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 1.52 2007/08/13 17:31:56 dillon Exp $ 41 */ 42 43 /* 44 * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide. 45 */ 46 #ifndef DIAGNOSTIC 47 #define DIAGNOSTIC 48 #endif 49 #ifndef DEBUG 50 #define DEBUG 51 #endif 52 53 #include <sys/param.h> 54 #include <sys/kernel.h> 55 #include <sys/systm.h> 56 #include <sys/buf.h> 57 #include <sys/malloc.h> 58 #include <sys/mount.h> 59 #include <sys/proc.h> 60 #include <sys/syslog.h> 61 #include <sys/vnode.h> 62 #include <sys/conf.h> 63 #include <sys/buf2.h> 64 #include <machine/inttypes.h> 65 #include "dir.h" 66 #include "quota.h" 67 #include "inode.h" 68 #include "ufsmount.h" 69 #include "fs.h" 70 #include "softdep.h" 71 #include "ffs_extern.h" 72 #include "ufs_extern.h" 73 74 #include <sys/thread2.h> 75 76 /* 77 * These definitions need to be adapted to the system to which 78 * this file is being ported. 79 */ 80 /* 81 * malloc types defined for the softdep system. 82 */ 83 MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies"); 84 MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies"); 85 MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation"); 86 MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map"); 87 MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode"); 88 MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies"); 89 MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block"); 90 MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode"); 91 MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode"); 92 MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated"); 93 MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry"); 94 MALLOC_DEFINE(M_MKDIR, "mkdir","New directory"); 95 MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted"); 96 97 #define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE) 98 99 #define D_PAGEDEP 0 100 #define D_INODEDEP 1 101 #define D_NEWBLK 2 102 #define D_BMSAFEMAP 3 103 #define D_ALLOCDIRECT 4 104 #define D_INDIRDEP 5 105 #define D_ALLOCINDIR 6 106 #define D_FREEFRAG 7 107 #define D_FREEBLKS 8 108 #define D_FREEFILE 9 109 #define D_DIRADD 10 110 #define D_MKDIR 11 111 #define D_DIRREM 12 112 #define D_LAST D_DIRREM 113 114 /* 115 * translate from workitem type to memory type 116 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 117 */ 118 static struct malloc_type *memtype[] = { 119 M_PAGEDEP, 120 M_INODEDEP, 121 M_NEWBLK, 122 M_BMSAFEMAP, 123 M_ALLOCDIRECT, 124 M_INDIRDEP, 125 M_ALLOCINDIR, 126 M_FREEFRAG, 127 M_FREEBLKS, 128 M_FREEFILE, 129 M_DIRADD, 130 M_MKDIR, 131 M_DIRREM 132 }; 133 134 #define DtoM(type) (memtype[type]) 135 136 /* 137 * Names of malloc types. 138 */ 139 #define TYPENAME(type) \ 140 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???") 141 /* 142 * End system adaptaion definitions. 143 */ 144 145 /* 146 * Internal function prototypes. 147 */ 148 static void softdep_error(char *, int); 149 static void drain_output(struct vnode *, int); 150 static int getdirtybuf(struct buf **, int); 151 static void clear_remove(struct thread *); 152 static void clear_inodedeps(struct thread *); 153 static int flush_pagedep_deps(struct vnode *, struct mount *, 154 struct diraddhd *); 155 static int flush_inodedep_deps(struct fs *, ino_t); 156 static int handle_written_filepage(struct pagedep *, struct buf *); 157 static void diradd_inode_written(struct diradd *, struct inodedep *); 158 static int handle_written_inodeblock(struct inodedep *, struct buf *); 159 static void handle_allocdirect_partdone(struct allocdirect *); 160 static void handle_allocindir_partdone(struct allocindir *); 161 static void initiate_write_filepage(struct pagedep *, struct buf *); 162 static void handle_written_mkdir(struct mkdir *, int); 163 static void initiate_write_inodeblock(struct inodedep *, struct buf *); 164 static void handle_workitem_freefile(struct freefile *); 165 static void handle_workitem_remove(struct dirrem *); 166 static struct dirrem *newdirrem(struct buf *, struct inode *, 167 struct inode *, int, struct dirrem **); 168 static void free_diradd(struct diradd *); 169 static void free_allocindir(struct allocindir *, struct inodedep *); 170 static int indir_trunc (struct inode *, off_t, int, ufs_lbn_t, long *); 171 static void deallocate_dependencies(struct buf *, struct inodedep *); 172 static void free_allocdirect(struct allocdirectlst *, 173 struct allocdirect *, int); 174 static int check_inode_unwritten(struct inodedep *); 175 static int free_inodedep(struct inodedep *); 176 static void handle_workitem_freeblocks(struct freeblks *); 177 static void merge_inode_lists(struct inodedep *); 178 static void setup_allocindir_phase2(struct buf *, struct inode *, 179 struct allocindir *); 180 static struct allocindir *newallocindir(struct inode *, int, ufs_daddr_t, 181 ufs_daddr_t); 182 static void handle_workitem_freefrag(struct freefrag *); 183 static struct freefrag *newfreefrag(struct inode *, ufs_daddr_t, long); 184 static void allocdirect_merge(struct allocdirectlst *, 185 struct allocdirect *, struct allocdirect *); 186 static struct bmsafemap *bmsafemap_lookup(struct buf *); 187 static int newblk_lookup(struct fs *, ufs_daddr_t, int, 188 struct newblk **); 189 static int inodedep_lookup(struct fs *, ino_t, int, struct inodedep **); 190 static int pagedep_lookup(struct inode *, ufs_lbn_t, int, 191 struct pagedep **); 192 static void pause_timer(void *); 193 static int request_cleanup(int, int); 194 static int process_worklist_item(struct mount *, int); 195 static void add_to_worklist(struct worklist *); 196 197 /* 198 * Exported softdep operations. 199 */ 200 static void softdep_disk_io_initiation(struct buf *); 201 static void softdep_disk_write_complete(struct buf *); 202 static void softdep_deallocate_dependencies(struct buf *); 203 static int softdep_fsync(struct vnode *); 204 static int softdep_process_worklist(struct mount *); 205 static void softdep_move_dependencies(struct buf *, struct buf *); 206 static int softdep_count_dependencies(struct buf *bp, int); 207 208 static struct bio_ops softdep_bioops = { 209 softdep_disk_io_initiation, /* io_start */ 210 softdep_disk_write_complete, /* io_complete */ 211 softdep_deallocate_dependencies, /* io_deallocate */ 212 softdep_fsync, /* io_fsync */ 213 softdep_process_worklist, /* io_sync */ 214 softdep_move_dependencies, /* io_movedeps */ 215 softdep_count_dependencies, /* io_countdeps */ 216 }; 217 218 /* 219 * Locking primitives. 220 * 221 * For a uniprocessor, all we need to do is protect against disk 222 * interrupts. For a multiprocessor, this lock would have to be 223 * a mutex. A single mutex is used throughout this file, though 224 * finer grain locking could be used if contention warranted it. 225 * 226 * For a multiprocessor, the sleep call would accept a lock and 227 * release it after the sleep processing was complete. In a uniprocessor 228 * implementation there is no such interlock, so we simple mark 229 * the places where it needs to be done with the `interlocked' form 230 * of the lock calls. Since the uniprocessor sleep already interlocks 231 * the spl, there is nothing that really needs to be done. 232 */ 233 #ifndef /* NOT */ DEBUG 234 static struct lockit { 235 } lk = { 0 }; 236 #define ACQUIRE_LOCK(lk) crit_enter_id("softupdates"); 237 #define FREE_LOCK(lk) crit_exit_id("softupdates"); 238 239 #else /* DEBUG */ 240 #define NOHOLDER ((struct thread *)-1) 241 #define SPECIAL_FLAG ((struct thread *)-2) 242 static struct lockit { 243 int lkt_spl; 244 struct thread *lkt_held; 245 } lk = { 0, NOHOLDER }; 246 static int lockcnt; 247 248 static void acquire_lock(struct lockit *); 249 static void free_lock(struct lockit *); 250 void softdep_panic(char *); 251 252 #define ACQUIRE_LOCK(lk) acquire_lock(lk) 253 #define FREE_LOCK(lk) free_lock(lk) 254 255 static void 256 acquire_lock(struct lockit *lk) 257 { 258 thread_t holder; 259 260 if (lk->lkt_held != NOHOLDER) { 261 holder = lk->lkt_held; 262 FREE_LOCK(lk); 263 if (holder == curthread) 264 panic("softdep_lock: locking against myself"); 265 else 266 panic("softdep_lock: lock held by %p", holder); 267 } 268 crit_enter_id("softupdates"); 269 lk->lkt_held = curthread; 270 lockcnt++; 271 } 272 273 static void 274 free_lock(struct lockit *lk) 275 { 276 277 if (lk->lkt_held == NOHOLDER) 278 panic("softdep_unlock: lock not held"); 279 lk->lkt_held = NOHOLDER; 280 crit_exit_id("softupdates"); 281 } 282 283 /* 284 * Function to release soft updates lock and panic. 285 */ 286 void 287 softdep_panic(char *msg) 288 { 289 290 if (lk.lkt_held != NOHOLDER) 291 FREE_LOCK(&lk); 292 panic(msg); 293 } 294 #endif /* DEBUG */ 295 296 static int interlocked_sleep(struct lockit *, int, void *, int, 297 const char *, int); 298 299 /* 300 * When going to sleep, we must save our SPL so that it does 301 * not get lost if some other process uses the lock while we 302 * are sleeping. We restore it after we have slept. This routine 303 * wraps the interlocking with functions that sleep. The list 304 * below enumerates the available set of operations. 305 */ 306 #define UNKNOWN 0 307 #define SLEEP 1 308 #define LOCKBUF 2 309 310 static int 311 interlocked_sleep(struct lockit *lk, int op, void *ident, int flags, 312 const char *wmesg, int timo) 313 { 314 thread_t holder; 315 int s, retval; 316 317 s = lk->lkt_spl; 318 # ifdef DEBUG 319 if (lk->lkt_held == NOHOLDER) 320 panic("interlocked_sleep: lock not held"); 321 lk->lkt_held = NOHOLDER; 322 # endif /* DEBUG */ 323 switch (op) { 324 case SLEEP: 325 retval = tsleep(ident, flags, wmesg, timo); 326 break; 327 case LOCKBUF: 328 retval = BUF_LOCK((struct buf *)ident, flags); 329 break; 330 default: 331 panic("interlocked_sleep: unknown operation"); 332 } 333 # ifdef DEBUG 334 if (lk->lkt_held != NOHOLDER) { 335 holder = lk->lkt_held; 336 FREE_LOCK(lk); 337 if (holder == curthread) 338 panic("interlocked_sleep: locking against self"); 339 else 340 panic("interlocked_sleep: lock held by %p", holder); 341 } 342 lk->lkt_held = curthread; 343 lockcnt++; 344 # endif /* DEBUG */ 345 lk->lkt_spl = s; 346 return (retval); 347 } 348 349 /* 350 * Place holder for real semaphores. 351 */ 352 struct sema { 353 int value; 354 thread_t holder; 355 char *name; 356 int prio; 357 int timo; 358 }; 359 static void sema_init(struct sema *, char *, int, int); 360 static int sema_get(struct sema *, struct lockit *); 361 static void sema_release(struct sema *); 362 363 static void 364 sema_init(struct sema *semap, char *name, int prio, int timo) 365 { 366 367 semap->holder = NOHOLDER; 368 semap->value = 0; 369 semap->name = name; 370 semap->prio = prio; 371 semap->timo = timo; 372 } 373 374 static int 375 sema_get(struct sema *semap, struct lockit *interlock) 376 { 377 378 if (semap->value++ > 0) { 379 if (interlock != NULL) { 380 interlocked_sleep(interlock, SLEEP, (caddr_t)semap, 381 semap->prio, semap->name, semap->timo); 382 FREE_LOCK(interlock); 383 } else { 384 tsleep((caddr_t)semap, semap->prio, semap->name, 385 semap->timo); 386 } 387 return (0); 388 } 389 semap->holder = curthread; 390 if (interlock != NULL) 391 FREE_LOCK(interlock); 392 return (1); 393 } 394 395 static void 396 sema_release(struct sema *semap) 397 { 398 399 if (semap->value <= 0 || semap->holder != curthread) { 400 if (lk.lkt_held != NOHOLDER) 401 FREE_LOCK(&lk); 402 panic("sema_release: not held"); 403 } 404 if (--semap->value > 0) { 405 semap->value = 0; 406 wakeup(semap); 407 } 408 semap->holder = NOHOLDER; 409 } 410 411 /* 412 * Worklist queue management. 413 * These routines require that the lock be held. 414 */ 415 #ifndef /* NOT */ DEBUG 416 #define WORKLIST_INSERT(head, item) do { \ 417 (item)->wk_state |= ONWORKLIST; \ 418 LIST_INSERT_HEAD(head, item, wk_list); \ 419 } while (0) 420 #define WORKLIST_REMOVE(item) do { \ 421 (item)->wk_state &= ~ONWORKLIST; \ 422 LIST_REMOVE(item, wk_list); \ 423 } while (0) 424 #define WORKITEM_FREE(item, type) FREE(item, DtoM(type)) 425 426 #else /* DEBUG */ 427 static void worklist_insert(struct workhead *, struct worklist *); 428 static void worklist_remove(struct worklist *); 429 static void workitem_free(struct worklist *, int); 430 431 #define WORKLIST_INSERT(head, item) worklist_insert(head, item) 432 #define WORKLIST_REMOVE(item) worklist_remove(item) 433 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type) 434 435 static void 436 worklist_insert(struct workhead *head, struct worklist *item) 437 { 438 439 if (lk.lkt_held == NOHOLDER) 440 panic("worklist_insert: lock not held"); 441 if (item->wk_state & ONWORKLIST) { 442 FREE_LOCK(&lk); 443 panic("worklist_insert: already on list"); 444 } 445 item->wk_state |= ONWORKLIST; 446 LIST_INSERT_HEAD(head, item, wk_list); 447 } 448 449 static void 450 worklist_remove(struct worklist *item) 451 { 452 453 if (lk.lkt_held == NOHOLDER) 454 panic("worklist_remove: lock not held"); 455 if ((item->wk_state & ONWORKLIST) == 0) { 456 FREE_LOCK(&lk); 457 panic("worklist_remove: not on list"); 458 } 459 item->wk_state &= ~ONWORKLIST; 460 LIST_REMOVE(item, wk_list); 461 } 462 463 static void 464 workitem_free(struct worklist *item, int type) 465 { 466 467 if (item->wk_state & ONWORKLIST) { 468 if (lk.lkt_held != NOHOLDER) 469 FREE_LOCK(&lk); 470 panic("workitem_free: still on list"); 471 } 472 if (item->wk_type != type) { 473 if (lk.lkt_held != NOHOLDER) 474 FREE_LOCK(&lk); 475 panic("workitem_free: type mismatch"); 476 } 477 FREE(item, DtoM(type)); 478 } 479 #endif /* DEBUG */ 480 481 /* 482 * Workitem queue management 483 */ 484 static struct workhead softdep_workitem_pending; 485 static int num_on_worklist; /* number of worklist items to be processed */ 486 static int softdep_worklist_busy; /* 1 => trying to do unmount */ 487 static int softdep_worklist_req; /* serialized waiters */ 488 static int max_softdeps; /* maximum number of structs before slowdown */ 489 static int tickdelay = 2; /* number of ticks to pause during slowdown */ 490 static int *stat_countp; /* statistic to count in proc_waiting timeout */ 491 static int proc_waiting; /* tracks whether we have a timeout posted */ 492 static struct callout handle; /* handle on posted proc_waiting timeout */ 493 static struct thread *filesys_syncer; /* proc of filesystem syncer process */ 494 static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 495 #define FLUSH_INODES 1 496 static int req_clear_remove; /* syncer process flush some freeblks */ 497 #define FLUSH_REMOVE 2 498 /* 499 * runtime statistics 500 */ 501 static int stat_worklist_push; /* number of worklist cleanups */ 502 static int stat_blk_limit_push; /* number of times block limit neared */ 503 static int stat_ino_limit_push; /* number of times inode limit neared */ 504 static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 505 static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 506 static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 507 static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 508 static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 509 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 510 static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 511 #ifdef DEBUG 512 #include <vm/vm.h> 513 #include <sys/sysctl.h> 514 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, ""); 515 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, ""); 516 SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,""); 517 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,""); 518 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,""); 519 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, ""); 520 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, ""); 521 SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, ""); 522 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, ""); 523 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, ""); 524 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, ""); 525 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, ""); 526 #endif /* DEBUG */ 527 528 /* 529 * Add an item to the end of the work queue. 530 * This routine requires that the lock be held. 531 * This is the only routine that adds items to the list. 532 * The following routine is the only one that removes items 533 * and does so in order from first to last. 534 */ 535 static void 536 add_to_worklist(struct worklist *wk) 537 { 538 static struct worklist *worklist_tail; 539 540 if (wk->wk_state & ONWORKLIST) { 541 if (lk.lkt_held != NOHOLDER) 542 FREE_LOCK(&lk); 543 panic("add_to_worklist: already on list"); 544 } 545 wk->wk_state |= ONWORKLIST; 546 if (LIST_FIRST(&softdep_workitem_pending) == NULL) 547 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list); 548 else 549 LIST_INSERT_AFTER(worklist_tail, wk, wk_list); 550 worklist_tail = wk; 551 num_on_worklist += 1; 552 } 553 554 /* 555 * Process that runs once per second to handle items in the background queue. 556 * 557 * Note that we ensure that everything is done in the order in which they 558 * appear in the queue. The code below depends on this property to ensure 559 * that blocks of a file are freed before the inode itself is freed. This 560 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 561 * until all the old ones have been purged from the dependency lists. 562 */ 563 static int 564 softdep_process_worklist(struct mount *matchmnt) 565 { 566 thread_t td = curthread; 567 int matchcnt, loopcount; 568 long starttime; 569 570 /* 571 * Record the process identifier of our caller so that we can give 572 * this process preferential treatment in request_cleanup below. 573 */ 574 filesys_syncer = td; 575 matchcnt = 0; 576 577 /* 578 * There is no danger of having multiple processes run this 579 * code, but we have to single-thread it when softdep_flushfiles() 580 * is in operation to get an accurate count of the number of items 581 * related to its mount point that are in the list. 582 */ 583 if (matchmnt == NULL) { 584 if (softdep_worklist_busy < 0) 585 return(-1); 586 softdep_worklist_busy += 1; 587 } 588 589 /* 590 * If requested, try removing inode or removal dependencies. 591 */ 592 if (req_clear_inodedeps) { 593 clear_inodedeps(td); 594 req_clear_inodedeps -= 1; 595 wakeup_one(&proc_waiting); 596 } 597 if (req_clear_remove) { 598 clear_remove(td); 599 req_clear_remove -= 1; 600 wakeup_one(&proc_waiting); 601 } 602 loopcount = 1; 603 starttime = time_second; 604 while (num_on_worklist > 0) { 605 matchcnt += process_worklist_item(matchmnt, 0); 606 607 /* 608 * If a umount operation wants to run the worklist 609 * accurately, abort. 610 */ 611 if (softdep_worklist_req && matchmnt == NULL) { 612 matchcnt = -1; 613 break; 614 } 615 616 /* 617 * If requested, try removing inode or removal dependencies. 618 */ 619 if (req_clear_inodedeps) { 620 clear_inodedeps(td); 621 req_clear_inodedeps -= 1; 622 wakeup_one(&proc_waiting); 623 } 624 if (req_clear_remove) { 625 clear_remove(td); 626 req_clear_remove -= 1; 627 wakeup_one(&proc_waiting); 628 } 629 /* 630 * We do not generally want to stop for buffer space, but if 631 * we are really being a buffer hog, we will stop and wait. 632 */ 633 if (loopcount++ % 128 == 0) 634 bwillwrite(); 635 /* 636 * Never allow processing to run for more than one 637 * second. Otherwise the other syncer tasks may get 638 * excessively backlogged. 639 */ 640 if (starttime != time_second && matchmnt == NULL) { 641 matchcnt = -1; 642 break; 643 } 644 } 645 if (matchmnt == NULL) { 646 --softdep_worklist_busy; 647 if (softdep_worklist_req && softdep_worklist_busy == 0) 648 wakeup(&softdep_worklist_req); 649 } 650 return (matchcnt); 651 } 652 653 /* 654 * Process one item on the worklist. 655 */ 656 static int 657 process_worklist_item(struct mount *matchmnt, int flags) 658 { 659 struct worklist *wk; 660 struct dirrem *dirrem; 661 struct fs *matchfs; 662 struct vnode *vp; 663 int matchcnt = 0; 664 665 matchfs = NULL; 666 if (matchmnt != NULL) 667 matchfs = VFSTOUFS(matchmnt)->um_fs; 668 ACQUIRE_LOCK(&lk); 669 /* 670 * Normally we just process each item on the worklist in order. 671 * However, if we are in a situation where we cannot lock any 672 * inodes, we have to skip over any dirrem requests whose 673 * vnodes are resident and locked. 674 */ 675 LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) { 676 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM) 677 break; 678 dirrem = WK_DIRREM(wk); 679 vp = ufs_ihashlookup(VFSTOUFS(dirrem->dm_mnt)->um_dev, 680 dirrem->dm_oldinum); 681 if (vp == NULL || !vn_islocked(vp)) 682 break; 683 } 684 if (wk == 0) { 685 FREE_LOCK(&lk); 686 return (0); 687 } 688 WORKLIST_REMOVE(wk); 689 num_on_worklist -= 1; 690 FREE_LOCK(&lk); 691 switch (wk->wk_type) { 692 693 case D_DIRREM: 694 /* removal of a directory entry */ 695 if (WK_DIRREM(wk)->dm_mnt == matchmnt) 696 matchcnt += 1; 697 handle_workitem_remove(WK_DIRREM(wk)); 698 break; 699 700 case D_FREEBLKS: 701 /* releasing blocks and/or fragments from a file */ 702 if (WK_FREEBLKS(wk)->fb_fs == matchfs) 703 matchcnt += 1; 704 handle_workitem_freeblocks(WK_FREEBLKS(wk)); 705 break; 706 707 case D_FREEFRAG: 708 /* releasing a fragment when replaced as a file grows */ 709 if (WK_FREEFRAG(wk)->ff_fs == matchfs) 710 matchcnt += 1; 711 handle_workitem_freefrag(WK_FREEFRAG(wk)); 712 break; 713 714 case D_FREEFILE: 715 /* releasing an inode when its link count drops to 0 */ 716 if (WK_FREEFILE(wk)->fx_fs == matchfs) 717 matchcnt += 1; 718 handle_workitem_freefile(WK_FREEFILE(wk)); 719 break; 720 721 default: 722 panic("%s_process_worklist: Unknown type %s", 723 "softdep", TYPENAME(wk->wk_type)); 724 /* NOTREACHED */ 725 } 726 return (matchcnt); 727 } 728 729 /* 730 * Move dependencies from one buffer to another. 731 */ 732 static void 733 softdep_move_dependencies(struct buf *oldbp, struct buf *newbp) 734 { 735 struct worklist *wk, *wktail; 736 737 if (LIST_FIRST(&newbp->b_dep) != NULL) 738 panic("softdep_move_dependencies: need merge code"); 739 wktail = 0; 740 ACQUIRE_LOCK(&lk); 741 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 742 LIST_REMOVE(wk, wk_list); 743 if (wktail == 0) 744 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 745 else 746 LIST_INSERT_AFTER(wktail, wk, wk_list); 747 wktail = wk; 748 } 749 FREE_LOCK(&lk); 750 } 751 752 /* 753 * Purge the work list of all items associated with a particular mount point. 754 */ 755 int 756 softdep_flushfiles(struct mount *oldmnt, int flags) 757 { 758 struct vnode *devvp; 759 int error, loopcnt; 760 761 /* 762 * Await our turn to clear out the queue, then serialize access. 763 */ 764 while (softdep_worklist_busy != 0) { 765 softdep_worklist_req += 1; 766 tsleep(&softdep_worklist_req, 0, "softflush", 0); 767 softdep_worklist_req -= 1; 768 } 769 softdep_worklist_busy = -1; 770 771 if ((error = ffs_flushfiles(oldmnt, flags)) != 0) { 772 softdep_worklist_busy = 0; 773 if (softdep_worklist_req) 774 wakeup(&softdep_worklist_req); 775 return (error); 776 } 777 /* 778 * Alternately flush the block device associated with the mount 779 * point and process any dependencies that the flushing 780 * creates. In theory, this loop can happen at most twice, 781 * but we give it a few extra just to be sure. 782 */ 783 devvp = VFSTOUFS(oldmnt)->um_devvp; 784 for (loopcnt = 10; loopcnt > 0; ) { 785 if (softdep_process_worklist(oldmnt) == 0) { 786 loopcnt--; 787 /* 788 * Do another flush in case any vnodes were brought in 789 * as part of the cleanup operations. 790 */ 791 if ((error = ffs_flushfiles(oldmnt, flags)) != 0) 792 break; 793 /* 794 * If we still found nothing to do, we are really done. 795 */ 796 if (softdep_process_worklist(oldmnt) == 0) 797 break; 798 } 799 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 800 error = VOP_FSYNC(devvp, MNT_WAIT); 801 vn_unlock(devvp); 802 if (error) 803 break; 804 } 805 softdep_worklist_busy = 0; 806 if (softdep_worklist_req) 807 wakeup(&softdep_worklist_req); 808 809 /* 810 * If we are unmounting then it is an error to fail. If we 811 * are simply trying to downgrade to read-only, then filesystem 812 * activity can keep us busy forever, so we just fail with EBUSY. 813 */ 814 if (loopcnt == 0) { 815 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 816 panic("softdep_flushfiles: looping"); 817 error = EBUSY; 818 } 819 return (error); 820 } 821 822 /* 823 * Structure hashing. 824 * 825 * There are three types of structures that can be looked up: 826 * 1) pagedep structures identified by mount point, inode number, 827 * and logical block. 828 * 2) inodedep structures identified by mount point and inode number. 829 * 3) newblk structures identified by mount point and 830 * physical block number. 831 * 832 * The "pagedep" and "inodedep" dependency structures are hashed 833 * separately from the file blocks and inodes to which they correspond. 834 * This separation helps when the in-memory copy of an inode or 835 * file block must be replaced. It also obviates the need to access 836 * an inode or file page when simply updating (or de-allocating) 837 * dependency structures. Lookup of newblk structures is needed to 838 * find newly allocated blocks when trying to associate them with 839 * their allocdirect or allocindir structure. 840 * 841 * The lookup routines optionally create and hash a new instance when 842 * an existing entry is not found. 843 */ 844 #define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 845 #define NODELAY 0x0002 /* cannot do background work */ 846 847 /* 848 * Structures and routines associated with pagedep caching. 849 */ 850 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl; 851 u_long pagedep_hash; /* size of hash table - 1 */ 852 #define PAGEDEP_HASH(mp, inum, lbn) \ 853 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \ 854 pagedep_hash]) 855 static struct sema pagedep_in_progress; 856 857 /* 858 * Helper routine for pagedep_lookup() 859 */ 860 static __inline 861 struct pagedep * 862 pagedep_find(struct pagedep_hashhead *pagedephd, ino_t ino, ufs_lbn_t lbn, 863 struct mount *mp) 864 { 865 struct pagedep *pagedep; 866 867 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 868 if (ino == pagedep->pd_ino && 869 lbn == pagedep->pd_lbn && 870 mp == pagedep->pd_mnt) { 871 return (pagedep); 872 } 873 } 874 return(NULL); 875 } 876 877 /* 878 * Look up a pagedep. Return 1 if found, 0 if not found. 879 * If not found, allocate if DEPALLOC flag is passed. 880 * Found or allocated entry is returned in pagedeppp. 881 * This routine must be called with splbio interrupts blocked. 882 */ 883 static int 884 pagedep_lookup(struct inode *ip, ufs_lbn_t lbn, int flags, 885 struct pagedep **pagedeppp) 886 { 887 struct pagedep *pagedep; 888 struct pagedep_hashhead *pagedephd; 889 struct mount *mp; 890 int i; 891 892 #ifdef DEBUG 893 if (lk.lkt_held == NOHOLDER) 894 panic("pagedep_lookup: lock not held"); 895 #endif 896 mp = ITOV(ip)->v_mount; 897 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn); 898 top: 899 *pagedeppp = pagedep_find(pagedephd, ip->i_number, lbn, mp); 900 if (*pagedeppp) 901 return(1); 902 if ((flags & DEPALLOC) == 0) 903 return (0); 904 if (sema_get(&pagedep_in_progress, &lk) == 0) { 905 ACQUIRE_LOCK(&lk); 906 goto top; 907 } 908 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP, 909 M_SOFTDEP_FLAGS | M_ZERO); 910 911 if (pagedep_find(pagedephd, ip->i_number, lbn, mp)) { 912 kprintf("pagedep_lookup: blocking race avoided\n"); 913 ACQUIRE_LOCK(&lk); 914 sema_release(&pagedep_in_progress); 915 kfree(pagedep, M_PAGEDEP); 916 goto top; 917 } 918 919 pagedep->pd_list.wk_type = D_PAGEDEP; 920 pagedep->pd_mnt = mp; 921 pagedep->pd_ino = ip->i_number; 922 pagedep->pd_lbn = lbn; 923 LIST_INIT(&pagedep->pd_dirremhd); 924 LIST_INIT(&pagedep->pd_pendinghd); 925 for (i = 0; i < DAHASHSZ; i++) 926 LIST_INIT(&pagedep->pd_diraddhd[i]); 927 ACQUIRE_LOCK(&lk); 928 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 929 sema_release(&pagedep_in_progress); 930 *pagedeppp = pagedep; 931 return (0); 932 } 933 934 /* 935 * Structures and routines associated with inodedep caching. 936 */ 937 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl; 938 static u_long inodedep_hash; /* size of hash table - 1 */ 939 static long num_inodedep; /* number of inodedep allocated */ 940 #define INODEDEP_HASH(fs, inum) \ 941 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash]) 942 static struct sema inodedep_in_progress; 943 944 /* 945 * Helper routine for inodedep_lookup() 946 */ 947 static __inline 948 struct inodedep * 949 inodedep_find(struct inodedep_hashhead *inodedephd, struct fs *fs, ino_t inum) 950 { 951 struct inodedep *inodedep; 952 953 LIST_FOREACH(inodedep, inodedephd, id_hash) { 954 if (inum == inodedep->id_ino && fs == inodedep->id_fs) 955 return(inodedep); 956 } 957 return (NULL); 958 } 959 960 /* 961 * Look up a inodedep. Return 1 if found, 0 if not found. 962 * If not found, allocate if DEPALLOC flag is passed. 963 * Found or allocated entry is returned in inodedeppp. 964 * This routine must be called with splbio interrupts blocked. 965 */ 966 static int 967 inodedep_lookup(struct fs *fs, ino_t inum, int flags, 968 struct inodedep **inodedeppp) 969 { 970 struct inodedep *inodedep; 971 struct inodedep_hashhead *inodedephd; 972 int firsttry; 973 974 #ifdef DEBUG 975 if (lk.lkt_held == NOHOLDER) 976 panic("inodedep_lookup: lock not held"); 977 #endif 978 firsttry = 1; 979 inodedephd = INODEDEP_HASH(fs, inum); 980 top: 981 *inodedeppp = inodedep_find(inodedephd, fs, inum); 982 if (*inodedeppp) 983 return (1); 984 if ((flags & DEPALLOC) == 0) 985 return (0); 986 /* 987 * If we are over our limit, try to improve the situation. 988 */ 989 if (num_inodedep > max_softdeps && firsttry && 990 speedup_syncer() == 0 && (flags & NODELAY) == 0 && 991 request_cleanup(FLUSH_INODES, 1)) { 992 firsttry = 0; 993 goto top; 994 } 995 if (sema_get(&inodedep_in_progress, &lk) == 0) { 996 ACQUIRE_LOCK(&lk); 997 goto top; 998 } 999 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep), 1000 M_INODEDEP, M_SOFTDEP_FLAGS | M_ZERO); 1001 if (inodedep_find(inodedephd, fs, inum)) { 1002 kprintf("inodedep_lookup: blocking race avoided\n"); 1003 ACQUIRE_LOCK(&lk); 1004 sema_release(&inodedep_in_progress); 1005 kfree(inodedep, M_INODEDEP); 1006 goto top; 1007 } 1008 inodedep->id_list.wk_type = D_INODEDEP; 1009 inodedep->id_fs = fs; 1010 inodedep->id_ino = inum; 1011 inodedep->id_state = ALLCOMPLETE; 1012 inodedep->id_nlinkdelta = 0; 1013 inodedep->id_savedino = NULL; 1014 inodedep->id_savedsize = -1; 1015 inodedep->id_buf = NULL; 1016 LIST_INIT(&inodedep->id_pendinghd); 1017 LIST_INIT(&inodedep->id_inowait); 1018 LIST_INIT(&inodedep->id_bufwait); 1019 TAILQ_INIT(&inodedep->id_inoupdt); 1020 TAILQ_INIT(&inodedep->id_newinoupdt); 1021 ACQUIRE_LOCK(&lk); 1022 num_inodedep += 1; 1023 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 1024 sema_release(&inodedep_in_progress); 1025 *inodedeppp = inodedep; 1026 return (0); 1027 } 1028 1029 /* 1030 * Structures and routines associated with newblk caching. 1031 */ 1032 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl; 1033 u_long newblk_hash; /* size of hash table - 1 */ 1034 #define NEWBLK_HASH(fs, inum) \ 1035 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash]) 1036 static struct sema newblk_in_progress; 1037 1038 /* 1039 * Helper routine for newblk_lookup() 1040 */ 1041 static __inline 1042 struct newblk * 1043 newblk_find(struct newblk_hashhead *newblkhd, struct fs *fs, 1044 ufs_daddr_t newblkno) 1045 { 1046 struct newblk *newblk; 1047 1048 LIST_FOREACH(newblk, newblkhd, nb_hash) { 1049 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs) 1050 return (newblk); 1051 } 1052 return(NULL); 1053 } 1054 1055 /* 1056 * Look up a newblk. Return 1 if found, 0 if not found. 1057 * If not found, allocate if DEPALLOC flag is passed. 1058 * Found or allocated entry is returned in newblkpp. 1059 */ 1060 static int 1061 newblk_lookup(struct fs *fs, ufs_daddr_t newblkno, int flags, 1062 struct newblk **newblkpp) 1063 { 1064 struct newblk *newblk; 1065 struct newblk_hashhead *newblkhd; 1066 1067 newblkhd = NEWBLK_HASH(fs, newblkno); 1068 top: 1069 *newblkpp = newblk_find(newblkhd, fs, newblkno); 1070 if (*newblkpp) 1071 return(1); 1072 if ((flags & DEPALLOC) == 0) 1073 return (0); 1074 if (sema_get(&newblk_in_progress, 0) == 0) 1075 goto top; 1076 MALLOC(newblk, struct newblk *, sizeof(struct newblk), 1077 M_NEWBLK, M_SOFTDEP_FLAGS | M_ZERO); 1078 1079 if (newblk_find(newblkhd, fs, newblkno)) { 1080 kprintf("newblk_lookup: blocking race avoided\n"); 1081 sema_release(&pagedep_in_progress); 1082 kfree(newblk, M_NEWBLK); 1083 goto top; 1084 } 1085 newblk->nb_state = 0; 1086 newblk->nb_fs = fs; 1087 newblk->nb_newblkno = newblkno; 1088 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 1089 sema_release(&newblk_in_progress); 1090 *newblkpp = newblk; 1091 return (0); 1092 } 1093 1094 /* 1095 * Executed during filesystem system initialization before 1096 * mounting any filesystems. 1097 */ 1098 void 1099 softdep_initialize(void) 1100 { 1101 callout_init(&handle); 1102 bioops = softdep_bioops; /* XXX hack */ 1103 1104 LIST_INIT(&mkdirlisthd); 1105 LIST_INIT(&softdep_workitem_pending); 1106 max_softdeps = min(desiredvnodes * 8, 1107 M_INODEDEP->ks_limit / (2 * sizeof(struct inodedep))); 1108 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 1109 &pagedep_hash); 1110 sema_init(&pagedep_in_progress, "pagedep", 0, 0); 1111 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash); 1112 sema_init(&inodedep_in_progress, "inodedep", 0, 0); 1113 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash); 1114 sema_init(&newblk_in_progress, "newblk", 0, 0); 1115 } 1116 1117 /* 1118 * Called at mount time to notify the dependency code that a 1119 * filesystem wishes to use it. 1120 */ 1121 int 1122 softdep_mount(struct vnode *devvp, struct mount *mp, struct fs *fs) 1123 { 1124 struct csum cstotal; 1125 struct cg *cgp; 1126 struct buf *bp; 1127 int error, cyl; 1128 1129 mp->mnt_flag &= ~MNT_ASYNC; 1130 mp->mnt_flag |= MNT_SOFTDEP; 1131 /* 1132 * When doing soft updates, the counters in the 1133 * superblock may have gotten out of sync, so we have 1134 * to scan the cylinder groups and recalculate them. 1135 */ 1136 if (fs->fs_clean != 0) 1137 return (0); 1138 bzero(&cstotal, sizeof cstotal); 1139 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 1140 if ((error = bread(devvp, fsbtodoff(fs, cgtod(fs, cyl)), 1141 fs->fs_cgsize, &bp)) != 0) { 1142 brelse(bp); 1143 return (error); 1144 } 1145 cgp = (struct cg *)bp->b_data; 1146 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 1147 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 1148 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 1149 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 1150 fs->fs_cs(fs, cyl) = cgp->cg_cs; 1151 brelse(bp); 1152 } 1153 #ifdef DEBUG 1154 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 1155 kprintf("ffs_mountfs: superblock updated for soft updates\n"); 1156 #endif 1157 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 1158 return (0); 1159 } 1160 1161 /* 1162 * Protecting the freemaps (or bitmaps). 1163 * 1164 * To eliminate the need to execute fsck before mounting a filesystem 1165 * after a power failure, one must (conservatively) guarantee that the 1166 * on-disk copy of the bitmaps never indicate that a live inode or block is 1167 * free. So, when a block or inode is allocated, the bitmap should be 1168 * updated (on disk) before any new pointers. When a block or inode is 1169 * freed, the bitmap should not be updated until all pointers have been 1170 * reset. The latter dependency is handled by the delayed de-allocation 1171 * approach described below for block and inode de-allocation. The former 1172 * dependency is handled by calling the following procedure when a block or 1173 * inode is allocated. When an inode is allocated an "inodedep" is created 1174 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 1175 * Each "inodedep" is also inserted into the hash indexing structure so 1176 * that any additional link additions can be made dependent on the inode 1177 * allocation. 1178 * 1179 * The ufs filesystem maintains a number of free block counts (e.g., per 1180 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 1181 * in addition to the bitmaps. These counts are used to improve efficiency 1182 * during allocation and therefore must be consistent with the bitmaps. 1183 * There is no convenient way to guarantee post-crash consistency of these 1184 * counts with simple update ordering, for two main reasons: (1) The counts 1185 * and bitmaps for a single cylinder group block are not in the same disk 1186 * sector. If a disk write is interrupted (e.g., by power failure), one may 1187 * be written and the other not. (2) Some of the counts are located in the 1188 * superblock rather than the cylinder group block. So, we focus our soft 1189 * updates implementation on protecting the bitmaps. When mounting a 1190 * filesystem, we recompute the auxiliary counts from the bitmaps. 1191 */ 1192 1193 /* 1194 * Called just after updating the cylinder group block to allocate an inode. 1195 * 1196 * Parameters: 1197 * bp: buffer for cylgroup block with inode map 1198 * ip: inode related to allocation 1199 * newinum: new inode number being allocated 1200 */ 1201 void 1202 softdep_setup_inomapdep(struct buf *bp, struct inode *ip, ino_t newinum) 1203 { 1204 struct inodedep *inodedep; 1205 struct bmsafemap *bmsafemap; 1206 1207 /* 1208 * Create a dependency for the newly allocated inode. 1209 * Panic if it already exists as something is seriously wrong. 1210 * Otherwise add it to the dependency list for the buffer holding 1211 * the cylinder group map from which it was allocated. 1212 */ 1213 ACQUIRE_LOCK(&lk); 1214 if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) { 1215 FREE_LOCK(&lk); 1216 panic("softdep_setup_inomapdep: found inode"); 1217 } 1218 inodedep->id_buf = bp; 1219 inodedep->id_state &= ~DEPCOMPLETE; 1220 bmsafemap = bmsafemap_lookup(bp); 1221 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 1222 FREE_LOCK(&lk); 1223 } 1224 1225 /* 1226 * Called just after updating the cylinder group block to 1227 * allocate block or fragment. 1228 * 1229 * Parameters: 1230 * bp: buffer for cylgroup block with block map 1231 * fs: filesystem doing allocation 1232 * newblkno: number of newly allocated block 1233 */ 1234 void 1235 softdep_setup_blkmapdep(struct buf *bp, struct fs *fs, 1236 ufs_daddr_t newblkno) 1237 { 1238 struct newblk *newblk; 1239 struct bmsafemap *bmsafemap; 1240 1241 /* 1242 * Create a dependency for the newly allocated block. 1243 * Add it to the dependency list for the buffer holding 1244 * the cylinder group map from which it was allocated. 1245 */ 1246 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0) 1247 panic("softdep_setup_blkmapdep: found block"); 1248 ACQUIRE_LOCK(&lk); 1249 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp); 1250 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 1251 FREE_LOCK(&lk); 1252 } 1253 1254 /* 1255 * Find the bmsafemap associated with a cylinder group buffer. 1256 * If none exists, create one. The buffer must be locked when 1257 * this routine is called and this routine must be called with 1258 * splbio interrupts blocked. 1259 */ 1260 static struct bmsafemap * 1261 bmsafemap_lookup(struct buf *bp) 1262 { 1263 struct bmsafemap *bmsafemap; 1264 struct worklist *wk; 1265 1266 #ifdef DEBUG 1267 if (lk.lkt_held == NOHOLDER) 1268 panic("bmsafemap_lookup: lock not held"); 1269 #endif 1270 LIST_FOREACH(wk, &bp->b_dep, wk_list) 1271 if (wk->wk_type == D_BMSAFEMAP) 1272 return (WK_BMSAFEMAP(wk)); 1273 FREE_LOCK(&lk); 1274 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap), 1275 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 1276 bmsafemap->sm_list.wk_type = D_BMSAFEMAP; 1277 bmsafemap->sm_list.wk_state = 0; 1278 bmsafemap->sm_buf = bp; 1279 LIST_INIT(&bmsafemap->sm_allocdirecthd); 1280 LIST_INIT(&bmsafemap->sm_allocindirhd); 1281 LIST_INIT(&bmsafemap->sm_inodedephd); 1282 LIST_INIT(&bmsafemap->sm_newblkhd); 1283 ACQUIRE_LOCK(&lk); 1284 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 1285 return (bmsafemap); 1286 } 1287 1288 /* 1289 * Direct block allocation dependencies. 1290 * 1291 * When a new block is allocated, the corresponding disk locations must be 1292 * initialized (with zeros or new data) before the on-disk inode points to 1293 * them. Also, the freemap from which the block was allocated must be 1294 * updated (on disk) before the inode's pointer. These two dependencies are 1295 * independent of each other and are needed for all file blocks and indirect 1296 * blocks that are pointed to directly by the inode. Just before the 1297 * "in-core" version of the inode is updated with a newly allocated block 1298 * number, a procedure (below) is called to setup allocation dependency 1299 * structures. These structures are removed when the corresponding 1300 * dependencies are satisfied or when the block allocation becomes obsolete 1301 * (i.e., the file is deleted, the block is de-allocated, or the block is a 1302 * fragment that gets upgraded). All of these cases are handled in 1303 * procedures described later. 1304 * 1305 * When a file extension causes a fragment to be upgraded, either to a larger 1306 * fragment or to a full block, the on-disk location may change (if the 1307 * previous fragment could not simply be extended). In this case, the old 1308 * fragment must be de-allocated, but not until after the inode's pointer has 1309 * been updated. In most cases, this is handled by later procedures, which 1310 * will construct a "freefrag" structure to be added to the workitem queue 1311 * when the inode update is complete (or obsolete). The main exception to 1312 * this is when an allocation occurs while a pending allocation dependency 1313 * (for the same block pointer) remains. This case is handled in the main 1314 * allocation dependency setup procedure by immediately freeing the 1315 * unreferenced fragments. 1316 * 1317 * Parameters: 1318 * ip: inode to which block is being added 1319 * lbn: block pointer within inode 1320 * newblkno: disk block number being added 1321 * oldblkno: previous block number, 0 unless frag 1322 * newsize: size of new block 1323 * oldsize: size of new block 1324 * bp: bp for allocated block 1325 */ 1326 void 1327 softdep_setup_allocdirect(struct inode *ip, ufs_lbn_t lbn, ufs_daddr_t newblkno, 1328 ufs_daddr_t oldblkno, long newsize, long oldsize, 1329 struct buf *bp) 1330 { 1331 struct allocdirect *adp, *oldadp; 1332 struct allocdirectlst *adphead; 1333 struct bmsafemap *bmsafemap; 1334 struct inodedep *inodedep; 1335 struct pagedep *pagedep; 1336 struct newblk *newblk; 1337 1338 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect), 1339 M_ALLOCDIRECT, M_SOFTDEP_FLAGS); 1340 bzero(adp, sizeof(struct allocdirect)); 1341 adp->ad_list.wk_type = D_ALLOCDIRECT; 1342 adp->ad_lbn = lbn; 1343 adp->ad_newblkno = newblkno; 1344 adp->ad_oldblkno = oldblkno; 1345 adp->ad_newsize = newsize; 1346 adp->ad_oldsize = oldsize; 1347 adp->ad_state = ATTACHED; 1348 if (newblkno == oldblkno) 1349 adp->ad_freefrag = NULL; 1350 else 1351 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1352 1353 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1354 panic("softdep_setup_allocdirect: lost block"); 1355 1356 ACQUIRE_LOCK(&lk); 1357 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep); 1358 adp->ad_inodedep = inodedep; 1359 1360 if (newblk->nb_state == DEPCOMPLETE) { 1361 adp->ad_state |= DEPCOMPLETE; 1362 adp->ad_buf = NULL; 1363 } else { 1364 bmsafemap = newblk->nb_bmsafemap; 1365 adp->ad_buf = bmsafemap->sm_buf; 1366 LIST_REMOVE(newblk, nb_deps); 1367 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1368 } 1369 LIST_REMOVE(newblk, nb_hash); 1370 FREE(newblk, M_NEWBLK); 1371 1372 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list); 1373 if (lbn >= NDADDR) { 1374 /* allocating an indirect block */ 1375 if (oldblkno != 0) { 1376 FREE_LOCK(&lk); 1377 panic("softdep_setup_allocdirect: non-zero indir"); 1378 } 1379 } else { 1380 /* 1381 * Allocating a direct block. 1382 * 1383 * If we are allocating a directory block, then we must 1384 * allocate an associated pagedep to track additions and 1385 * deletions. 1386 */ 1387 if ((ip->i_mode & IFMT) == IFDIR && 1388 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1389 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 1390 } 1391 /* 1392 * The list of allocdirects must be kept in sorted and ascending 1393 * order so that the rollback routines can quickly determine the 1394 * first uncommitted block (the size of the file stored on disk 1395 * ends at the end of the lowest committed fragment, or if there 1396 * are no fragments, at the end of the highest committed block). 1397 * Since files generally grow, the typical case is that the new 1398 * block is to be added at the end of the list. We speed this 1399 * special case by checking against the last allocdirect in the 1400 * list before laboriously traversing the list looking for the 1401 * insertion point. 1402 */ 1403 adphead = &inodedep->id_newinoupdt; 1404 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1405 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1406 /* insert at end of list */ 1407 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1408 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1409 allocdirect_merge(adphead, adp, oldadp); 1410 FREE_LOCK(&lk); 1411 return; 1412 } 1413 TAILQ_FOREACH(oldadp, adphead, ad_next) { 1414 if (oldadp->ad_lbn >= lbn) 1415 break; 1416 } 1417 if (oldadp == NULL) { 1418 FREE_LOCK(&lk); 1419 panic("softdep_setup_allocdirect: lost entry"); 1420 } 1421 /* insert in middle of list */ 1422 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1423 if (oldadp->ad_lbn == lbn) 1424 allocdirect_merge(adphead, adp, oldadp); 1425 FREE_LOCK(&lk); 1426 } 1427 1428 /* 1429 * Replace an old allocdirect dependency with a newer one. 1430 * This routine must be called with splbio interrupts blocked. 1431 * 1432 * Parameters: 1433 * adphead: head of list holding allocdirects 1434 * newadp: allocdirect being added 1435 * oldadp: existing allocdirect being checked 1436 */ 1437 static void 1438 allocdirect_merge(struct allocdirectlst *adphead, 1439 struct allocdirect *newadp, 1440 struct allocdirect *oldadp) 1441 { 1442 struct freefrag *freefrag; 1443 1444 #ifdef DEBUG 1445 if (lk.lkt_held == NOHOLDER) 1446 panic("allocdirect_merge: lock not held"); 1447 #endif 1448 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 1449 newadp->ad_oldsize != oldadp->ad_newsize || 1450 newadp->ad_lbn >= NDADDR) { 1451 FREE_LOCK(&lk); 1452 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d", 1453 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn, 1454 NDADDR); 1455 } 1456 newadp->ad_oldblkno = oldadp->ad_oldblkno; 1457 newadp->ad_oldsize = oldadp->ad_oldsize; 1458 /* 1459 * If the old dependency had a fragment to free or had never 1460 * previously had a block allocated, then the new dependency 1461 * can immediately post its freefrag and adopt the old freefrag. 1462 * This action is done by swapping the freefrag dependencies. 1463 * The new dependency gains the old one's freefrag, and the 1464 * old one gets the new one and then immediately puts it on 1465 * the worklist when it is freed by free_allocdirect. It is 1466 * not possible to do this swap when the old dependency had a 1467 * non-zero size but no previous fragment to free. This condition 1468 * arises when the new block is an extension of the old block. 1469 * Here, the first part of the fragment allocated to the new 1470 * dependency is part of the block currently claimed on disk by 1471 * the old dependency, so cannot legitimately be freed until the 1472 * conditions for the new dependency are fulfilled. 1473 */ 1474 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 1475 freefrag = newadp->ad_freefrag; 1476 newadp->ad_freefrag = oldadp->ad_freefrag; 1477 oldadp->ad_freefrag = freefrag; 1478 } 1479 free_allocdirect(adphead, oldadp, 0); 1480 } 1481 1482 /* 1483 * Allocate a new freefrag structure if needed. 1484 */ 1485 static struct freefrag * 1486 newfreefrag(struct inode *ip, ufs_daddr_t blkno, long size) 1487 { 1488 struct freefrag *freefrag; 1489 struct fs *fs; 1490 1491 if (blkno == 0) 1492 return (NULL); 1493 fs = ip->i_fs; 1494 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 1495 panic("newfreefrag: frag size"); 1496 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag), 1497 M_FREEFRAG, M_SOFTDEP_FLAGS); 1498 freefrag->ff_list.wk_type = D_FREEFRAG; 1499 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */ 1500 freefrag->ff_inum = ip->i_number; 1501 freefrag->ff_fs = fs; 1502 freefrag->ff_devvp = ip->i_devvp; 1503 freefrag->ff_blkno = blkno; 1504 freefrag->ff_fragsize = size; 1505 return (freefrag); 1506 } 1507 1508 /* 1509 * This workitem de-allocates fragments that were replaced during 1510 * file block allocation. 1511 */ 1512 static void 1513 handle_workitem_freefrag(struct freefrag *freefrag) 1514 { 1515 struct inode tip; 1516 1517 tip.i_fs = freefrag->ff_fs; 1518 tip.i_devvp = freefrag->ff_devvp; 1519 tip.i_dev = freefrag->ff_devvp->v_rdev; 1520 tip.i_number = freefrag->ff_inum; 1521 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */ 1522 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize); 1523 FREE(freefrag, M_FREEFRAG); 1524 } 1525 1526 /* 1527 * Indirect block allocation dependencies. 1528 * 1529 * The same dependencies that exist for a direct block also exist when 1530 * a new block is allocated and pointed to by an entry in a block of 1531 * indirect pointers. The undo/redo states described above are also 1532 * used here. Because an indirect block contains many pointers that 1533 * may have dependencies, a second copy of the entire in-memory indirect 1534 * block is kept. The buffer cache copy is always completely up-to-date. 1535 * The second copy, which is used only as a source for disk writes, 1536 * contains only the safe pointers (i.e., those that have no remaining 1537 * update dependencies). The second copy is freed when all pointers 1538 * are safe. The cache is not allowed to replace indirect blocks with 1539 * pending update dependencies. If a buffer containing an indirect 1540 * block with dependencies is written, these routines will mark it 1541 * dirty again. It can only be successfully written once all the 1542 * dependencies are removed. The ffs_fsync routine in conjunction with 1543 * softdep_sync_metadata work together to get all the dependencies 1544 * removed so that a file can be successfully written to disk. Three 1545 * procedures are used when setting up indirect block pointer 1546 * dependencies. The division is necessary because of the organization 1547 * of the "balloc" routine and because of the distinction between file 1548 * pages and file metadata blocks. 1549 */ 1550 1551 /* 1552 * Allocate a new allocindir structure. 1553 * 1554 * Parameters: 1555 * ip: inode for file being extended 1556 * ptrno: offset of pointer in indirect block 1557 * newblkno: disk block number being added 1558 * oldblkno: previous block number, 0 if none 1559 */ 1560 static struct allocindir * 1561 newallocindir(struct inode *ip, int ptrno, ufs_daddr_t newblkno, 1562 ufs_daddr_t oldblkno) 1563 { 1564 struct allocindir *aip; 1565 1566 MALLOC(aip, struct allocindir *, sizeof(struct allocindir), 1567 M_ALLOCINDIR, M_SOFTDEP_FLAGS); 1568 bzero(aip, sizeof(struct allocindir)); 1569 aip->ai_list.wk_type = D_ALLOCINDIR; 1570 aip->ai_state = ATTACHED; 1571 aip->ai_offset = ptrno; 1572 aip->ai_newblkno = newblkno; 1573 aip->ai_oldblkno = oldblkno; 1574 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize); 1575 return (aip); 1576 } 1577 1578 /* 1579 * Called just before setting an indirect block pointer 1580 * to a newly allocated file page. 1581 * 1582 * Parameters: 1583 * ip: inode for file being extended 1584 * lbn: allocated block number within file 1585 * bp: buffer with indirect blk referencing page 1586 * ptrno: offset of pointer in indirect block 1587 * newblkno: disk block number being added 1588 * oldblkno: previous block number, 0 if none 1589 * nbp: buffer holding allocated page 1590 */ 1591 void 1592 softdep_setup_allocindir_page(struct inode *ip, ufs_lbn_t lbn, 1593 struct buf *bp, int ptrno, 1594 ufs_daddr_t newblkno, ufs_daddr_t oldblkno, 1595 struct buf *nbp) 1596 { 1597 struct allocindir *aip; 1598 struct pagedep *pagedep; 1599 1600 aip = newallocindir(ip, ptrno, newblkno, oldblkno); 1601 ACQUIRE_LOCK(&lk); 1602 /* 1603 * If we are allocating a directory page, then we must 1604 * allocate an associated pagedep to track additions and 1605 * deletions. 1606 */ 1607 if ((ip->i_mode & IFMT) == IFDIR && 1608 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1609 WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list); 1610 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1611 FREE_LOCK(&lk); 1612 setup_allocindir_phase2(bp, ip, aip); 1613 } 1614 1615 /* 1616 * Called just before setting an indirect block pointer to a 1617 * newly allocated indirect block. 1618 * Parameters: 1619 * nbp: newly allocated indirect block 1620 * ip: inode for file being extended 1621 * bp: indirect block referencing allocated block 1622 * ptrno: offset of pointer in indirect block 1623 * newblkno: disk block number being added 1624 */ 1625 void 1626 softdep_setup_allocindir_meta(struct buf *nbp, struct inode *ip, 1627 struct buf *bp, int ptrno, 1628 ufs_daddr_t newblkno) 1629 { 1630 struct allocindir *aip; 1631 1632 aip = newallocindir(ip, ptrno, newblkno, 0); 1633 ACQUIRE_LOCK(&lk); 1634 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1635 FREE_LOCK(&lk); 1636 setup_allocindir_phase2(bp, ip, aip); 1637 } 1638 1639 /* 1640 * Called to finish the allocation of the "aip" allocated 1641 * by one of the two routines above. 1642 * 1643 * Parameters: 1644 * bp: in-memory copy of the indirect block 1645 * ip: inode for file being extended 1646 * aip: allocindir allocated by the above routines 1647 */ 1648 static void 1649 setup_allocindir_phase2(struct buf *bp, struct inode *ip, 1650 struct allocindir *aip) 1651 { 1652 struct worklist *wk; 1653 struct indirdep *indirdep, *newindirdep; 1654 struct bmsafemap *bmsafemap; 1655 struct allocindir *oldaip; 1656 struct freefrag *freefrag; 1657 struct newblk *newblk; 1658 1659 if (bp->b_loffset >= 0) 1660 panic("setup_allocindir_phase2: not indir blk"); 1661 for (indirdep = NULL, newindirdep = NULL; ; ) { 1662 ACQUIRE_LOCK(&lk); 1663 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 1664 if (wk->wk_type != D_INDIRDEP) 1665 continue; 1666 indirdep = WK_INDIRDEP(wk); 1667 break; 1668 } 1669 if (indirdep == NULL && newindirdep) { 1670 indirdep = newindirdep; 1671 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 1672 newindirdep = NULL; 1673 } 1674 FREE_LOCK(&lk); 1675 if (indirdep) { 1676 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0, 1677 &newblk) == 0) 1678 panic("setup_allocindir: lost block"); 1679 ACQUIRE_LOCK(&lk); 1680 if (newblk->nb_state == DEPCOMPLETE) { 1681 aip->ai_state |= DEPCOMPLETE; 1682 aip->ai_buf = NULL; 1683 } else { 1684 bmsafemap = newblk->nb_bmsafemap; 1685 aip->ai_buf = bmsafemap->sm_buf; 1686 LIST_REMOVE(newblk, nb_deps); 1687 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd, 1688 aip, ai_deps); 1689 } 1690 LIST_REMOVE(newblk, nb_hash); 1691 FREE(newblk, M_NEWBLK); 1692 aip->ai_indirdep = indirdep; 1693 /* 1694 * Check to see if there is an existing dependency 1695 * for this block. If there is, merge the old 1696 * dependency into the new one. 1697 */ 1698 if (aip->ai_oldblkno == 0) 1699 oldaip = NULL; 1700 else 1701 1702 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) 1703 if (oldaip->ai_offset == aip->ai_offset) 1704 break; 1705 if (oldaip != NULL) { 1706 if (oldaip->ai_newblkno != aip->ai_oldblkno) { 1707 FREE_LOCK(&lk); 1708 panic("setup_allocindir_phase2: blkno"); 1709 } 1710 aip->ai_oldblkno = oldaip->ai_oldblkno; 1711 freefrag = oldaip->ai_freefrag; 1712 oldaip->ai_freefrag = aip->ai_freefrag; 1713 aip->ai_freefrag = freefrag; 1714 free_allocindir(oldaip, NULL); 1715 } 1716 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 1717 ((ufs_daddr_t *)indirdep->ir_savebp->b_data) 1718 [aip->ai_offset] = aip->ai_oldblkno; 1719 FREE_LOCK(&lk); 1720 } 1721 if (newindirdep) { 1722 /* 1723 * Avoid any possibility of data corruption by 1724 * ensuring that our old version is thrown away. 1725 */ 1726 newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 1727 brelse(newindirdep->ir_savebp); 1728 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP); 1729 } 1730 if (indirdep) 1731 break; 1732 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep), 1733 M_INDIRDEP, M_SOFTDEP_FLAGS); 1734 newindirdep->ir_list.wk_type = D_INDIRDEP; 1735 newindirdep->ir_state = ATTACHED; 1736 LIST_INIT(&newindirdep->ir_deplisthd); 1737 LIST_INIT(&newindirdep->ir_donehd); 1738 if (bp->b_bio2.bio_offset == NOOFFSET) { 1739 VOP_BMAP(bp->b_vp, bp->b_bio1.bio_offset, 1740 &bp->b_bio2.bio_offset, NULL, NULL); 1741 } 1742 KKASSERT(bp->b_bio2.bio_offset != NOOFFSET); 1743 newindirdep->ir_savebp = getblk(ip->i_devvp, 1744 bp->b_bio2.bio_offset, 1745 bp->b_bcount, 0, 0); 1746 BUF_KERNPROC(newindirdep->ir_savebp); 1747 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 1748 } 1749 } 1750 1751 /* 1752 * Block de-allocation dependencies. 1753 * 1754 * When blocks are de-allocated, the on-disk pointers must be nullified before 1755 * the blocks are made available for use by other files. (The true 1756 * requirement is that old pointers must be nullified before new on-disk 1757 * pointers are set. We chose this slightly more stringent requirement to 1758 * reduce complexity.) Our implementation handles this dependency by updating 1759 * the inode (or indirect block) appropriately but delaying the actual block 1760 * de-allocation (i.e., freemap and free space count manipulation) until 1761 * after the updated versions reach stable storage. After the disk is 1762 * updated, the blocks can be safely de-allocated whenever it is convenient. 1763 * This implementation handles only the common case of reducing a file's 1764 * length to zero. Other cases are handled by the conventional synchronous 1765 * write approach. 1766 * 1767 * The ffs implementation with which we worked double-checks 1768 * the state of the block pointers and file size as it reduces 1769 * a file's length. Some of this code is replicated here in our 1770 * soft updates implementation. The freeblks->fb_chkcnt field is 1771 * used to transfer a part of this information to the procedure 1772 * that eventually de-allocates the blocks. 1773 * 1774 * This routine should be called from the routine that shortens 1775 * a file's length, before the inode's size or block pointers 1776 * are modified. It will save the block pointer information for 1777 * later release and zero the inode so that the calling routine 1778 * can release it. 1779 */ 1780 struct softdep_setup_freeblocks_info { 1781 struct fs *fs; 1782 struct inode *ip; 1783 }; 1784 1785 static int softdep_setup_freeblocks_bp(struct buf *bp, void *data); 1786 1787 /* 1788 * Parameters: 1789 * ip: The inode whose length is to be reduced 1790 * length: The new length for the file 1791 */ 1792 void 1793 softdep_setup_freeblocks(struct inode *ip, off_t length) 1794 { 1795 struct softdep_setup_freeblocks_info info; 1796 struct freeblks *freeblks; 1797 struct inodedep *inodedep; 1798 struct allocdirect *adp; 1799 struct vnode *vp; 1800 struct buf *bp; 1801 struct fs *fs; 1802 int i, error, delay; 1803 int count; 1804 1805 fs = ip->i_fs; 1806 if (length != 0) 1807 panic("softde_setup_freeblocks: non-zero length"); 1808 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks), 1809 M_FREEBLKS, M_SOFTDEP_FLAGS); 1810 bzero(freeblks, sizeof(struct freeblks)); 1811 freeblks->fb_list.wk_type = D_FREEBLKS; 1812 freeblks->fb_state = ATTACHED; 1813 freeblks->fb_uid = ip->i_uid; 1814 freeblks->fb_previousinum = ip->i_number; 1815 freeblks->fb_devvp = ip->i_devvp; 1816 freeblks->fb_fs = fs; 1817 freeblks->fb_oldsize = ip->i_size; 1818 freeblks->fb_newsize = length; 1819 freeblks->fb_chkcnt = ip->i_blocks; 1820 for (i = 0; i < NDADDR; i++) { 1821 freeblks->fb_dblks[i] = ip->i_db[i]; 1822 ip->i_db[i] = 0; 1823 } 1824 for (i = 0; i < NIADDR; i++) { 1825 freeblks->fb_iblks[i] = ip->i_ib[i]; 1826 ip->i_ib[i] = 0; 1827 } 1828 ip->i_blocks = 0; 1829 ip->i_size = 0; 1830 /* 1831 * Push the zero'ed inode to to its disk buffer so that we are free 1832 * to delete its dependencies below. Once the dependencies are gone 1833 * the buffer can be safely released. 1834 */ 1835 if ((error = bread(ip->i_devvp, 1836 fsbtodoff(fs, ino_to_fsba(fs, ip->i_number)), 1837 (int)fs->fs_bsize, &bp)) != 0) 1838 softdep_error("softdep_setup_freeblocks", error); 1839 *((struct ufs1_dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) = 1840 ip->i_din; 1841 /* 1842 * Find and eliminate any inode dependencies. 1843 */ 1844 ACQUIRE_LOCK(&lk); 1845 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep); 1846 if ((inodedep->id_state & IOSTARTED) != 0) { 1847 FREE_LOCK(&lk); 1848 panic("softdep_setup_freeblocks: inode busy"); 1849 } 1850 /* 1851 * Add the freeblks structure to the list of operations that 1852 * must await the zero'ed inode being written to disk. If we 1853 * still have a bitmap dependency (delay == 0), then the inode 1854 * has never been written to disk, so we can process the 1855 * freeblks below once we have deleted the dependencies. 1856 */ 1857 delay = (inodedep->id_state & DEPCOMPLETE); 1858 if (delay) 1859 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list); 1860 /* 1861 * Because the file length has been truncated to zero, any 1862 * pending block allocation dependency structures associated 1863 * with this inode are obsolete and can simply be de-allocated. 1864 * We must first merge the two dependency lists to get rid of 1865 * any duplicate freefrag structures, then purge the merged list. 1866 */ 1867 merge_inode_lists(inodedep); 1868 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 1869 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 1870 FREE_LOCK(&lk); 1871 bdwrite(bp); 1872 /* 1873 * We must wait for any I/O in progress to finish so that 1874 * all potential buffers on the dirty list will be visible. 1875 * Once they are all there, walk the list and get rid of 1876 * any dependencies. 1877 */ 1878 vp = ITOV(ip); 1879 ACQUIRE_LOCK(&lk); 1880 drain_output(vp, 1); 1881 1882 info.fs = fs; 1883 info.ip = ip; 1884 do { 1885 count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 1886 softdep_setup_freeblocks_bp, &info); 1887 } while (count != 0); 1888 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0) 1889 (void)free_inodedep(inodedep); 1890 1891 if (delay) { 1892 freeblks->fb_state |= DEPCOMPLETE; 1893 /* 1894 * If the inode with zeroed block pointers is now on disk 1895 * we can start freeing blocks. Add freeblks to the worklist 1896 * instead of calling handle_workitem_freeblocks directly as 1897 * it is more likely that additional IO is needed to complete 1898 * the request here than in the !delay case. 1899 */ 1900 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 1901 add_to_worklist(&freeblks->fb_list); 1902 } 1903 1904 FREE_LOCK(&lk); 1905 /* 1906 * If the inode has never been written to disk (delay == 0), 1907 * then we can process the freeblks now that we have deleted 1908 * the dependencies. 1909 */ 1910 if (!delay) 1911 handle_workitem_freeblocks(freeblks); 1912 } 1913 1914 static int 1915 softdep_setup_freeblocks_bp(struct buf *bp, void *data) 1916 { 1917 struct softdep_setup_freeblocks_info *info = data; 1918 struct inodedep *inodedep; 1919 1920 if (getdirtybuf(&bp, MNT_WAIT) == 0) { 1921 kprintf("softdep_setup_freeblocks_bp(1): caught bp %p going away\n", bp); 1922 return(-1); 1923 } 1924 if (bp->b_vp != ITOV(info->ip) || (bp->b_flags & B_DELWRI) == 0) { 1925 kprintf("softdep_setup_freeblocks_bp(2): caught bp %p going away\n", bp); 1926 BUF_UNLOCK(bp); 1927 return(-1); 1928 } 1929 (void) inodedep_lookup(info->fs, info->ip->i_number, 0, &inodedep); 1930 deallocate_dependencies(bp, inodedep); 1931 bp->b_flags |= B_INVAL | B_NOCACHE; 1932 FREE_LOCK(&lk); 1933 brelse(bp); 1934 ACQUIRE_LOCK(&lk); 1935 return(1); 1936 } 1937 1938 /* 1939 * Reclaim any dependency structures from a buffer that is about to 1940 * be reallocated to a new vnode. The buffer must be locked, thus, 1941 * no I/O completion operations can occur while we are manipulating 1942 * its associated dependencies. The mutex is held so that other I/O's 1943 * associated with related dependencies do not occur. 1944 */ 1945 static void 1946 deallocate_dependencies(struct buf *bp, struct inodedep *inodedep) 1947 { 1948 struct worklist *wk; 1949 struct indirdep *indirdep; 1950 struct allocindir *aip; 1951 struct pagedep *pagedep; 1952 struct dirrem *dirrem; 1953 struct diradd *dap; 1954 int i; 1955 1956 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 1957 switch (wk->wk_type) { 1958 1959 case D_INDIRDEP: 1960 indirdep = WK_INDIRDEP(wk); 1961 /* 1962 * None of the indirect pointers will ever be visible, 1963 * so they can simply be tossed. GOINGAWAY ensures 1964 * that allocated pointers will be saved in the buffer 1965 * cache until they are freed. Note that they will 1966 * only be able to be found by their physical address 1967 * since the inode mapping the logical address will 1968 * be gone. The save buffer used for the safe copy 1969 * was allocated in setup_allocindir_phase2 using 1970 * the physical address so it could be used for this 1971 * purpose. Hence we swap the safe copy with the real 1972 * copy, allowing the safe copy to be freed and holding 1973 * on to the real copy for later use in indir_trunc. 1974 * 1975 * NOTE: ir_savebp is relative to the block device 1976 * so b_bio1 contains the device block number. 1977 */ 1978 if (indirdep->ir_state & GOINGAWAY) { 1979 FREE_LOCK(&lk); 1980 panic("deallocate_dependencies: already gone"); 1981 } 1982 indirdep->ir_state |= GOINGAWAY; 1983 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 1984 free_allocindir(aip, inodedep); 1985 if (bp->b_bio1.bio_offset >= 0 || 1986 bp->b_bio2.bio_offset != indirdep->ir_savebp->b_bio1.bio_offset) { 1987 FREE_LOCK(&lk); 1988 panic("deallocate_dependencies: not indir"); 1989 } 1990 bcopy(bp->b_data, indirdep->ir_savebp->b_data, 1991 bp->b_bcount); 1992 WORKLIST_REMOVE(wk); 1993 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk); 1994 continue; 1995 1996 case D_PAGEDEP: 1997 pagedep = WK_PAGEDEP(wk); 1998 /* 1999 * None of the directory additions will ever be 2000 * visible, so they can simply be tossed. 2001 */ 2002 for (i = 0; i < DAHASHSZ; i++) 2003 while ((dap = 2004 LIST_FIRST(&pagedep->pd_diraddhd[i]))) 2005 free_diradd(dap); 2006 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0) 2007 free_diradd(dap); 2008 /* 2009 * Copy any directory remove dependencies to the list 2010 * to be processed after the zero'ed inode is written. 2011 * If the inode has already been written, then they 2012 * can be dumped directly onto the work list. 2013 */ 2014 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 2015 LIST_REMOVE(dirrem, dm_next); 2016 dirrem->dm_dirinum = pagedep->pd_ino; 2017 if (inodedep == NULL || 2018 (inodedep->id_state & ALLCOMPLETE) == 2019 ALLCOMPLETE) 2020 add_to_worklist(&dirrem->dm_list); 2021 else 2022 WORKLIST_INSERT(&inodedep->id_bufwait, 2023 &dirrem->dm_list); 2024 } 2025 WORKLIST_REMOVE(&pagedep->pd_list); 2026 LIST_REMOVE(pagedep, pd_hash); 2027 WORKITEM_FREE(pagedep, D_PAGEDEP); 2028 continue; 2029 2030 case D_ALLOCINDIR: 2031 free_allocindir(WK_ALLOCINDIR(wk), inodedep); 2032 continue; 2033 2034 case D_ALLOCDIRECT: 2035 case D_INODEDEP: 2036 FREE_LOCK(&lk); 2037 panic("deallocate_dependencies: Unexpected type %s", 2038 TYPENAME(wk->wk_type)); 2039 /* NOTREACHED */ 2040 2041 default: 2042 FREE_LOCK(&lk); 2043 panic("deallocate_dependencies: Unknown type %s", 2044 TYPENAME(wk->wk_type)); 2045 /* NOTREACHED */ 2046 } 2047 } 2048 } 2049 2050 /* 2051 * Free an allocdirect. Generate a new freefrag work request if appropriate. 2052 * This routine must be called with splbio interrupts blocked. 2053 */ 2054 static void 2055 free_allocdirect(struct allocdirectlst *adphead, 2056 struct allocdirect *adp, int delay) 2057 { 2058 2059 #ifdef DEBUG 2060 if (lk.lkt_held == NOHOLDER) 2061 panic("free_allocdirect: lock not held"); 2062 #endif 2063 if ((adp->ad_state & DEPCOMPLETE) == 0) 2064 LIST_REMOVE(adp, ad_deps); 2065 TAILQ_REMOVE(adphead, adp, ad_next); 2066 if ((adp->ad_state & COMPLETE) == 0) 2067 WORKLIST_REMOVE(&adp->ad_list); 2068 if (adp->ad_freefrag != NULL) { 2069 if (delay) 2070 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 2071 &adp->ad_freefrag->ff_list); 2072 else 2073 add_to_worklist(&adp->ad_freefrag->ff_list); 2074 } 2075 WORKITEM_FREE(adp, D_ALLOCDIRECT); 2076 } 2077 2078 /* 2079 * Prepare an inode to be freed. The actual free operation is not 2080 * done until the zero'ed inode has been written to disk. 2081 */ 2082 void 2083 softdep_freefile(struct vnode *pvp, ino_t ino, int mode) 2084 { 2085 struct inode *ip = VTOI(pvp); 2086 struct inodedep *inodedep; 2087 struct freefile *freefile; 2088 2089 /* 2090 * This sets up the inode de-allocation dependency. 2091 */ 2092 MALLOC(freefile, struct freefile *, sizeof(struct freefile), 2093 M_FREEFILE, M_SOFTDEP_FLAGS); 2094 freefile->fx_list.wk_type = D_FREEFILE; 2095 freefile->fx_list.wk_state = 0; 2096 freefile->fx_mode = mode; 2097 freefile->fx_oldinum = ino; 2098 freefile->fx_devvp = ip->i_devvp; 2099 freefile->fx_fs = ip->i_fs; 2100 2101 /* 2102 * If the inodedep does not exist, then the zero'ed inode has 2103 * been written to disk. If the allocated inode has never been 2104 * written to disk, then the on-disk inode is zero'ed. In either 2105 * case we can free the file immediately. 2106 */ 2107 ACQUIRE_LOCK(&lk); 2108 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 || 2109 check_inode_unwritten(inodedep)) { 2110 FREE_LOCK(&lk); 2111 handle_workitem_freefile(freefile); 2112 return; 2113 } 2114 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 2115 FREE_LOCK(&lk); 2116 } 2117 2118 /* 2119 * Check to see if an inode has never been written to disk. If 2120 * so free the inodedep and return success, otherwise return failure. 2121 * This routine must be called with splbio interrupts blocked. 2122 * 2123 * If we still have a bitmap dependency, then the inode has never 2124 * been written to disk. Drop the dependency as it is no longer 2125 * necessary since the inode is being deallocated. We set the 2126 * ALLCOMPLETE flags since the bitmap now properly shows that the 2127 * inode is not allocated. Even if the inode is actively being 2128 * written, it has been rolled back to its zero'ed state, so we 2129 * are ensured that a zero inode is what is on the disk. For short 2130 * lived files, this change will usually result in removing all the 2131 * dependencies from the inode so that it can be freed immediately. 2132 */ 2133 static int 2134 check_inode_unwritten(struct inodedep *inodedep) 2135 { 2136 2137 if ((inodedep->id_state & DEPCOMPLETE) != 0 || 2138 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2139 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2140 LIST_FIRST(&inodedep->id_inowait) != NULL || 2141 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2142 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2143 inodedep->id_nlinkdelta != 0) 2144 return (0); 2145 2146 /* 2147 * Another process might be in initiate_write_inodeblock 2148 * trying to allocate memory without holding "Softdep Lock". 2149 */ 2150 if ((inodedep->id_state & IOSTARTED) != 0 && 2151 inodedep->id_savedino == NULL) 2152 return(0); 2153 2154 inodedep->id_state |= ALLCOMPLETE; 2155 LIST_REMOVE(inodedep, id_deps); 2156 inodedep->id_buf = NULL; 2157 if (inodedep->id_state & ONWORKLIST) 2158 WORKLIST_REMOVE(&inodedep->id_list); 2159 if (inodedep->id_savedino != NULL) { 2160 FREE(inodedep->id_savedino, M_INODEDEP); 2161 inodedep->id_savedino = NULL; 2162 } 2163 if (free_inodedep(inodedep) == 0) { 2164 FREE_LOCK(&lk); 2165 panic("check_inode_unwritten: busy inode"); 2166 } 2167 return (1); 2168 } 2169 2170 /* 2171 * Try to free an inodedep structure. Return 1 if it could be freed. 2172 */ 2173 static int 2174 free_inodedep(struct inodedep *inodedep) 2175 { 2176 2177 if ((inodedep->id_state & ONWORKLIST) != 0 || 2178 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 2179 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2180 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2181 LIST_FIRST(&inodedep->id_inowait) != NULL || 2182 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2183 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2184 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL) 2185 return (0); 2186 LIST_REMOVE(inodedep, id_hash); 2187 WORKITEM_FREE(inodedep, D_INODEDEP); 2188 num_inodedep -= 1; 2189 return (1); 2190 } 2191 2192 /* 2193 * This workitem routine performs the block de-allocation. 2194 * The workitem is added to the pending list after the updated 2195 * inode block has been written to disk. As mentioned above, 2196 * checks regarding the number of blocks de-allocated (compared 2197 * to the number of blocks allocated for the file) are also 2198 * performed in this function. 2199 */ 2200 static void 2201 handle_workitem_freeblocks(struct freeblks *freeblks) 2202 { 2203 struct inode tip; 2204 ufs_daddr_t bn; 2205 struct fs *fs; 2206 int i, level, bsize; 2207 long nblocks, blocksreleased = 0; 2208 int error, allerror = 0; 2209 ufs_lbn_t baselbns[NIADDR], tmpval; 2210 2211 tip.i_number = freeblks->fb_previousinum; 2212 tip.i_devvp = freeblks->fb_devvp; 2213 tip.i_dev = freeblks->fb_devvp->v_rdev; 2214 tip.i_fs = freeblks->fb_fs; 2215 tip.i_size = freeblks->fb_oldsize; 2216 tip.i_uid = freeblks->fb_uid; 2217 fs = freeblks->fb_fs; 2218 tmpval = 1; 2219 baselbns[0] = NDADDR; 2220 for (i = 1; i < NIADDR; i++) { 2221 tmpval *= NINDIR(fs); 2222 baselbns[i] = baselbns[i - 1] + tmpval; 2223 } 2224 nblocks = btodb(fs->fs_bsize); 2225 blocksreleased = 0; 2226 /* 2227 * Indirect blocks first. 2228 */ 2229 for (level = (NIADDR - 1); level >= 0; level--) { 2230 if ((bn = freeblks->fb_iblks[level]) == 0) 2231 continue; 2232 if ((error = indir_trunc(&tip, fsbtodoff(fs, bn), level, 2233 baselbns[level], &blocksreleased)) == 0) 2234 allerror = error; 2235 ffs_blkfree(&tip, bn, fs->fs_bsize); 2236 blocksreleased += nblocks; 2237 } 2238 /* 2239 * All direct blocks or frags. 2240 */ 2241 for (i = (NDADDR - 1); i >= 0; i--) { 2242 if ((bn = freeblks->fb_dblks[i]) == 0) 2243 continue; 2244 bsize = blksize(fs, &tip, i); 2245 ffs_blkfree(&tip, bn, bsize); 2246 blocksreleased += btodb(bsize); 2247 } 2248 2249 #ifdef DIAGNOSTIC 2250 if (freeblks->fb_chkcnt != blocksreleased) 2251 kprintf("handle_workitem_freeblocks: block count\n"); 2252 if (allerror) 2253 softdep_error("handle_workitem_freeblks", allerror); 2254 #endif /* DIAGNOSTIC */ 2255 WORKITEM_FREE(freeblks, D_FREEBLKS); 2256 } 2257 2258 /* 2259 * Release blocks associated with the inode ip and stored in the indirect 2260 * block at doffset. If level is greater than SINGLE, the block is an 2261 * indirect block and recursive calls to indirtrunc must be used to 2262 * cleanse other indirect blocks. 2263 */ 2264 static int 2265 indir_trunc(struct inode *ip, off_t doffset, int level, ufs_lbn_t lbn, 2266 long *countp) 2267 { 2268 struct buf *bp; 2269 ufs_daddr_t *bap; 2270 ufs_daddr_t nb; 2271 struct fs *fs; 2272 struct worklist *wk; 2273 struct indirdep *indirdep; 2274 int i, lbnadd, nblocks; 2275 int error, allerror = 0; 2276 2277 fs = ip->i_fs; 2278 lbnadd = 1; 2279 for (i = level; i > 0; i--) 2280 lbnadd *= NINDIR(fs); 2281 /* 2282 * Get buffer of block pointers to be freed. This routine is not 2283 * called until the zero'ed inode has been written, so it is safe 2284 * to free blocks as they are encountered. Because the inode has 2285 * been zero'ed, calls to bmap on these blocks will fail. So, we 2286 * have to use the on-disk address and the block device for the 2287 * filesystem to look them up. If the file was deleted before its 2288 * indirect blocks were all written to disk, the routine that set 2289 * us up (deallocate_dependencies) will have arranged to leave 2290 * a complete copy of the indirect block in memory for our use. 2291 * Otherwise we have to read the blocks in from the disk. 2292 */ 2293 ACQUIRE_LOCK(&lk); 2294 if ((bp = findblk(ip->i_devvp, doffset)) != NULL && 2295 (wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2296 /* 2297 * bp must be ir_savebp, which is held locked for our use. 2298 */ 2299 if (wk->wk_type != D_INDIRDEP || 2300 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp || 2301 (indirdep->ir_state & GOINGAWAY) == 0) { 2302 FREE_LOCK(&lk); 2303 panic("indir_trunc: lost indirdep"); 2304 } 2305 WORKLIST_REMOVE(wk); 2306 WORKITEM_FREE(indirdep, D_INDIRDEP); 2307 if (LIST_FIRST(&bp->b_dep) != NULL) { 2308 FREE_LOCK(&lk); 2309 panic("indir_trunc: dangling dep"); 2310 } 2311 FREE_LOCK(&lk); 2312 } else { 2313 FREE_LOCK(&lk); 2314 error = bread(ip->i_devvp, doffset, (int)fs->fs_bsize, &bp); 2315 if (error) 2316 return (error); 2317 } 2318 /* 2319 * Recursively free indirect blocks. 2320 */ 2321 bap = (ufs_daddr_t *)bp->b_data; 2322 nblocks = btodb(fs->fs_bsize); 2323 for (i = NINDIR(fs) - 1; i >= 0; i--) { 2324 if ((nb = bap[i]) == 0) 2325 continue; 2326 if (level != 0) { 2327 if ((error = indir_trunc(ip, fsbtodoff(fs, nb), 2328 level - 1, lbn + (i * lbnadd), countp)) != 0) 2329 allerror = error; 2330 } 2331 ffs_blkfree(ip, nb, fs->fs_bsize); 2332 *countp += nblocks; 2333 } 2334 bp->b_flags |= B_INVAL | B_NOCACHE; 2335 brelse(bp); 2336 return (allerror); 2337 } 2338 2339 /* 2340 * Free an allocindir. 2341 * This routine must be called with splbio interrupts blocked. 2342 */ 2343 static void 2344 free_allocindir(struct allocindir *aip, struct inodedep *inodedep) 2345 { 2346 struct freefrag *freefrag; 2347 2348 #ifdef DEBUG 2349 if (lk.lkt_held == NOHOLDER) 2350 panic("free_allocindir: lock not held"); 2351 #endif 2352 if ((aip->ai_state & DEPCOMPLETE) == 0) 2353 LIST_REMOVE(aip, ai_deps); 2354 if (aip->ai_state & ONWORKLIST) 2355 WORKLIST_REMOVE(&aip->ai_list); 2356 LIST_REMOVE(aip, ai_next); 2357 if ((freefrag = aip->ai_freefrag) != NULL) { 2358 if (inodedep == NULL) 2359 add_to_worklist(&freefrag->ff_list); 2360 else 2361 WORKLIST_INSERT(&inodedep->id_bufwait, 2362 &freefrag->ff_list); 2363 } 2364 WORKITEM_FREE(aip, D_ALLOCINDIR); 2365 } 2366 2367 /* 2368 * Directory entry addition dependencies. 2369 * 2370 * When adding a new directory entry, the inode (with its incremented link 2371 * count) must be written to disk before the directory entry's pointer to it. 2372 * Also, if the inode is newly allocated, the corresponding freemap must be 2373 * updated (on disk) before the directory entry's pointer. These requirements 2374 * are met via undo/redo on the directory entry's pointer, which consists 2375 * simply of the inode number. 2376 * 2377 * As directory entries are added and deleted, the free space within a 2378 * directory block can become fragmented. The ufs filesystem will compact 2379 * a fragmented directory block to make space for a new entry. When this 2380 * occurs, the offsets of previously added entries change. Any "diradd" 2381 * dependency structures corresponding to these entries must be updated with 2382 * the new offsets. 2383 */ 2384 2385 /* 2386 * This routine is called after the in-memory inode's link 2387 * count has been incremented, but before the directory entry's 2388 * pointer to the inode has been set. 2389 * 2390 * Parameters: 2391 * bp: buffer containing directory block 2392 * dp: inode for directory 2393 * diroffset: offset of new entry in directory 2394 * newinum: inode referenced by new directory entry 2395 * newdirbp: non-NULL => contents of new mkdir 2396 */ 2397 void 2398 softdep_setup_directory_add(struct buf *bp, struct inode *dp, off_t diroffset, 2399 ino_t newinum, struct buf *newdirbp) 2400 { 2401 int offset; /* offset of new entry within directory block */ 2402 ufs_lbn_t lbn; /* block in directory containing new entry */ 2403 struct fs *fs; 2404 struct diradd *dap; 2405 struct pagedep *pagedep; 2406 struct inodedep *inodedep; 2407 struct mkdir *mkdir1, *mkdir2; 2408 2409 /* 2410 * Whiteouts have no dependencies. 2411 */ 2412 if (newinum == WINO) { 2413 if (newdirbp != NULL) 2414 bdwrite(newdirbp); 2415 return; 2416 } 2417 2418 fs = dp->i_fs; 2419 lbn = lblkno(fs, diroffset); 2420 offset = blkoff(fs, diroffset); 2421 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, 2422 M_SOFTDEP_FLAGS); 2423 bzero(dap, sizeof(struct diradd)); 2424 dap->da_list.wk_type = D_DIRADD; 2425 dap->da_offset = offset; 2426 dap->da_newinum = newinum; 2427 dap->da_state = ATTACHED; 2428 if (newdirbp == NULL) { 2429 dap->da_state |= DEPCOMPLETE; 2430 ACQUIRE_LOCK(&lk); 2431 } else { 2432 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 2433 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2434 M_SOFTDEP_FLAGS); 2435 mkdir1->md_list.wk_type = D_MKDIR; 2436 mkdir1->md_state = MKDIR_BODY; 2437 mkdir1->md_diradd = dap; 2438 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2439 M_SOFTDEP_FLAGS); 2440 mkdir2->md_list.wk_type = D_MKDIR; 2441 mkdir2->md_state = MKDIR_PARENT; 2442 mkdir2->md_diradd = dap; 2443 /* 2444 * Dependency on "." and ".." being written to disk. 2445 */ 2446 mkdir1->md_buf = newdirbp; 2447 ACQUIRE_LOCK(&lk); 2448 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 2449 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list); 2450 FREE_LOCK(&lk); 2451 bdwrite(newdirbp); 2452 /* 2453 * Dependency on link count increase for parent directory 2454 */ 2455 ACQUIRE_LOCK(&lk); 2456 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0 2457 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2458 dap->da_state &= ~MKDIR_PARENT; 2459 WORKITEM_FREE(mkdir2, D_MKDIR); 2460 } else { 2461 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 2462 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 2463 } 2464 } 2465 /* 2466 * Link into parent directory pagedep to await its being written. 2467 */ 2468 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2469 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2470 dap->da_pagedep = pagedep; 2471 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 2472 da_pdlist); 2473 /* 2474 * Link into its inodedep. Put it on the id_bufwait list if the inode 2475 * is not yet written. If it is written, do the post-inode write 2476 * processing to put it on the id_pendinghd list. 2477 */ 2478 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep); 2479 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 2480 diradd_inode_written(dap, inodedep); 2481 else 2482 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2483 FREE_LOCK(&lk); 2484 } 2485 2486 /* 2487 * This procedure is called to change the offset of a directory 2488 * entry when compacting a directory block which must be owned 2489 * exclusively by the caller. Note that the actual entry movement 2490 * must be done in this procedure to ensure that no I/O completions 2491 * occur while the move is in progress. 2492 * 2493 * Parameters: 2494 * dp: inode for directory 2495 * base: address of dp->i_offset 2496 * oldloc: address of old directory location 2497 * newloc: address of new directory location 2498 * entrysize: size of directory entry 2499 */ 2500 void 2501 softdep_change_directoryentry_offset(struct inode *dp, caddr_t base, 2502 caddr_t oldloc, caddr_t newloc, 2503 int entrysize) 2504 { 2505 int offset, oldoffset, newoffset; 2506 struct pagedep *pagedep; 2507 struct diradd *dap; 2508 ufs_lbn_t lbn; 2509 2510 ACQUIRE_LOCK(&lk); 2511 lbn = lblkno(dp->i_fs, dp->i_offset); 2512 offset = blkoff(dp->i_fs, dp->i_offset); 2513 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) 2514 goto done; 2515 oldoffset = offset + (oldloc - base); 2516 newoffset = offset + (newloc - base); 2517 2518 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) { 2519 if (dap->da_offset != oldoffset) 2520 continue; 2521 dap->da_offset = newoffset; 2522 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) 2523 break; 2524 LIST_REMOVE(dap, da_pdlist); 2525 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], 2526 dap, da_pdlist); 2527 break; 2528 } 2529 if (dap == NULL) { 2530 2531 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) { 2532 if (dap->da_offset == oldoffset) { 2533 dap->da_offset = newoffset; 2534 break; 2535 } 2536 } 2537 } 2538 done: 2539 bcopy(oldloc, newloc, entrysize); 2540 FREE_LOCK(&lk); 2541 } 2542 2543 /* 2544 * Free a diradd dependency structure. This routine must be called 2545 * with splbio interrupts blocked. 2546 */ 2547 static void 2548 free_diradd(struct diradd *dap) 2549 { 2550 struct dirrem *dirrem; 2551 struct pagedep *pagedep; 2552 struct inodedep *inodedep; 2553 struct mkdir *mkdir, *nextmd; 2554 2555 #ifdef DEBUG 2556 if (lk.lkt_held == NOHOLDER) 2557 panic("free_diradd: lock not held"); 2558 #endif 2559 WORKLIST_REMOVE(&dap->da_list); 2560 LIST_REMOVE(dap, da_pdlist); 2561 if ((dap->da_state & DIRCHG) == 0) { 2562 pagedep = dap->da_pagedep; 2563 } else { 2564 dirrem = dap->da_previous; 2565 pagedep = dirrem->dm_pagedep; 2566 dirrem->dm_dirinum = pagedep->pd_ino; 2567 add_to_worklist(&dirrem->dm_list); 2568 } 2569 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum, 2570 0, &inodedep) != 0) 2571 (void) free_inodedep(inodedep); 2572 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2573 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 2574 nextmd = LIST_NEXT(mkdir, md_mkdirs); 2575 if (mkdir->md_diradd != dap) 2576 continue; 2577 dap->da_state &= ~mkdir->md_state; 2578 WORKLIST_REMOVE(&mkdir->md_list); 2579 LIST_REMOVE(mkdir, md_mkdirs); 2580 WORKITEM_FREE(mkdir, D_MKDIR); 2581 } 2582 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2583 FREE_LOCK(&lk); 2584 panic("free_diradd: unfound ref"); 2585 } 2586 } 2587 WORKITEM_FREE(dap, D_DIRADD); 2588 } 2589 2590 /* 2591 * Directory entry removal dependencies. 2592 * 2593 * When removing a directory entry, the entry's inode pointer must be 2594 * zero'ed on disk before the corresponding inode's link count is decremented 2595 * (possibly freeing the inode for re-use). This dependency is handled by 2596 * updating the directory entry but delaying the inode count reduction until 2597 * after the directory block has been written to disk. After this point, the 2598 * inode count can be decremented whenever it is convenient. 2599 */ 2600 2601 /* 2602 * This routine should be called immediately after removing 2603 * a directory entry. The inode's link count should not be 2604 * decremented by the calling procedure -- the soft updates 2605 * code will do this task when it is safe. 2606 * 2607 * Parameters: 2608 * bp: buffer containing directory block 2609 * dp: inode for the directory being modified 2610 * ip: inode for directory entry being removed 2611 * isrmdir: indicates if doing RMDIR 2612 */ 2613 void 2614 softdep_setup_remove(struct buf *bp, struct inode *dp, struct inode *ip, 2615 int isrmdir) 2616 { 2617 struct dirrem *dirrem, *prevdirrem; 2618 2619 /* 2620 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 2621 */ 2622 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2623 2624 /* 2625 * If the COMPLETE flag is clear, then there were no active 2626 * entries and we want to roll back to a zeroed entry until 2627 * the new inode is committed to disk. If the COMPLETE flag is 2628 * set then we have deleted an entry that never made it to 2629 * disk. If the entry we deleted resulted from a name change, 2630 * then the old name still resides on disk. We cannot delete 2631 * its inode (returned to us in prevdirrem) until the zeroed 2632 * directory entry gets to disk. The new inode has never been 2633 * referenced on the disk, so can be deleted immediately. 2634 */ 2635 if ((dirrem->dm_state & COMPLETE) == 0) { 2636 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 2637 dm_next); 2638 FREE_LOCK(&lk); 2639 } else { 2640 if (prevdirrem != NULL) 2641 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 2642 prevdirrem, dm_next); 2643 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 2644 FREE_LOCK(&lk); 2645 handle_workitem_remove(dirrem); 2646 } 2647 } 2648 2649 /* 2650 * Allocate a new dirrem if appropriate and return it along with 2651 * its associated pagedep. Called without a lock, returns with lock. 2652 */ 2653 static long num_dirrem; /* number of dirrem allocated */ 2654 2655 /* 2656 * Parameters: 2657 * bp: buffer containing directory block 2658 * dp: inode for the directory being modified 2659 * ip: inode for directory entry being removed 2660 * isrmdir: indicates if doing RMDIR 2661 * prevdirremp: previously referenced inode, if any 2662 */ 2663 static struct dirrem * 2664 newdirrem(struct buf *bp, struct inode *dp, struct inode *ip, 2665 int isrmdir, struct dirrem **prevdirremp) 2666 { 2667 int offset; 2668 ufs_lbn_t lbn; 2669 struct diradd *dap; 2670 struct dirrem *dirrem; 2671 struct pagedep *pagedep; 2672 2673 /* 2674 * Whiteouts have no deletion dependencies. 2675 */ 2676 if (ip == NULL) 2677 panic("newdirrem: whiteout"); 2678 /* 2679 * If we are over our limit, try to improve the situation. 2680 * Limiting the number of dirrem structures will also limit 2681 * the number of freefile and freeblks structures. 2682 */ 2683 if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0) 2684 (void) request_cleanup(FLUSH_REMOVE, 0); 2685 num_dirrem += 1; 2686 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem), 2687 M_DIRREM, M_SOFTDEP_FLAGS); 2688 bzero(dirrem, sizeof(struct dirrem)); 2689 dirrem->dm_list.wk_type = D_DIRREM; 2690 dirrem->dm_state = isrmdir ? RMDIR : 0; 2691 dirrem->dm_mnt = ITOV(ip)->v_mount; 2692 dirrem->dm_oldinum = ip->i_number; 2693 *prevdirremp = NULL; 2694 2695 ACQUIRE_LOCK(&lk); 2696 lbn = lblkno(dp->i_fs, dp->i_offset); 2697 offset = blkoff(dp->i_fs, dp->i_offset); 2698 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2699 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2700 dirrem->dm_pagedep = pagedep; 2701 /* 2702 * Check for a diradd dependency for the same directory entry. 2703 * If present, then both dependencies become obsolete and can 2704 * be de-allocated. Check for an entry on both the pd_dirraddhd 2705 * list and the pd_pendinghd list. 2706 */ 2707 2708 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 2709 if (dap->da_offset == offset) 2710 break; 2711 if (dap == NULL) { 2712 2713 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 2714 if (dap->da_offset == offset) 2715 break; 2716 if (dap == NULL) 2717 return (dirrem); 2718 } 2719 /* 2720 * Must be ATTACHED at this point. 2721 */ 2722 if ((dap->da_state & ATTACHED) == 0) { 2723 FREE_LOCK(&lk); 2724 panic("newdirrem: not ATTACHED"); 2725 } 2726 if (dap->da_newinum != ip->i_number) { 2727 FREE_LOCK(&lk); 2728 panic("newdirrem: inum %"PRId64" should be %"PRId64, 2729 ip->i_number, dap->da_newinum); 2730 } 2731 /* 2732 * If we are deleting a changed name that never made it to disk, 2733 * then return the dirrem describing the previous inode (which 2734 * represents the inode currently referenced from this entry on disk). 2735 */ 2736 if ((dap->da_state & DIRCHG) != 0) { 2737 *prevdirremp = dap->da_previous; 2738 dap->da_state &= ~DIRCHG; 2739 dap->da_pagedep = pagedep; 2740 } 2741 /* 2742 * We are deleting an entry that never made it to disk. 2743 * Mark it COMPLETE so we can delete its inode immediately. 2744 */ 2745 dirrem->dm_state |= COMPLETE; 2746 free_diradd(dap); 2747 return (dirrem); 2748 } 2749 2750 /* 2751 * Directory entry change dependencies. 2752 * 2753 * Changing an existing directory entry requires that an add operation 2754 * be completed first followed by a deletion. The semantics for the addition 2755 * are identical to the description of adding a new entry above except 2756 * that the rollback is to the old inode number rather than zero. Once 2757 * the addition dependency is completed, the removal is done as described 2758 * in the removal routine above. 2759 */ 2760 2761 /* 2762 * This routine should be called immediately after changing 2763 * a directory entry. The inode's link count should not be 2764 * decremented by the calling procedure -- the soft updates 2765 * code will perform this task when it is safe. 2766 * 2767 * Parameters: 2768 * bp: buffer containing directory block 2769 * dp: inode for the directory being modified 2770 * ip: inode for directory entry being removed 2771 * newinum: new inode number for changed entry 2772 * isrmdir: indicates if doing RMDIR 2773 */ 2774 void 2775 softdep_setup_directory_change(struct buf *bp, struct inode *dp, 2776 struct inode *ip, ino_t newinum, 2777 int isrmdir) 2778 { 2779 int offset; 2780 struct diradd *dap = NULL; 2781 struct dirrem *dirrem, *prevdirrem; 2782 struct pagedep *pagedep; 2783 struct inodedep *inodedep; 2784 2785 offset = blkoff(dp->i_fs, dp->i_offset); 2786 2787 /* 2788 * Whiteouts do not need diradd dependencies. 2789 */ 2790 if (newinum != WINO) { 2791 MALLOC(dap, struct diradd *, sizeof(struct diradd), 2792 M_DIRADD, M_SOFTDEP_FLAGS); 2793 bzero(dap, sizeof(struct diradd)); 2794 dap->da_list.wk_type = D_DIRADD; 2795 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 2796 dap->da_offset = offset; 2797 dap->da_newinum = newinum; 2798 } 2799 2800 /* 2801 * Allocate a new dirrem and ACQUIRE_LOCK. 2802 */ 2803 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2804 pagedep = dirrem->dm_pagedep; 2805 /* 2806 * The possible values for isrmdir: 2807 * 0 - non-directory file rename 2808 * 1 - directory rename within same directory 2809 * inum - directory rename to new directory of given inode number 2810 * When renaming to a new directory, we are both deleting and 2811 * creating a new directory entry, so the link count on the new 2812 * directory should not change. Thus we do not need the followup 2813 * dirrem which is usually done in handle_workitem_remove. We set 2814 * the DIRCHG flag to tell handle_workitem_remove to skip the 2815 * followup dirrem. 2816 */ 2817 if (isrmdir > 1) 2818 dirrem->dm_state |= DIRCHG; 2819 2820 /* 2821 * Whiteouts have no additional dependencies, 2822 * so just put the dirrem on the correct list. 2823 */ 2824 if (newinum == WINO) { 2825 if ((dirrem->dm_state & COMPLETE) == 0) { 2826 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 2827 dm_next); 2828 } else { 2829 dirrem->dm_dirinum = pagedep->pd_ino; 2830 add_to_worklist(&dirrem->dm_list); 2831 } 2832 FREE_LOCK(&lk); 2833 return; 2834 } 2835 2836 /* 2837 * If the COMPLETE flag is clear, then there were no active 2838 * entries and we want to roll back to the previous inode until 2839 * the new inode is committed to disk. If the COMPLETE flag is 2840 * set, then we have deleted an entry that never made it to disk. 2841 * If the entry we deleted resulted from a name change, then the old 2842 * inode reference still resides on disk. Any rollback that we do 2843 * needs to be to that old inode (returned to us in prevdirrem). If 2844 * the entry we deleted resulted from a create, then there is 2845 * no entry on the disk, so we want to roll back to zero rather 2846 * than the uncommitted inode. In either of the COMPLETE cases we 2847 * want to immediately free the unwritten and unreferenced inode. 2848 */ 2849 if ((dirrem->dm_state & COMPLETE) == 0) { 2850 dap->da_previous = dirrem; 2851 } else { 2852 if (prevdirrem != NULL) { 2853 dap->da_previous = prevdirrem; 2854 } else { 2855 dap->da_state &= ~DIRCHG; 2856 dap->da_pagedep = pagedep; 2857 } 2858 dirrem->dm_dirinum = pagedep->pd_ino; 2859 add_to_worklist(&dirrem->dm_list); 2860 } 2861 /* 2862 * Link into its inodedep. Put it on the id_bufwait list if the inode 2863 * is not yet written. If it is written, do the post-inode write 2864 * processing to put it on the id_pendinghd list. 2865 */ 2866 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 || 2867 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2868 dap->da_state |= COMPLETE; 2869 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 2870 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 2871 } else { 2872 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 2873 dap, da_pdlist); 2874 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2875 } 2876 FREE_LOCK(&lk); 2877 } 2878 2879 /* 2880 * Called whenever the link count on an inode is changed. 2881 * It creates an inode dependency so that the new reference(s) 2882 * to the inode cannot be committed to disk until the updated 2883 * inode has been written. 2884 * 2885 * Parameters: 2886 * ip: the inode with the increased link count 2887 */ 2888 void 2889 softdep_change_linkcnt(struct inode *ip) 2890 { 2891 struct inodedep *inodedep; 2892 2893 ACQUIRE_LOCK(&lk); 2894 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 2895 if (ip->i_nlink < ip->i_effnlink) { 2896 FREE_LOCK(&lk); 2897 panic("softdep_change_linkcnt: bad delta"); 2898 } 2899 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2900 FREE_LOCK(&lk); 2901 } 2902 2903 /* 2904 * This workitem decrements the inode's link count. 2905 * If the link count reaches zero, the file is removed. 2906 */ 2907 static void 2908 handle_workitem_remove(struct dirrem *dirrem) 2909 { 2910 struct inodedep *inodedep; 2911 struct vnode *vp; 2912 struct inode *ip; 2913 ino_t oldinum; 2914 int error; 2915 2916 if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) { 2917 softdep_error("handle_workitem_remove: vget", error); 2918 return; 2919 } 2920 ip = VTOI(vp); 2921 ACQUIRE_LOCK(&lk); 2922 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){ 2923 FREE_LOCK(&lk); 2924 panic("handle_workitem_remove: lost inodedep"); 2925 } 2926 /* 2927 * Normal file deletion. 2928 */ 2929 if ((dirrem->dm_state & RMDIR) == 0) { 2930 ip->i_nlink--; 2931 ip->i_flag |= IN_CHANGE; 2932 if (ip->i_nlink < ip->i_effnlink) { 2933 FREE_LOCK(&lk); 2934 panic("handle_workitem_remove: bad file delta"); 2935 } 2936 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2937 FREE_LOCK(&lk); 2938 vput(vp); 2939 num_dirrem -= 1; 2940 WORKITEM_FREE(dirrem, D_DIRREM); 2941 return; 2942 } 2943 /* 2944 * Directory deletion. Decrement reference count for both the 2945 * just deleted parent directory entry and the reference for ".". 2946 * Next truncate the directory to length zero. When the 2947 * truncation completes, arrange to have the reference count on 2948 * the parent decremented to account for the loss of "..". 2949 */ 2950 ip->i_nlink -= 2; 2951 ip->i_flag |= IN_CHANGE; 2952 if (ip->i_nlink < ip->i_effnlink) { 2953 FREE_LOCK(&lk); 2954 panic("handle_workitem_remove: bad dir delta"); 2955 } 2956 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2957 FREE_LOCK(&lk); 2958 if ((error = ffs_truncate(vp, (off_t)0, 0, proc0.p_ucred)) != 0) 2959 softdep_error("handle_workitem_remove: truncate", error); 2960 /* 2961 * Rename a directory to a new parent. Since, we are both deleting 2962 * and creating a new directory entry, the link count on the new 2963 * directory should not change. Thus we skip the followup dirrem. 2964 */ 2965 if (dirrem->dm_state & DIRCHG) { 2966 vput(vp); 2967 num_dirrem -= 1; 2968 WORKITEM_FREE(dirrem, D_DIRREM); 2969 return; 2970 } 2971 /* 2972 * If the inodedep does not exist, then the zero'ed inode has 2973 * been written to disk. If the allocated inode has never been 2974 * written to disk, then the on-disk inode is zero'ed. In either 2975 * case we can remove the file immediately. 2976 */ 2977 ACQUIRE_LOCK(&lk); 2978 dirrem->dm_state = 0; 2979 oldinum = dirrem->dm_oldinum; 2980 dirrem->dm_oldinum = dirrem->dm_dirinum; 2981 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 || 2982 check_inode_unwritten(inodedep)) { 2983 FREE_LOCK(&lk); 2984 vput(vp); 2985 handle_workitem_remove(dirrem); 2986 return; 2987 } 2988 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 2989 FREE_LOCK(&lk); 2990 ip->i_flag |= IN_CHANGE; 2991 ffs_update(vp, 0); 2992 vput(vp); 2993 } 2994 2995 /* 2996 * Inode de-allocation dependencies. 2997 * 2998 * When an inode's link count is reduced to zero, it can be de-allocated. We 2999 * found it convenient to postpone de-allocation until after the inode is 3000 * written to disk with its new link count (zero). At this point, all of the 3001 * on-disk inode's block pointers are nullified and, with careful dependency 3002 * list ordering, all dependencies related to the inode will be satisfied and 3003 * the corresponding dependency structures de-allocated. So, if/when the 3004 * inode is reused, there will be no mixing of old dependencies with new 3005 * ones. This artificial dependency is set up by the block de-allocation 3006 * procedure above (softdep_setup_freeblocks) and completed by the 3007 * following procedure. 3008 */ 3009 static void 3010 handle_workitem_freefile(struct freefile *freefile) 3011 { 3012 struct vnode vp; 3013 struct inode tip; 3014 struct inodedep *idp; 3015 int error; 3016 3017 #ifdef DEBUG 3018 ACQUIRE_LOCK(&lk); 3019 error = inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp); 3020 FREE_LOCK(&lk); 3021 if (error) 3022 panic("handle_workitem_freefile: inodedep survived"); 3023 #endif 3024 tip.i_devvp = freefile->fx_devvp; 3025 tip.i_dev = freefile->fx_devvp->v_rdev; 3026 tip.i_fs = freefile->fx_fs; 3027 vp.v_data = &tip; 3028 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0) 3029 softdep_error("handle_workitem_freefile", error); 3030 WORKITEM_FREE(freefile, D_FREEFILE); 3031 } 3032 3033 /* 3034 * Helper function which unlinks marker element from work list and returns 3035 * the next element on the list. 3036 */ 3037 static __inline struct worklist * 3038 markernext(struct worklist *marker) 3039 { 3040 struct worklist *next; 3041 3042 next = LIST_NEXT(marker, wk_list); 3043 LIST_REMOVE(marker, wk_list); 3044 return next; 3045 } 3046 3047 /* 3048 * Disk writes. 3049 * 3050 * The dependency structures constructed above are most actively used when file 3051 * system blocks are written to disk. No constraints are placed on when a 3052 * block can be written, but unsatisfied update dependencies are made safe by 3053 * modifying (or replacing) the source memory for the duration of the disk 3054 * write. When the disk write completes, the memory block is again brought 3055 * up-to-date. 3056 * 3057 * In-core inode structure reclamation. 3058 * 3059 * Because there are a finite number of "in-core" inode structures, they are 3060 * reused regularly. By transferring all inode-related dependencies to the 3061 * in-memory inode block and indexing them separately (via "inodedep"s), we 3062 * can allow "in-core" inode structures to be reused at any time and avoid 3063 * any increase in contention. 3064 * 3065 * Called just before entering the device driver to initiate a new disk I/O. 3066 * The buffer must be locked, thus, no I/O completion operations can occur 3067 * while we are manipulating its associated dependencies. 3068 * 3069 * Parameters: 3070 * bp: structure describing disk write to occur 3071 */ 3072 static void 3073 softdep_disk_io_initiation(struct buf *bp) 3074 { 3075 struct worklist *wk; 3076 struct worklist marker; 3077 struct indirdep *indirdep; 3078 3079 /* 3080 * We only care about write operations. There should never 3081 * be dependencies for reads. 3082 */ 3083 if (bp->b_cmd == BUF_CMD_READ) 3084 panic("softdep_disk_io_initiation: read"); 3085 3086 marker.wk_type = D_LAST + 1; /* Not a normal workitem */ 3087 3088 /* 3089 * Do any necessary pre-I/O processing. 3090 */ 3091 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = markernext(&marker)) { 3092 LIST_INSERT_AFTER(wk, &marker, wk_list); 3093 3094 switch (wk->wk_type) { 3095 3096 case D_PAGEDEP: 3097 initiate_write_filepage(WK_PAGEDEP(wk), bp); 3098 continue; 3099 3100 case D_INODEDEP: 3101 initiate_write_inodeblock(WK_INODEDEP(wk), bp); 3102 continue; 3103 3104 case D_INDIRDEP: 3105 indirdep = WK_INDIRDEP(wk); 3106 if (indirdep->ir_state & GOINGAWAY) 3107 panic("disk_io_initiation: indirdep gone"); 3108 /* 3109 * If there are no remaining dependencies, this 3110 * will be writing the real pointers, so the 3111 * dependency can be freed. 3112 */ 3113 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) { 3114 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 3115 brelse(indirdep->ir_savebp); 3116 /* inline expand WORKLIST_REMOVE(wk); */ 3117 wk->wk_state &= ~ONWORKLIST; 3118 LIST_REMOVE(wk, wk_list); 3119 WORKITEM_FREE(indirdep, D_INDIRDEP); 3120 continue; 3121 } 3122 /* 3123 * Replace up-to-date version with safe version. 3124 */ 3125 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount, 3126 M_INDIRDEP, M_SOFTDEP_FLAGS); 3127 ACQUIRE_LOCK(&lk); 3128 indirdep->ir_state &= ~ATTACHED; 3129 indirdep->ir_state |= UNDONE; 3130 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 3131 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 3132 bp->b_bcount); 3133 FREE_LOCK(&lk); 3134 continue; 3135 3136 case D_MKDIR: 3137 case D_BMSAFEMAP: 3138 case D_ALLOCDIRECT: 3139 case D_ALLOCINDIR: 3140 continue; 3141 3142 default: 3143 panic("handle_disk_io_initiation: Unexpected type %s", 3144 TYPENAME(wk->wk_type)); 3145 /* NOTREACHED */ 3146 } 3147 } 3148 } 3149 3150 /* 3151 * Called from within the procedure above to deal with unsatisfied 3152 * allocation dependencies in a directory. The buffer must be locked, 3153 * thus, no I/O completion operations can occur while we are 3154 * manipulating its associated dependencies. 3155 */ 3156 static void 3157 initiate_write_filepage(struct pagedep *pagedep, struct buf *bp) 3158 { 3159 struct diradd *dap; 3160 struct direct *ep; 3161 int i; 3162 3163 if (pagedep->pd_state & IOSTARTED) { 3164 /* 3165 * This can only happen if there is a driver that does not 3166 * understand chaining. Here biodone will reissue the call 3167 * to strategy for the incomplete buffers. 3168 */ 3169 kprintf("initiate_write_filepage: already started\n"); 3170 return; 3171 } 3172 pagedep->pd_state |= IOSTARTED; 3173 ACQUIRE_LOCK(&lk); 3174 for (i = 0; i < DAHASHSZ; i++) { 3175 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 3176 ep = (struct direct *) 3177 ((char *)bp->b_data + dap->da_offset); 3178 if (ep->d_ino != dap->da_newinum) { 3179 FREE_LOCK(&lk); 3180 panic("%s: dir inum %d != new %"PRId64, 3181 "initiate_write_filepage", 3182 ep->d_ino, dap->da_newinum); 3183 } 3184 if (dap->da_state & DIRCHG) 3185 ep->d_ino = dap->da_previous->dm_oldinum; 3186 else 3187 ep->d_ino = 0; 3188 dap->da_state &= ~ATTACHED; 3189 dap->da_state |= UNDONE; 3190 } 3191 } 3192 FREE_LOCK(&lk); 3193 } 3194 3195 /* 3196 * Called from within the procedure above to deal with unsatisfied 3197 * allocation dependencies in an inodeblock. The buffer must be 3198 * locked, thus, no I/O completion operations can occur while we 3199 * are manipulating its associated dependencies. 3200 * 3201 * Parameters: 3202 * bp: The inode block 3203 */ 3204 static void 3205 initiate_write_inodeblock(struct inodedep *inodedep, struct buf *bp) 3206 { 3207 struct allocdirect *adp, *lastadp; 3208 struct ufs1_dinode *dp; 3209 struct ufs1_dinode *sip; 3210 struct fs *fs; 3211 ufs_lbn_t prevlbn = 0; 3212 int i, deplist; 3213 3214 if (inodedep->id_state & IOSTARTED) 3215 panic("initiate_write_inodeblock: already started"); 3216 inodedep->id_state |= IOSTARTED; 3217 fs = inodedep->id_fs; 3218 dp = (struct ufs1_dinode *)bp->b_data + 3219 ino_to_fsbo(fs, inodedep->id_ino); 3220 /* 3221 * If the bitmap is not yet written, then the allocated 3222 * inode cannot be written to disk. 3223 */ 3224 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 3225 if (inodedep->id_savedino != NULL) 3226 panic("initiate_write_inodeblock: already doing I/O"); 3227 MALLOC(sip, struct ufs1_dinode *, 3228 sizeof(struct ufs1_dinode), M_INODEDEP, M_SOFTDEP_FLAGS); 3229 inodedep->id_savedino = sip; 3230 *inodedep->id_savedino = *dp; 3231 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 3232 dp->di_gen = inodedep->id_savedino->di_gen; 3233 return; 3234 } 3235 /* 3236 * If no dependencies, then there is nothing to roll back. 3237 */ 3238 inodedep->id_savedsize = dp->di_size; 3239 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 3240 return; 3241 /* 3242 * Set the dependencies to busy. 3243 */ 3244 ACQUIRE_LOCK(&lk); 3245 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3246 adp = TAILQ_NEXT(adp, ad_next)) { 3247 #ifdef DIAGNOSTIC 3248 if (deplist != 0 && prevlbn >= adp->ad_lbn) { 3249 FREE_LOCK(&lk); 3250 panic("softdep_write_inodeblock: lbn order"); 3251 } 3252 prevlbn = adp->ad_lbn; 3253 if (adp->ad_lbn < NDADDR && 3254 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) { 3255 FREE_LOCK(&lk); 3256 panic("%s: direct pointer #%ld mismatch %d != %d", 3257 "softdep_write_inodeblock", adp->ad_lbn, 3258 dp->di_db[adp->ad_lbn], adp->ad_newblkno); 3259 } 3260 if (adp->ad_lbn >= NDADDR && 3261 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) { 3262 FREE_LOCK(&lk); 3263 panic("%s: indirect pointer #%ld mismatch %d != %d", 3264 "softdep_write_inodeblock", adp->ad_lbn - NDADDR, 3265 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno); 3266 } 3267 deplist |= 1 << adp->ad_lbn; 3268 if ((adp->ad_state & ATTACHED) == 0) { 3269 FREE_LOCK(&lk); 3270 panic("softdep_write_inodeblock: Unknown state 0x%x", 3271 adp->ad_state); 3272 } 3273 #endif /* DIAGNOSTIC */ 3274 adp->ad_state &= ~ATTACHED; 3275 adp->ad_state |= UNDONE; 3276 } 3277 /* 3278 * The on-disk inode cannot claim to be any larger than the last 3279 * fragment that has been written. Otherwise, the on-disk inode 3280 * might have fragments that were not the last block in the file 3281 * which would corrupt the filesystem. 3282 */ 3283 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3284 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 3285 if (adp->ad_lbn >= NDADDR) 3286 break; 3287 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 3288 /* keep going until hitting a rollback to a frag */ 3289 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 3290 continue; 3291 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 3292 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 3293 #ifdef DIAGNOSTIC 3294 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) { 3295 FREE_LOCK(&lk); 3296 panic("softdep_write_inodeblock: lost dep1"); 3297 } 3298 #endif /* DIAGNOSTIC */ 3299 dp->di_db[i] = 0; 3300 } 3301 for (i = 0; i < NIADDR; i++) { 3302 #ifdef DIAGNOSTIC 3303 if (dp->di_ib[i] != 0 && 3304 (deplist & ((1 << NDADDR) << i)) == 0) { 3305 FREE_LOCK(&lk); 3306 panic("softdep_write_inodeblock: lost dep2"); 3307 } 3308 #endif /* DIAGNOSTIC */ 3309 dp->di_ib[i] = 0; 3310 } 3311 FREE_LOCK(&lk); 3312 return; 3313 } 3314 /* 3315 * If we have zero'ed out the last allocated block of the file, 3316 * roll back the size to the last currently allocated block. 3317 * We know that this last allocated block is a full-sized as 3318 * we already checked for fragments in the loop above. 3319 */ 3320 if (lastadp != NULL && 3321 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 3322 for (i = lastadp->ad_lbn; i >= 0; i--) 3323 if (dp->di_db[i] != 0) 3324 break; 3325 dp->di_size = (i + 1) * fs->fs_bsize; 3326 } 3327 /* 3328 * The only dependencies are for indirect blocks. 3329 * 3330 * The file size for indirect block additions is not guaranteed. 3331 * Such a guarantee would be non-trivial to achieve. The conventional 3332 * synchronous write implementation also does not make this guarantee. 3333 * Fsck should catch and fix discrepancies. Arguably, the file size 3334 * can be over-estimated without destroying integrity when the file 3335 * moves into the indirect blocks (i.e., is large). If we want to 3336 * postpone fsck, we are stuck with this argument. 3337 */ 3338 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 3339 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 3340 FREE_LOCK(&lk); 3341 } 3342 3343 /* 3344 * This routine is called during the completion interrupt 3345 * service routine for a disk write (from the procedure called 3346 * by the device driver to inform the filesystem caches of 3347 * a request completion). It should be called early in this 3348 * procedure, before the block is made available to other 3349 * processes or other routines are called. 3350 * 3351 * Parameters: 3352 * bp: describes the completed disk write 3353 */ 3354 static void 3355 softdep_disk_write_complete(struct buf *bp) 3356 { 3357 struct worklist *wk; 3358 struct workhead reattach; 3359 struct newblk *newblk; 3360 struct allocindir *aip; 3361 struct allocdirect *adp; 3362 struct indirdep *indirdep; 3363 struct inodedep *inodedep; 3364 struct bmsafemap *bmsafemap; 3365 3366 #ifdef DEBUG 3367 if (lk.lkt_held != NOHOLDER) 3368 panic("softdep_disk_write_complete: lock is held"); 3369 lk.lkt_held = SPECIAL_FLAG; 3370 #endif 3371 LIST_INIT(&reattach); 3372 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 3373 WORKLIST_REMOVE(wk); 3374 switch (wk->wk_type) { 3375 3376 case D_PAGEDEP: 3377 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 3378 WORKLIST_INSERT(&reattach, wk); 3379 continue; 3380 3381 case D_INODEDEP: 3382 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 3383 WORKLIST_INSERT(&reattach, wk); 3384 continue; 3385 3386 case D_BMSAFEMAP: 3387 bmsafemap = WK_BMSAFEMAP(wk); 3388 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) { 3389 newblk->nb_state |= DEPCOMPLETE; 3390 newblk->nb_bmsafemap = NULL; 3391 LIST_REMOVE(newblk, nb_deps); 3392 } 3393 while ((adp = 3394 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) { 3395 adp->ad_state |= DEPCOMPLETE; 3396 adp->ad_buf = NULL; 3397 LIST_REMOVE(adp, ad_deps); 3398 handle_allocdirect_partdone(adp); 3399 } 3400 while ((aip = 3401 LIST_FIRST(&bmsafemap->sm_allocindirhd))) { 3402 aip->ai_state |= DEPCOMPLETE; 3403 aip->ai_buf = NULL; 3404 LIST_REMOVE(aip, ai_deps); 3405 handle_allocindir_partdone(aip); 3406 } 3407 while ((inodedep = 3408 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { 3409 inodedep->id_state |= DEPCOMPLETE; 3410 LIST_REMOVE(inodedep, id_deps); 3411 inodedep->id_buf = NULL; 3412 } 3413 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 3414 continue; 3415 3416 case D_MKDIR: 3417 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 3418 continue; 3419 3420 case D_ALLOCDIRECT: 3421 adp = WK_ALLOCDIRECT(wk); 3422 adp->ad_state |= COMPLETE; 3423 handle_allocdirect_partdone(adp); 3424 continue; 3425 3426 case D_ALLOCINDIR: 3427 aip = WK_ALLOCINDIR(wk); 3428 aip->ai_state |= COMPLETE; 3429 handle_allocindir_partdone(aip); 3430 continue; 3431 3432 case D_INDIRDEP: 3433 indirdep = WK_INDIRDEP(wk); 3434 if (indirdep->ir_state & GOINGAWAY) { 3435 lk.lkt_held = NOHOLDER; 3436 panic("disk_write_complete: indirdep gone"); 3437 } 3438 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 3439 FREE(indirdep->ir_saveddata, M_INDIRDEP); 3440 indirdep->ir_saveddata = 0; 3441 indirdep->ir_state &= ~UNDONE; 3442 indirdep->ir_state |= ATTACHED; 3443 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 3444 handle_allocindir_partdone(aip); 3445 if (aip == LIST_FIRST(&indirdep->ir_donehd)) { 3446 lk.lkt_held = NOHOLDER; 3447 panic("disk_write_complete: not gone"); 3448 } 3449 } 3450 WORKLIST_INSERT(&reattach, wk); 3451 if ((bp->b_flags & B_DELWRI) == 0) 3452 stat_indir_blk_ptrs++; 3453 bdirty(bp); 3454 continue; 3455 3456 default: 3457 lk.lkt_held = NOHOLDER; 3458 panic("handle_disk_write_complete: Unknown type %s", 3459 TYPENAME(wk->wk_type)); 3460 /* NOTREACHED */ 3461 } 3462 } 3463 /* 3464 * Reattach any requests that must be redone. 3465 */ 3466 while ((wk = LIST_FIRST(&reattach)) != NULL) { 3467 WORKLIST_REMOVE(wk); 3468 WORKLIST_INSERT(&bp->b_dep, wk); 3469 } 3470 #ifdef DEBUG 3471 if (lk.lkt_held != SPECIAL_FLAG) 3472 panic("softdep_disk_write_complete: lock lost"); 3473 lk.lkt_held = NOHOLDER; 3474 #endif 3475 } 3476 3477 /* 3478 * Called from within softdep_disk_write_complete above. Note that 3479 * this routine is always called from interrupt level with further 3480 * splbio interrupts blocked. 3481 * 3482 * Parameters: 3483 * adp: the completed allocdirect 3484 */ 3485 static void 3486 handle_allocdirect_partdone(struct allocdirect *adp) 3487 { 3488 struct allocdirect *listadp; 3489 struct inodedep *inodedep; 3490 long bsize; 3491 3492 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3493 return; 3494 if (adp->ad_buf != NULL) { 3495 lk.lkt_held = NOHOLDER; 3496 panic("handle_allocdirect_partdone: dangling dep"); 3497 } 3498 /* 3499 * The on-disk inode cannot claim to be any larger than the last 3500 * fragment that has been written. Otherwise, the on-disk inode 3501 * might have fragments that were not the last block in the file 3502 * which would corrupt the filesystem. Thus, we cannot free any 3503 * allocdirects after one whose ad_oldblkno claims a fragment as 3504 * these blocks must be rolled back to zero before writing the inode. 3505 * We check the currently active set of allocdirects in id_inoupdt. 3506 */ 3507 inodedep = adp->ad_inodedep; 3508 bsize = inodedep->id_fs->fs_bsize; 3509 TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) { 3510 /* found our block */ 3511 if (listadp == adp) 3512 break; 3513 /* continue if ad_oldlbn is not a fragment */ 3514 if (listadp->ad_oldsize == 0 || 3515 listadp->ad_oldsize == bsize) 3516 continue; 3517 /* hit a fragment */ 3518 return; 3519 } 3520 /* 3521 * If we have reached the end of the current list without 3522 * finding the just finished dependency, then it must be 3523 * on the future dependency list. Future dependencies cannot 3524 * be freed until they are moved to the current list. 3525 */ 3526 if (listadp == NULL) { 3527 #ifdef DEBUG 3528 TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next) 3529 /* found our block */ 3530 if (listadp == adp) 3531 break; 3532 if (listadp == NULL) { 3533 lk.lkt_held = NOHOLDER; 3534 panic("handle_allocdirect_partdone: lost dep"); 3535 } 3536 #endif /* DEBUG */ 3537 return; 3538 } 3539 /* 3540 * If we have found the just finished dependency, then free 3541 * it along with anything that follows it that is complete. 3542 */ 3543 for (; adp; adp = listadp) { 3544 listadp = TAILQ_NEXT(adp, ad_next); 3545 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3546 return; 3547 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 3548 } 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 * aip: the completed allocindir 3558 */ 3559 static void 3560 handle_allocindir_partdone(struct allocindir *aip) 3561 { 3562 struct indirdep *indirdep; 3563 3564 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 3565 return; 3566 if (aip->ai_buf != NULL) { 3567 lk.lkt_held = NOHOLDER; 3568 panic("handle_allocindir_partdone: dangling dependency"); 3569 } 3570 indirdep = aip->ai_indirdep; 3571 if (indirdep->ir_state & UNDONE) { 3572 LIST_REMOVE(aip, ai_next); 3573 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 3574 return; 3575 } 3576 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 3577 aip->ai_newblkno; 3578 LIST_REMOVE(aip, ai_next); 3579 if (aip->ai_freefrag != NULL) 3580 add_to_worklist(&aip->ai_freefrag->ff_list); 3581 WORKITEM_FREE(aip, D_ALLOCINDIR); 3582 } 3583 3584 /* 3585 * Called from within softdep_disk_write_complete above to restore 3586 * in-memory inode block contents to their most up-to-date state. Note 3587 * that this routine is always called from interrupt level with further 3588 * splbio interrupts blocked. 3589 * 3590 * Parameters: 3591 * bp: buffer containing the inode block 3592 */ 3593 static int 3594 handle_written_inodeblock(struct inodedep *inodedep, struct buf *bp) 3595 { 3596 struct worklist *wk, *filefree; 3597 struct allocdirect *adp, *nextadp; 3598 struct ufs1_dinode *dp; 3599 int hadchanges; 3600 3601 if ((inodedep->id_state & IOSTARTED) == 0) { 3602 lk.lkt_held = NOHOLDER; 3603 panic("handle_written_inodeblock: not started"); 3604 } 3605 inodedep->id_state &= ~IOSTARTED; 3606 dp = (struct ufs1_dinode *)bp->b_data + 3607 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 3608 /* 3609 * If we had to rollback the inode allocation because of 3610 * bitmaps being incomplete, then simply restore it. 3611 * Keep the block dirty so that it will not be reclaimed until 3612 * all associated dependencies have been cleared and the 3613 * corresponding updates written to disk. 3614 */ 3615 if (inodedep->id_savedino != NULL) { 3616 *dp = *inodedep->id_savedino; 3617 FREE(inodedep->id_savedino, M_INODEDEP); 3618 inodedep->id_savedino = NULL; 3619 if ((bp->b_flags & B_DELWRI) == 0) 3620 stat_inode_bitmap++; 3621 bdirty(bp); 3622 return (1); 3623 } 3624 inodedep->id_state |= COMPLETE; 3625 /* 3626 * Roll forward anything that had to be rolled back before 3627 * the inode could be updated. 3628 */ 3629 hadchanges = 0; 3630 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 3631 nextadp = TAILQ_NEXT(adp, ad_next); 3632 if (adp->ad_state & ATTACHED) { 3633 lk.lkt_held = NOHOLDER; 3634 panic("handle_written_inodeblock: new entry"); 3635 } 3636 if (adp->ad_lbn < NDADDR) { 3637 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) { 3638 lk.lkt_held = NOHOLDER; 3639 panic("%s: %s #%ld mismatch %d != %d", 3640 "handle_written_inodeblock", 3641 "direct pointer", adp->ad_lbn, 3642 dp->di_db[adp->ad_lbn], adp->ad_oldblkno); 3643 } 3644 dp->di_db[adp->ad_lbn] = adp->ad_newblkno; 3645 } else { 3646 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) { 3647 lk.lkt_held = NOHOLDER; 3648 panic("%s: %s #%ld allocated as %d", 3649 "handle_written_inodeblock", 3650 "indirect pointer", adp->ad_lbn - NDADDR, 3651 dp->di_ib[adp->ad_lbn - NDADDR]); 3652 } 3653 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno; 3654 } 3655 adp->ad_state &= ~UNDONE; 3656 adp->ad_state |= ATTACHED; 3657 hadchanges = 1; 3658 } 3659 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 3660 stat_direct_blk_ptrs++; 3661 /* 3662 * Reset the file size to its most up-to-date value. 3663 */ 3664 if (inodedep->id_savedsize == -1) { 3665 lk.lkt_held = NOHOLDER; 3666 panic("handle_written_inodeblock: bad size"); 3667 } 3668 if (dp->di_size != inodedep->id_savedsize) { 3669 dp->di_size = inodedep->id_savedsize; 3670 hadchanges = 1; 3671 } 3672 inodedep->id_savedsize = -1; 3673 /* 3674 * If there were any rollbacks in the inode block, then it must be 3675 * marked dirty so that its will eventually get written back in 3676 * its correct form. 3677 */ 3678 if (hadchanges) 3679 bdirty(bp); 3680 /* 3681 * Process any allocdirects that completed during the update. 3682 */ 3683 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 3684 handle_allocdirect_partdone(adp); 3685 /* 3686 * Process deallocations that were held pending until the 3687 * inode had been written to disk. Freeing of the inode 3688 * is delayed until after all blocks have been freed to 3689 * avoid creation of new <vfsid, inum, lbn> triples 3690 * before the old ones have been deleted. 3691 */ 3692 filefree = NULL; 3693 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 3694 WORKLIST_REMOVE(wk); 3695 switch (wk->wk_type) { 3696 3697 case D_FREEFILE: 3698 /* 3699 * We defer adding filefree to the worklist until 3700 * all other additions have been made to ensure 3701 * that it will be done after all the old blocks 3702 * have been freed. 3703 */ 3704 if (filefree != NULL) { 3705 lk.lkt_held = NOHOLDER; 3706 panic("handle_written_inodeblock: filefree"); 3707 } 3708 filefree = wk; 3709 continue; 3710 3711 case D_MKDIR: 3712 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 3713 continue; 3714 3715 case D_DIRADD: 3716 diradd_inode_written(WK_DIRADD(wk), inodedep); 3717 continue; 3718 3719 case D_FREEBLKS: 3720 wk->wk_state |= COMPLETE; 3721 if ((wk->wk_state & ALLCOMPLETE) != ALLCOMPLETE) 3722 continue; 3723 /* -- fall through -- */ 3724 case D_FREEFRAG: 3725 case D_DIRREM: 3726 add_to_worklist(wk); 3727 continue; 3728 3729 default: 3730 lk.lkt_held = NOHOLDER; 3731 panic("handle_written_inodeblock: Unknown type %s", 3732 TYPENAME(wk->wk_type)); 3733 /* NOTREACHED */ 3734 } 3735 } 3736 if (filefree != NULL) { 3737 if (free_inodedep(inodedep) == 0) { 3738 lk.lkt_held = NOHOLDER; 3739 panic("handle_written_inodeblock: live inodedep"); 3740 } 3741 add_to_worklist(filefree); 3742 return (0); 3743 } 3744 3745 /* 3746 * If no outstanding dependencies, free it. 3747 */ 3748 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0) 3749 return (0); 3750 return (hadchanges); 3751 } 3752 3753 /* 3754 * Process a diradd entry after its dependent inode has been written. 3755 * This routine must be called with splbio interrupts blocked. 3756 */ 3757 static void 3758 diradd_inode_written(struct diradd *dap, struct inodedep *inodedep) 3759 { 3760 struct pagedep *pagedep; 3761 3762 dap->da_state |= COMPLETE; 3763 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3764 if (dap->da_state & DIRCHG) 3765 pagedep = dap->da_previous->dm_pagedep; 3766 else 3767 pagedep = dap->da_pagedep; 3768 LIST_REMOVE(dap, da_pdlist); 3769 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3770 } 3771 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 3772 } 3773 3774 /* 3775 * Handle the completion of a mkdir dependency. 3776 */ 3777 static void 3778 handle_written_mkdir(struct mkdir *mkdir, int type) 3779 { 3780 struct diradd *dap; 3781 struct pagedep *pagedep; 3782 3783 if (mkdir->md_state != type) { 3784 lk.lkt_held = NOHOLDER; 3785 panic("handle_written_mkdir: bad type"); 3786 } 3787 dap = mkdir->md_diradd; 3788 dap->da_state &= ~type; 3789 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 3790 dap->da_state |= DEPCOMPLETE; 3791 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3792 if (dap->da_state & DIRCHG) 3793 pagedep = dap->da_previous->dm_pagedep; 3794 else 3795 pagedep = dap->da_pagedep; 3796 LIST_REMOVE(dap, da_pdlist); 3797 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3798 } 3799 LIST_REMOVE(mkdir, md_mkdirs); 3800 WORKITEM_FREE(mkdir, D_MKDIR); 3801 } 3802 3803 /* 3804 * Called from within softdep_disk_write_complete above. 3805 * A write operation was just completed. Removed inodes can 3806 * now be freed and associated block pointers may be committed. 3807 * Note that this routine is always called from interrupt level 3808 * with further splbio interrupts blocked. 3809 * 3810 * Parameters: 3811 * bp: buffer containing the written page 3812 */ 3813 static int 3814 handle_written_filepage(struct pagedep *pagedep, struct buf *bp) 3815 { 3816 struct dirrem *dirrem; 3817 struct diradd *dap, *nextdap; 3818 struct direct *ep; 3819 int i, chgs; 3820 3821 if ((pagedep->pd_state & IOSTARTED) == 0) { 3822 lk.lkt_held = NOHOLDER; 3823 panic("handle_written_filepage: not started"); 3824 } 3825 pagedep->pd_state &= ~IOSTARTED; 3826 /* 3827 * Process any directory removals that have been committed. 3828 */ 3829 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 3830 LIST_REMOVE(dirrem, dm_next); 3831 dirrem->dm_dirinum = pagedep->pd_ino; 3832 add_to_worklist(&dirrem->dm_list); 3833 } 3834 /* 3835 * Free any directory additions that have been committed. 3836 */ 3837 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 3838 free_diradd(dap); 3839 /* 3840 * Uncommitted directory entries must be restored. 3841 */ 3842 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 3843 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 3844 dap = nextdap) { 3845 nextdap = LIST_NEXT(dap, da_pdlist); 3846 if (dap->da_state & ATTACHED) { 3847 lk.lkt_held = NOHOLDER; 3848 panic("handle_written_filepage: attached"); 3849 } 3850 ep = (struct direct *) 3851 ((char *)bp->b_data + dap->da_offset); 3852 ep->d_ino = dap->da_newinum; 3853 dap->da_state &= ~UNDONE; 3854 dap->da_state |= ATTACHED; 3855 chgs = 1; 3856 /* 3857 * If the inode referenced by the directory has 3858 * been written out, then the dependency can be 3859 * moved to the pending list. 3860 */ 3861 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3862 LIST_REMOVE(dap, da_pdlist); 3863 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 3864 da_pdlist); 3865 } 3866 } 3867 } 3868 /* 3869 * If there were any rollbacks in the directory, then it must be 3870 * marked dirty so that its will eventually get written back in 3871 * its correct form. 3872 */ 3873 if (chgs) { 3874 if ((bp->b_flags & B_DELWRI) == 0) 3875 stat_dir_entry++; 3876 bdirty(bp); 3877 } 3878 /* 3879 * If no dependencies remain, the pagedep will be freed. 3880 * Otherwise it will remain to update the page before it 3881 * is written back to disk. 3882 */ 3883 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) { 3884 for (i = 0; i < DAHASHSZ; i++) 3885 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL) 3886 break; 3887 if (i == DAHASHSZ) { 3888 LIST_REMOVE(pagedep, pd_hash); 3889 WORKITEM_FREE(pagedep, D_PAGEDEP); 3890 return (0); 3891 } 3892 } 3893 return (1); 3894 } 3895 3896 /* 3897 * Writing back in-core inode structures. 3898 * 3899 * The filesystem only accesses an inode's contents when it occupies an 3900 * "in-core" inode structure. These "in-core" structures are separate from 3901 * the page frames used to cache inode blocks. Only the latter are 3902 * transferred to/from the disk. So, when the updated contents of the 3903 * "in-core" inode structure are copied to the corresponding in-memory inode 3904 * block, the dependencies are also transferred. The following procedure is 3905 * called when copying a dirty "in-core" inode to a cached inode block. 3906 */ 3907 3908 /* 3909 * Called when an inode is loaded from disk. If the effective link count 3910 * differed from the actual link count when it was last flushed, then we 3911 * need to ensure that the correct effective link count is put back. 3912 * 3913 * Parameters: 3914 * ip: the "in_core" copy of the inode 3915 */ 3916 void 3917 softdep_load_inodeblock(struct inode *ip) 3918 { 3919 struct inodedep *inodedep; 3920 3921 /* 3922 * Check for alternate nlink count. 3923 */ 3924 ip->i_effnlink = ip->i_nlink; 3925 ACQUIRE_LOCK(&lk); 3926 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3927 FREE_LOCK(&lk); 3928 return; 3929 } 3930 ip->i_effnlink -= inodedep->id_nlinkdelta; 3931 FREE_LOCK(&lk); 3932 } 3933 3934 /* 3935 * This routine is called just before the "in-core" inode 3936 * information is to be copied to the in-memory inode block. 3937 * Recall that an inode block contains several inodes. If 3938 * the force flag is set, then the dependencies will be 3939 * cleared so that the update can always be made. Note that 3940 * the buffer is locked when this routine is called, so we 3941 * will never be in the middle of writing the inode block 3942 * to disk. 3943 * 3944 * Parameters: 3945 * ip: the "in_core" copy of the inode 3946 * bp: the buffer containing the inode block 3947 * waitfor: nonzero => update must be allowed 3948 */ 3949 void 3950 softdep_update_inodeblock(struct inode *ip, struct buf *bp, 3951 int waitfor) 3952 { 3953 struct inodedep *inodedep; 3954 struct worklist *wk; 3955 int error, gotit; 3956 3957 /* 3958 * If the effective link count is not equal to the actual link 3959 * count, then we must track the difference in an inodedep while 3960 * the inode is (potentially) tossed out of the cache. Otherwise, 3961 * if there is no existing inodedep, then there are no dependencies 3962 * to track. 3963 */ 3964 ACQUIRE_LOCK(&lk); 3965 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3966 FREE_LOCK(&lk); 3967 if (ip->i_effnlink != ip->i_nlink) 3968 panic("softdep_update_inodeblock: bad link count"); 3969 return; 3970 } 3971 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) { 3972 FREE_LOCK(&lk); 3973 panic("softdep_update_inodeblock: bad delta"); 3974 } 3975 /* 3976 * Changes have been initiated. Anything depending on these 3977 * changes cannot occur until this inode has been written. 3978 */ 3979 inodedep->id_state &= ~COMPLETE; 3980 if ((inodedep->id_state & ONWORKLIST) == 0) 3981 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 3982 /* 3983 * Any new dependencies associated with the incore inode must 3984 * now be moved to the list associated with the buffer holding 3985 * the in-memory copy of the inode. Once merged process any 3986 * allocdirects that are completed by the merger. 3987 */ 3988 merge_inode_lists(inodedep); 3989 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL) 3990 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); 3991 /* 3992 * Now that the inode has been pushed into the buffer, the 3993 * operations dependent on the inode being written to disk 3994 * can be moved to the id_bufwait so that they will be 3995 * processed when the buffer I/O completes. 3996 */ 3997 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 3998 WORKLIST_REMOVE(wk); 3999 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 4000 } 4001 /* 4002 * Newly allocated inodes cannot be written until the bitmap 4003 * that allocates them have been written (indicated by 4004 * DEPCOMPLETE being set in id_state). If we are doing a 4005 * forced sync (e.g., an fsync on a file), we force the bitmap 4006 * to be written so that the update can be done. 4007 */ 4008 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) { 4009 FREE_LOCK(&lk); 4010 return; 4011 } 4012 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 4013 FREE_LOCK(&lk); 4014 if (gotit && 4015 (error = bwrite(inodedep->id_buf)) != 0) 4016 softdep_error("softdep_update_inodeblock: bwrite", error); 4017 } 4018 4019 /* 4020 * Merge the new inode dependency list (id_newinoupdt) into the old 4021 * inode dependency list (id_inoupdt). This routine must be called 4022 * with splbio interrupts blocked. 4023 */ 4024 static void 4025 merge_inode_lists(struct inodedep *inodedep) 4026 { 4027 struct allocdirect *listadp, *newadp; 4028 4029 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 4030 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) { 4031 if (listadp->ad_lbn < newadp->ad_lbn) { 4032 listadp = TAILQ_NEXT(listadp, ad_next); 4033 continue; 4034 } 4035 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 4036 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 4037 if (listadp->ad_lbn == newadp->ad_lbn) { 4038 allocdirect_merge(&inodedep->id_inoupdt, newadp, 4039 listadp); 4040 listadp = newadp; 4041 } 4042 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 4043 } 4044 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) { 4045 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 4046 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next); 4047 } 4048 } 4049 4050 /* 4051 * If we are doing an fsync, then we must ensure that any directory 4052 * entries for the inode have been written after the inode gets to disk. 4053 * 4054 * Parameters: 4055 * vp: the "in_core" copy of the inode 4056 */ 4057 static int 4058 softdep_fsync(struct vnode *vp) 4059 { 4060 struct inodedep *inodedep; 4061 struct pagedep *pagedep; 4062 struct worklist *wk; 4063 struct diradd *dap; 4064 struct mount *mnt; 4065 struct vnode *pvp; 4066 struct inode *ip; 4067 struct buf *bp; 4068 struct fs *fs; 4069 int error, flushparent; 4070 ino_t parentino; 4071 ufs_lbn_t lbn; 4072 4073 ip = VTOI(vp); 4074 fs = ip->i_fs; 4075 ACQUIRE_LOCK(&lk); 4076 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) { 4077 FREE_LOCK(&lk); 4078 return (0); 4079 } 4080 if (LIST_FIRST(&inodedep->id_inowait) != NULL || 4081 LIST_FIRST(&inodedep->id_bufwait) != NULL || 4082 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 4083 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) { 4084 FREE_LOCK(&lk); 4085 panic("softdep_fsync: pending ops"); 4086 } 4087 for (error = 0, flushparent = 0; ; ) { 4088 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 4089 break; 4090 if (wk->wk_type != D_DIRADD) { 4091 FREE_LOCK(&lk); 4092 panic("softdep_fsync: Unexpected type %s", 4093 TYPENAME(wk->wk_type)); 4094 } 4095 dap = WK_DIRADD(wk); 4096 /* 4097 * Flush our parent if this directory entry 4098 * has a MKDIR_PARENT dependency. 4099 */ 4100 if (dap->da_state & DIRCHG) 4101 pagedep = dap->da_previous->dm_pagedep; 4102 else 4103 pagedep = dap->da_pagedep; 4104 mnt = pagedep->pd_mnt; 4105 parentino = pagedep->pd_ino; 4106 lbn = pagedep->pd_lbn; 4107 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) { 4108 FREE_LOCK(&lk); 4109 panic("softdep_fsync: dirty"); 4110 } 4111 flushparent = dap->da_state & MKDIR_PARENT; 4112 /* 4113 * If we are being fsync'ed as part of vgone'ing this vnode, 4114 * then we will not be able to release and recover the 4115 * vnode below, so we just have to give up on writing its 4116 * directory entry out. It will eventually be written, just 4117 * not now, but then the user was not asking to have it 4118 * written, so we are not breaking any promises. 4119 */ 4120 if (vp->v_flag & VRECLAIMED) 4121 break; 4122 /* 4123 * We prevent deadlock by always fetching inodes from the 4124 * root, moving down the directory tree. Thus, when fetching 4125 * our parent directory, we must unlock ourselves before 4126 * requesting the lock on our parent. See the comment in 4127 * ufs_lookup for details on possible races. 4128 */ 4129 FREE_LOCK(&lk); 4130 vn_unlock(vp); 4131 error = VFS_VGET(mnt, parentino, &pvp); 4132 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4133 if (error != 0) 4134 return (error); 4135 if (flushparent) { 4136 if ((error = ffs_update(pvp, 1)) != 0) { 4137 vput(pvp); 4138 return (error); 4139 } 4140 } 4141 /* 4142 * Flush directory page containing the inode's name. 4143 */ 4144 error = bread(pvp, lblktodoff(fs, lbn), blksize(fs, VTOI(pvp), lbn), &bp); 4145 if (error == 0) 4146 error = bwrite(bp); 4147 vput(pvp); 4148 if (error != 0) 4149 return (error); 4150 ACQUIRE_LOCK(&lk); 4151 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) 4152 break; 4153 } 4154 FREE_LOCK(&lk); 4155 return (0); 4156 } 4157 4158 /* 4159 * Flush all the dirty bitmaps associated with the block device 4160 * before flushing the rest of the dirty blocks so as to reduce 4161 * the number of dependencies that will have to be rolled back. 4162 */ 4163 static int softdep_fsync_mountdev_bp(struct buf *bp, void *data); 4164 4165 void 4166 softdep_fsync_mountdev(struct vnode *vp) 4167 { 4168 if (!vn_isdisk(vp, NULL)) 4169 panic("softdep_fsync_mountdev: vnode not a disk"); 4170 ACQUIRE_LOCK(&lk); 4171 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 4172 softdep_fsync_mountdev_bp, vp); 4173 drain_output(vp, 1); 4174 FREE_LOCK(&lk); 4175 } 4176 4177 static int 4178 softdep_fsync_mountdev_bp(struct buf *bp, void *data) 4179 { 4180 struct worklist *wk; 4181 struct vnode *vp = data; 4182 4183 /* 4184 * If it is already scheduled, skip to the next buffer. 4185 */ 4186 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) 4187 return(0); 4188 if (bp->b_vp != vp || (bp->b_flags & B_DELWRI) == 0) { 4189 BUF_UNLOCK(bp); 4190 kprintf("softdep_fsync_mountdev_bp: warning, buffer %p ripped out from under vnode %p\n", bp, vp); 4191 return(0); 4192 } 4193 /* 4194 * We are only interested in bitmaps with outstanding 4195 * dependencies. 4196 */ 4197 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 4198 wk->wk_type != D_BMSAFEMAP) { 4199 BUF_UNLOCK(bp); 4200 return(0); 4201 } 4202 bremfree(bp); 4203 FREE_LOCK(&lk); 4204 (void) bawrite(bp); 4205 ACQUIRE_LOCK(&lk); 4206 return(0); 4207 } 4208 4209 /* 4210 * This routine is called when we are trying to synchronously flush a 4211 * file. This routine must eliminate any filesystem metadata dependencies 4212 * so that the syncing routine can succeed by pushing the dirty blocks 4213 * associated with the file. If any I/O errors occur, they are returned. 4214 */ 4215 struct softdep_sync_metadata_info { 4216 struct vnode *vp; 4217 int waitfor; 4218 }; 4219 4220 static int softdep_sync_metadata_bp(struct buf *bp, void *data); 4221 4222 int 4223 softdep_sync_metadata(struct vnode *vp, struct thread *td) 4224 { 4225 struct softdep_sync_metadata_info info; 4226 int error, waitfor; 4227 4228 /* 4229 * Check whether this vnode is involved in a filesystem 4230 * that is doing soft dependency processing. 4231 */ 4232 if (!vn_isdisk(vp, NULL)) { 4233 if (!DOINGSOFTDEP(vp)) 4234 return (0); 4235 } else 4236 if (vp->v_rdev->si_mountpoint == NULL || 4237 (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0) 4238 return (0); 4239 /* 4240 * Ensure that any direct block dependencies have been cleared. 4241 */ 4242 ACQUIRE_LOCK(&lk); 4243 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) { 4244 FREE_LOCK(&lk); 4245 return (error); 4246 } 4247 /* 4248 * For most files, the only metadata dependencies are the 4249 * cylinder group maps that allocate their inode or blocks. 4250 * The block allocation dependencies can be found by traversing 4251 * the dependency lists for any buffers that remain on their 4252 * dirty buffer list. The inode allocation dependency will 4253 * be resolved when the inode is updated with MNT_WAIT. 4254 * This work is done in two passes. The first pass grabs most 4255 * of the buffers and begins asynchronously writing them. The 4256 * only way to wait for these asynchronous writes is to sleep 4257 * on the filesystem vnode which may stay busy for a long time 4258 * if the filesystem is active. So, instead, we make a second 4259 * pass over the dependencies blocking on each write. In the 4260 * usual case we will be blocking against a write that we 4261 * initiated, so when it is done the dependency will have been 4262 * resolved. Thus the second pass is expected to end quickly. 4263 */ 4264 waitfor = MNT_NOWAIT; 4265 top: 4266 /* 4267 * We must wait for any I/O in progress to finish so that 4268 * all potential buffers on the dirty list will be visible. 4269 */ 4270 drain_output(vp, 1); 4271 info.vp = vp; 4272 info.waitfor = waitfor; 4273 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 4274 softdep_sync_metadata_bp, &info); 4275 if (error < 0) { 4276 FREE_LOCK(&lk); 4277 return(-error); /* error code */ 4278 } 4279 4280 /* 4281 * The brief unlock is to allow any pent up dependency 4282 * processing to be done. Then proceed with the second pass. 4283 */ 4284 if (waitfor == MNT_NOWAIT) { 4285 waitfor = MNT_WAIT; 4286 FREE_LOCK(&lk); 4287 ACQUIRE_LOCK(&lk); 4288 goto top; 4289 } 4290 4291 /* 4292 * If we have managed to get rid of all the dirty buffers, 4293 * then we are done. For certain directories and block 4294 * devices, we may need to do further work. 4295 * 4296 * We must wait for any I/O in progress to finish so that 4297 * all potential buffers on the dirty list will be visible. 4298 */ 4299 drain_output(vp, 1); 4300 if (RB_EMPTY(&vp->v_rbdirty_tree)) { 4301 FREE_LOCK(&lk); 4302 return (0); 4303 } 4304 4305 FREE_LOCK(&lk); 4306 /* 4307 * If we are trying to sync a block device, some of its buffers may 4308 * contain metadata that cannot be written until the contents of some 4309 * partially written files have been written to disk. The only easy 4310 * way to accomplish this is to sync the entire filesystem (luckily 4311 * this happens rarely). 4312 */ 4313 if (vn_isdisk(vp, NULL) && 4314 vp->v_rdev && 4315 vp->v_rdev->si_mountpoint && !vn_islocked(vp) && 4316 (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT)) != 0) 4317 return (error); 4318 return (0); 4319 } 4320 4321 static int 4322 softdep_sync_metadata_bp(struct buf *bp, void *data) 4323 { 4324 struct softdep_sync_metadata_info *info = data; 4325 struct pagedep *pagedep; 4326 struct allocdirect *adp; 4327 struct allocindir *aip; 4328 struct worklist *wk; 4329 struct buf *nbp; 4330 int error; 4331 int i; 4332 4333 if (getdirtybuf(&bp, MNT_WAIT) == 0) { 4334 kprintf("softdep_sync_metadata_bp(1): caught buf %p going away\n", bp); 4335 return (1); 4336 } 4337 if (bp->b_vp != info->vp || (bp->b_flags & B_DELWRI) == 0) { 4338 kprintf("softdep_sync_metadata_bp(2): caught buf %p going away vp %p\n", bp, info->vp); 4339 BUF_UNLOCK(bp); 4340 return(1); 4341 } 4342 4343 /* 4344 * As we hold the buffer locked, none of its dependencies 4345 * will disappear. 4346 */ 4347 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 4348 switch (wk->wk_type) { 4349 4350 case D_ALLOCDIRECT: 4351 adp = WK_ALLOCDIRECT(wk); 4352 if (adp->ad_state & DEPCOMPLETE) 4353 break; 4354 nbp = adp->ad_buf; 4355 if (getdirtybuf(&nbp, info->waitfor) == 0) 4356 break; 4357 FREE_LOCK(&lk); 4358 if (info->waitfor == MNT_NOWAIT) { 4359 bawrite(nbp); 4360 } else if ((error = bwrite(nbp)) != 0) { 4361 bawrite(bp); 4362 ACQUIRE_LOCK(&lk); 4363 return (-error); 4364 } 4365 ACQUIRE_LOCK(&lk); 4366 break; 4367 4368 case D_ALLOCINDIR: 4369 aip = WK_ALLOCINDIR(wk); 4370 if (aip->ai_state & DEPCOMPLETE) 4371 break; 4372 nbp = aip->ai_buf; 4373 if (getdirtybuf(&nbp, info->waitfor) == 0) 4374 break; 4375 FREE_LOCK(&lk); 4376 if (info->waitfor == MNT_NOWAIT) { 4377 bawrite(nbp); 4378 } else if ((error = bwrite(nbp)) != 0) { 4379 bawrite(bp); 4380 ACQUIRE_LOCK(&lk); 4381 return (-error); 4382 } 4383 ACQUIRE_LOCK(&lk); 4384 break; 4385 4386 case D_INDIRDEP: 4387 restart: 4388 4389 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) { 4390 if (aip->ai_state & DEPCOMPLETE) 4391 continue; 4392 nbp = aip->ai_buf; 4393 if (getdirtybuf(&nbp, MNT_WAIT) == 0) 4394 goto restart; 4395 FREE_LOCK(&lk); 4396 if ((error = bwrite(nbp)) != 0) { 4397 bawrite(bp); 4398 ACQUIRE_LOCK(&lk); 4399 return (-error); 4400 } 4401 ACQUIRE_LOCK(&lk); 4402 goto restart; 4403 } 4404 break; 4405 4406 case D_INODEDEP: 4407 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs, 4408 WK_INODEDEP(wk)->id_ino)) != 0) { 4409 FREE_LOCK(&lk); 4410 bawrite(bp); 4411 ACQUIRE_LOCK(&lk); 4412 return (-error); 4413 } 4414 break; 4415 4416 case D_PAGEDEP: 4417 /* 4418 * We are trying to sync a directory that may 4419 * have dependencies on both its own metadata 4420 * and/or dependencies on the inodes of any 4421 * recently allocated files. We walk its diradd 4422 * lists pushing out the associated inode. 4423 */ 4424 pagedep = WK_PAGEDEP(wk); 4425 for (i = 0; i < DAHASHSZ; i++) { 4426 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 4427 continue; 4428 if ((error = 4429 flush_pagedep_deps(info->vp, 4430 pagedep->pd_mnt, 4431 &pagedep->pd_diraddhd[i]))) { 4432 FREE_LOCK(&lk); 4433 bawrite(bp); 4434 ACQUIRE_LOCK(&lk); 4435 return (-error); 4436 } 4437 } 4438 break; 4439 4440 case D_MKDIR: 4441 /* 4442 * This case should never happen if the vnode has 4443 * been properly sync'ed. However, if this function 4444 * is used at a place where the vnode has not yet 4445 * been sync'ed, this dependency can show up. So, 4446 * rather than panic, just flush it. 4447 */ 4448 nbp = WK_MKDIR(wk)->md_buf; 4449 if (getdirtybuf(&nbp, info->waitfor) == 0) 4450 break; 4451 FREE_LOCK(&lk); 4452 if (info->waitfor == MNT_NOWAIT) { 4453 bawrite(nbp); 4454 } else if ((error = bwrite(nbp)) != 0) { 4455 bawrite(bp); 4456 ACQUIRE_LOCK(&lk); 4457 return (-error); 4458 } 4459 ACQUIRE_LOCK(&lk); 4460 break; 4461 4462 case D_BMSAFEMAP: 4463 /* 4464 * This case should never happen if the vnode has 4465 * been properly sync'ed. However, if this function 4466 * is used at a place where the vnode has not yet 4467 * been sync'ed, this dependency can show up. So, 4468 * rather than panic, just flush it. 4469 * 4470 * nbp can wind up == bp if a device node for the 4471 * same filesystem is being fsynced at the same time, 4472 * leading to a panic if we don't catch the case. 4473 */ 4474 nbp = WK_BMSAFEMAP(wk)->sm_buf; 4475 if (nbp == bp) 4476 break; 4477 if (getdirtybuf(&nbp, info->waitfor) == 0) 4478 break; 4479 FREE_LOCK(&lk); 4480 if (info->waitfor == MNT_NOWAIT) { 4481 bawrite(nbp); 4482 } else if ((error = bwrite(nbp)) != 0) { 4483 bawrite(bp); 4484 ACQUIRE_LOCK(&lk); 4485 return (-error); 4486 } 4487 ACQUIRE_LOCK(&lk); 4488 break; 4489 4490 default: 4491 FREE_LOCK(&lk); 4492 panic("softdep_sync_metadata: Unknown type %s", 4493 TYPENAME(wk->wk_type)); 4494 /* NOTREACHED */ 4495 } 4496 } 4497 FREE_LOCK(&lk); 4498 bawrite(bp); 4499 ACQUIRE_LOCK(&lk); 4500 return(0); 4501 } 4502 4503 /* 4504 * Flush the dependencies associated with an inodedep. 4505 * Called with splbio blocked. 4506 */ 4507 static int 4508 flush_inodedep_deps(struct fs *fs, ino_t ino) 4509 { 4510 struct inodedep *inodedep; 4511 struct allocdirect *adp; 4512 int error, waitfor; 4513 struct buf *bp; 4514 4515 /* 4516 * This work is done in two passes. The first pass grabs most 4517 * of the buffers and begins asynchronously writing them. The 4518 * only way to wait for these asynchronous writes is to sleep 4519 * on the filesystem vnode which may stay busy for a long time 4520 * if the filesystem is active. So, instead, we make a second 4521 * pass over the dependencies blocking on each write. In the 4522 * usual case we will be blocking against a write that we 4523 * initiated, so when it is done the dependency will have been 4524 * resolved. Thus the second pass is expected to end quickly. 4525 * We give a brief window at the top of the loop to allow 4526 * any pending I/O to complete. 4527 */ 4528 for (waitfor = MNT_NOWAIT; ; ) { 4529 FREE_LOCK(&lk); 4530 ACQUIRE_LOCK(&lk); 4531 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4532 return (0); 4533 TAILQ_FOREACH(adp, &inodedep->id_inoupdt, ad_next) { 4534 if (adp->ad_state & DEPCOMPLETE) 4535 continue; 4536 bp = adp->ad_buf; 4537 if (getdirtybuf(&bp, waitfor) == 0) { 4538 if (waitfor == MNT_NOWAIT) 4539 continue; 4540 break; 4541 } 4542 FREE_LOCK(&lk); 4543 if (waitfor == MNT_NOWAIT) { 4544 bawrite(bp); 4545 } else if ((error = bwrite(bp)) != 0) { 4546 ACQUIRE_LOCK(&lk); 4547 return (error); 4548 } 4549 ACQUIRE_LOCK(&lk); 4550 break; 4551 } 4552 if (adp != NULL) 4553 continue; 4554 TAILQ_FOREACH(adp, &inodedep->id_newinoupdt, ad_next) { 4555 if (adp->ad_state & DEPCOMPLETE) 4556 continue; 4557 bp = adp->ad_buf; 4558 if (getdirtybuf(&bp, waitfor) == 0) { 4559 if (waitfor == MNT_NOWAIT) 4560 continue; 4561 break; 4562 } 4563 FREE_LOCK(&lk); 4564 if (waitfor == MNT_NOWAIT) { 4565 bawrite(bp); 4566 } else if ((error = bwrite(bp)) != 0) { 4567 ACQUIRE_LOCK(&lk); 4568 return (error); 4569 } 4570 ACQUIRE_LOCK(&lk); 4571 break; 4572 } 4573 if (adp != NULL) 4574 continue; 4575 /* 4576 * If pass2, we are done, otherwise do pass 2. 4577 */ 4578 if (waitfor == MNT_WAIT) 4579 break; 4580 waitfor = MNT_WAIT; 4581 } 4582 /* 4583 * Try freeing inodedep in case all dependencies have been removed. 4584 */ 4585 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0) 4586 (void) free_inodedep(inodedep); 4587 return (0); 4588 } 4589 4590 /* 4591 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 4592 * Called with splbio blocked. 4593 */ 4594 static int 4595 flush_pagedep_deps(struct vnode *pvp, struct mount *mp, 4596 struct diraddhd *diraddhdp) 4597 { 4598 struct inodedep *inodedep; 4599 struct ufsmount *ump; 4600 struct diradd *dap; 4601 struct vnode *vp; 4602 int gotit, error = 0; 4603 struct buf *bp; 4604 ino_t inum; 4605 4606 ump = VFSTOUFS(mp); 4607 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 4608 /* 4609 * Flush ourselves if this directory entry 4610 * has a MKDIR_PARENT dependency. 4611 */ 4612 if (dap->da_state & MKDIR_PARENT) { 4613 FREE_LOCK(&lk); 4614 if ((error = ffs_update(pvp, 1)) != 0) 4615 break; 4616 ACQUIRE_LOCK(&lk); 4617 /* 4618 * If that cleared dependencies, go on to next. 4619 */ 4620 if (dap != LIST_FIRST(diraddhdp)) 4621 continue; 4622 if (dap->da_state & MKDIR_PARENT) { 4623 FREE_LOCK(&lk); 4624 panic("flush_pagedep_deps: MKDIR_PARENT"); 4625 } 4626 } 4627 /* 4628 * A newly allocated directory must have its "." and 4629 * ".." entries written out before its name can be 4630 * committed in its parent. We do not want or need 4631 * the full semantics of a synchronous VOP_FSYNC as 4632 * that may end up here again, once for each directory 4633 * level in the filesystem. Instead, we push the blocks 4634 * and wait for them to clear. We have to fsync twice 4635 * because the first call may choose to defer blocks 4636 * that still have dependencies, but deferral will 4637 * happen at most once. 4638 */ 4639 inum = dap->da_newinum; 4640 if (dap->da_state & MKDIR_BODY) { 4641 FREE_LOCK(&lk); 4642 if ((error = VFS_VGET(mp, inum, &vp)) != 0) 4643 break; 4644 if ((error=VOP_FSYNC(vp, MNT_NOWAIT)) || 4645 (error=VOP_FSYNC(vp, MNT_NOWAIT))) { 4646 vput(vp); 4647 break; 4648 } 4649 drain_output(vp, 0); 4650 vput(vp); 4651 ACQUIRE_LOCK(&lk); 4652 /* 4653 * If that cleared dependencies, go on to next. 4654 */ 4655 if (dap != LIST_FIRST(diraddhdp)) 4656 continue; 4657 if (dap->da_state & MKDIR_BODY) { 4658 FREE_LOCK(&lk); 4659 panic("flush_pagedep_deps: MKDIR_BODY"); 4660 } 4661 } 4662 /* 4663 * Flush the inode on which the directory entry depends. 4664 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 4665 * the only remaining dependency is that the updated inode 4666 * count must get pushed to disk. The inode has already 4667 * been pushed into its inode buffer (via VOP_UPDATE) at 4668 * the time of the reference count change. So we need only 4669 * locate that buffer, ensure that there will be no rollback 4670 * caused by a bitmap dependency, then write the inode buffer. 4671 */ 4672 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) { 4673 FREE_LOCK(&lk); 4674 panic("flush_pagedep_deps: lost inode"); 4675 } 4676 /* 4677 * If the inode still has bitmap dependencies, 4678 * push them to disk. 4679 */ 4680 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4681 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 4682 FREE_LOCK(&lk); 4683 if (gotit && (error = bwrite(inodedep->id_buf)) != 0) 4684 break; 4685 ACQUIRE_LOCK(&lk); 4686 if (dap != LIST_FIRST(diraddhdp)) 4687 continue; 4688 } 4689 /* 4690 * If the inode is still sitting in a buffer waiting 4691 * to be written, push it to disk. 4692 */ 4693 FREE_LOCK(&lk); 4694 if ((error = bread(ump->um_devvp, 4695 fsbtodoff(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), 4696 (int)ump->um_fs->fs_bsize, &bp)) != 0) 4697 break; 4698 if ((error = bwrite(bp)) != 0) 4699 break; 4700 ACQUIRE_LOCK(&lk); 4701 /* 4702 * If we have failed to get rid of all the dependencies 4703 * then something is seriously wrong. 4704 */ 4705 if (dap == LIST_FIRST(diraddhdp)) { 4706 FREE_LOCK(&lk); 4707 panic("flush_pagedep_deps: flush failed"); 4708 } 4709 } 4710 if (error) 4711 ACQUIRE_LOCK(&lk); 4712 return (error); 4713 } 4714 4715 /* 4716 * A large burst of file addition or deletion activity can drive the 4717 * memory load excessively high. First attempt to slow things down 4718 * using the techniques below. If that fails, this routine requests 4719 * the offending operations to fall back to running synchronously 4720 * until the memory load returns to a reasonable level. 4721 */ 4722 int 4723 softdep_slowdown(struct vnode *vp) 4724 { 4725 int max_softdeps_hard; 4726 4727 max_softdeps_hard = max_softdeps * 11 / 10; 4728 if (num_dirrem < max_softdeps_hard / 2 && 4729 num_inodedep < max_softdeps_hard) 4730 return (0); 4731 stat_sync_limit_hit += 1; 4732 return (1); 4733 } 4734 4735 /* 4736 * If memory utilization has gotten too high, deliberately slow things 4737 * down and speed up the I/O processing. 4738 */ 4739 static int 4740 request_cleanup(int resource, int islocked) 4741 { 4742 struct thread *td = curthread; /* XXX */ 4743 4744 /* 4745 * We never hold up the filesystem syncer process. 4746 */ 4747 if (td == filesys_syncer) 4748 return (0); 4749 /* 4750 * First check to see if the work list has gotten backlogged. 4751 * If it has, co-opt this process to help clean up two entries. 4752 * Because this process may hold inodes locked, we cannot 4753 * handle any remove requests that might block on a locked 4754 * inode as that could lead to deadlock. 4755 */ 4756 if (num_on_worklist > max_softdeps / 10) { 4757 if (islocked) 4758 FREE_LOCK(&lk); 4759 process_worklist_item(NULL, LK_NOWAIT); 4760 process_worklist_item(NULL, LK_NOWAIT); 4761 stat_worklist_push += 2; 4762 if (islocked) 4763 ACQUIRE_LOCK(&lk); 4764 return(1); 4765 } 4766 4767 /* 4768 * If we are resource constrained on inode dependencies, try 4769 * flushing some dirty inodes. Otherwise, we are constrained 4770 * by file deletions, so try accelerating flushes of directories 4771 * with removal dependencies. We would like to do the cleanup 4772 * here, but we probably hold an inode locked at this point and 4773 * that might deadlock against one that we try to clean. So, 4774 * the best that we can do is request the syncer daemon to do 4775 * the cleanup for us. 4776 */ 4777 switch (resource) { 4778 4779 case FLUSH_INODES: 4780 stat_ino_limit_push += 1; 4781 req_clear_inodedeps += 1; 4782 stat_countp = &stat_ino_limit_hit; 4783 break; 4784 4785 case FLUSH_REMOVE: 4786 stat_blk_limit_push += 1; 4787 req_clear_remove += 1; 4788 stat_countp = &stat_blk_limit_hit; 4789 break; 4790 4791 default: 4792 if (islocked) 4793 FREE_LOCK(&lk); 4794 panic("request_cleanup: unknown type"); 4795 } 4796 /* 4797 * Hopefully the syncer daemon will catch up and awaken us. 4798 * We wait at most tickdelay before proceeding in any case. 4799 */ 4800 if (islocked == 0) 4801 ACQUIRE_LOCK(&lk); 4802 proc_waiting += 1; 4803 if (!callout_active(&handle)) 4804 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2, 4805 pause_timer, NULL); 4806 interlocked_sleep(&lk, SLEEP, (caddr_t)&proc_waiting, 0, 4807 "softupdate", 0); 4808 proc_waiting -= 1; 4809 if (islocked == 0) 4810 FREE_LOCK(&lk); 4811 return (1); 4812 } 4813 4814 /* 4815 * Awaken processes pausing in request_cleanup and clear proc_waiting 4816 * to indicate that there is no longer a timer running. 4817 */ 4818 void 4819 pause_timer(void *arg) 4820 { 4821 *stat_countp += 1; 4822 wakeup_one(&proc_waiting); 4823 if (proc_waiting > 0) 4824 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2, 4825 pause_timer, NULL); 4826 else 4827 callout_deactivate(&handle); 4828 } 4829 4830 /* 4831 * Flush out a directory with at least one removal dependency in an effort to 4832 * reduce the number of dirrem, freefile, and freeblks dependency structures. 4833 */ 4834 static void 4835 clear_remove(struct thread *td) 4836 { 4837 struct pagedep_hashhead *pagedephd; 4838 struct pagedep *pagedep; 4839 static int next = 0; 4840 struct mount *mp; 4841 struct vnode *vp; 4842 int error, cnt; 4843 ino_t ino; 4844 4845 ACQUIRE_LOCK(&lk); 4846 for (cnt = 0; cnt < pagedep_hash; cnt++) { 4847 pagedephd = &pagedep_hashtbl[next++]; 4848 if (next >= pagedep_hash) 4849 next = 0; 4850 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 4851 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL) 4852 continue; 4853 mp = pagedep->pd_mnt; 4854 ino = pagedep->pd_ino; 4855 FREE_LOCK(&lk); 4856 if ((error = VFS_VGET(mp, ino, &vp)) != 0) { 4857 softdep_error("clear_remove: vget", error); 4858 return; 4859 } 4860 if ((error = VOP_FSYNC(vp, MNT_NOWAIT))) 4861 softdep_error("clear_remove: fsync", error); 4862 drain_output(vp, 0); 4863 vput(vp); 4864 return; 4865 } 4866 } 4867 FREE_LOCK(&lk); 4868 } 4869 4870 /* 4871 * Clear out a block of dirty inodes in an effort to reduce 4872 * the number of inodedep dependency structures. 4873 */ 4874 struct clear_inodedeps_info { 4875 struct fs *fs; 4876 struct mount *mp; 4877 }; 4878 4879 static int 4880 clear_inodedeps_mountlist_callback(struct mount *mp, void *data) 4881 { 4882 struct clear_inodedeps_info *info = data; 4883 4884 if ((mp->mnt_flag & MNT_SOFTDEP) && info->fs == VFSTOUFS(mp)->um_fs) { 4885 info->mp = mp; 4886 return(-1); 4887 } 4888 return(0); 4889 } 4890 4891 static void 4892 clear_inodedeps(struct thread *td) 4893 { 4894 struct clear_inodedeps_info info; 4895 struct inodedep_hashhead *inodedephd; 4896 struct inodedep *inodedep; 4897 static int next = 0; 4898 struct vnode *vp; 4899 struct fs *fs; 4900 int error, cnt; 4901 ino_t firstino, lastino, ino; 4902 4903 ACQUIRE_LOCK(&lk); 4904 /* 4905 * Pick a random inode dependency to be cleared. 4906 * We will then gather up all the inodes in its block 4907 * that have dependencies and flush them out. 4908 */ 4909 for (cnt = 0; cnt < inodedep_hash; cnt++) { 4910 inodedephd = &inodedep_hashtbl[next++]; 4911 if (next >= inodedep_hash) 4912 next = 0; 4913 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 4914 break; 4915 } 4916 if (inodedep == NULL) { 4917 FREE_LOCK(&lk); 4918 return; 4919 } 4920 /* 4921 * Ugly code to find mount point given pointer to superblock. 4922 */ 4923 fs = inodedep->id_fs; 4924 info.mp = NULL; 4925 info.fs = fs; 4926 mountlist_scan(clear_inodedeps_mountlist_callback, 4927 &info, MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 4928 /* 4929 * Find the last inode in the block with dependencies. 4930 */ 4931 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 4932 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 4933 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0) 4934 break; 4935 /* 4936 * Asynchronously push all but the last inode with dependencies. 4937 * Synchronously push the last inode with dependencies to ensure 4938 * that the inode block gets written to free up the inodedeps. 4939 */ 4940 for (ino = firstino; ino <= lastino; ino++) { 4941 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4942 continue; 4943 FREE_LOCK(&lk); 4944 if ((error = VFS_VGET(info.mp, ino, &vp)) != 0) { 4945 softdep_error("clear_inodedeps: vget", error); 4946 return; 4947 } 4948 if (ino == lastino) { 4949 if ((error = VOP_FSYNC(vp, MNT_WAIT))) 4950 softdep_error("clear_inodedeps: fsync1", error); 4951 } else { 4952 if ((error = VOP_FSYNC(vp, MNT_NOWAIT))) 4953 softdep_error("clear_inodedeps: fsync2", error); 4954 drain_output(vp, 0); 4955 } 4956 vput(vp); 4957 ACQUIRE_LOCK(&lk); 4958 } 4959 FREE_LOCK(&lk); 4960 } 4961 4962 /* 4963 * Function to determine if the buffer has outstanding dependencies 4964 * that will cause a roll-back if the buffer is written. If wantcount 4965 * is set, return number of dependencies, otherwise just yes or no. 4966 */ 4967 static int 4968 softdep_count_dependencies(struct buf *bp, int wantcount) 4969 { 4970 struct worklist *wk; 4971 struct inodedep *inodedep; 4972 struct indirdep *indirdep; 4973 struct allocindir *aip; 4974 struct pagedep *pagedep; 4975 struct diradd *dap; 4976 int i, retval; 4977 4978 retval = 0; 4979 ACQUIRE_LOCK(&lk); 4980 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 4981 switch (wk->wk_type) { 4982 4983 case D_INODEDEP: 4984 inodedep = WK_INODEDEP(wk); 4985 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4986 /* bitmap allocation dependency */ 4987 retval += 1; 4988 if (!wantcount) 4989 goto out; 4990 } 4991 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 4992 /* direct block pointer dependency */ 4993 retval += 1; 4994 if (!wantcount) 4995 goto out; 4996 } 4997 continue; 4998 4999 case D_INDIRDEP: 5000 indirdep = WK_INDIRDEP(wk); 5001 5002 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 5003 /* indirect block pointer dependency */ 5004 retval += 1; 5005 if (!wantcount) 5006 goto out; 5007 } 5008 continue; 5009 5010 case D_PAGEDEP: 5011 pagedep = WK_PAGEDEP(wk); 5012 for (i = 0; i < DAHASHSZ; i++) { 5013 5014 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 5015 /* directory entry dependency */ 5016 retval += 1; 5017 if (!wantcount) 5018 goto out; 5019 } 5020 } 5021 continue; 5022 5023 case D_BMSAFEMAP: 5024 case D_ALLOCDIRECT: 5025 case D_ALLOCINDIR: 5026 case D_MKDIR: 5027 /* never a dependency on these blocks */ 5028 continue; 5029 5030 default: 5031 FREE_LOCK(&lk); 5032 panic("softdep_check_for_rollback: Unexpected type %s", 5033 TYPENAME(wk->wk_type)); 5034 /* NOTREACHED */ 5035 } 5036 } 5037 out: 5038 FREE_LOCK(&lk); 5039 return retval; 5040 } 5041 5042 /* 5043 * Acquire exclusive access to a buffer. 5044 * Must be called with splbio blocked. 5045 * Return 1 if buffer was acquired. 5046 */ 5047 static int 5048 getdirtybuf(struct buf **bpp, int waitfor) 5049 { 5050 struct buf *bp; 5051 int error; 5052 5053 for (;;) { 5054 if ((bp = *bpp) == NULL) 5055 return (0); 5056 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) 5057 break; 5058 if (waitfor != MNT_WAIT) 5059 return (0); 5060 error = interlocked_sleep(&lk, LOCKBUF, bp, 5061 LK_EXCLUSIVE | LK_SLEEPFAIL, 0, 0); 5062 if (error != ENOLCK) { 5063 FREE_LOCK(&lk); 5064 panic("getdirtybuf: inconsistent lock"); 5065 } 5066 } 5067 if ((bp->b_flags & B_DELWRI) == 0) { 5068 BUF_UNLOCK(bp); 5069 return (0); 5070 } 5071 bremfree(bp); 5072 return (1); 5073 } 5074 5075 /* 5076 * Wait for pending output on a vnode to complete. 5077 * Must be called with vnode locked. 5078 */ 5079 static void 5080 drain_output(struct vnode *vp, int islocked) 5081 { 5082 5083 if (!islocked) 5084 ACQUIRE_LOCK(&lk); 5085 while (vp->v_track_write.bk_active) { 5086 vp->v_track_write.bk_waitflag = 1; 5087 interlocked_sleep(&lk, SLEEP, &vp->v_track_write, 5088 0, "drainvp", 0); 5089 } 5090 if (!islocked) 5091 FREE_LOCK(&lk); 5092 } 5093 5094 /* 5095 * Called whenever a buffer that is being invalidated or reallocated 5096 * contains dependencies. This should only happen if an I/O error has 5097 * occurred. The routine is called with the buffer locked. 5098 */ 5099 static void 5100 softdep_deallocate_dependencies(struct buf *bp) 5101 { 5102 5103 if ((bp->b_flags & B_ERROR) == 0) 5104 panic("softdep_deallocate_dependencies: dangling deps"); 5105 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntfromname, bp->b_error); 5106 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 5107 } 5108 5109 /* 5110 * Function to handle asynchronous write errors in the filesystem. 5111 */ 5112 void 5113 softdep_error(char *func, int error) 5114 { 5115 5116 /* XXX should do something better! */ 5117 kprintf("%s: got error %d while accessing filesystem\n", func, error); 5118 } 5119