREE(mkdir2, D_MKDIR); 2506 } else { 2507 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 2508 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 2509 } 2510 } 2511 /* 2512 * Link into parent directory pagedep to await its being written. 2513 */ 2514 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2515 WORKLIST_INSERT_BP(bp, &pagedep->pd_list); 2516 dap->da_pagedep = pagedep; 2517 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 2518 da_pdlist); 2519 /* 2520 * Link into its inodedep. Put it on the id_bufwait list if the inode 2521 * is not yet written. If it is written, do the post-inode write 2522 * processing to put it on the id_pendinghd list. 2523 */ 2524 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep); 2525 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 2526 diradd_inode_written(dap, inodedep); 2527 else 2528 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2529 FREE_LOCK(&lk); 2530 } 2531 2532 /* 2533 * This procedure is called to change the offset of a directory 2534 * entry when compacting a directory block which must be owned 2535 * exclusively by the caller. Note that the actual entry movement 2536 * must be done in this procedure to ensure that no I/O completions 2537 * occur while the move is in progress. 2538 * 2539 * Parameters: 2540 * dp: inode for directory 2541 * base: address of dp->i_offset 2542 * oldloc: address of old directory location 2543 * newloc: address of new directory location 2544 * entrysize: size of directory entry 2545 */ 2546 void 2547 softdep_change_directoryentry_offset(struct inode *dp, caddr_t base, 2548 caddr_t oldloc, caddr_t newloc, 2549 int entrysize) 2550 { 2551 int offset, oldoffset, newoffset; 2552 struct pagedep *pagedep; 2553 struct diradd *dap; 2554 ufs_lbn_t lbn; 2555 2556 ACQUIRE_LOCK(&lk); 2557 lbn = lblkno(dp->i_fs, dp->i_offset); 2558 offset = blkoff(dp->i_fs, dp->i_offset); 2559 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) 2560 goto done; 2561 oldoffset = offset + (oldloc - base); 2562 newoffset = offset + (newloc - base); 2563 2564 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) { 2565 if (dap->da_offset != oldoffset) 2566 continue; 2567 dap->da_offset = newoffset; 2568 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) 2569 break; 2570 LIST_REMOVE(dap, da_pdlist); 2571 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], 2572 dap, da_pdlist); 2573 break; 2574 } 2575 if (dap == NULL) { 2576 2577 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) { 2578 if (dap->da_offset == oldoffset) { 2579 dap->da_offset = newoffset; 2580 break; 2581 } 2582 } 2583 } 2584 done: 2585 bcopy(oldloc, newloc, entrysize); 2586 FREE_LOCK(&lk); 2587 } 2588 2589 /* 2590 * Free a diradd dependency structure. This routine must be called 2591 * with splbio interrupts blocked. 2592 */ 2593 static void 2594 free_diradd(struct diradd *dap) 2595 { 2596 struct dirrem *dirrem; 2597 struct pagedep *pagedep; 2598 struct inodedep *inodedep; 2599 struct mkdir *mkdir, *nextmd; 2600 2601 #ifdef DEBUG 2602 if (lk.lkt_held == NOHOLDER) 2603 panic("free_diradd: lock not held"); 2604 #endif 2605 WORKLIST_REMOVE(&dap->da_list); 2606 LIST_REMOVE(dap, da_pdlist); 2607 if ((dap->da_state & DIRCHG) == 0) { 2608 pagedep = dap->da_pagedep; 2609 } else { 2610 dirrem = dap->da_previous; 2611 pagedep = dirrem->dm_pagedep; 2612 dirrem->dm_dirinum = pagedep->pd_ino; 2613 add_to_worklist(&dirrem->dm_list); 2614 } 2615 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum, 2616 0, &inodedep) != 0) 2617 (void) free_inodedep(inodedep); 2618 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2619 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 2620 nextmd = LIST_NEXT(mkdir, md_mkdirs); 2621 if (mkdir->md_diradd != dap) 2622 continue; 2623 dap->da_state &= ~mkdir->md_state; 2624 WORKLIST_REMOVE(&mkdir->md_list); 2625 LIST_REMOVE(mkdir, md_mkdirs); 2626 WORKITEM_FREE(mkdir, D_MKDIR); 2627 } 2628 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2629 FREE_LOCK(&lk); 2630 panic("free_diradd: unfound ref"); 2631 } 2632 } 2633 WORKITEM_FREE(dap, D_DIRADD); 2634 } 2635 2636 /* 2637 * Directory entry removal dependencies. 2638 * 2639 * When removing a directory entry, the entry's inode pointer must be 2640 * zero'ed on disk before the corresponding inode's link count is decremented 2641 * (possibly freeing the inode for re-use). This dependency is handled by 2642 * updating the directory entry but delaying the inode count reduction until 2643 * after the directory block has been written to disk. After this point, the 2644 * inode count can be decremented whenever it is convenient. 2645 */ 2646 2647 /* 2648 * This routine should be called immediately after removing 2649 * a directory entry. The inode's link count should not be 2650 * decremented by the calling procedure -- the soft updates 2651 * code will do this task when it is safe. 2652 * 2653 * Parameters: 2654 * bp: buffer containing directory block 2655 * dp: inode for the directory being modified 2656 * ip: inode for directory entry being removed 2657 * isrmdir: indicates if doing RMDIR 2658 */ 2659 void 2660 softdep_setup_remove(struct buf *bp, struct inode *dp, struct inode *ip, 2661 int isrmdir) 2662 { 2663 struct dirrem *dirrem, *prevdirrem; 2664 2665 /* 2666 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 2667 */ 2668 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2669 2670 /* 2671 * If the COMPLETE flag is clear, then there were no active 2672 * entries and we want to roll back to a zeroed entry until 2673 * the new inode is committed to disk. If the COMPLETE flag is 2674 * set then we have deleted an entry that never made it to 2675 * disk. If the entry we deleted resulted from a name change, 2676 * then the old name still resides on disk. We cannot delete 2677 * its inode (returned to us in prevdirrem) until the zeroed 2678 * directory entry gets to disk. The new inode has never been 2679 * referenced on the disk, so can be deleted immediately. 2680 */ 2681 if ((dirrem->dm_state & COMPLETE) == 0) { 2682 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 2683 dm_next); 2684 FREE_LOCK(&lk); 2685 } else { 2686 if (prevdirrem != NULL) 2687 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 2688 prevdirrem, dm_next); 2689 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 2690 FREE_LOCK(&lk); 2691 handle_workitem_remove(dirrem); 2692 } 2693 } 2694 2695 /* 2696 * Allocate a new dirrem if appropriate and return it along with 2697 * its associated pagedep. Called without a lock, returns with lock. 2698 */ 2699 static long num_dirrem; /* number of dirrem allocated */ 2700 2701 /* 2702 * Parameters: 2703 * bp: buffer containing directory block 2704 * dp: inode for the directory being modified 2705 * ip: inode for directory entry being removed 2706 * isrmdir: indicates if doing RMDIR 2707 * prevdirremp: previously referenced inode, if any 2708 */ 2709 static struct dirrem * 2710 newdirrem(struct buf *bp, struct inode *dp, struct inode *ip, 2711 int isrmdir, struct dirrem **prevdirremp) 2712 { 2713 int offset; 2714 ufs_lbn_t lbn; 2715 struct diradd *dap; 2716 struct dirrem *dirrem; 2717 struct pagedep *pagedep; 2718 2719 /* 2720 * Whiteouts have no deletion dependencies. 2721 */ 2722 if (ip == NULL) 2723 panic("newdirrem: whiteout"); 2724 /* 2725 * If we are over our limit, try to improve the situation. 2726 * Limiting the number of dirrem structures will also limit 2727 * the number of freefile and freeblks structures. 2728 */ 2729 if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0) 2730 (void) request_cleanup(FLUSH_REMOVE, 0); 2731 num_dirrem += 1; 2732 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem), 2733 M_DIRREM, M_SOFTDEP_FLAGS | M_ZERO); 2734 dirrem->dm_list.wk_type = D_DIRREM; 2735 dirrem->dm_state = isrmdir ? RMDIR : 0; 2736 dirrem->dm_mnt = ITOV(ip)->v_mount; 2737 dirrem->dm_oldinum = ip->i_number; 2738 *prevdirremp = NULL; 2739 2740 ACQUIRE_LOCK(&lk); 2741 lbn = lblkno(dp->i_fs, dp->i_offset); 2742 offset = blkoff(dp->i_fs, dp->i_offset); 2743 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2744 WORKLIST_INSERT_BP(bp, &pagedep->pd_list); 2745 dirrem->dm_pagedep = pagedep; 2746 /* 2747 * Check for a diradd dependency for the same directory entry. 2748 * If present, then both dependencies become obsolete and can 2749 * be de-allocated. Check for an entry on both the pd_dirraddhd 2750 * list and the pd_pendinghd list. 2751 */ 2752 2753 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 2754 if (dap->da_offset == offset) 2755 break; 2756 if (dap == NULL) { 2757 2758 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 2759 if (dap->da_offset == offset) 2760 break; 2761 if (dap == NULL) 2762 return (dirrem); 2763 } 2764 /* 2765 * Must be ATTACHED at this point. 2766 */ 2767 if ((dap->da_state & ATTACHED) == 0) { 2768 FREE_LOCK(&lk); 2769 panic("newdirrem: not ATTACHED"); 2770 } 2771 if (dap->da_newinum != ip->i_number) { 2772 FREE_LOCK(&lk); 2773 panic("newdirrem: inum %"PRId64" should be %"PRId64, 2774 ip->i_number, dap->da_newinum); 2775 } 2776 /* 2777 * If we are deleting a changed name that never made it to disk, 2778 * then return the dirrem describing the previous inode (which 2779 * represents the inode currently referenced from this entry on disk). 2780 */ 2781 if ((dap->da_state & DIRCHG) != 0) { 2782 *prevdirremp = dap->da_previous; 2783 dap->da_state &= ~DIRCHG; 2784 dap->da_pagedep = pagedep; 2785 } 2786 /* 2787 * We are deleting an entry that never made it to disk. 2788 * Mark it COMPLETE so we can delete its inode immediately. 2789 */ 2790 dirrem->dm_state |= COMPLETE; 2791 free_diradd(dap); 2792 return (dirrem); 2793 } 2794 2795 /* 2796 * Directory entry change dependencies. 2797 * 2798 * Changing an existing directory entry requires that an add operation 2799 * be completed first followed by a deletion. The semantics for the addition 2800 * are identical to the description of adding a new entry above except 2801 * that the rollback is to the old inode number rather than zero. Once 2802 * the addition dependency is completed, the removal is done as described 2803 * in the removal routine above. 2804 */ 2805 2806 /* 2807 * This routine should be called immediately after changing 2808 * a directory entry. The inode's link count should not be 2809 * decremented by the calling procedure -- the soft updates 2810 * code will perform this task when it is safe. 2811 * 2812 * Parameters: 2813 * bp: buffer containing directory block 2814 * dp: inode for the directory being modified 2815 * ip: inode for directory entry being removed 2816 * newinum: new inode number for changed entry 2817 * isrmdir: indicates if doing RMDIR 2818 */ 2819 void 2820 softdep_setup_directory_change(struct buf *bp, struct inode *dp, 2821 struct inode *ip, ino_t newinum, 2822 int isrmdir) 2823 { 2824 int offset; 2825 struct diradd *dap = NULL; 2826 struct dirrem *dirrem, *prevdirrem; 2827 struct pagedep *pagedep; 2828 struct inodedep *inodedep; 2829 2830 offset = blkoff(dp->i_fs, dp->i_offset); 2831 2832 /* 2833 * Whiteouts do not need diradd dependencies. 2834 */ 2835 if (newinum != WINO) { 2836 MALLOC(dap, struct diradd *, sizeof(struct diradd), 2837 M_DIRADD, M_SOFTDEP_FLAGS | M_ZERO); 2838 dap->da_list.wk_type = D_DIRADD; 2839 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 2840 dap->da_offset = offset; 2841 dap->da_newinum = newinum; 2842 } 2843 2844 /* 2845 * Allocate a new dirrem and ACQUIRE_LOCK. 2846 */ 2847 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2848 pagedep = dirrem->dm_pagedep; 2849 /* 2850 * The possible values for isrmdir: 2851 * 0 - non-directory file rename 2852 * 1 - directory rename within same directory 2853 * inum - directory rename to new directory of given inode number 2854 * When renaming to a new directory, we are both deleting and 2855 * creating a new directory entry, so the link count on the new 2856 * directory should not change. Thus we do not need the followup 2857 * dirrem which is usually done in handle_workitem_remove. We set 2858 * the DIRCHG flag to tell handle_workitem_remove to skip the 2859 * followup dirrem. 2860 */ 2861 if (isrmdir > 1) 2862 dirrem->dm_state |= DIRCHG; 2863 2864 /* 2865 * Whiteouts have no additional dependencies, 2866 * so just put the dirrem on the correct list. 2867 */ 2868 if (newinum == WINO) { 2869 if ((dirrem->dm_state & COMPLETE) == 0) { 2870 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 2871 dm_next); 2872 } else { 2873 dirrem->dm_dirinum = pagedep->pd_ino; 2874 add_to_worklist(&dirrem->dm_list); 2875 } 2876 FREE_LOCK(&lk); 2877 return; 2878 } 2879 2880 /* 2881 * If the COMPLETE flag is clear, then there were no active 2882 * entries and we want to roll back to the previous inode until 2883 * the new inode is committed to disk. If the COMPLETE flag is 2884 * set, then we have deleted an entry that never made it to disk. 2885 * If the entry we deleted resulted from a name change, then the old 2886 * inode reference still resides on disk. Any rollback that we do 2887 * needs to be to that old inode (returned to us in prevdirrem). If 2888 * the entry we deleted resulted from a create, then there is 2889 * no entry on the disk, so we want to roll back to zero rather 2890 * than the uncommitted inode. In either of the COMPLETE cases we 2891 * want to immediately free the unwritten and unreferenced inode. 2892 */ 2893 if ((dirrem->dm_state & COMPLETE) == 0) { 2894 dap->da_previous = dirrem; 2895 } else { 2896 if (prevdirrem != NULL) { 2897 dap->da_previous = prevdirrem; 2898 } else { 2899 dap->da_state &= ~DIRCHG; 2900 dap->da_pagedep = pagedep; 2901 } 2902 dirrem->dm_dirinum = pagedep->pd_ino; 2903 add_to_worklist(&dirrem->dm_list); 2904 } 2905 /* 2906 * Link into its inodedep. Put it on the id_bufwait list if the inode 2907 * is not yet written. If it is written, do the post-inode write 2908 * processing to put it on the id_pendinghd list. 2909 */ 2910 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 || 2911 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2912 dap->da_state |= COMPLETE; 2913 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 2914 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 2915 } else { 2916 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 2917 dap, da_pdlist); 2918 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2919 } 2920 FREE_LOCK(&lk); 2921 } 2922 2923 /* 2924 * Called whenever the link count on an inode is changed. 2925 * It creates an inode dependency so that the new reference(s) 2926 * to the inode cannot be committed to disk until the updated 2927 * inode has been written. 2928 * 2929 * Parameters: 2930 * ip: the inode with the increased link count 2931 */ 2932 void 2933 softdep_change_linkcnt(struct inode *ip) 2934 { 2935 struct inodedep *inodedep; 2936 2937 ACQUIRE_LOCK(&lk); 2938 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 2939 if (ip->i_nlink < ip->i_effnlink) { 2940 FREE_LOCK(&lk); 2941 panic("softdep_change_linkcnt: bad delta"); 2942 } 2943 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2944 FREE_LOCK(&lk); 2945 } 2946 2947 /* 2948 * This workitem decrements the inode's link count. 2949 * If the link count reaches zero, the file is removed. 2950 */ 2951 static void 2952 handle_workitem_remove(struct dirrem *dirrem) 2953 { 2954 struct inodedep *inodedep; 2955 struct vnode *vp; 2956 struct inode *ip; 2957 ino_t oldinum; 2958 int error; 2959 2960 error = VFS_VGET(dirrem->dm_mnt, NULL, dirrem->dm_oldinum, &vp); 2961 if (error) { 2962 softdep_error("handle_workitem_remove: vget", error); 2963 return; 2964 } 2965 ip = VTOI(vp); 2966 ACQUIRE_LOCK(&lk); 2967 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){ 2968 FREE_LOCK(&lk); 2969 panic("handle_workitem_remove: lost inodedep"); 2970 } 2971 /* 2972 * Normal file deletion. 2973 */ 2974 if ((dirrem->dm_state & RMDIR) == 0) { 2975 ip->i_nlink--; 2976 ip->i_flag |= IN_CHANGE; 2977 if (ip->i_nlink < ip->i_effnlink) { 2978 FREE_LOCK(&lk); 2979 panic("handle_workitem_remove: bad file delta"); 2980 } 2981 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2982 FREE_LOCK(&lk); 2983 vput(vp); 2984 num_dirrem -= 1; 2985 WORKITEM_FREE(dirrem, D_DIRREM); 2986 return; 2987 } 2988 /* 2989 * Directory deletion. Decrement reference count for both the 2990 * just deleted parent directory entry and the reference for ".". 2991 * Next truncate the directory to length zero. When the 2992 * truncation completes, arrange to have the reference count on 2993 * the parent decremented to account for the loss of "..". 2994 */ 2995 ip->i_nlink -= 2; 2996 ip->i_flag |= IN_CHANGE; 2997 if (ip->i_nlink < ip->i_effnlink) { 2998 FREE_LOCK(&lk); 2999 panic("handle_workitem_remove: bad dir delta"); 3000 } 3001 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3002 FREE_LOCK(&lk); 3003 if ((error = ffs_truncate(vp, (off_t)0, 0, proc0.p_ucred)) != 0) 3004 softdep_error("handle_workitem_remove: truncate", error); 3005 /* 3006 * Rename a directory to a new parent. Since, we are both deleting 3007 * and creating a new directory entry, the link count on the new 3008 * directory should not change. Thus we skip the followup dirrem. 3009 */ 3010 if (dirrem->dm_state & DIRCHG) { 3011 vput(vp); 3012 num_dirrem -= 1; 3013 WORKITEM_FREE(dirrem, D_DIRREM); 3014 return; 3015 } 3016 /* 3017 * If the inodedep does not exist, then the zero'ed inode has 3018 * been written to disk. If the allocated inode has never been 3019 * written to disk, then the on-disk inode is zero'ed. In either 3020 * case we can remove the file immediately. 3021 */ 3022 ACQUIRE_LOCK(&lk); 3023 dirrem->dm_state = 0; 3024 oldinum = dirrem->dm_oldinum; 3025 dirrem->dm_oldinum = dirrem->dm_dirinum; 3026 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 || 3027 check_inode_unwritten(inodedep)) { 3028 FREE_LOCK(&lk); 3029 vput(vp); 3030 handle_workitem_remove(dirrem); 3031 return; 3032 } 3033 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 3034 FREE_LOCK(&lk); 3035 ip->i_flag |= IN_CHANGE; 3036 ffs_update(vp, 0); 3037 vput(vp); 3038 } 3039 3040 /* 3041 * Inode de-allocation dependencies. 3042 * 3043 * When an inode's link count is reduced to zero, it can be de-allocated. We 3044 * found it convenient to postpone de-allocation until after the inode is 3045 * written to disk with its new link count (zero). At this point, all of the 3046 * on-disk inode's block pointers are nullified and, with careful dependency 3047 * list ordering, all dependencies related to the inode will be satisfied and 3048 * the corresponding dependency structures de-allocated. So, if/when the 3049 * inode is reused, there will be no mixing of old dependencies with new 3050 * ones. This artificial dependency is set up by the block de-allocation 3051 * procedure above (softdep_setup_freeblocks) and completed by the 3052 * following procedure. 3053 */ 3054 static void 3055 handle_workitem_freefile(struct freefile *freefile) 3056 { 3057 struct vnode vp; 3058 struct inode tip; 3059 struct inodedep *idp; 3060 int error; 3061 3062 #ifdef DEBUG 3063 ACQUIRE_LOCK(&lk); 3064 error = inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp); 3065 FREE_LOCK(&lk); 3066 if (error) 3067 panic("handle_workitem_freefile: inodedep survived"); 3068 #endif 3069 tip.i_devvp = freefile->fx_devvp; 3070 tip.i_dev = freefile->fx_devvp->v_rdev; 3071 tip.i_fs = freefile->fx_fs; 3072 vp.v_data = &tip; 3073 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0) 3074 softdep_error("handle_workitem_freefile", error); 3075 WORKITEM_FREE(freefile, D_FREEFILE); 3076 } 3077 3078 /* 3079 * Helper function which unlinks marker element from work list and returns 3080 * the next element on the list. 3081 */ 3082 static __inline struct worklist * 3083 markernext(struct worklist *marker) 3084 { 3085 struct worklist *next; 3086 3087 next = LIST_NEXT(marker, wk_list); 3088 LIST_REMOVE(marker, wk_list); 3089 return next; 3090 } 3091 3092 /* 3093 * checkread, checkwrite 3094 * 3095 * bioops callback - hold io_token 3096 */ 3097 static int 3098 softdep_checkread(struct buf *bp) 3099 { 3100 /* nothing to do, mp lock not needed */ 3101 return(0); 3102 } 3103 3104 /* 3105 * bioops callback - hold io_token 3106 */ 3107 static int 3108 softdep_checkwrite(struct buf *bp) 3109 { 3110 /* nothing to do, mp lock not needed */ 3111 return(0); 3112 } 3113 3114 /* 3115 * Disk writes. 3116 * 3117 * The dependency structures constructed above are most actively used when file 3118 * system blocks are written to disk. No constraints are placed on when a 3119 * block can be written, but unsatisfied update dependencies are made safe by 3120 * modifying (or replacing) the source memory for the duration of the disk 3121 * write. When the disk write completes, the memory block is again brought 3122 * up-to-date. 3123 * 3124 * In-core inode structure reclamation. 3125 * 3126 * Because there are a finite number of "in-core" inode structures, they are 3127 * reused regularly. By transferring all inode-related dependencies to the 3128 * in-memory inode block and indexing them separately (via "inodedep"s), we 3129 * can allow "in-core" inode structures to be reused at any time and avoid 3130 * any increase in contention. 3131 * 3132 * Called just before entering the device driver to initiate a new disk I/O. 3133 * The buffer must be locked, thus, no I/O completion operations can occur 3134 * while we are manipulating its associated dependencies. 3135 * 3136 * bioops callback - hold io_token 3137 * 3138 * Parameters: 3139 * bp: structure describing disk write to occur 3140 */ 3141 static void 3142 softdep_disk_io_initiation(struct buf *bp) 3143 { 3144 struct worklist *wk; 3145 struct worklist marker; 3146 struct indirdep *indirdep; 3147 3148 /* 3149 * We only care about write operations. There should never 3150 * be dependencies for reads. 3151 */ 3152 if (bp->b_cmd == BUF_CMD_READ) 3153 panic("softdep_disk_io_initiation: read"); 3154 3155 get_mplock(); 3156 marker.wk_type = D_LAST + 1; /* Not a normal workitem */ 3157 3158 /* 3159 * Do any necessary pre-I/O processing. 3160 */ 3161 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = markernext(&marker)) { 3162 LIST_INSERT_AFTER(wk, &marker, wk_list); 3163 3164 switch (wk->wk_type) { 3165 case D_PAGEDEP: 3166 initiate_write_filepage(WK_PAGEDEP(wk), bp); 3167 continue; 3168 3169 case D_INODEDEP: 3170 initiate_write_inodeblock(WK_INODEDEP(wk), bp); 3171 continue; 3172 3173 case D_INDIRDEP: 3174 indirdep = WK_INDIRDEP(wk); 3175 if (indirdep->ir_state & GOINGAWAY) 3176 panic("disk_io_initiation: indirdep gone"); 3177 /* 3178 * If there are no remaining dependencies, this 3179 * will be writing the real pointers, so the 3180 * dependency can be freed. 3181 */ 3182 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) { 3183 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 3184 brelse(indirdep->ir_savebp); 3185 /* inline expand WORKLIST_REMOVE(wk); */ 3186 wk->wk_state &= ~ONWORKLIST; 3187 LIST_REMOVE(wk, wk_list); 3188 WORKITEM_FREE(indirdep, D_INDIRDEP); 3189 continue; 3190 } 3191 /* 3192 * Replace up-to-date version with safe version. 3193 */ 3194 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount, 3195 M_INDIRDEP, M_SOFTDEP_FLAGS); 3196 ACQUIRE_LOCK(&lk); 3197 indirdep->ir_state &= ~ATTACHED; 3198 indirdep->ir_state |= UNDONE; 3199 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 3200 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 3201 bp->b_bcount); 3202 FREE_LOCK(&lk); 3203 continue; 3204 3205 case D_MKDIR: 3206 case D_BMSAFEMAP: 3207 case D_ALLOCDIRECT: 3208 case D_ALLOCINDIR: 3209 continue; 3210 3211 default: 3212 panic("handle_disk_io_initiation: Unexpected type %s", 3213 TYPENAME(wk->wk_type)); 3214 /* NOTREACHED */ 3215 } 3216 } 3217 rel_mplock(); 3218 } 3219 3220 /* 3221 * Called from within the procedure above to deal with unsatisfied 3222 * allocation dependencies in a directory. The buffer must be locked, 3223 * thus, no I/O completion operations can occur while we are 3224 * manipulating its associated dependencies. 3225 */ 3226 static void 3227 initiate_write_filepage(struct pagedep *pagedep, struct buf *bp) 3228 { 3229 struct diradd *dap; 3230 struct direct *ep; 3231 int i; 3232 3233 if (pagedep->pd_state & IOSTARTED) { 3234 /* 3235 * This can only happen if there is a driver that does not 3236 * understand chaining. Here biodone will reissue the call 3237 * to strategy for the incomplete buffers. 3238 */ 3239 kprintf("initiate_write_filepage: already started\n"); 3240 return; 3241 } 3242 pagedep->pd_state |= IOSTARTED; 3243 ACQUIRE_LOCK(&lk); 3244 for (i = 0; i < DAHASHSZ; i++) { 3245 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 3246 ep = (struct direct *) 3247 ((char *)bp->b_data + dap->da_offset); 3248 if (ep->d_ino != dap->da_newinum) { 3249 FREE_LOCK(&lk); 3250 panic("%s: dir inum %d != new %"PRId64, 3251 "initiate_write_filepage", 3252 ep->d_ino, dap->da_newinum); 3253 } 3254 if (dap->da_state & DIRCHG) 3255 ep->d_ino = dap->da_previous->dm_oldinum; 3256 else 3257 ep->d_ino = 0; 3258 dap->da_state &= ~ATTACHED; 3259 dap->da_state |= UNDONE; 3260 } 3261 } 3262 FREE_LOCK(&lk); 3263 } 3264 3265 /* 3266 * Called from within the procedure above to deal with unsatisfied 3267 * allocation dependencies in an inodeblock. The buffer must be 3268 * locked, thus, no I/O completion operations can occur while we 3269 * are manipulating its associated dependencies. 3270 * 3271 * Parameters: 3272 * bp: The inode block 3273 */ 3274 static void 3275 initiate_write_inodeblock(struct inodedep *inodedep, struct buf *bp) 3276 { 3277 struct allocdirect *adp, *lastadp; 3278 struct ufs1_dinode *dp; 3279 struct ufs1_dinode *sip; 3280 struct fs *fs; 3281 ufs_lbn_t prevlbn = 0; 3282 int i, deplist; 3283 3284 if (inodedep->id_state & IOSTARTED) 3285 panic("initiate_write_inodeblock: already started"); 3286 inodedep->id_state |= IOSTARTED; 3287 fs = inodedep->id_fs; 3288 dp = (struct ufs1_dinode *)bp->b_data + 3289 ino_to_fsbo(fs, inodedep->id_ino); 3290 /* 3291 * If the bitmap is not yet written, then the allocated 3292 * inode cannot be written to disk. 3293 */ 3294 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 3295 if (inodedep->id_savedino != NULL) 3296 panic("initiate_write_inodeblock: already doing I/O"); 3297 MALLOC(sip, struct ufs1_dinode *, 3298 sizeof(struct ufs1_dinode), M_INODEDEP, M_SOFTDEP_FLAGS); 3299 inodedep->id_savedino = sip; 3300 *inodedep->id_savedino = *dp; 3301 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 3302 dp->di_gen = inodedep->id_savedino->di_gen; 3303 return; 3304 } 3305 /* 3306 * If no dependencies, then there is nothing to roll back. 3307 */ 3308 inodedep->id_savedsize = dp->di_size; 3309 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 3310 return; 3311 /* 3312 * Set the dependencies to busy. 3313 */ 3314 ACQUIRE_LOCK(&lk); 3315 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3316 adp = TAILQ_NEXT(adp, ad_next)) { 3317 #ifdef DIAGNOSTIC 3318 if (deplist != 0 && prevlbn >= adp->ad_lbn) { 3319 FREE_LOCK(&lk); 3320 panic("softdep_write_inodeblock: lbn order"); 3321 } 3322 prevlbn = adp->ad_lbn; 3323 if (adp->ad_lbn < NDADDR && 3324 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) { 3325 FREE_LOCK(&lk); 3326 panic("%s: direct pointer #%ld mismatch %d != %d", 3327 "softdep_write_inodeblock", adp->ad_lbn, 3328 dp->di_db[adp->ad_lbn], adp->ad_newblkno); 3329 } 3330 if (adp->ad_lbn >= NDADDR && 3331 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) { 3332 FREE_LOCK(&lk); 3333 panic("%s: indirect pointer #%ld mismatch %d != %d", 3334 "softdep_write_inodeblock", adp->ad_lbn - NDADDR, 3335 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno); 3336 } 3337 deplist |= 1 << adp->ad_lbn; 3338 if ((adp->ad_state & ATTACHED) == 0) { 3339 FREE_LOCK(&lk); 3340 panic("softdep_write_inodeblock: Unknown state 0x%x", 3341 adp->ad_state); 3342 } 3343 #endif /* DIAGNOSTIC */ 3344 adp->ad_state &= ~ATTACHED; 3345 adp->ad_state |= UNDONE; 3346 } 3347 /* 3348 * The on-disk inode cannot claim to be any larger than the last 3349 * fragment that has been written. Otherwise, the on-disk inode 3350 * might have fragments that were not the last block in the file 3351 * which would corrupt the filesystem. 3352 */ 3353 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3354 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 3355 if (adp->ad_lbn >= NDADDR) 3356 break; 3357 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 3358 /* keep going until hitting a rollback to a frag */ 3359 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 3360 continue; 3361 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 3362 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 3363 #ifdef DIAGNOSTIC 3364 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) { 3365 FREE_LOCK(&lk); 3366 panic("softdep_write_inodeblock: lost dep1"); 3367 } 3368 #endif /* DIAGNOSTIC */ 3369 dp->di_db[i] = 0; 3370 } 3371 for (i = 0; i < NIADDR; i++) { 3372 #ifdef DIAGNOSTIC 3373 if (dp->di_ib[i] != 0 && 3374 (deplist & ((1 << NDADDR) << i)) == 0) { 3375 FREE_LOCK(&lk); 3376 panic("softdep_write_inodeblock: lost dep2"); 3377 } 3378 #endif /* DIAGNOSTIC */ 3379 dp->di_ib[i] = 0; 3380 } 3381 FREE_LOCK(&lk); 3382 return; 3383 } 3384 /* 3385 * If we have zero'ed out the last allocated block of the file, 3386 * roll back the size to the last currently allocated block. 3387 * We know that this last allocated block is a full-sized as 3388 * we already checked for fragments in the loop above. 3389 */ 3390 if (lastadp != NULL && 3391 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 3392 for (i = lastadp->ad_lbn; i >= 0; i--) 3393 if (dp->di_db[i] != 0) 3394 break; 3395 dp->di_size = (i + 1) * fs->fs_bsize; 3396 } 3397 /* 3398 * The only dependencies are for indirect blocks. 3399 * 3400 * The file size for indirect block additions is not guaranteed. 3401 * Such a guarantee would be non-trivial to achieve. The conventional 3402 * synchronous write implementation also does not make this guarantee. 3403 * Fsck should catch and fix discrepancies. Arguably, the file size 3404 * can be over-estimated without destroying integrity when the file 3405 * moves into the indirect blocks (i.e., is large). If we want to 3406 * postpone fsck, we are stuck with this argument. 3407 */ 3408 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 3409 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 3410 FREE_LOCK(&lk); 3411 } 3412 3413 /* 3414 * This routine is called during the completion interrupt 3415 * service routine for a disk write (from the procedure called 3416 * by the device driver to inform the filesystem caches of 3417 * a request completion). It should be called early in this 3418 * procedure, before the block is made available to other 3419 * processes or other routines are called. 3420 * 3421 * bioops callback - hold io_token 3422 * 3423 * Parameters: 3424 * bp: describes the completed disk write 3425 */ 3426 static void 3427 softdep_disk_write_complete(struct buf *bp) 3428 { 3429 struct worklist *wk; 3430 struct workhead reattach; 3431 struct newblk *newblk; 3432 struct allocindir *aip; 3433 struct allocdirect *adp; 3434 struct indirdep *indirdep; 3435 struct inodedep *inodedep; 3436 struct bmsafemap *bmsafemap; 3437 3438 get_mplock(); 3439 #ifdef DEBUG 3440 if (lk.lkt_held != NOHOLDER) 3441 panic("softdep_disk_write_complete: lock is held"); 3442 lk.lkt_held = SPECIAL_FLAG; 3443 #endif 3444 LIST_INIT(&reattach); 3445 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 3446 WORKLIST_REMOVE(wk); 3447 switch (wk->wk_type) { 3448 3449 case D_PAGEDEP: 3450 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 3451 WORKLIST_INSERT(&reattach, wk); 3452 continue; 3453 3454 case D_INODEDEP: 3455 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 3456 WORKLIST_INSERT(&reattach, wk); 3457 continue; 3458 3459 case D_BMSAFEMAP: 3460 bmsafemap = WK_BMSAFEMAP(wk); 3461 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) { 3462 newblk->nb_state |= DEPCOMPLETE; 3463 newblk->nb_bmsafemap = NULL; 3464 LIST_REMOVE(newblk, nb_deps); 3465 } 3466 while ((adp = 3467 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) { 3468 adp->ad_state |= DEPCOMPLETE; 3469 adp->ad_buf = NULL; 3470 LIST_REMOVE(adp, ad_deps); 3471 handle_allocdirect_partdone(adp); 3472 } 3473 while ((aip = 3474 LIST_FIRST(&bmsafemap->sm_allocindirhd))) { 3475 aip->ai_state |= DEPCOMPLETE; 3476 aip->ai_buf = NULL; 3477 LIST_REMOVE(aip, ai_deps); 3478 handle_allocindir_partdone(aip); 3479 } 3480 while ((inodedep = 3481 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { 3482 inodedep->id_state |= DEPCOMPLETE; 3483 LIST_REMOVE(inodedep, id_deps); 3484 inodedep->id_buf = NULL; 3485 } 3486 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 3487 continue; 3488 3489 case D_MKDIR: 3490 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 3491 continue; 3492 3493 case D_ALLOCDIRECT: 3494 adp = WK_ALLOCDIRECT(wk); 3495 adp->ad_state |= COMPLETE; 3496 handle_allocdirect_partdone(adp); 3497 continue; 3498 3499 case D_ALLOCINDIR: 3500 aip = WK_ALLOCINDIR(wk); 3501 aip->ai_state |= COMPLETE; 3502 handle_allocindir_partdone(aip); 3503 continue; 3504 3505 case D_INDIRDEP: 3506 indirdep = WK_INDIRDEP(wk); 3507 if (indirdep->ir_state & GOINGAWAY) { 3508 lk.lkt_held = NOHOLDER; 3509 panic("disk_write_complete: indirdep gone"); 3510 } 3511 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 3512 FREE(indirdep->ir_saveddata, M_INDIRDEP); 3513 indirdep->ir_saveddata = 0; 3514 indirdep->ir_state &= ~UNDONE; 3515 indirdep->ir_state |= ATTACHED; 3516 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 3517 handle_allocindir_partdone(aip); 3518 if (aip == LIST_FIRST(&indirdep->ir_donehd)) { 3519 lk.lkt_held = NOHOLDER; 3520 panic("disk_write_complete: not gone"); 3521 } 3522 } 3523 WORKLIST_INSERT(&reattach, wk); 3524 if ((bp->b_flags & B_DELWRI) == 0) 3525 stat_indir_blk_ptrs++; 3526 bdirty(bp); 3527 continue; 3528 3529 default: 3530 lk.lkt_held = NOHOLDER; 3531 panic("handle_disk_write_complete: Unknown type %s", 3532 TYPENAME(wk->wk_type)); 3533 /* NOTREACHED */ 3534 } 3535 } 3536 /* 3537 * Reattach any requests that must be redone. 3538 */ 3539 while ((wk = LIST_FIRST(&reattach)) != NULL) { 3540 WORKLIST_REMOVE(wk); 3541 WORKLIST_INSERT_BP(bp, wk); 3542 } 3543 #ifdef DEBUG 3544 if (lk.lkt_held != SPECIAL_FLAG) 3545 panic("softdep_disk_write_complete: lock lost"); 3546 lk.lkt_held = NOHOLDER; 3547 #endif 3548 rel_mplock(); 3549 } 3550 3551 /* 3552 * Called from within softdep_disk_write_complete above. Note that 3553 * this routine is always called from interrupt level with further 3554 * splbio interrupts blocked. 3555 * 3556 * Parameters: 3557 * adp: the completed allocdirect 3558 */ 3559 static void 3560 handle_allocdirect_partdone(struct allocdirect *adp) 3561 { 3562 struct allocdirect *listadp; 3563 struct inodedep *inodedep; 3564 long bsize; 3565 3566 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3567 return; 3568 if (adp->ad_buf != NULL) { 3569 lk.lkt_held = NOHOLDER; 3570 panic("handle_allocdirect_partdone: dangling dep"); 3571 } 3572 /* 3573 * The on-disk inode cannot claim to be any larger than the last 3574 * fragment that has been written. Otherwise, the on-disk inode 3575 * might have fragments that were not the last block in the file 3576 * which would corrupt the filesystem. Thus, we cannot free any 3577 * allocdirects after one whose ad_oldblkno claims a fragment as 3578 * these blocks must be rolled back to zero before writing the inode. 3579 * We check the currently active set of allocdirects in id_inoupdt. 3580 */ 3581 inodedep = adp->ad_inodedep; 3582 bsize = inodedep->id_fs->fs_bsize; 3583 TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) { 3584 /* found our block */ 3585 if (listadp == adp) 3586 break; 3587 /* continue if ad_oldlbn is not a fragment */ 3588 if (listadp->ad_oldsize == 0 || 3589 listadp->ad_oldsize == bsize) 3590 continue; 3591 /* hit a fragment */ 3592 return; 3593 } 3594 /* 3595 * If we have reached the end of the current list without 3596 * finding the just finished dependency, then it must be 3597 * on the future dependency list. Future dependencies cannot 3598 * be freed until they are moved to the current list. 3599 */ 3600 if (listadp == NULL) { 3601 #ifdef DEBUG 3602 TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next) 3603 /* found our block */ 3604 if (listadp == adp) 3605 break; 3606 if (listadp == NULL) { 3607 lk.lkt_held = NOHOLDER; 3608 panic("handle_allocdirect_partdone: lost dep"); 3609 } 3610 #endif /* DEBUG */ 3611 return; 3612 } 3613 /* 3614 * If we have found the just finished dependency, then free 3615 * it along with anything that follows it that is complete. 3616 */ 3617 for (; adp; adp = listadp) { 3618 listadp = TAILQ_NEXT(adp, ad_next); 3619 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3620 return; 3621 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 3622 } 3623 } 3624 3625 /* 3626 * Called from within softdep_disk_write_complete above. Note that 3627 * this routine is always called from interrupt level with further 3628 * splbio interrupts blocked. 3629 * 3630 * Parameters: 3631 * aip: the completed allocindir 3632 */ 3633 static void 3634 handle_allocindir_partdone(struct allocindir *aip) 3635 { 3636 struct indirdep *indirdep; 3637 3638 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 3639 return; 3640 if (aip->ai_buf != NULL) { 3641 lk.lkt_held = NOHOLDER; 3642 panic("handle_allocindir_partdone: dangling dependency"); 3643 } 3644 indirdep = aip->ai_indirdep; 3645 if (indirdep->ir_state & UNDONE) { 3646 LIST_REMOVE(aip, ai_next); 3647 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 3648 return; 3649 } 3650 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 3651 aip->ai_newblkno; 3652 LIST_REMOVE(aip, ai_next); 3653 if (aip->ai_freefrag != NULL) 3654 add_to_worklist(&aip->ai_freefrag->ff_list); 3655 WORKITEM_FREE(aip, D_ALLOCINDIR); 3656 } 3657 3658 /* 3659 * Called from within softdep_disk_write_complete above to restore 3660 * in-memory inode block contents to their most up-to-date state. Note 3661 * that this routine is always called from interrupt level with further 3662 * splbio interrupts blocked. 3663 * 3664 * Parameters: 3665 * bp: buffer containing the inode block 3666 */ 3667 static int 3668 handle_written_inodeblock(struct inodedep *inodedep, struct buf *bp) 3669 { 3670 struct worklist *wk, *filefree; 3671 struct allocdirect *adp, *nextadp; 3672 struct ufs1_dinode *dp; 3673 int hadchanges; 3674 3675 if ((inodedep->id_state & IOSTARTED) == 0) { 3676 lk.lkt_held = NOHOLDER; 3677 panic("handle_written_inodeblock: not started"); 3678 } 3679 inodedep->id_state &= ~IOSTARTED; 3680 dp = (struct ufs1_dinode *)bp->b_data + 3681 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 3682 /* 3683 * If we had to rollback the inode allocation because of 3684 * bitmaps being incomplete, then simply restore it. 3685 * Keep the block dirty so that it will not be reclaimed until 3686 * all associated dependencies have been cleared and the 3687 * corresponding updates written to disk. 3688 */ 3689 if (inodedep->id_savedino != NULL) { 3690 *dp = *inodedep->id_savedino; 3691 FREE(inodedep->id_savedino, M_INODEDEP); 3692 inodedep->id_savedino = NULL; 3693 if ((bp->b_flags & B_DELWRI) == 0) 3694 stat_inode_bitmap++; 3695 bdirty(bp); 3696 return (1); 3697 } 3698 inodedep->id_state |= COMPLETE; 3699 /* 3700 * Roll forward anything that had to be rolled back before 3701 * the inode could be updated. 3702 */ 3703 hadchanges = 0; 3704 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 3705 nextadp = TAILQ_NEXT(adp, ad_next); 3706 if (adp->ad_state & ATTACHED) { 3707 lk.lkt_held = NOHOLDER; 3708 panic("handle_written_inodeblock: new entry"); 3709 } 3710 if (adp->ad_lbn < NDADDR) { 3711 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) { 3712 lk.lkt_held = NOHOLDER; 3713 panic("%s: %s #%ld mismatch %d != %d", 3714 "handle_written_inodeblock", 3715 "direct pointer", adp->ad_lbn, 3716 dp->di_db[adp->ad_lbn], adp->ad_oldblkno); 3717 } 3718 dp->di_db[adp->ad_lbn] = adp->ad_newblkno; 3719 } else { 3720 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) { 3721 lk.lkt_held = NOHOLDER; 3722 panic("%s: %s #%ld allocated as %d", 3723 "handle_written_inodeblock", 3724 "indirect pointer", adp->ad_lbn - NDADDR, 3725 dp->di_ib[adp->ad_lbn - NDADDR]); 3726 } 3727 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno; 3728 } 3729 adp->ad_state &= ~UNDONE; 3730 adp->ad_state |= ATTACHED; 3731 hadchanges = 1; 3732 } 3733 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 3734 stat_direct_blk_ptrs++; 3735 /* 3736 * Reset the file size to its most up-to-date value. 3737 */ 3738 if (inodedep->id_savedsize == -1) { 3739 lk.lkt_held = NOHOLDER; 3740 panic("handle_written_inodeblock: bad size"); 3741 } 3742 if (dp->di_size != inodedep->id_savedsize) { 3743 dp->di_size = inodedep->id_savedsize; 3744 hadchanges = 1; 3745 } 3746 inodedep->id_savedsize = -1; 3747 /* 3748 * If there were any rollbacks in the inode block, then it must be 3749 * marked dirty so that its will eventually get written back in 3750 * its correct form. 3751 */ 3752 if (hadchanges) 3753 bdirty(bp); 3754 /* 3755 * Process any allocdirects that completed during the update. 3756 */ 3757 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 3758 handle_allocdirect_partdone(adp); 3759 /* 3760 * Process deallocations that were held pending until the 3761 * inode had been written to disk. Freeing of the inode 3762 * is delayed until after all blocks have been freed to 3763 * avoid creation of new <vfsid, inum, lbn> triples 3764 * before the old ones have been deleted. 3765 */ 3766 filefree = NULL; 3767 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 3768 WORKLIST_REMOVE(wk); 3769 switch (wk->wk_type) { 3770 3771 case D_FREEFILE: 3772 /* 3773 * We defer adding filefree to the worklist until 3774 * all other additions have been made to ensure 3775 * that it will be done after all the old blocks 3776 * have been freed. 3777 */ 3778 if (filefree != NULL) { 3779 lk.lkt_held = NOHOLDER; 3780 panic("handle_written_inodeblock: filefree"); 3781 } 3782 filefree = wk; 3783 continue; 3784 3785 case D_MKDIR: 3786 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 3787 continue; 3788 3789 case D_DIRADD: 3790 diradd_inode_written(WK_DIRADD(wk), inodedep); 3791 continue; 3792 3793 case D_FREEBLKS: 3794 wk->wk_state |= COMPLETE; 3795 if ((wk->wk_state & ALLCOMPLETE) != ALLCOMPLETE) 3796 continue; 3797 /* -- fall through -- */ 3798 case D_FREEFRAG: 3799 case D_DIRREM: 3800 add_to_worklist(wk); 3801 continue; 3802 3803 default: 3804 lk.lkt_held = NOHOLDER; 3805 panic("handle_written_inodeblock: Unknown type %s", 3806 TYPENAME(wk->wk_type)); 3807 /* NOTREACHED */ 3808 } 3809 } 3810 if (filefree != NULL) { 3811 if (free_inodedep(inodedep) == 0) { 3812 lk.lkt_held = NOHOLDER; 3813 panic("handle_written_inodeblock: live inodedep"); 3814 } 3815 add_to_worklist(filefree); 3816 return (0); 3817 } 3818 3819 /* 3820 * If no outstanding dependencies, free it. 3821 */ 3822 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0) 3823 return (0); 3824 return (hadchanges); 3825 } 3826 3827 /* 3828 * Process a diradd entry after its dependent inode has been written. 3829 * This routine must be called with splbio interrupts blocked. 3830 */ 3831 static void 3832 diradd_inode_written(struct diradd *dap, struct inodedep *inodedep) 3833 { 3834 struct pagedep *pagedep; 3835 3836 dap->da_state |= COMPLETE; 3837 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3838 if (dap->da_state & DIRCHG) 3839 pagedep = dap->da_previous->dm_pagedep; 3840 else 3841 pagedep = dap->da_pagedep; 3842 LIST_REMOVE(dap, da_pdlist); 3843 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3844 } 3845 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 3846 } 3847 3848 /* 3849 * Handle the completion of a mkdir dependency. 3850 */ 3851 static void 3852 handle_written_mkdir(struct mkdir *mkdir, int type) 3853 { 3854 struct diradd *dap; 3855 struct pagedep *pagedep; 3856 3857 if (mkdir->md_state != type) { 3858 lk.lkt_held = NOHOLDER; 3859 panic("handle_written_mkdir: bad type"); 3860 } 3861 dap = mkdir->md_diradd; 3862 dap->da_state &= ~type; 3863 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 3864 dap->da_state |= DEPCOMPLETE; 3865 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3866 if (dap->da_state & DIRCHG) 3867 pagedep = dap->da_previous->dm_pagedep; 3868 else 3869 pagedep = dap->da_pagedep; 3870 LIST_REMOVE(dap, da_pdlist); 3871 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3872 } 3873 LIST_REMOVE(mkdir, md_mkdirs); 3874 WORKITEM_FREE(mkdir, D_MKDIR); 3875 } 3876 3877 /* 3878 * Called from within softdep_disk_write_complete above. 3879 * A write operation was just completed. Removed inodes can 3880 * now be freed and associated block pointers may be committed. 3881 * Note that this routine is always called from interrupt level 3882 * with further splbio interrupts blocked. 3883 * 3884 * Parameters: 3885 * bp: buffer containing the written page 3886 */ 3887 static int 3888 handle_written_filepage(struct pagedep *pagedep, struct buf *bp) 3889 { 3890 struct dirrem *dirrem; 3891 struct diradd *dap, *nextdap; 3892 struct direct *ep; 3893 int i, chgs; 3894 3895 if ((pagedep->pd_state & IOSTARTED) == 0) { 3896 lk.lkt_held = NOHOLDER; 3897 panic("handle_written_filepage: not started"); 3898 } 3899 pagedep->pd_state &= ~IOSTARTED; 3900 /* 3901 * Process any directory removals that have been committed. 3902 */ 3903 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 3904 LIST_REMOVE(dirrem, dm_next); 3905 dirrem->dm_dirinum = pagedep->pd_ino; 3906 add_to_worklist(&dirrem->dm_list); 3907 } 3908 /* 3909 * Free any directory additions that have been committed. 3910 */ 3911 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 3912 free_diradd(dap); 3913 /* 3914 * Uncommitted directory entries must be restored. 3915 */ 3916 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 3917 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 3918 dap = nextdap) { 3919 nextdap = LIST_NEXT(dap, da_pdlist); 3920 if (dap->da_state & ATTACHED) { 3921 lk.lkt_held = NOHOLDER; 3922 panic("handle_written_filepage: attached"); 3923 } 3924 ep = (struct direct *) 3925 ((char *)bp->b_data + dap->da_offset); 3926 ep->d_ino = dap->da_newinum; 3927 dap->da_state &= ~UNDONE; 3928 dap->da_state |= ATTACHED; 3929 chgs = 1; 3930 /* 3931 * If the inode referenced by the directory has 3932 * been written out, then the dependency can be 3933 * moved to the pending list. 3934 */ 3935 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3936 LIST_REMOVE(dap, da_pdlist); 3937 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 3938 da_pdlist); 3939 } 3940 } 3941 } 3942 /* 3943 * If there were any rollbacks in the directory, then it must be 3944 * marked dirty so that its will eventually get written back in 3945 * its correct form. 3946 */ 3947 if (chgs) { 3948 if ((bp->b_flags & B_DELWRI) == 0) 3949 stat_dir_entry++; 3950 bdirty(bp); 3951 } 3952 /* 3953 * If no dependencies remain, the pagedep will be freed. 3954 * Otherwise it will remain to update the page before it 3955 * is written back to disk. 3956 */ 3957 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) { 3958 for (i = 0; i < DAHASHSZ; i++) 3959 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL) 3960 break; 3961 if (i == DAHASHSZ) { 3962 LIST_REMOVE(pagedep, pd_hash); 3963 WORKITEM_FREE(pagedep, D_PAGEDEP); 3964 return (0); 3965 } 3966 } 3967 return (1); 3968 } 3969 3970 /* 3971 * Writing back in-core inode structures. 3972 * 3973 * The filesystem only accesses an inode's contents when it occupies an 3974 * "in-core" inode structure. These "in-core" structures are separate from 3975 * the page frames used to cache inode blocks. Only the latter are 3976 * transferred to/from the disk. So, when the updated contents of the 3977 * "in-core" inode structure are copied to the corresponding in-memory inode 3978 * block, the dependencies are also transferred. The following procedure is 3979 * called when copying a dirty "in-core" inode to a cached inode block. 3980 */ 3981 3982 /* 3983 * Called when an inode is loaded from disk. If the effective link count 3984 * differed from the actual link count when it was last flushed, then we 3985 * need to ensure that the correct effective link count is put back. 3986 * 3987 * Parameters: 3988 * ip: the "in_core" copy of the inode 3989 */ 3990 void 3991 softdep_load_inodeblock(struct inode *ip) 3992 { 3993 struct inodedep *inodedep; 3994 3995 /* 3996 * Check for alternate nlink count. 3997 */ 3998 ip->i_effnlink = ip->i_nlink; 3999 ACQUIRE_LOCK(&lk); 4000 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 4001 FREE_LOCK(&lk); 4002 return; 4003 } 4004 ip->i_effnlink -= inodedep->id_nlinkdelta; 4005 FREE_LOCK(&lk); 4006 } 4007 4008 /* 4009 * This routine is called just before the "in-core" inode 4010 * information is to be copied to the in-memory inode block. 4011 * Recall that an inode block contains several inodes. If 4012 * the force flag is set, then the dependencies will be 4013 * cleared so that the update can always be made. Note that 4014 * the buffer is locked when this routine is called, so we 4015 * will never be in the middle of writing the inode block 4016 * to disk. 4017 * 4018 * Parameters: 4019 * ip: the "in_core" copy of the inode 4020 * bp: the buffer containing the inode block 4021 * waitfor: nonzero => update must be allowed 4022 */ 4023 void 4024 softdep_update_inodeblock(struct inode *ip, struct buf *bp, 4025 int waitfor) 4026 { 4027 struct inodedep *inodedep; 4028 struct worklist *wk; 4029 int error, gotit; 4030 4031 /* 4032 * If the effective link count is not equal to the actual link 4033 * count, then we must track the difference in an inodedep while 4034 * the inode is (potentially) tossed out of the cache. Otherwise, 4035 * if there is no existing inodedep, then there are no dependencies 4036 * to track. 4037 */ 4038 ACQUIRE_LOCK(&lk); 4039 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 4040 FREE_LOCK(&lk); 4041 if (ip->i_effnlink != ip->i_nlink) 4042 panic("softdep_update_inodeblock: bad link count"); 4043 return; 4044 } 4045 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) { 4046 FREE_LOCK(&lk); 4047 panic("softdep_update_inodeblock: bad delta"); 4048 } 4049 /* 4050 * Changes have been initiated. Anything depending on these 4051 * changes cannot occur until this inode has been written. 4052 */ 4053 inodedep->id_state &= ~COMPLETE; 4054 if ((inodedep->id_state & ONWORKLIST) == 0) 4055 WORKLIST_INSERT_BP(bp, &inodedep->id_list); 4056 /* 4057 * Any new dependencies associated with the incore inode must 4058 * now be moved to the list associated with the buffer holding 4059 * the in-memory copy of the inode. Once merged process any 4060 * allocdirects that are completed by the merger. 4061 */ 4062 merge_inode_lists(inodedep); 4063 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL) 4064 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); 4065 /* 4066 * Now that the inode has been pushed into the buffer, the 4067 * operations dependent on the inode being written to disk 4068 * can be moved to the id_bufwait so that they will be 4069 * processed when the buffer I/O completes. 4070 */ 4071 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 4072 WORKLIST_REMOVE(wk); 4073 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 4074 } 4075 /* 4076 * Newly allocated inodes cannot be written until the bitmap 4077 * that allocates them have been written (indicated by 4078 * DEPCOMPLETE being set in id_state). If we are doing a 4079 * forced sync (e.g., an fsync on a file), we force the bitmap 4080 * to be written so that the update can be done. 4081 */ 4082 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) { 4083 FREE_LOCK(&lk); 4084 return; 4085 } 4086 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 4087 FREE_LOCK(&lk); 4088 if (gotit && 4089 (error = bwrite(inodedep->id_buf)) != 0) 4090 softdep_error("softdep_update_inodeblock: bwrite", error); 4091 } 4092 4093 /* 4094 * Merge the new inode dependency list (id_newinoupdt) into the old 4095 * inode dependency list (id_inoupdt). This routine must be called 4096 * with splbio interrupts blocked. 4097 */ 4098 static void 4099 merge_inode_lists(struct inodedep *inodedep) 4100 { 4101 struct allocdirect *listadp, *newadp; 4102 4103 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 4104 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) { 4105 if (listadp->ad_lbn < newadp->ad_lbn) { 4106 listadp = TAILQ_NEXT(listadp, ad_next); 4107 continue; 4108 } 4109 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 4110 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 4111 if (listadp->ad_lbn == newadp->ad_lbn) { 4112 allocdirect_merge(&inodedep->id_inoupdt, newadp, 4113 listadp); 4114 listadp = newadp; 4115 } 4116 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 4117 } 4118 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) { 4119 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 4120 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next); 4121 } 4122 } 4123 4124 /* 4125 * If we are doing an fsync, then we must ensure that any directory 4126 * entries for the inode have been written after the inode gets to disk. 4127 * 4128 * bioops callback - hold io_token 4129 * 4130 * Parameters: 4131 * vp: the "in_core" copy of the inode 4132 */ 4133 static int 4134 softdep_fsync(struct vnode *vp) 4135 { 4136 struct inodedep *inodedep; 4137 struct pagedep *pagedep; 4138 struct worklist *wk; 4139 struct diradd *dap; 4140 struct mount *mnt; 4141 struct vnode *pvp; 4142 struct inode *ip; 4143 struct buf *bp; 4144 struct fs *fs; 4145 int error, flushparent; 4146 ino_t parentino; 4147 ufs_lbn_t lbn; 4148 4149 /* 4150 * Move check from original kernel code, possibly not needed any 4151 * more with the per-mount bioops. 4152 */ 4153 if ((vp->v_mount->mnt_flag & MNT_SOFTDEP) == 0) 4154 return (0); 4155 4156 get_mplock(); 4157 ip = VTOI(vp); 4158 fs = ip->i_fs; 4159 ACQUIRE_LOCK(&lk); 4160 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) { 4161 FREE_LOCK(&lk); 4162 rel_mplock(); 4163 return (0); 4164 } 4165 if (LIST_FIRST(&inodedep->id_inowait) != NULL || 4166 LIST_FIRST(&inodedep->id_bufwait) != NULL || 4167 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 4168 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) { 4169 FREE_LOCK(&lk); 4170 panic("softdep_fsync: pending ops"); 4171 } 4172 for (error = 0, flushparent = 0; ; ) { 4173 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 4174 break; 4175 if (wk->wk_type != D_DIRADD) { 4176 FREE_LOCK(&lk); 4177 panic("softdep_fsync: Unexpected type %s", 4178 TYPENAME(wk->wk_type)); 4179 } 4180 dap = WK_DIRADD(wk); 4181 /* 4182 * Flush our parent if this directory entry 4183 * has a MKDIR_PARENT dependency. 4184 */ 4185 if (dap->da_state & DIRCHG) 4186 pagedep = dap->da_previous->dm_pagedep; 4187 else 4188 pagedep = dap->da_pagedep; 4189 mnt = pagedep->pd_mnt; 4190 parentino = pagedep->pd_ino; 4191 lbn = pagedep->pd_lbn; 4192 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) { 4193 FREE_LOCK(&lk); 4194 panic("softdep_fsync: dirty"); 4195 } 4196 flushparent = dap->da_state & MKDIR_PARENT; 4197 /* 4198 * If we are being fsync'ed as part of vgone'ing this vnode, 4199 * then we will not be able to release and recover the 4200 * vnode below, so we just have to give up on writing its 4201 * directory entry out. It will eventually be written, just 4202 * not now, but then the user was not asking to have it 4203 * written, so we are not breaking any promises. 4204 */ 4205 if (vp->v_flag & VRECLAIMED) 4206 break; 4207 /* 4208 * We prevent deadlock by always fetching inodes from the 4209 * root, moving down the directory tree. Thus, when fetching 4210 * our parent directory, we must unlock ourselves before 4211 * requesting the lock on our parent. See the comment in 4212 * ufs_lookup for details on possible races. 4213 */ 4214 FREE_LOCK(&lk); 4215 vn_unlock(vp); 4216 error = VFS_VGET(mnt, NULL, parentino, &pvp); 4217 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4218 if (error != 0) { 4219 rel_mplock(); 4220 return (error); 4221 } 4222 if (flushparent) { 4223 if ((error = ffs_update(pvp, 1)) != 0) { 4224 vput(pvp); 4225 rel_mplock(); 4226 return (error); 4227 } 4228 } 4229 /* 4230 * Flush directory page containing the inode's name. 4231 */ 4232 error = bread(pvp, lblktodoff(fs, lbn), blksize(fs, VTOI(pvp), lbn), &bp); 4233 if (error == 0) 4234 error = bwrite(bp); 4235 vput(pvp); 4236 if (error != 0) { 4237 rel_mplock(); 4238 return (error); 4239 } 4240 ACQUIRE_LOCK(&lk); 4241 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) 4242 break; 4243 } 4244 FREE_LOCK(&lk); 4245 rel_mplock(); 4246 return (0); 4247 } 4248 4249 /* 4250 * Flush all the dirty bitmaps associated with the block device 4251 * before flushing the rest of the dirty blocks so as to reduce 4252 * the number of dependencies that will have to be rolled back. 4253 */ 4254 static int softdep_fsync_mountdev_bp(struct buf *bp, void *data); 4255 4256 void 4257 softdep_fsync_mountdev(struct vnode *vp) 4258 { 4259 if (!vn_isdisk(vp, NULL)) 4260 panic("softdep_fsync_mountdev: vnode not a disk"); 4261 ACQUIRE_LOCK(&lk); 4262 lwkt_gettoken(&vp->v_token); 4263 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 4264 softdep_fsync_mountdev_bp, vp); 4265 lwkt_reltoken(&vp->v_token); 4266 drain_output(vp, 1); 4267 FREE_LOCK(&lk); 4268 } 4269 4270 static int 4271 softdep_fsync_mountdev_bp(struct buf *bp, void *data) 4272 { 4273 struct worklist *wk; 4274 struct vnode *vp = data; 4275 4276 /* 4277 * If it is already scheduled, skip to the next buffer. 4278 */ 4279 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) 4280 return(0); 4281 if (bp->b_vp != vp || (bp->b_flags & B_DELWRI) == 0) { 4282 BUF_UNLOCK(bp); 4283 kprintf("softdep_fsync_mountdev_bp: warning, buffer %p ripped out from under vnode %p\n", bp, vp); 4284 return(0); 4285 } 4286 /* 4287 * We are only interested in bitmaps with outstanding 4288 * dependencies. 4289 */ 4290 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 4291 wk->wk_type != D_BMSAFEMAP) { 4292 BUF_UNLOCK(bp); 4293 return(0); 4294 } 4295 bremfree(bp); 4296 FREE_LOCK(&lk); 4297 (void) bawrite(bp); 4298 ACQUIRE_LOCK(&lk); 4299 return(0); 4300 } 4301 4302 /* 4303 * This routine is called when we are trying to synchronously flush a 4304 * file. This routine must eliminate any filesystem metadata dependencies 4305 * so that the syncing routine can succeed by pushing the dirty blocks 4306 * associated with the file. If any I/O errors occur, they are returned. 4307 */ 4308 struct softdep_sync_metadata_info { 4309 struct vnode *vp; 4310 int waitfor; 4311 }; 4312 4313 static int softdep_sync_metadata_bp(struct buf *bp, void *data); 4314 4315 int 4316 softdep_sync_metadata(struct vnode *vp, struct thread *td) 4317 { 4318 struct softdep_sync_metadata_info info; 4319 int error, waitfor; 4320 4321 /* 4322 * Check whether this vnode is involved in a filesystem 4323 * that is doing soft dependency processing. 4324 */ 4325 if (!vn_isdisk(vp, NULL)) { 4326 if (!DOINGSOFTDEP(vp)) 4327 return (0); 4328 } else 4329 if (vp->v_rdev->si_mountpoint == NULL || 4330 (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0) 4331 return (0); 4332 /* 4333 * Ensure that any direct block dependencies have been cleared. 4334 */ 4335 ACQUIRE_LOCK(&lk); 4336 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) { 4337 FREE_LOCK(&lk); 4338 return (error); 4339 } 4340 /* 4341 * For most files, the only metadata dependencies are the 4342 * cylinder group maps that allocate their inode or blocks. 4343 * The block allocation dependencies can be found by traversing 4344 * the dependency lists for any buffers that remain on their 4345 * dirty buffer list. The inode allocation dependency will 4346 * be resolved when the inode is updated with MNT_WAIT. 4347 * This work is done in two passes. The first pass grabs most 4348 * of the buffers and begins asynchronously writing them. The 4349 * only way to wait for these asynchronous writes is to sleep 4350 * on the filesystem vnode which may stay busy for a long time 4351 * if the filesystem is active. So, instead, we make a second 4352 * pass over the dependencies blocking on each write. In the 4353 * usual case we will be blocking against a write that we 4354 * initiated, so when it is done the dependency will have been 4355 * resolved. Thus the second pass is expected to end quickly. 4356 */ 4357 waitfor = MNT_NOWAIT; 4358 top: 4359 /* 4360 * We must wait for any I/O in progress to finish so that 4361 * all potential buffers on the dirty list will be visible. 4362 */ 4363 drain_output(vp, 1); 4364 4365 info.vp = vp; 4366 info.waitfor = waitfor; 4367 lwkt_gettoken(&vp->v_token); 4368 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 4369 softdep_sync_metadata_bp, &info); 4370 lwkt_reltoken(&vp->v_token); 4371 if (error < 0) { 4372 FREE_LOCK(&lk); 4373 return(-error); /* error code */ 4374 } 4375 4376 /* 4377 * The brief unlock is to allow any pent up dependency 4378 * processing to be done. Then proceed with the second pass. 4379 */ 4380 if (waitfor & MNT_NOWAIT) { 4381 waitfor = MNT_WAIT; 4382 FREE_LOCK(&lk); 4383 ACQUIRE_LOCK(&lk); 4384 goto top; 4385 } 4386 4387 /* 4388 * If we have managed to get rid of all the dirty buffers, 4389 * then we are done. For certain directories and block 4390 * devices, we may need to do further work. 4391 * 4392 * We must wait for any I/O in progress to finish so that 4393 * all potential buffers on the dirty list will be visible. 4394 */ 4395 drain_output(vp, 1); 4396 if (RB_EMPTY(&vp->v_rbdirty_tree)) { 4397 FREE_LOCK(&lk); 4398 return (0); 4399 } 4400 4401 FREE_LOCK(&lk); 4402 /* 4403 * If we are trying to sync a block device, some of its buffers may 4404 * contain metadata that cannot be written until the contents of some 4405 * partially written files have been written to disk. The only easy 4406 * way to accomplish this is to sync the entire filesystem (luckily 4407 * this happens rarely). 4408 */ 4409 if (vn_isdisk(vp, NULL) && 4410 vp->v_rdev && 4411 vp->v_rdev->si_mountpoint && !vn_islocked(vp) && 4412 (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT)) != 0) 4413 return (error); 4414 return (0); 4415 } 4416 4417 static int 4418 softdep_sync_metadata_bp(struct buf *bp, void *data) 4419 { 4420 struct softdep_sync_metadata_info *info = data; 4421 struct pagedep *pagedep; 4422 struct allocdirect *adp; 4423 struct allocindir *aip; 4424 struct worklist *wk; 4425 struct buf *nbp; 4426 int error; 4427 int i; 4428 4429 if (getdirtybuf(&bp, MNT_WAIT) == 0) { 4430 kprintf("softdep_sync_metadata_bp(1): caught buf %p going away\n", bp); 4431 return (1); 4432 } 4433 if (bp->b_vp != info->vp || (bp->b_flags & B_DELWRI) == 0) { 4434 kprintf("softdep_sync_metadata_bp(2): caught buf %p going away vp %p\n", bp, info->vp); 4435 BUF_UNLOCK(bp); 4436 return(1); 4437 } 4438 4439 /* 4440 * As we hold the buffer locked, none of its dependencies 4441 * will disappear. 4442 */ 4443 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 4444 switch (wk->wk_type) { 4445 4446 case D_ALLOCDIRECT: 4447 adp = WK_ALLOCDIRECT(wk); 4448 if (adp->ad_state & DEPCOMPLETE) 4449 break; 4450 nbp = adp->ad_buf; 4451 if (getdirtybuf(&nbp, info->waitfor) == 0) 4452 break; 4453 FREE_LOCK(&lk); 4454 if (info->waitfor & MNT_NOWAIT) { 4455 bawrite(nbp); 4456 } else if ((error = bwrite(nbp)) != 0) { 4457 bawrite(bp); 4458 ACQUIRE_LOCK(&lk); 4459 return (-error); 4460 } 4461 ACQUIRE_LOCK(&lk); 4462 break; 4463 4464 case D_ALLOCINDIR: 4465 aip = WK_ALLOCINDIR(wk); 4466 if (aip->ai_state & DEPCOMPLETE) 4467 break; 4468 nbp = aip->ai_buf; 4469 if (getdirtybuf(&nbp, info->waitfor) == 0) 4470 break; 4471 FREE_LOCK(&lk); 4472 if (info->waitfor & MNT_NOWAIT) { 4473 bawrite(nbp); 4474 } else if ((error = bwrite(nbp)) != 0) { 4475 bawrite(bp); 4476 ACQUIRE_LOCK(&lk); 4477 return (-error); 4478 } 4479 ACQUIRE_LOCK(&lk); 4480 break; 4481 4482 case D_INDIRDEP: 4483 restart: 4484 4485 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) { 4486 if (aip->ai_state & DEPCOMPLETE) 4487 continue; 4488 nbp = aip->ai_buf; 4489 if (getdirtybuf(&nbp, MNT_WAIT) == 0) 4490 goto restart; 4491 FREE_LOCK(&lk); 4492 if ((error = bwrite(nbp)) != 0) { 4493 bawrite(bp); 4494 ACQUIRE_LOCK(&lk); 4495 return (-error); 4496 } 4497 ACQUIRE_LOCK(&lk); 4498 goto restart; 4499 } 4500 break; 4501 4502 case D_INODEDEP: 4503 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs, 4504 WK_INODEDEP(wk)->id_ino)) != 0) { 4505 FREE_LOCK(&lk); 4506 bawrite(bp); 4507 ACQUIRE_LOCK(&lk); 4508 return (-error); 4509 } 4510 break; 4511 4512 case D_PAGEDEP: 4513 /* 4514 * We are trying to sync a directory that may 4515 * have dependencies on both its own metadata 4516 * and/or dependencies on the inodes of any 4517 * recently allocated files. We walk its diradd 4518 * lists pushing out the associated inode. 4519 */ 4520 pagedep = WK_PAGEDEP(wk); 4521 for (i = 0; i < DAHASHSZ; i++) { 4522 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 4523 continue; 4524 if ((error = 4525 flush_pagedep_deps(info->vp, 4526 pagedep->pd_mnt, 4527 &pagedep->pd_diraddhd[i]))) { 4528 FREE_LOCK(&lk); 4529 bawrite(bp); 4530 ACQUIRE_LOCK(&lk); 4531 return (-error); 4532 } 4533 } 4534 break; 4535 4536 case D_MKDIR: 4537 /* 4538 * This case should never happen if the vnode has 4539 * been properly sync'ed. However, if this function 4540 * is used at a place where the vnode has not yet 4541 * been sync'ed, this dependency can show up. So, 4542 * rather than panic, just flush it. 4543 */ 4544 nbp = WK_MKDIR(wk)->md_buf; 4545 if (getdirtybuf(&nbp, info->waitfor) == 0) 4546 break; 4547 FREE_LOCK(&lk); 4548 if (info->waitfor & MNT_NOWAIT) { 4549 bawrite(nbp); 4550 } else if ((error = bwrite(nbp)) != 0) { 4551 bawrite(bp); 4552 ACQUIRE_LOCK(&lk); 4553 return (-error); 4554 } 4555 ACQUIRE_LOCK(&lk); 4556 break; 4557 4558 case D_BMSAFEMAP: 4559 /* 4560 * This case should never happen if the vnode has 4561 * been properly sync'ed. However, if this function 4562 * is used at a place where the vnode has not yet 4563 * been sync'ed, this dependency can show up. So, 4564 * rather than panic, just flush it. 4565 * 4566 * nbp can wind up == bp if a device node for the 4567 * same filesystem is being fsynced at the same time, 4568 * leading to a panic if we don't catch the case. 4569 */ 4570 nbp = WK_BMSAFEMAP(wk)->sm_buf; 4571 if (nbp == bp) 4572 break; 4573 if (getdirtybuf(&nbp, info->waitfor) == 0) 4574 break; 4575 FREE_LOCK(&lk); 4576 if (info->waitfor & MNT_NOWAIT) { 4577 bawrite(nbp); 4578 } else if ((error = bwrite(nbp)) != 0) { 4579 bawrite(bp); 4580 ACQUIRE_LOCK(&lk); 4581 return (-error); 4582 } 4583 ACQUIRE_LOCK(&lk); 4584 break; 4585 4586 default: 4587 FREE_LOCK(&lk); 4588 panic("softdep_sync_metadata: Unknown type %s", 4589 TYPENAME(wk->wk_type)); 4590 /* NOTREACHED */ 4591 } 4592 } 4593 FREE_LOCK(&lk); 4594 bawrite(bp); 4595 ACQUIRE_LOCK(&lk); 4596 return(0); 4597 } 4598 4599 /* 4600 * Flush the dependencies associated with an inodedep. 4601 * Called with splbio blocked. 4602 */ 4603 static int 4604 flush_inodedep_deps(struct fs *fs, ino_t ino) 4605 { 4606 struct inodedep *inodedep; 4607 struct allocdirect *adp; 4608 int error, waitfor; 4609 struct buf *bp; 4610 4611 /* 4612 * This work is done in two passes. The first pass grabs most 4613 * of the buffers and begins asynchronously writing them. The 4614 * only way to wait for these asynchronous writes is to sleep 4615 * on the filesystem vnode which may stay busy for a long time 4616 * if the filesystem is active. So, instead, we make a second 4617 * pass over the dependencies blocking on each write. In the 4618 * usual case we will be blocking against a write that we 4619 * initiated, so when it is done the dependency will have been 4620 * resolved. Thus the second pass is expected to end quickly. 4621 * We give a brief window at the top of the loop to allow 4622 * any pending I/O to complete. 4623 */ 4624 for (waitfor = MNT_NOWAIT; ; ) { 4625 FREE_LOCK(&lk); 4626 ACQUIRE_LOCK(&lk); 4627 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4628 return (0); 4629 TAILQ_FOREACH(adp, &inodedep->id_inoupdt, ad_next) { 4630 if (adp->ad_state & DEPCOMPLETE) 4631 continue; 4632 bp = adp->ad_buf; 4633 if (getdirtybuf(&bp, waitfor) == 0) { 4634 if (waitfor & MNT_NOWAIT) 4635 continue; 4636 break; 4637 } 4638 FREE_LOCK(&lk); 4639 if (waitfor & MNT_NOWAIT) { 4640 bawrite(bp); 4641 } else if ((error = bwrite(bp)) != 0) { 4642 ACQUIRE_LOCK(&lk); 4643 return (error); 4644 } 4645 ACQUIRE_LOCK(&lk); 4646 break; 4647 } 4648 if (adp != NULL) 4649 continue; 4650 TAILQ_FOREACH(adp, &inodedep->id_newinoupdt, ad_next) { 4651 if (adp->ad_state & DEPCOMPLETE) 4652 continue; 4653 bp = adp->ad_buf; 4654 if (getdirtybuf(&bp, waitfor) == 0) { 4655 if (waitfor & MNT_NOWAIT) 4656 continue; 4657 break; 4658 } 4659 FREE_LOCK(&lk); 4660 if (waitfor & MNT_NOWAIT) { 4661 bawrite(bp); 4662 } else if ((error = bwrite(bp)) != 0) { 4663 ACQUIRE_LOCK(&lk); 4664 return (error); 4665 } 4666 ACQUIRE_LOCK(&lk); 4667 break; 4668 } 4669 if (adp != NULL) 4670 continue; 4671 /* 4672 * If pass2, we are done, otherwise do pass 2. 4673 */ 4674 if (waitfor == MNT_WAIT) 4675 break; 4676 waitfor = MNT_WAIT; 4677 } 4678 /* 4679 * Try freeing inodedep in case all dependencies have been removed. 4680 */ 4681 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0) 4682 (void) free_inodedep(inodedep); 4683 return (0); 4684 } 4685 4686 /* 4687 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 4688 * Called with splbio blocked. 4689 */ 4690 static int 4691 flush_pagedep_deps(struct vnode *pvp, struct mount *mp, 4692 struct diraddhd *diraddhdp) 4693 { 4694 struct inodedep *inodedep; 4695 struct ufsmount *ump; 4696 struct diradd *dap; 4697 struct vnode *vp; 4698 int gotit, error = 0; 4699 struct buf *bp; 4700 ino_t inum; 4701 4702 ump = VFSTOUFS(mp); 4703 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 4704 /* 4705 * Flush ourselves if this directory entry 4706 * has a MKDIR_PARENT dependency. 4707 */ 4708 if (dap->da_state & MKDIR_PARENT) { 4709 FREE_LOCK(&lk); 4710 if ((error = ffs_update(pvp, 1)) != 0) 4711 break; 4712 ACQUIRE_LOCK(&lk); 4713 /* 4714 * If that cleared dependencies, go on to next. 4715 */ 4716 if (dap != LIST_FIRST(diraddhdp)) 4717 continue; 4718 if (dap->da_state & MKDIR_PARENT) { 4719 FREE_LOCK(&lk); 4720 panic("flush_pagedep_deps: MKDIR_PARENT"); 4721 } 4722 } 4723 /* 4724 * A newly allocated directory must have its "." and 4725 * ".." entries written out before its name can be 4726 * committed in its parent. We do not want or need 4727 * the full semantics of a synchronous VOP_FSYNC as 4728 * that may end up here again, once for each directory 4729 * level in the filesystem. Instead, we push the blocks 4730 * and wait for them to clear. We have to fsync twice 4731 * because the first call may choose to defer blocks 4732 * that still have dependencies, but deferral will 4733 * happen at most once. 4734 */ 4735 inum = dap->da_newinum; 4736 if (dap->da_state & MKDIR_BODY) { 4737 FREE_LOCK(&lk); 4738 if ((error = VFS_VGET(mp, NULL, inum, &vp)) != 0) 4739 break; 4740 if ((error=VOP_FSYNC(vp, MNT_NOWAIT, 0)) || 4741 (error=VOP_FSYNC(vp, MNT_NOWAIT, 0))) { 4742 vput(vp); 4743 break; 4744 } 4745 drain_output(vp, 0); 4746 vput(vp); 4747 ACQUIRE_LOCK(&lk); 4748 /* 4749 * If that cleared dependencies, go on to next. 4750 */ 4751 if (dap != LIST_FIRST(diraddhdp)) 4752 continue; 4753 if (dap->da_state & MKDIR_BODY) { 4754 FREE_LOCK(&lk); 4755 panic("flush_pagedep_deps: MKDIR_BODY"); 4756 } 4757 } 4758 /* 4759 * Flush the inode on which the directory entry depends. 4760 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 4761 * the only remaining dependency is that the updated inode 4762 * count must get pushed to disk. The inode has already 4763 * been pushed into its inode buffer (via VOP_UPDATE) at 4764 * the time of the reference count change. So we need only 4765 * locate that buffer, ensure that there will be no rollback 4766 * caused by a bitmap dependency, then write the inode buffer. 4767 */ 4768 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) { 4769 FREE_LOCK(&lk); 4770 panic("flush_pagedep_deps: lost inode"); 4771 } 4772 /* 4773 * If the inode still has bitmap dependencies, 4774 * push them to disk. 4775 */ 4776 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4777 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 4778 FREE_LOCK(&lk); 4779 if (gotit && (error = bwrite(inodedep->id_buf)) != 0) 4780 break; 4781 ACQUIRE_LOCK(&lk); 4782 if (dap != LIST_FIRST(diraddhdp)) 4783 continue; 4784 } 4785 /* 4786 * If the inode is still sitting in a buffer waiting 4787 * to be written, push it to disk. 4788 */ 4789 FREE_LOCK(&lk); 4790 if ((error = bread(ump->um_devvp, 4791 fsbtodoff(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), 4792 (int)ump->um_fs->fs_bsize, &bp)) != 0) 4793 break; 4794 if ((error = bwrite(bp)) != 0) 4795 break; 4796 ACQUIRE_LOCK(&lk); 4797 /* 4798 * If we have failed to get rid of all the dependencies 4799 * then something is seriously wrong. 4800 */ 4801 if (dap == LIST_FIRST(diraddhdp)) { 4802 FREE_LOCK(&lk); 4803 panic("flush_pagedep_deps: flush failed"); 4804 } 4805 } 4806 if (error) 4807 ACQUIRE_LOCK(&lk); 4808 return (error); 4809 } 4810 4811 /* 4812 * A large burst of file addition or deletion activity can drive the 4813 * memory load excessively high. First attempt to slow things down 4814 * using the techniques below. If that fails, this routine requests 4815 * the offending operations to fall back to running synchronously 4816 * until the memory load returns to a reasonable level. 4817 */ 4818 int 4819 softdep_slowdown(struct vnode *vp) 4820 { 4821 int max_softdeps_hard; 4822 4823 max_softdeps_hard = max_softdeps * 11 / 10; 4824 if (num_dirrem < max_softdeps_hard / 2 && 4825 num_inodedep < max_softdeps_hard) 4826 return (0); 4827 stat_sync_limit_hit += 1; 4828 return (1); 4829 } 4830 4831 /* 4832 * If memory utilization has gotten too high, deliberately slow things 4833 * down and speed up the I/O processing. 4834 */ 4835 static int 4836 request_cleanup(int resource, int islocked) 4837 { 4838 struct thread *td = curthread; /* XXX */ 4839 4840 /* 4841 * We never hold up the filesystem syncer process. 4842 */ 4843 if (td == filesys_syncer) 4844 return (0); 4845 /* 4846 * First check to see if the work list has gotten backlogged. 4847 * If it has, co-opt this process to help clean up two entries. 4848 * Because this process may hold inodes locked, we cannot 4849 * handle any remove requests that might block on a locked 4850 * inode as that could lead to deadlock. 4851 */ 4852 if (num_on_worklist > max_softdeps / 10) { 4853 if (islocked) 4854 FREE_LOCK(&lk); 4855 process_worklist_item(NULL, LK_NOWAIT); 4856 process_worklist_item(NULL, LK_NOWAIT); 4857 stat_worklist_push += 2; 4858 if (islocked) 4859 ACQUIRE_LOCK(&lk); 4860 return(1); 4861 } 4862 4863 /* 4864 * If we are resource constrained on inode dependencies, try 4865 * flushing some dirty inodes. Otherwise, we are constrained 4866 * by file deletions, so try accelerating flushes of directories 4867 * with removal dependencies. We would like to do the cleanup 4868 * here, but we probably hold an inode locked at this point and 4869 * that might deadlock against one that we try to clean. So, 4870 * the best that we can do is request the syncer daemon to do 4871 * the cleanup for us. 4872 */ 4873 switch (resource) { 4874 4875 case FLUSH_INODES: 4876 stat_ino_limit_push += 1; 4877 req_clear_inodedeps += 1; 4878 stat_countp = &stat_ino_limit_hit; 4879 break; 4880 4881 case FLUSH_REMOVE: 4882 stat_blk_limit_push += 1; 4883 req_clear_remove += 1; 4884 stat_countp = &stat_blk_limit_hit; 4885 break; 4886 4887 default: 4888 if (islocked) 4889 FREE_LOCK(&lk); 4890 panic("request_cleanup: unknown type"); 4891 } 4892 /* 4893 * Hopefully the syncer daemon will catch up and awaken us. 4894 * We wait at most tickdelay before proceeding in any case. 4895 */ 4896 if (islocked == 0) 4897 ACQUIRE_LOCK(&lk); 4898 proc_waiting += 1; 4899 if (!callout_active(&handle)) 4900 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2, 4901 pause_timer, NULL); 4902 interlocked_sleep(&lk, SLEEP, (caddr_t)&proc_waiting, 0, 4903 "softupdate", 0); 4904 proc_waiting -= 1; 4905 if (islocked == 0) 4906 FREE_LOCK(&lk); 4907 return (1); 4908 } 4909 4910 /* 4911 * Awaken processes pausing in request_cleanup and clear proc_waiting 4912 * to indicate that there is no longer a timer running. 4913 */ 4914 void 4915 pause_timer(void *arg) 4916 { 4917 *stat_countp += 1; 4918 wakeup_one(&proc_waiting); 4919 if (proc_waiting > 0) 4920 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2, 4921 pause_timer, NULL); 4922 else 4923 callout_deactivate(&handle); 4924 } 4925 4926 /* 4927 * Flush out a directory with at least one removal dependency in an effort to 4928 * reduce the number of dirrem, freefile, and freeblks dependency structures. 4929 */ 4930 static void 4931 clear_remove(struct thread *td) 4932 { 4933 struct pagedep_hashhead *pagedephd; 4934 struct pagedep *pagedep; 4935 static int next = 0; 4936 struct mount *mp; 4937 struct vnode *vp; 4938 int error, cnt; 4939 ino_t ino; 4940 4941 ACQUIRE_LOCK(&lk); 4942 for (cnt = 0; cnt < pagedep_hash; cnt++) { 4943 pagedephd = &pagedep_hashtbl[next++]; 4944 if (next >= pagedep_hash) 4945 next = 0; 4946 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 4947 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL) 4948 continue; 4949 mp = pagedep->pd_mnt; 4950 ino = pagedep->pd_ino; 4951 FREE_LOCK(&lk); 4952 if ((error = VFS_VGET(mp, NULL, ino, &vp)) != 0) { 4953 softdep_error("clear_remove: vget", error); 4954 return; 4955 } 4956 if ((error = VOP_FSYNC(vp, MNT_NOWAIT, 0))) 4957 softdep_error("clear_remove: fsync", error); 4958 drain_output(vp, 0); 4959 vput(vp); 4960 return; 4961 } 4962 } 4963 FREE_LOCK(&lk); 4964 } 4965 4966 /* 4967 * Clear out a block of dirty inodes in an effort to reduce 4968 * the number of inodedep dependency structures. 4969 */ 4970 struct clear_inodedeps_info { 4971 struct fs *fs; 4972 struct mount *mp; 4973 }; 4974 4975 static int 4976 clear_inodedeps_mountlist_callback(struct mount *mp, void *data) 4977 { 4978 struct clear_inodedeps_info *info = data; 4979 4980 if ((mp->mnt_flag & MNT_SOFTDEP) && info->fs == VFSTOUFS(mp)->um_fs) { 4981 info->mp = mp; 4982 return(-1); 4983 } 4984 return(0); 4985 } 4986 4987 static void 4988 clear_inodedeps(struct thread *td) 4989 { 4990 struct clear_inodedeps_info info; 4991 struct inodedep_hashhead *inodedephd; 4992 struct inodedep *inodedep; 4993 static int next = 0; 4994 struct vnode *vp; 4995 struct fs *fs; 4996 int error, cnt; 4997 ino_t firstino, lastino, ino; 4998 4999 ACQUIRE_LOCK(&lk); 5000 /* 5001 * Pick a random inode dependency to be cleared. 5002 * We will then gather up all the inodes in its block 5003 * that have dependencies and flush them out. 5004 */ 5005 for (cnt = 0; cnt < inodedep_hash; cnt++) { 5006 inodedephd = &inodedep_hashtbl[next++]; 5007 if (next >= inodedep_hash) 5008 next = 0; 5009 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 5010 break; 5011 } 5012 if (inodedep == NULL) { 5013 FREE_LOCK(&lk); 5014 return; 5015 } 5016 /* 5017 * Ugly code to find mount point given pointer to superblock. 5018 */ 5019 fs = inodedep->id_fs; 5020 info.mp = NULL; 5021 info.fs = fs; 5022 mountlist_scan(clear_inodedeps_mountlist_callback, 5023 &info, MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 5024 /* 5025 * Find the last inode in the block with dependencies. 5026 */ 5027 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 5028 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 5029 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0) 5030 break; 5031 /* 5032 * Asynchronously push all but the last inode with dependencies. 5033 * Synchronously push the last inode with dependencies to ensure 5034 * that the inode block gets written to free up the inodedeps. 5035 */ 5036 for (ino = firstino; ino <= lastino; ino++) { 5037 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 5038 continue; 5039 FREE_LOCK(&lk); 5040 if ((error = VFS_VGET(info.mp, NULL, ino, &vp)) != 0) { 5041 softdep_error("clear_inodedeps: vget", error); 5042 return; 5043 } 5044 if (ino == lastino) { 5045 if ((error = VOP_FSYNC(vp, MNT_WAIT, 0))) 5046 softdep_error("clear_inodedeps: fsync1", error); 5047 } else { 5048 if ((error = VOP_FSYNC(vp, MNT_NOWAIT, 0))) 5049 softdep_error("clear_inodedeps: fsync2", error); 5050 drain_output(vp, 0); 5051 } 5052 vput(vp); 5053 ACQUIRE_LOCK(&lk); 5054 } 5055 FREE_LOCK(&lk); 5056 } 5057 5058 /* 5059 * Function to determine if the buffer has outstanding dependencies 5060 * that will cause a roll-back if the buffer is written. If wantcount 5061 * is set, return number of dependencies, otherwise just yes or no. 5062 * 5063 * bioops callback - hold io_token 5064 */ 5065 static int 5066 softdep_count_dependencies(struct buf *bp, int wantcount) 5067 { 5068 struct worklist *wk; 5069 struct inodedep *inodedep; 5070 struct indirdep *indirdep; 5071 struct allocindir *aip; 5072 struct pagedep *pagedep; 5073 struct diradd *dap; 5074 int i, retval; 5075 5076 get_mplock(); 5077 5078 retval = 0; 5079 ACQUIRE_LOCK(&lk); 5080 5081 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5082 switch (wk->wk_type) { 5083 5084 case D_INODEDEP: 5085 inodedep = WK_INODEDEP(wk); 5086 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 5087 /* bitmap allocation dependency */ 5088 retval += 1; 5089 if (!wantcount) 5090 goto out; 5091 } 5092 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 5093 /* direct block pointer dependency */ 5094 retval += 1; 5095 if (!wantcount) 5096 goto out; 5097 } 5098 continue; 5099 5100 case D_INDIRDEP: 5101 indirdep = WK_INDIRDEP(wk); 5102 5103 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 5104 /* indirect block pointer dependency */ 5105 retval += 1; 5106 if (!wantcount) 5107 goto out; 5108 } 5109 continue; 5110 5111 case D_PAGEDEP: 5112 pagedep = WK_PAGEDEP(wk); 5113 for (i = 0; i < DAHASHSZ; i++) { 5114 5115 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 5116 /* directory entry dependency */ 5117 retval += 1; 5118 if (!wantcount) 5119 goto out; 5120 } 5121 } 5122 continue; 5123 5124 case D_BMSAFEMAP: 5125 case D_ALLOCDIRECT: 5126 case D_ALLOCINDIR: 5127 case D_MKDIR: 5128 /* never a dependency on these blocks */ 5129 continue; 5130 5131 default: 5132 FREE_LOCK(&lk); 5133 panic("softdep_check_for_rollback: Unexpected type %s", 5134 TYPENAME(wk->wk_type)); 5135 /* NOTREACHED */ 5136 } 5137 } 5138 out: 5139 FREE_LOCK(&lk); 5140 rel_mplock(); 5141 5142 return retval; 5143 } 5144 5145 /* 5146 * Acquire exclusive access to a buffer. 5147 * Must be called with splbio blocked. 5148 * Return 1 if buffer was acquired. 5149 */ 5150 static int 5151 getdirtybuf(struct buf **bpp, int waitfor) 5152 { 5153 struct buf *bp; 5154 int error; 5155 5156 for (;;) { 5157 if ((bp = *bpp) == NULL) 5158 return (0); 5159 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) 5160 break; 5161 if (waitfor != MNT_WAIT) 5162 return (0); 5163 error = interlocked_sleep(&lk, LOCKBUF, bp, 5164 LK_EXCLUSIVE | LK_SLEEPFAIL, 0, 0); 5165 if (error != ENOLCK) { 5166 FREE_LOCK(&lk); 5167 panic("getdirtybuf: inconsistent lock"); 5168 } 5169 } 5170 if ((bp->b_flags & B_DELWRI) == 0) { 5171 BUF_UNLOCK(bp); 5172 return (0); 5173 } 5174 bremfree(bp); 5175 return (1); 5176 } 5177 5178 /* 5179 * Wait for pending output on a vnode to complete. 5180 * Must be called with vnode locked. 5181 */ 5182 static void 5183 drain_output(struct vnode *vp, int islocked) 5184 { 5185 5186 if (!islocked) 5187 ACQUIRE_LOCK(&lk); 5188 while (bio_track_active(&vp->v_track_write)) { 5189 FREE_LOCK(&lk); 5190 bio_track_wait(&vp->v_track_write, 0, 0); 5191 ACQUIRE_LOCK(&lk); 5192 } 5193 if (!islocked) 5194 FREE_LOCK(&lk); 5195 } 5196 5197 /* 5198 * Called whenever a buffer that is being invalidated or reallocated 5199 * contains dependencies. This should only happen if an I/O error has 5200 * occurred. The routine is called with the buffer locked. 5201 * 5202 * bioops callback - hold io_token 5203 */ 5204 static void 5205 softdep_deallocate_dependencies(struct buf *bp) 5206 { 5207 /* nothing to do, mp lock not needed */ 5208 if ((bp->b_flags & B_ERROR) == 0) 5209 panic("softdep_deallocate_dependencies: dangling deps"); 5210 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntfromname, bp->b_error); 5211 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 5212 } 5213 5214 /* 5215 * Function to handle asynchronous write errors in the filesystem. 5216 */ 5217 void 5218 softdep_error(char *func, int error) 5219 { 5220 5221 /* XXX should do something better! */ 5222 kprintf("%s: got error %d while accessing filesystem\n", func, error); 5223 } 5224