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 1.51 2007/04/12 19:50:20 swildner 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 NULL, &bp->b_bio2.bio_offset, 1741 NULL, NULL); 1742 } 1743 KKASSERT(bp->b_bio2.bio_offset != NOOFFSET); 1744 newindirdep->ir_savebp = getblk(ip->i_devvp, 1745 bp->b_bio2.bio_offset, 1746 bp->b_bcount, 0, 0); 1747 BUF_KERNPROC(newindirdep->ir_savebp); 1748 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 1749 } 1750 } 1751 1752 /* 1753 * Block de-allocation dependencies. 1754 * 1755 * When blocks are de-allocated, the on-disk pointers must be nullified before 1756 * the blocks are made available for use by other files. (The true 1757 * requirement is that old pointers must be nullified before new on-disk 1758 * pointers are set. We chose this slightly more stringent requirement to 1759 * reduce complexity.) Our implementation handles this dependency by updating 1760 * the inode (or indirect block) appropriately but delaying the actual block 1761 * de-allocation (i.e., freemap and free space count manipulation) until 1762 * after the updated versions reach stable storage. After the disk is 1763 * updated, the blocks can be safely de-allocated whenever it is convenient. 1764 * This implementation handles only the common case of reducing a file's 1765 * length to zero. Other cases are handled by the conventional synchronous 1766 * write approach. 1767 * 1768 * The ffs implementation with which we worked double-checks 1769 * the state of the block pointers and file size as it reduces 1770 * a file's length. Some of this code is replicated here in our 1771 * soft updates implementation. The freeblks->fb_chkcnt field is 1772 * used to transfer a part of this information to the procedure 1773 * that eventually de-allocates the blocks. 1774 * 1775 * This routine should be called from the routine that shortens 1776 * a file's length, before the inode's size or block pointers 1777 * are modified. It will save the block pointer information for 1778 * later release and zero the inode so that the calling routine 1779 * can release it. 1780 */ 1781 struct softdep_setup_freeblocks_info { 1782 struct fs *fs; 1783 struct inode *ip; 1784 }; 1785 1786 static int softdep_setup_freeblocks_bp(struct buf *bp, void *data); 1787 1788 /* 1789 * Parameters: 1790 * ip: The inode whose length is to be reduced 1791 * length: The new length for the file 1792 */ 1793 void 1794 softdep_setup_freeblocks(struct inode *ip, off_t length) 1795 { 1796 struct softdep_setup_freeblocks_info info; 1797 struct freeblks *freeblks; 1798 struct inodedep *inodedep; 1799 struct allocdirect *adp; 1800 struct vnode *vp; 1801 struct buf *bp; 1802 struct fs *fs; 1803 int i, error, delay; 1804 int count; 1805 1806 fs = ip->i_fs; 1807 if (length != 0) 1808 panic("softde_setup_freeblocks: non-zero length"); 1809 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks), 1810 M_FREEBLKS, M_SOFTDEP_FLAGS); 1811 bzero(freeblks, sizeof(struct freeblks)); 1812 freeblks->fb_list.wk_type = D_FREEBLKS; 1813 freeblks->fb_state = ATTACHED; 1814 freeblks->fb_uid = ip->i_uid; 1815 freeblks->fb_previousinum = ip->i_number; 1816 freeblks->fb_devvp = ip->i_devvp; 1817 freeblks->fb_fs = fs; 1818 freeblks->fb_oldsize = ip->i_size; 1819 freeblks->fb_newsize = length; 1820 freeblks->fb_chkcnt = ip->i_blocks; 1821 for (i = 0; i < NDADDR; i++) { 1822 freeblks->fb_dblks[i] = ip->i_db[i]; 1823 ip->i_db[i] = 0; 1824 } 1825 for (i = 0; i < NIADDR; i++) { 1826 freeblks->fb_iblks[i] = ip->i_ib[i]; 1827 ip->i_ib[i] = 0; 1828 } 1829 ip->i_blocks = 0; 1830 ip->i_size = 0; 1831 /* 1832 * Push the zero'ed inode to to its disk buffer so that we are free 1833 * to delete its dependencies below. Once the dependencies are gone 1834 * the buffer can be safely released. 1835 */ 1836 if ((error = bread(ip->i_devvp, 1837 fsbtodoff(fs, ino_to_fsba(fs, ip->i_number)), 1838 (int)fs->fs_bsize, &bp)) != 0) 1839 softdep_error("softdep_setup_freeblocks", error); 1840 *((struct ufs1_dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) = 1841 ip->i_din; 1842 /* 1843 * Find and eliminate any inode dependencies. 1844 */ 1845 ACQUIRE_LOCK(&lk); 1846 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep); 1847 if ((inodedep->id_state & IOSTARTED) != 0) { 1848 FREE_LOCK(&lk); 1849 panic("softdep_setup_freeblocks: inode busy"); 1850 } 1851 /* 1852 * Add the freeblks structure to the list of operations that 1853 * must await the zero'ed inode being written to disk. If we 1854 * still have a bitmap dependency (delay == 0), then the inode 1855 * has never been written to disk, so we can process the 1856 * freeblks below once we have deleted the dependencies. 1857 */ 1858 delay = (inodedep->id_state & DEPCOMPLETE); 1859 if (delay) 1860 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list); 1861 /* 1862 * Because the file length has been truncated to zero, any 1863 * pending block allocation dependency structures associated 1864 * with this inode are obsolete and can simply be de-allocated. 1865 * We must first merge the two dependency lists to get rid of 1866 * any duplicate freefrag structures, then purge the merged list. 1867 */ 1868 merge_inode_lists(inodedep); 1869 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 1870 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 1871 FREE_LOCK(&lk); 1872 bdwrite(bp); 1873 /* 1874 * We must wait for any I/O in progress to finish so that 1875 * all potential buffers on the dirty list will be visible. 1876 * Once they are all there, walk the list and get rid of 1877 * any dependencies. 1878 */ 1879 vp = ITOV(ip); 1880 ACQUIRE_LOCK(&lk); 1881 drain_output(vp, 1); 1882 1883 info.fs = fs; 1884 info.ip = ip; 1885 do { 1886 count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 1887 softdep_setup_freeblocks_bp, &info); 1888 } while (count != 0); 1889 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0) 1890 (void)free_inodedep(inodedep); 1891 1892 if (delay) { 1893 freeblks->fb_state |= DEPCOMPLETE; 1894 /* 1895 * If the inode with zeroed block pointers is now on disk 1896 * we can start freeing blocks. Add freeblks to the worklist 1897 * instead of calling handle_workitem_freeblocks directly as 1898 * it is more likely that additional IO is needed to complete 1899 * the request here than in the !delay case. 1900 */ 1901 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 1902 add_to_worklist(&freeblks->fb_list); 1903 } 1904 1905 FREE_LOCK(&lk); 1906 /* 1907 * If the inode has never been written to disk (delay == 0), 1908 * then we can process the freeblks now that we have deleted 1909 * the dependencies. 1910 */ 1911 if (!delay) 1912 handle_workitem_freeblocks(freeblks); 1913 } 1914 1915 static int 1916 softdep_setup_freeblocks_bp(struct buf *bp, void *data) 1917 { 1918 struct softdep_setup_freeblocks_info *info = data; 1919 struct inodedep *inodedep; 1920 1921 if (getdirtybuf(&bp, MNT_WAIT) == 0) { 1922 kprintf("softdep_setup_freeblocks_bp(1): caught bp %p going away\n", bp); 1923 return(-1); 1924 } 1925 if (bp->b_vp != ITOV(info->ip) || (bp->b_flags & B_DELWRI) == 0) { 1926 kprintf("softdep_setup_freeblocks_bp(2): caught bp %p going away\n", bp); 1927 BUF_UNLOCK(bp); 1928 return(-1); 1929 } 1930 (void) inodedep_lookup(info->fs, info->ip->i_number, 0, &inodedep); 1931 deallocate_dependencies(bp, inodedep); 1932 bp->b_flags |= B_INVAL | B_NOCACHE; 1933 FREE_LOCK(&lk); 1934 brelse(bp); 1935 ACQUIRE_LOCK(&lk); 1936 return(1); 1937 } 1938 1939 /* 1940 * Reclaim any dependency structures from a buffer that is about to 1941 * be reallocated to a new vnode. The buffer must be locked, thus, 1942 * no I/O completion operations can occur while we are manipulating 1943 * its associated dependencies. The mutex is held so that other I/O's 1944 * associated with related dependencies do not occur. 1945 */ 1946 static void 1947 deallocate_dependencies(struct buf *bp, struct inodedep *inodedep) 1948 { 1949 struct worklist *wk; 1950 struct indirdep *indirdep; 1951 struct allocindir *aip; 1952 struct pagedep *pagedep; 1953 struct dirrem *dirrem; 1954 struct diradd *dap; 1955 int i; 1956 1957 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 1958 switch (wk->wk_type) { 1959 1960 case D_INDIRDEP: 1961 indirdep = WK_INDIRDEP(wk); 1962 /* 1963 * None of the indirect pointers will ever be visible, 1964 * so they can simply be tossed. GOINGAWAY ensures 1965 * that allocated pointers will be saved in the buffer 1966 * cache until they are freed. Note that they will 1967 * only be able to be found by their physical address 1968 * since the inode mapping the logical address will 1969 * be gone. The save buffer used for the safe copy 1970 * was allocated in setup_allocindir_phase2 using 1971 * the physical address so it could be used for this 1972 * purpose. Hence we swap the safe copy with the real 1973 * copy, allowing the safe copy to be freed and holding 1974 * on to the real copy for later use in indir_trunc. 1975 * 1976 * NOTE: ir_savebp is relative to the block device 1977 * so b_bio1 contains the device block number. 1978 */ 1979 if (indirdep->ir_state & GOINGAWAY) { 1980 FREE_LOCK(&lk); 1981 panic("deallocate_dependencies: already gone"); 1982 } 1983 indirdep->ir_state |= GOINGAWAY; 1984 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 1985 free_allocindir(aip, inodedep); 1986 if (bp->b_bio1.bio_offset >= 0 || 1987 bp->b_bio2.bio_offset != indirdep->ir_savebp->b_bio1.bio_offset) { 1988 FREE_LOCK(&lk); 1989 panic("deallocate_dependencies: not indir"); 1990 } 1991 bcopy(bp->b_data, indirdep->ir_savebp->b_data, 1992 bp->b_bcount); 1993 WORKLIST_REMOVE(wk); 1994 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk); 1995 continue; 1996 1997 case D_PAGEDEP: 1998 pagedep = WK_PAGEDEP(wk); 1999 /* 2000 * None of the directory additions will ever be 2001 * visible, so they can simply be tossed. 2002 */ 2003 for (i = 0; i < DAHASHSZ; i++) 2004 while ((dap = 2005 LIST_FIRST(&pagedep->pd_diraddhd[i]))) 2006 free_diradd(dap); 2007 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0) 2008 free_diradd(dap); 2009 /* 2010 * Copy any directory remove dependencies to the list 2011 * to be processed after the zero'ed inode is written. 2012 * If the inode has already been written, then they 2013 * can be dumped directly onto the work list. 2014 */ 2015 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 2016 LIST_REMOVE(dirrem, dm_next); 2017 dirrem->dm_dirinum = pagedep->pd_ino; 2018 if (inodedep == NULL || 2019 (inodedep->id_state & ALLCOMPLETE) == 2020 ALLCOMPLETE) 2021 add_to_worklist(&dirrem->dm_list); 2022 else 2023 WORKLIST_INSERT(&inodedep->id_bufwait, 2024 &dirrem->dm_list); 2025 } 2026 WORKLIST_REMOVE(&pagedep->pd_list); 2027 LIST_REMOVE(pagedep, pd_hash); 2028 WORKITEM_FREE(pagedep, D_PAGEDEP); 2029 continue; 2030 2031 case D_ALLOCINDIR: 2032 free_allocindir(WK_ALLOCINDIR(wk), inodedep); 2033 continue; 2034 2035 case D_ALLOCDIRECT: 2036 case D_INODEDEP: 2037 FREE_LOCK(&lk); 2038 panic("deallocate_dependencies: Unexpected type %s", 2039 TYPENAME(wk->wk_type)); 2040 /* NOTREACHED */ 2041 2042 default: 2043 FREE_LOCK(&lk); 2044 panic("deallocate_dependencies: Unknown type %s", 2045 TYPENAME(wk->wk_type)); 2046 /* NOTREACHED */ 2047 } 2048 } 2049 } 2050 2051 /* 2052 * Free an allocdirect. Generate a new freefrag work request if appropriate. 2053 * This routine must be called with splbio interrupts blocked. 2054 */ 2055 static void 2056 free_allocdirect(struct allocdirectlst *adphead, 2057 struct allocdirect *adp, int delay) 2058 { 2059 2060 #ifdef DEBUG 2061 if (lk.lkt_held == NOHOLDER) 2062 panic("free_allocdirect: lock not held"); 2063 #endif 2064 if ((adp->ad_state & DEPCOMPLETE) == 0) 2065 LIST_REMOVE(adp, ad_deps); 2066 TAILQ_REMOVE(adphead, adp, ad_next); 2067 if ((adp->ad_state & COMPLETE) == 0) 2068 WORKLIST_REMOVE(&adp->ad_list); 2069 if (adp->ad_freefrag != NULL) { 2070 if (delay) 2071 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 2072 &adp->ad_freefrag->ff_list); 2073 else 2074 add_to_worklist(&adp->ad_freefrag->ff_list); 2075 } 2076 WORKITEM_FREE(adp, D_ALLOCDIRECT); 2077 } 2078 2079 /* 2080 * Prepare an inode to be freed. The actual free operation is not 2081 * done until the zero'ed inode has been written to disk. 2082 */ 2083 void 2084 softdep_freefile(struct vnode *pvp, ino_t ino, int mode) 2085 { 2086 struct inode *ip = VTOI(pvp); 2087 struct inodedep *inodedep; 2088 struct freefile *freefile; 2089 2090 /* 2091 * This sets up the inode de-allocation dependency. 2092 */ 2093 MALLOC(freefile, struct freefile *, sizeof(struct freefile), 2094 M_FREEFILE, M_SOFTDEP_FLAGS); 2095 freefile->fx_list.wk_type = D_FREEFILE; 2096 freefile->fx_list.wk_state = 0; 2097 freefile->fx_mode = mode; 2098 freefile->fx_oldinum = ino; 2099 freefile->fx_devvp = ip->i_devvp; 2100 freefile->fx_fs = ip->i_fs; 2101 2102 /* 2103 * If the inodedep does not exist, then the zero'ed inode has 2104 * been written to disk. If the allocated inode has never been 2105 * written to disk, then the on-disk inode is zero'ed. In either 2106 * case we can free the file immediately. 2107 */ 2108 ACQUIRE_LOCK(&lk); 2109 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 || 2110 check_inode_unwritten(inodedep)) { 2111 FREE_LOCK(&lk); 2112 handle_workitem_freefile(freefile); 2113 return; 2114 } 2115 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 2116 FREE_LOCK(&lk); 2117 } 2118 2119 /* 2120 * Check to see if an inode has never been written to disk. If 2121 * so free the inodedep and return success, otherwise return failure. 2122 * This routine must be called with splbio interrupts blocked. 2123 * 2124 * If we still have a bitmap dependency, then the inode has never 2125 * been written to disk. Drop the dependency as it is no longer 2126 * necessary since the inode is being deallocated. We set the 2127 * ALLCOMPLETE flags since the bitmap now properly shows that the 2128 * inode is not allocated. Even if the inode is actively being 2129 * written, it has been rolled back to its zero'ed state, so we 2130 * are ensured that a zero inode is what is on the disk. For short 2131 * lived files, this change will usually result in removing all the 2132 * dependencies from the inode so that it can be freed immediately. 2133 */ 2134 static int 2135 check_inode_unwritten(struct inodedep *inodedep) 2136 { 2137 2138 if ((inodedep->id_state & DEPCOMPLETE) != 0 || 2139 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2140 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2141 LIST_FIRST(&inodedep->id_inowait) != NULL || 2142 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2143 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2144 inodedep->id_nlinkdelta != 0) 2145 return (0); 2146 2147 /* 2148 * Another process might be in initiate_write_inodeblock 2149 * trying to allocate memory without holding "Softdep Lock". 2150 */ 2151 if ((inodedep->id_state & IOSTARTED) != 0 && 2152 inodedep->id_savedino == NULL) 2153 return(0); 2154 2155 inodedep->id_state |= ALLCOMPLETE; 2156 LIST_REMOVE(inodedep, id_deps); 2157 inodedep->id_buf = NULL; 2158 if (inodedep->id_state & ONWORKLIST) 2159 WORKLIST_REMOVE(&inodedep->id_list); 2160 if (inodedep->id_savedino != NULL) { 2161 FREE(inodedep->id_savedino, M_INODEDEP); 2162 inodedep->id_savedino = NULL; 2163 } 2164 if (free_inodedep(inodedep) == 0) { 2165 FREE_LOCK(&lk); 2166 panic("check_inode_unwritten: busy inode"); 2167 } 2168 return (1); 2169 } 2170 2171 /* 2172 * Try to free an inodedep structure. Return 1 if it could be freed. 2173 */ 2174 static int 2175 free_inodedep(struct inodedep *inodedep) 2176 { 2177 2178 if ((inodedep->id_state & ONWORKLIST) != 0 || 2179 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 2180 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2181 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2182 LIST_FIRST(&inodedep->id_inowait) != NULL || 2183 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2184 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2185 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL) 2186 return (0); 2187 LIST_REMOVE(inodedep, id_hash); 2188 WORKITEM_FREE(inodedep, D_INODEDEP); 2189 num_inodedep -= 1; 2190 return (1); 2191 } 2192 2193 /* 2194 * This workitem routine performs the block de-allocation. 2195 * The workitem is added to the pending list after the updated 2196 * inode block has been written to disk. As mentioned above, 2197 * checks regarding the number of blocks de-allocated (compared 2198 * to the number of blocks allocated for the file) are also 2199 * performed in this function. 2200 */ 2201 static void 2202 handle_workitem_freeblocks(struct freeblks *freeblks) 2203 { 2204 struct inode tip; 2205 ufs_daddr_t bn; 2206 struct fs *fs; 2207 int i, level, bsize; 2208 long nblocks, blocksreleased = 0; 2209 int error, allerror = 0; 2210 ufs_lbn_t baselbns[NIADDR], tmpval; 2211 2212 tip.i_number = freeblks->fb_previousinum; 2213 tip.i_devvp = freeblks->fb_devvp; 2214 tip.i_dev = freeblks->fb_devvp->v_rdev; 2215 tip.i_fs = freeblks->fb_fs; 2216 tip.i_size = freeblks->fb_oldsize; 2217 tip.i_uid = freeblks->fb_uid; 2218 fs = freeblks->fb_fs; 2219 tmpval = 1; 2220 baselbns[0] = NDADDR; 2221 for (i = 1; i < NIADDR; i++) { 2222 tmpval *= NINDIR(fs); 2223 baselbns[i] = baselbns[i - 1] + tmpval; 2224 } 2225 nblocks = btodb(fs->fs_bsize); 2226 blocksreleased = 0; 2227 /* 2228 * Indirect blocks first. 2229 */ 2230 for (level = (NIADDR - 1); level >= 0; level--) { 2231 if ((bn = freeblks->fb_iblks[level]) == 0) 2232 continue; 2233 if ((error = indir_trunc(&tip, fsbtodoff(fs, bn), level, 2234 baselbns[level], &blocksreleased)) == 0) 2235 allerror = error; 2236 ffs_blkfree(&tip, bn, fs->fs_bsize); 2237 blocksreleased += nblocks; 2238 } 2239 /* 2240 * All direct blocks or frags. 2241 */ 2242 for (i = (NDADDR - 1); i >= 0; i--) { 2243 if ((bn = freeblks->fb_dblks[i]) == 0) 2244 continue; 2245 bsize = blksize(fs, &tip, i); 2246 ffs_blkfree(&tip, bn, bsize); 2247 blocksreleased += btodb(bsize); 2248 } 2249 2250 #ifdef DIAGNOSTIC 2251 if (freeblks->fb_chkcnt != blocksreleased) 2252 kprintf("handle_workitem_freeblocks: block count\n"); 2253 if (allerror) 2254 softdep_error("handle_workitem_freeblks", allerror); 2255 #endif /* DIAGNOSTIC */ 2256 WORKITEM_FREE(freeblks, D_FREEBLKS); 2257 } 2258 2259 /* 2260 * Release blocks associated with the inode ip and stored in the indirect 2261 * block at doffset. If level is greater than SINGLE, the block is an 2262 * indirect block and recursive calls to indirtrunc must be used to 2263 * cleanse other indirect blocks. 2264 */ 2265 static int 2266 indir_trunc(struct inode *ip, off_t doffset, int level, ufs_lbn_t lbn, 2267 long *countp) 2268 { 2269 struct buf *bp; 2270 ufs_daddr_t *bap; 2271 ufs_daddr_t nb; 2272 struct fs *fs; 2273 struct worklist *wk; 2274 struct indirdep *indirdep; 2275 int i, lbnadd, nblocks; 2276 int error, allerror = 0; 2277 2278 fs = ip->i_fs; 2279 lbnadd = 1; 2280 for (i = level; i > 0; i--) 2281 lbnadd *= NINDIR(fs); 2282 /* 2283 * Get buffer of block pointers to be freed. This routine is not 2284 * called until the zero'ed inode has been written, so it is safe 2285 * to free blocks as they are encountered. Because the inode has 2286 * been zero'ed, calls to bmap on these blocks will fail. So, we 2287 * have to use the on-disk address and the block device for the 2288 * filesystem to look them up. If the file was deleted before its 2289 * indirect blocks were all written to disk, the routine that set 2290 * us up (deallocate_dependencies) will have arranged to leave 2291 * a complete copy of the indirect block in memory for our use. 2292 * Otherwise we have to read the blocks in from the disk. 2293 */ 2294 ACQUIRE_LOCK(&lk); 2295 if ((bp = findblk(ip->i_devvp, doffset)) != NULL && 2296 (wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2297 /* 2298 * bp must be ir_savebp, which is held locked for our use. 2299 */ 2300 if (wk->wk_type != D_INDIRDEP || 2301 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp || 2302 (indirdep->ir_state & GOINGAWAY) == 0) { 2303 FREE_LOCK(&lk); 2304 panic("indir_trunc: lost indirdep"); 2305 } 2306 WORKLIST_REMOVE(wk); 2307 WORKITEM_FREE(indirdep, D_INDIRDEP); 2308 if (LIST_FIRST(&bp->b_dep) != NULL) { 2309 FREE_LOCK(&lk); 2310 panic("indir_trunc: dangling dep"); 2311 } 2312 FREE_LOCK(&lk); 2313 } else { 2314 FREE_LOCK(&lk); 2315 error = bread(ip->i_devvp, doffset, (int)fs->fs_bsize, &bp); 2316 if (error) 2317 return (error); 2318 } 2319 /* 2320 * Recursively free indirect blocks. 2321 */ 2322 bap = (ufs_daddr_t *)bp->b_data; 2323 nblocks = btodb(fs->fs_bsize); 2324 for (i = NINDIR(fs) - 1; i >= 0; i--) { 2325 if ((nb = bap[i]) == 0) 2326 continue; 2327 if (level != 0) { 2328 if ((error = indir_trunc(ip, fsbtodoff(fs, nb), 2329 level - 1, lbn + (i * lbnadd), countp)) != 0) 2330 allerror = error; 2331 } 2332 ffs_blkfree(ip, nb, fs->fs_bsize); 2333 *countp += nblocks; 2334 } 2335 bp->b_flags |= B_INVAL | B_NOCACHE; 2336 brelse(bp); 2337 return (allerror); 2338 } 2339 2340 /* 2341 * Free an allocindir. 2342 * This routine must be called with splbio interrupts blocked. 2343 */ 2344 static void 2345 free_allocindir(struct allocindir *aip, struct inodedep *inodedep) 2346 { 2347 struct freefrag *freefrag; 2348 2349 #ifdef DEBUG 2350 if (lk.lkt_held == NOHOLDER) 2351 panic("free_allocindir: lock not held"); 2352 #endif 2353 if ((aip->ai_state & DEPCOMPLETE) == 0) 2354 LIST_REMOVE(aip, ai_deps); 2355 if (aip->ai_state & ONWORKLIST) 2356 WORKLIST_REMOVE(&aip->ai_list); 2357 LIST_REMOVE(aip, ai_next); 2358 if ((freefrag = aip->ai_freefrag) != NULL) { 2359 if (inodedep == NULL) 2360 add_to_worklist(&freefrag->ff_list); 2361 else 2362 WORKLIST_INSERT(&inodedep->id_bufwait, 2363 &freefrag->ff_list); 2364 } 2365 WORKITEM_FREE(aip, D_ALLOCINDIR); 2366 } 2367 2368 /* 2369 * Directory entry addition dependencies. 2370 * 2371 * When adding a new directory entry, the inode (with its incremented link 2372 * count) must be written to disk before the directory entry's pointer to it. 2373 * Also, if the inode is newly allocated, the corresponding freemap must be 2374 * updated (on disk) before the directory entry's pointer. These requirements 2375 * are met via undo/redo on the directory entry's pointer, which consists 2376 * simply of the inode number. 2377 * 2378 * As directory entries are added and deleted, the free space within a 2379 * directory block can become fragmented. The ufs filesystem will compact 2380 * a fragmented directory block to make space for a new entry. When this 2381 * occurs, the offsets of previously added entries change. Any "diradd" 2382 * dependency structures corresponding to these entries must be updated with 2383 * the new offsets. 2384 */ 2385 2386 /* 2387 * This routine is called after the in-memory inode's link 2388 * count has been incremented, but before the directory entry's 2389 * pointer to the inode has been set. 2390 * 2391 * Parameters: 2392 * bp: buffer containing directory block 2393 * dp: inode for directory 2394 * diroffset: offset of new entry in directory 2395 * newinum: inode referenced by new directory entry 2396 * newdirbp: non-NULL => contents of new mkdir 2397 */ 2398 void 2399 softdep_setup_directory_add(struct buf *bp, struct inode *dp, off_t diroffset, 2400 ino_t newinum, struct buf *newdirbp) 2401 { 2402 int offset; /* offset of new entry within directory block */ 2403 ufs_lbn_t lbn; /* block in directory containing new entry */ 2404 struct fs *fs; 2405 struct diradd *dap; 2406 struct pagedep *pagedep; 2407 struct inodedep *inodedep; 2408 struct mkdir *mkdir1, *mkdir2; 2409 2410 /* 2411 * Whiteouts have no dependencies. 2412 */ 2413 if (newinum == WINO) { 2414 if (newdirbp != NULL) 2415 bdwrite(newdirbp); 2416 return; 2417 } 2418 2419 fs = dp->i_fs; 2420 lbn = lblkno(fs, diroffset); 2421 offset = blkoff(fs, diroffset); 2422 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, 2423 M_SOFTDEP_FLAGS); 2424 bzero(dap, sizeof(struct diradd)); 2425 dap->da_list.wk_type = D_DIRADD; 2426 dap->da_offset = offset; 2427 dap->da_newinum = newinum; 2428 dap->da_state = ATTACHED; 2429 if (newdirbp == NULL) { 2430 dap->da_state |= DEPCOMPLETE; 2431 ACQUIRE_LOCK(&lk); 2432 } else { 2433 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 2434 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2435 M_SOFTDEP_FLAGS); 2436 mkdir1->md_list.wk_type = D_MKDIR; 2437 mkdir1->md_state = MKDIR_BODY; 2438 mkdir1->md_diradd = dap; 2439 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2440 M_SOFTDEP_FLAGS); 2441 mkdir2->md_list.wk_type = D_MKDIR; 2442 mkdir2->md_state = MKDIR_PARENT; 2443 mkdir2->md_diradd = dap; 2444 /* 2445 * Dependency on "." and ".." being written to disk. 2446 */ 2447 mkdir1->md_buf = newdirbp; 2448 ACQUIRE_LOCK(&lk); 2449 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 2450 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list); 2451 FREE_LOCK(&lk); 2452 bdwrite(newdirbp); 2453 /* 2454 * Dependency on link count increase for parent directory 2455 */ 2456 ACQUIRE_LOCK(&lk); 2457 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0 2458 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2459 dap->da_state &= ~MKDIR_PARENT; 2460 WORKITEM_FREE(mkdir2, D_MKDIR); 2461 } else { 2462 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 2463 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 2464 } 2465 } 2466 /* 2467 * Link into parent directory pagedep to await its being written. 2468 */ 2469 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2470 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2471 dap->da_pagedep = pagedep; 2472 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 2473 da_pdlist); 2474 /* 2475 * Link into its inodedep. Put it on the id_bufwait list if the inode 2476 * is not yet written. If it is written, do the post-inode write 2477 * processing to put it on the id_pendinghd list. 2478 */ 2479 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep); 2480 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 2481 diradd_inode_written(dap, inodedep); 2482 else 2483 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2484 FREE_LOCK(&lk); 2485 } 2486 2487 /* 2488 * This procedure is called to change the offset of a directory 2489 * entry when compacting a directory block which must be owned 2490 * exclusively by the caller. Note that the actual entry movement 2491 * must be done in this procedure to ensure that no I/O completions 2492 * occur while the move is in progress. 2493 * 2494 * Parameters: 2495 * dp: inode for directory 2496 * base: address of dp->i_offset 2497 * oldloc: address of old directory location 2498 * newloc: address of new directory location 2499 * entrysize: size of directory entry 2500 */ 2501 void 2502 softdep_change_directoryentry_offset(struct inode *dp, caddr_t base, 2503 caddr_t oldloc, caddr_t newloc, 2504 int entrysize) 2505 { 2506 int offset, oldoffset, newoffset; 2507 struct pagedep *pagedep; 2508 struct diradd *dap; 2509 ufs_lbn_t lbn; 2510 2511 ACQUIRE_LOCK(&lk); 2512 lbn = lblkno(dp->i_fs, dp->i_offset); 2513 offset = blkoff(dp->i_fs, dp->i_offset); 2514 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) 2515 goto done; 2516 oldoffset = offset + (oldloc - base); 2517 newoffset = offset + (newloc - base); 2518 2519 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) { 2520 if (dap->da_offset != oldoffset) 2521 continue; 2522 dap->da_offset = newoffset; 2523 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) 2524 break; 2525 LIST_REMOVE(dap, da_pdlist); 2526 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], 2527 dap, da_pdlist); 2528 break; 2529 } 2530 if (dap == NULL) { 2531 2532 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) { 2533 if (dap->da_offset == oldoffset) { 2534 dap->da_offset = newoffset; 2535 break; 2536 } 2537 } 2538 } 2539 done: 2540 bcopy(oldloc, newloc, entrysize); 2541 FREE_LOCK(&lk); 2542 } 2543 2544 /* 2545 * Free a diradd dependency structure. This routine must be called 2546 * with splbio interrupts blocked. 2547 */ 2548 static void 2549 free_diradd(struct diradd *dap) 2550 { 2551 struct dirrem *dirrem; 2552 struct pagedep *pagedep; 2553 struct inodedep *inodedep; 2554 struct mkdir *mkdir, *nextmd; 2555 2556 #ifdef DEBUG 2557 if (lk.lkt_held == NOHOLDER) 2558 panic("free_diradd: lock not held"); 2559 #endif 2560 WORKLIST_REMOVE(&dap->da_list); 2561 LIST_REMOVE(dap, da_pdlist); 2562 if ((dap->da_state & DIRCHG) == 0) { 2563 pagedep = dap->da_pagedep; 2564 } else { 2565 dirrem = dap->da_previous; 2566 pagedep = dirrem->dm_pagedep; 2567 dirrem->dm_dirinum = pagedep->pd_ino; 2568 add_to_worklist(&dirrem->dm_list); 2569 } 2570 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum, 2571 0, &inodedep) != 0) 2572 (void) free_inodedep(inodedep); 2573 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2574 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 2575 nextmd = LIST_NEXT(mkdir, md_mkdirs); 2576 if (mkdir->md_diradd != dap) 2577 continue; 2578 dap->da_state &= ~mkdir->md_state; 2579 WORKLIST_REMOVE(&mkdir->md_list); 2580 LIST_REMOVE(mkdir, md_mkdirs); 2581 WORKITEM_FREE(mkdir, D_MKDIR); 2582 } 2583 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2584 FREE_LOCK(&lk); 2585 panic("free_diradd: unfound ref"); 2586 } 2587 } 2588 WORKITEM_FREE(dap, D_DIRADD); 2589 } 2590 2591 /* 2592 * Directory entry removal dependencies. 2593 * 2594 * When removing a directory entry, the entry's inode pointer must be 2595 * zero'ed on disk before the corresponding inode's link count is decremented 2596 * (possibly freeing the inode for re-use). This dependency is handled by 2597 * updating the directory entry but delaying the inode count reduction until 2598 * after the directory block has been written to disk. After this point, the 2599 * inode count can be decremented whenever it is convenient. 2600 */ 2601 2602 /* 2603 * This routine should be called immediately after removing 2604 * a directory entry. The inode's link count should not be 2605 * decremented by the calling procedure -- the soft updates 2606 * code will do this task when it is safe. 2607 * 2608 * Parameters: 2609 * bp: buffer containing directory block 2610 * dp: inode for the directory being modified 2611 * ip: inode for directory entry being removed 2612 * isrmdir: indicates if doing RMDIR 2613 */ 2614 void 2615 softdep_setup_remove(struct buf *bp, struct inode *dp, struct inode *ip, 2616 int isrmdir) 2617 { 2618 struct dirrem *dirrem, *prevdirrem; 2619 2620 /* 2621 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 2622 */ 2623 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2624 2625 /* 2626 * If the COMPLETE flag is clear, then there were no active 2627 * entries and we want to roll back to a zeroed entry until 2628 * the new inode is committed to disk. If the COMPLETE flag is 2629 * set then we have deleted an entry that never made it to 2630 * disk. If the entry we deleted resulted from a name change, 2631 * then the old name still resides on disk. We cannot delete 2632 * its inode (returned to us in prevdirrem) until the zeroed 2633 * directory entry gets to disk. The new inode has never been 2634 * referenced on the disk, so can be deleted immediately. 2635 */ 2636 if ((dirrem->dm_state & COMPLETE) == 0) { 2637 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 2638 dm_next); 2639 FREE_LOCK(&lk); 2640 } else { 2641 if (prevdirrem != NULL) 2642 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 2643 prevdirrem, dm_next); 2644 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 2645 FREE_LOCK(&lk); 2646 handle_workitem_remove(dirrem); 2647 } 2648 } 2649 2650 /* 2651 * Allocate a new dirrem if appropriate and return it along with 2652 * its associated pagedep. Called without a lock, returns with lock. 2653 */ 2654 static long num_dirrem; /* number of dirrem allocated */ 2655 2656 /* 2657 * Parameters: 2658 * bp: buffer containing directory block 2659 * dp: inode for the directory being modified 2660 * ip: inode for directory entry being removed 2661 * isrmdir: indicates if doing RMDIR 2662 * prevdirremp: previously referenced inode, if any 2663 */ 2664 static struct dirrem * 2665 newdirrem(struct buf *bp, struct inode *dp, struct inode *ip, 2666 int isrmdir, struct dirrem **prevdirremp) 2667 { 2668 int offset; 2669 ufs_lbn_t lbn; 2670 struct diradd *dap; 2671 struct dirrem *dirrem; 2672 struct pagedep *pagedep; 2673 2674 /* 2675 * Whiteouts have no deletion dependencies. 2676 */ 2677 if (ip == NULL) 2678 panic("newdirrem: whiteout"); 2679 /* 2680 * If we are over our limit, try to improve the situation. 2681 * Limiting the number of dirrem structures will also limit 2682 * the number of freefile and freeblks structures. 2683 */ 2684 if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0) 2685 (void) request_cleanup(FLUSH_REMOVE, 0); 2686 num_dirrem += 1; 2687 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem), 2688 M_DIRREM, M_SOFTDEP_FLAGS); 2689 bzero(dirrem, sizeof(struct dirrem)); 2690 dirrem->dm_list.wk_type = D_DIRREM; 2691 dirrem->dm_state = isrmdir ? RMDIR : 0; 2692 dirrem->dm_mnt = ITOV(ip)->v_mount; 2693 dirrem->dm_oldinum = ip->i_number; 2694 *prevdirremp = NULL; 2695 2696 ACQUIRE_LOCK(&lk); 2697 lbn = lblkno(dp->i_fs, dp->i_offset); 2698 offset = blkoff(dp->i_fs, dp->i_offset); 2699 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2700 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2701 dirrem->dm_pagedep = pagedep; 2702 /* 2703 * Check for a diradd dependency for the same directory entry. 2704 * If present, then both dependencies become obsolete and can 2705 * be de-allocated. Check for an entry on both the pd_dirraddhd 2706 * list and the pd_pendinghd list. 2707 */ 2708 2709 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 2710 if (dap->da_offset == offset) 2711 break; 2712 if (dap == NULL) { 2713 2714 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 2715 if (dap->da_offset == offset) 2716 break; 2717 if (dap == NULL) 2718 return (dirrem); 2719 } 2720 /* 2721 * Must be ATTACHED at this point. 2722 */ 2723 if ((dap->da_state & ATTACHED) == 0) { 2724 FREE_LOCK(&lk); 2725 panic("newdirrem: not ATTACHED"); 2726 } 2727 if (dap->da_newinum != ip->i_number) { 2728 FREE_LOCK(&lk); 2729 panic("newdirrem: inum %"PRId64" should be %"PRId64, 2730 ip->i_number, dap->da_newinum); 2731 } 2732 /* 2733 * If we are deleting a changed name that never made it to disk, 2734 * then return the dirrem describing the previous inode (which 2735 * represents the inode currently referenced from this entry on disk). 2736 */ 2737 if ((dap->da_state & DIRCHG) != 0) { 2738 *prevdirremp = dap->da_previous; 2739 dap->da_state &= ~DIRCHG; 2740 dap->da_pagedep = pagedep; 2741 } 2742 /* 2743 * We are deleting an entry that never made it to disk. 2744 * Mark it COMPLETE so we can delete its inode immediately. 2745 */ 2746 dirrem->dm_state |= COMPLETE; 2747 free_diradd(dap); 2748 return (dirrem); 2749 } 2750 2751 /* 2752 * Directory entry change dependencies. 2753 * 2754 * Changing an existing directory entry requires that an add operation 2755 * be completed first followed by a deletion. The semantics for the addition 2756 * are identical to the description of adding a new entry above except 2757 * that the rollback is to the old inode number rather than zero. Once 2758 * the addition dependency is completed, the removal is done as described 2759 * in the removal routine above. 2760 */ 2761 2762 /* 2763 * This routine should be called immediately after changing 2764 * a directory entry. The inode's link count should not be 2765 * decremented by the calling procedure -- the soft updates 2766 * code will perform this task when it is safe. 2767 * 2768 * Parameters: 2769 * bp: buffer containing directory block 2770 * dp: inode for the directory being modified 2771 * ip: inode for directory entry being removed 2772 * newinum: new inode number for changed entry 2773 * isrmdir: indicates if doing RMDIR 2774 */ 2775 void 2776 softdep_setup_directory_change(struct buf *bp, struct inode *dp, 2777 struct inode *ip, ino_t newinum, 2778 int isrmdir) 2779 { 2780 int offset; 2781 struct diradd *dap = NULL; 2782 struct dirrem *dirrem, *prevdirrem; 2783 struct pagedep *pagedep; 2784 struct inodedep *inodedep; 2785 2786 offset = blkoff(dp->i_fs, dp->i_offset); 2787 2788 /* 2789 * Whiteouts do not need diradd dependencies. 2790 */ 2791 if (newinum != WINO) { 2792 MALLOC(dap, struct diradd *, sizeof(struct diradd), 2793 M_DIRADD, M_SOFTDEP_FLAGS); 2794 bzero(dap, sizeof(struct diradd)); 2795 dap->da_list.wk_type = D_DIRADD; 2796 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 2797 dap->da_offset = offset; 2798 dap->da_newinum = newinum; 2799 } 2800 2801 /* 2802 * Allocate a new dirrem and ACQUIRE_LOCK. 2803 */ 2804 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2805 pagedep = dirrem->dm_pagedep; 2806 /* 2807 * The possible values for isrmdir: 2808 * 0 - non-directory file rename 2809 * 1 - directory rename within same directory 2810 * inum - directory rename to new directory of given inode number 2811 * When renaming to a new directory, we are both deleting and 2812 * creating a new directory entry, so the link count on the new 2813 * directory should not change. Thus we do not need the followup 2814 * dirrem which is usually done in handle_workitem_remove. We set 2815 * the DIRCHG flag to tell handle_workitem_remove to skip the 2816 * followup dirrem. 2817 */ 2818 if (isrmdir > 1) 2819 dirrem->dm_state |= DIRCHG; 2820 2821 /* 2822 * Whiteouts have no additional dependencies, 2823 * so just put the dirrem on the correct list. 2824 */ 2825 if (newinum == WINO) { 2826 if ((dirrem->dm_state & COMPLETE) == 0) { 2827 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 2828 dm_next); 2829 } else { 2830 dirrem->dm_dirinum = pagedep->pd_ino; 2831 add_to_worklist(&dirrem->dm_list); 2832 } 2833 FREE_LOCK(&lk); 2834 return; 2835 } 2836 2837 /* 2838 * If the COMPLETE flag is clear, then there were no active 2839 * entries and we want to roll back to the previous inode until 2840 * the new inode is committed to disk. If the COMPLETE flag is 2841 * set, then we have deleted an entry that never made it to disk. 2842 * If the entry we deleted resulted from a name change, then the old 2843 * inode reference still resides on disk. Any rollback that we do 2844 * needs to be to that old inode (returned to us in prevdirrem). If 2845 * the entry we deleted resulted from a create, then there is 2846 * no entry on the disk, so we want to roll back to zero rather 2847 * than the uncommitted inode. In either of the COMPLETE cases we 2848 * want to immediately free the unwritten and unreferenced inode. 2849 */ 2850 if ((dirrem->dm_state & COMPLETE) == 0) { 2851 dap->da_previous = dirrem; 2852 } else { 2853 if (prevdirrem != NULL) { 2854 dap->da_previous = prevdirrem; 2855 } else { 2856 dap->da_state &= ~DIRCHG; 2857 dap->da_pagedep = pagedep; 2858 } 2859 dirrem->dm_dirinum = pagedep->pd_ino; 2860 add_to_worklist(&dirrem->dm_list); 2861 } 2862 /* 2863 * Link into its inodedep. Put it on the id_bufwait list if the inode 2864 * is not yet written. If it is written, do the post-inode write 2865 * processing to put it on the id_pendinghd list. 2866 */ 2867 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 || 2868 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2869 dap->da_state |= COMPLETE; 2870 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 2871 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 2872 } else { 2873 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 2874 dap, da_pdlist); 2875 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2876 } 2877 FREE_LOCK(&lk); 2878 } 2879 2880 /* 2881 * Called whenever the link count on an inode is changed. 2882 * It creates an inode dependency so that the new reference(s) 2883 * to the inode cannot be committed to disk until the updated 2884 * inode has been written. 2885 * 2886 * Parameters: 2887 * ip: the inode with the increased link count 2888 */ 2889 void 2890 softdep_change_linkcnt(struct inode *ip) 2891 { 2892 struct inodedep *inodedep; 2893 2894 ACQUIRE_LOCK(&lk); 2895 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 2896 if (ip->i_nlink < ip->i_effnlink) { 2897 FREE_LOCK(&lk); 2898 panic("softdep_change_linkcnt: bad delta"); 2899 } 2900 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2901 FREE_LOCK(&lk); 2902 } 2903 2904 /* 2905 * This workitem decrements the inode's link count. 2906 * If the link count reaches zero, the file is removed. 2907 */ 2908 static void 2909 handle_workitem_remove(struct dirrem *dirrem) 2910 { 2911 struct inodedep *inodedep; 2912 struct vnode *vp; 2913 struct inode *ip; 2914 ino_t oldinum; 2915 int error; 2916 2917 if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) { 2918 softdep_error("handle_workitem_remove: vget", error); 2919 return; 2920 } 2921 ip = VTOI(vp); 2922 ACQUIRE_LOCK(&lk); 2923 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){ 2924 FREE_LOCK(&lk); 2925 panic("handle_workitem_remove: lost inodedep"); 2926 } 2927 /* 2928 * Normal file deletion. 2929 */ 2930 if ((dirrem->dm_state & RMDIR) == 0) { 2931 ip->i_nlink--; 2932 ip->i_flag |= IN_CHANGE; 2933 if (ip->i_nlink < ip->i_effnlink) { 2934 FREE_LOCK(&lk); 2935 panic("handle_workitem_remove: bad file delta"); 2936 } 2937 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2938 FREE_LOCK(&lk); 2939 vput(vp); 2940 num_dirrem -= 1; 2941 WORKITEM_FREE(dirrem, D_DIRREM); 2942 return; 2943 } 2944 /* 2945 * Directory deletion. Decrement reference count for both the 2946 * just deleted parent directory entry and the reference for ".". 2947 * Next truncate the directory to length zero. When the 2948 * truncation completes, arrange to have the reference count on 2949 * the parent decremented to account for the loss of "..". 2950 */ 2951 ip->i_nlink -= 2; 2952 ip->i_flag |= IN_CHANGE; 2953 if (ip->i_nlink < ip->i_effnlink) { 2954 FREE_LOCK(&lk); 2955 panic("handle_workitem_remove: bad dir delta"); 2956 } 2957 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2958 FREE_LOCK(&lk); 2959 if ((error = ffs_truncate(vp, (off_t)0, 0, proc0.p_ucred)) != 0) 2960 softdep_error("handle_workitem_remove: truncate", error); 2961 /* 2962 * Rename a directory to a new parent. Since, we are both deleting 2963 * and creating a new directory entry, the link count on the new 2964 * directory should not change. Thus we skip the followup dirrem. 2965 */ 2966 if (dirrem->dm_state & DIRCHG) { 2967 vput(vp); 2968 num_dirrem -= 1; 2969 WORKITEM_FREE(dirrem, D_DIRREM); 2970 return; 2971 } 2972 /* 2973 * If the inodedep does not exist, then the zero'ed inode has 2974 * been written to disk. If the allocated inode has never been 2975 * written to disk, then the on-disk inode is zero'ed. In either 2976 * case we can remove the file immediately. 2977 */ 2978 ACQUIRE_LOCK(&lk); 2979 dirrem->dm_state = 0; 2980 oldinum = dirrem->dm_oldinum; 2981 dirrem->dm_oldinum = dirrem->dm_dirinum; 2982 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 || 2983 check_inode_unwritten(inodedep)) { 2984 FREE_LOCK(&lk); 2985 vput(vp); 2986 handle_workitem_remove(dirrem); 2987 return; 2988 } 2989 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 2990 FREE_LOCK(&lk); 2991 ip->i_flag |= IN_CHANGE; 2992 ffs_update(vp, 0); 2993 vput(vp); 2994 } 2995 2996 /* 2997 * Inode de-allocation dependencies. 2998 * 2999 * When an inode's link count is reduced to zero, it can be de-allocated. We 3000 * found it convenient to postpone de-allocation until after the inode is 3001 * written to disk with its new link count (zero). At this point, all of the 3002 * on-disk inode's block pointers are nullified and, with careful dependency 3003 * list ordering, all dependencies related to the inode will be satisfied and 3004 * the corresponding dependency structures de-allocated. So, if/when the 3005 * inode is reused, there will be no mixing of old dependencies with new 3006 * ones. This artificial dependency is set up by the block de-allocation 3007 * procedure above (softdep_setup_freeblocks) and completed by the 3008 * following procedure. 3009 */ 3010 static void 3011 handle_workitem_freefile(struct freefile *freefile) 3012 { 3013 struct vnode vp; 3014 struct inode tip; 3015 struct inodedep *idp; 3016 int error; 3017 3018 #ifdef DEBUG 3019 ACQUIRE_LOCK(&lk); 3020 error = inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp); 3021 FREE_LOCK(&lk); 3022 if (error) 3023 panic("handle_workitem_freefile: inodedep survived"); 3024 #endif 3025 tip.i_devvp = freefile->fx_devvp; 3026 tip.i_dev = freefile->fx_devvp->v_rdev; 3027 tip.i_fs = freefile->fx_fs; 3028 vp.v_data = &tip; 3029 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0) 3030 softdep_error("handle_workitem_freefile", error); 3031 WORKITEM_FREE(freefile, D_FREEFILE); 3032 } 3033 3034 /* 3035 * Helper function which unlinks marker element from work list and returns 3036 * the next element on the list. 3037 */ 3038 static __inline struct worklist * 3039 markernext(struct worklist *marker) 3040 { 3041 struct worklist *next; 3042 3043 next = LIST_NEXT(marker, wk_list); 3044 LIST_REMOVE(marker, wk_list); 3045 return next; 3046 } 3047 3048 /* 3049 * Disk writes. 3050 * 3051 * The dependency structures constructed above are most actively used when file 3052 * system blocks are written to disk. No constraints are placed on when a 3053 * block can be written, but unsatisfied update dependencies are made safe by 3054 * modifying (or replacing) the source memory for the duration of the disk 3055 * write. When the disk write completes, the memory block is again brought 3056 * up-to-date. 3057 * 3058 * In-core inode structure reclamation. 3059 * 3060 * Because there are a finite number of "in-core" inode structures, they are 3061 * reused regularly. By transferring all inode-related dependencies to the 3062 * in-memory inode block and indexing them separately (via "inodedep"s), we 3063 * can allow "in-core" inode structures to be reused at any time and avoid 3064 * any increase in contention. 3065 * 3066 * Called just before entering the device driver to initiate a new disk I/O. 3067 * The buffer must be locked, thus, no I/O completion operations can occur 3068 * while we are manipulating its associated dependencies. 3069 * 3070 * Parameters: 3071 * bp: structure describing disk write to occur 3072 */ 3073 static void 3074 softdep_disk_io_initiation(struct buf *bp) 3075 { 3076 struct worklist *wk; 3077 struct worklist marker; 3078 struct indirdep *indirdep; 3079 3080 /* 3081 * We only care about write operations. There should never 3082 * be dependencies for reads. 3083 */ 3084 if (bp->b_cmd == BUF_CMD_READ) 3085 panic("softdep_disk_io_initiation: read"); 3086 3087 marker.wk_type = D_LAST + 1; /* Not a normal workitem */ 3088 3089 /* 3090 * Do any necessary pre-I/O processing. 3091 */ 3092 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = markernext(&marker)) { 3093 LIST_INSERT_AFTER(wk, &marker, wk_list); 3094 3095 switch (wk->wk_type) { 3096 3097 case D_PAGEDEP: 3098 initiate_write_filepage(WK_PAGEDEP(wk), bp); 3099 continue; 3100 3101 case D_INODEDEP: 3102 initiate_write_inodeblock(WK_INODEDEP(wk), bp); 3103 continue; 3104 3105 case D_INDIRDEP: 3106 indirdep = WK_INDIRDEP(wk); 3107 if (indirdep->ir_state & GOINGAWAY) 3108 panic("disk_io_initiation: indirdep gone"); 3109 /* 3110 * If there are no remaining dependencies, this 3111 * will be writing the real pointers, so the 3112 * dependency can be freed. 3113 */ 3114 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) { 3115 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 3116 brelse(indirdep->ir_savebp); 3117 /* inline expand WORKLIST_REMOVE(wk); */ 3118 wk->wk_state &= ~ONWORKLIST; 3119 LIST_REMOVE(wk, wk_list); 3120 WORKITEM_FREE(indirdep, D_INDIRDEP); 3121 continue; 3122 } 3123 /* 3124 * Replace up-to-date version with safe version. 3125 */ 3126 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount, 3127 M_INDIRDEP, M_SOFTDEP_FLAGS); 3128 ACQUIRE_LOCK(&lk); 3129 indirdep->ir_state &= ~ATTACHED; 3130 indirdep->ir_state |= UNDONE; 3131 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 3132 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 3133 bp->b_bcount); 3134 FREE_LOCK(&lk); 3135 continue; 3136 3137 case D_MKDIR: 3138 case D_BMSAFEMAP: 3139 case D_ALLOCDIRECT: 3140 case D_ALLOCINDIR: 3141 continue; 3142 3143 default: 3144 panic("handle_disk_io_initiation: Unexpected type %s", 3145 TYPENAME(wk->wk_type)); 3146 /* NOTREACHED */ 3147 } 3148 } 3149 } 3150 3151 /* 3152 * Called from within the procedure above to deal with unsatisfied 3153 * allocation dependencies in a directory. The buffer must be locked, 3154 * thus, no I/O completion operations can occur while we are 3155 * manipulating its associated dependencies. 3156 */ 3157 static void 3158 initiate_write_filepage(struct pagedep *pagedep, struct buf *bp) 3159 { 3160 struct diradd *dap; 3161 struct direct *ep; 3162 int i; 3163 3164 if (pagedep->pd_state & IOSTARTED) { 3165 /* 3166 * This can only happen if there is a driver that does not 3167 * understand chaining. Here biodone will reissue the call 3168 * to strategy for the incomplete buffers. 3169 */ 3170 kprintf("initiate_write_filepage: already started\n"); 3171 return; 3172 } 3173 pagedep->pd_state |= IOSTARTED; 3174 ACQUIRE_LOCK(&lk); 3175 for (i = 0; i < DAHASHSZ; i++) { 3176 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 3177 ep = (struct direct *) 3178 ((char *)bp->b_data + dap->da_offset); 3179 if (ep->d_ino != dap->da_newinum) { 3180 FREE_LOCK(&lk); 3181 panic("%s: dir inum %d != new %"PRId64, 3182 "initiate_write_filepage", 3183 ep->d_ino, dap->da_newinum); 3184 } 3185 if (dap->da_state & DIRCHG) 3186 ep->d_ino = dap->da_previous->dm_oldinum; 3187 else 3188 ep->d_ino = 0; 3189 dap->da_state &= ~ATTACHED; 3190 dap->da_state |= UNDONE; 3191 } 3192 } 3193 FREE_LOCK(&lk); 3194 } 3195 3196 /* 3197 * Called from within the procedure above to deal with unsatisfied 3198 * allocation dependencies in an inodeblock. The buffer must be 3199 * locked, thus, no I/O completion operations can occur while we 3200 * are manipulating its associated dependencies. 3201 * 3202 * Parameters: 3203 * bp: The inode block 3204 */ 3205 static void 3206 initiate_write_inodeblock(struct inodedep *inodedep, struct buf *bp) 3207 { 3208 struct allocdirect *adp, *lastadp; 3209 struct ufs1_dinode *dp; 3210 struct ufs1_dinode *sip; 3211 struct fs *fs; 3212 ufs_lbn_t prevlbn = 0; 3213 int i, deplist; 3214 3215 if (inodedep->id_state & IOSTARTED) 3216 panic("initiate_write_inodeblock: already started"); 3217 inodedep->id_state |= IOSTARTED; 3218 fs = inodedep->id_fs; 3219 dp = (struct ufs1_dinode *)bp->b_data + 3220 ino_to_fsbo(fs, inodedep->id_ino); 3221 /* 3222 * If the bitmap is not yet written, then the allocated 3223 * inode cannot be written to disk. 3224 */ 3225 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 3226 if (inodedep->id_savedino != NULL) 3227 panic("initiate_write_inodeblock: already doing I/O"); 3228 MALLOC(sip, struct ufs1_dinode *, 3229 sizeof(struct ufs1_dinode), M_INODEDEP, M_SOFTDEP_FLAGS); 3230 inodedep->id_savedino = sip; 3231 *inodedep->id_savedino = *dp; 3232 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 3233 dp->di_gen = inodedep->id_savedino->di_gen; 3234 return; 3235 } 3236 /* 3237 * If no dependencies, then there is nothing to roll back. 3238 */ 3239 inodedep->id_savedsize = dp->di_size; 3240 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 3241 return; 3242 /* 3243 * Set the dependencies to busy. 3244 */ 3245 ACQUIRE_LOCK(&lk); 3246 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3247 adp = TAILQ_NEXT(adp, ad_next)) { 3248 #ifdef DIAGNOSTIC 3249 if (deplist != 0 && prevlbn >= adp->ad_lbn) { 3250 FREE_LOCK(&lk); 3251 panic("softdep_write_inodeblock: lbn order"); 3252 } 3253 prevlbn = adp->ad_lbn; 3254 if (adp->ad_lbn < NDADDR && 3255 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) { 3256 FREE_LOCK(&lk); 3257 panic("%s: direct pointer #%ld mismatch %d != %d", 3258 "softdep_write_inodeblock", adp->ad_lbn, 3259 dp->di_db[adp->ad_lbn], adp->ad_newblkno); 3260 } 3261 if (adp->ad_lbn >= NDADDR && 3262 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) { 3263 FREE_LOCK(&lk); 3264 panic("%s: indirect pointer #%ld mismatch %d != %d", 3265 "softdep_write_inodeblock", adp->ad_lbn - NDADDR, 3266 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno); 3267 } 3268 deplist |= 1 << adp->ad_lbn; 3269 if ((adp->ad_state & ATTACHED) == 0) { 3270 FREE_LOCK(&lk); 3271 panic("softdep_write_inodeblock: Unknown state 0x%x", 3272 adp->ad_state); 3273 } 3274 #endif /* DIAGNOSTIC */ 3275 adp->ad_state &= ~ATTACHED; 3276 adp->ad_state |= UNDONE; 3277 } 3278 /* 3279 * The on-disk inode cannot claim to be any larger than the last 3280 * fragment that has been written. Otherwise, the on-disk inode 3281 * might have fragments that were not the last block in the file 3282 * which would corrupt the filesystem. 3283 */ 3284 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3285 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 3286 if (adp->ad_lbn >= NDADDR) 3287 break; 3288 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 3289 /* keep going until hitting a rollback to a frag */ 3290 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 3291 continue; 3292 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 3293 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 3294 #ifdef DIAGNOSTIC 3295 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) { 3296 FREE_LOCK(&lk); 3297 panic("softdep_write_inodeblock: lost dep1"); 3298 } 3299 #endif /* DIAGNOSTIC */ 3300 dp->di_db[i] = 0; 3301 } 3302 for (i = 0; i < NIADDR; i++) { 3303 #ifdef DIAGNOSTIC 3304 if (dp->di_ib[i] != 0 && 3305 (deplist & ((1 << NDADDR) << i)) == 0) { 3306 FREE_LOCK(&lk); 3307 panic("softdep_write_inodeblock: lost dep2"); 3308 } 3309 #endif /* DIAGNOSTIC */ 3310 dp->di_ib[i] = 0; 3311 } 3312 FREE_LOCK(&lk); 3313 return; 3314 } 3315 /* 3316 * If we have zero'ed out the last allocated block of the file, 3317 * roll back the size to the last currently allocated block. 3318 * We know that this last allocated block is a full-sized as 3319 * we already checked for fragments in the loop above. 3320 */ 3321 if (lastadp != NULL && 3322 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 3323 for (i = lastadp->ad_lbn; i >= 0; i--) 3324 if (dp->di_db[i] != 0) 3325 break; 3326 dp->di_size = (i + 1) * fs->fs_bsize; 3327 } 3328 /* 3329 * The only dependencies are for indirect blocks. 3330 * 3331 * The file size for indirect block additions is not guaranteed. 3332 * Such a guarantee would be non-trivial to achieve. The conventional 3333 * synchronous write implementation also does not make this guarantee. 3334 * Fsck should catch and fix discrepancies. Arguably, the file size 3335 * can be over-estimated without destroying integrity when the file 3336 * moves into the indirect blocks (i.e., is large). If we want to 3337 * postpone fsck, we are stuck with this argument. 3338 */ 3339 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 3340 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 3341 FREE_LOCK(&lk); 3342 } 3343 3344 /* 3345 * This routine is called during the completion interrupt 3346 * service routine for a disk write (from the procedure called 3347 * by the device driver to inform the filesystem caches of 3348 * a request completion). It should be called early in this 3349 * procedure, before the block is made available to other 3350 * processes or other routines are called. 3351 * 3352 * Parameters: 3353 * bp: describes the completed disk write 3354 */ 3355 static void 3356 softdep_disk_write_complete(struct buf *bp) 3357 { 3358 struct worklist *wk; 3359 struct workhead reattach; 3360 struct newblk *newblk; 3361 struct allocindir *aip; 3362 struct allocdirect *adp; 3363 struct indirdep *indirdep; 3364 struct inodedep *inodedep; 3365 struct bmsafemap *bmsafemap; 3366 3367 #ifdef DEBUG 3368 if (lk.lkt_held != NOHOLDER) 3369 panic("softdep_disk_write_complete: lock is held"); 3370 lk.lkt_held = SPECIAL_FLAG; 3371 #endif 3372 LIST_INIT(&reattach); 3373 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 3374 WORKLIST_REMOVE(wk); 3375 switch (wk->wk_type) { 3376 3377 case D_PAGEDEP: 3378 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 3379 WORKLIST_INSERT(&reattach, wk); 3380 continue; 3381 3382 case D_INODEDEP: 3383 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 3384 WORKLIST_INSERT(&reattach, wk); 3385 continue; 3386 3387 case D_BMSAFEMAP: 3388 bmsafemap = WK_BMSAFEMAP(wk); 3389 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) { 3390 newblk->nb_state |= DEPCOMPLETE; 3391 newblk->nb_bmsafemap = NULL; 3392 LIST_REMOVE(newblk, nb_deps); 3393 } 3394 while ((adp = 3395 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) { 3396 adp->ad_state |= DEPCOMPLETE; 3397 adp->ad_buf = NULL; 3398 LIST_REMOVE(adp, ad_deps); 3399 handle_allocdirect_partdone(adp); 3400 } 3401 while ((aip = 3402 LIST_FIRST(&bmsafemap->sm_allocindirhd))) { 3403 aip->ai_state |= DEPCOMPLETE; 3404 aip->ai_buf = NULL; 3405 LIST_REMOVE(aip, ai_deps); 3406 handle_allocindir_partdone(aip); 3407 } 3408 while ((inodedep = 3409 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { 3410 inodedep->id_state |= DEPCOMPLETE; 3411 LIST_REMOVE(inodedep, id_deps); 3412 inodedep->id_buf = NULL; 3413 } 3414 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 3415 continue; 3416 3417 case D_MKDIR: 3418 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 3419 continue; 3420 3421 case D_ALLOCDIRECT: 3422 adp = WK_ALLOCDIRECT(wk); 3423 adp->ad_state |= COMPLETE; 3424 handle_allocdirect_partdone(adp); 3425 continue; 3426 3427 case D_ALLOCINDIR: 3428 aip = WK_ALLOCINDIR(wk); 3429 aip->ai_state |= COMPLETE; 3430 handle_allocindir_partdone(aip); 3431 continue; 3432 3433 case D_INDIRDEP: 3434 indirdep = WK_INDIRDEP(wk); 3435 if (indirdep->ir_state & GOINGAWAY) { 3436 lk.lkt_held = NOHOLDER; 3437 panic("disk_write_complete: indirdep gone"); 3438 } 3439 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 3440 FREE(indirdep->ir_saveddata, M_INDIRDEP); 3441 indirdep->ir_saveddata = 0; 3442 indirdep->ir_state &= ~UNDONE; 3443 indirdep->ir_state |= ATTACHED; 3444 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 3445 handle_allocindir_partdone(aip); 3446 if (aip == LIST_FIRST(&indirdep->ir_donehd)) { 3447 lk.lkt_held = NOHOLDER; 3448 panic("disk_write_complete: not gone"); 3449 } 3450 } 3451 WORKLIST_INSERT(&reattach, wk); 3452 if ((bp->b_flags & B_DELWRI) == 0) 3453 stat_indir_blk_ptrs++; 3454 bdirty(bp); 3455 continue; 3456 3457 default: 3458 lk.lkt_held = NOHOLDER; 3459 panic("handle_disk_write_complete: Unknown type %s", 3460 TYPENAME(wk->wk_type)); 3461 /* NOTREACHED */ 3462 } 3463 } 3464 /* 3465 * Reattach any requests that must be redone. 3466 */ 3467 while ((wk = LIST_FIRST(&reattach)) != NULL) { 3468 WORKLIST_REMOVE(wk); 3469 WORKLIST_INSERT(&bp->b_dep, wk); 3470 } 3471 #ifdef DEBUG 3472 if (lk.lkt_held != SPECIAL_FLAG) 3473 panic("softdep_disk_write_complete: lock lost"); 3474 lk.lkt_held = NOHOLDER; 3475 #endif 3476 } 3477 3478 /* 3479 * Called from within softdep_disk_write_complete above. Note that 3480 * this routine is always called from interrupt level with further 3481 * splbio interrupts blocked. 3482 * 3483 * Parameters: 3484 * adp: the completed allocdirect 3485 */ 3486 static void 3487 handle_allocdirect_partdone(struct allocdirect *adp) 3488 { 3489 struct allocdirect *listadp; 3490 struct inodedep *inodedep; 3491 long bsize; 3492 3493 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3494 return; 3495 if (adp->ad_buf != NULL) { 3496 lk.lkt_held = NOHOLDER; 3497 panic("handle_allocdirect_partdone: dangling dep"); 3498 } 3499 /* 3500 * The on-disk inode cannot claim to be any larger than the last 3501 * fragment that has been written. Otherwise, the on-disk inode 3502 * might have fragments that were not the last block in the file 3503 * which would corrupt the filesystem. Thus, we cannot free any 3504 * allocdirects after one whose ad_oldblkno claims a fragment as 3505 * these blocks must be rolled back to zero before writing the inode. 3506 * We check the currently active set of allocdirects in id_inoupdt. 3507 */ 3508 inodedep = adp->ad_inodedep; 3509 bsize = inodedep->id_fs->fs_bsize; 3510 TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) { 3511 /* found our block */ 3512 if (listadp == adp) 3513 break; 3514 /* continue if ad_oldlbn is not a fragment */ 3515 if (listadp->ad_oldsize == 0 || 3516 listadp->ad_oldsize == bsize) 3517 continue; 3518 /* hit a fragment */ 3519 return; 3520 } 3521 /* 3522 * If we have reached the end of the current list without 3523 * finding the just finished dependency, then it must be 3524 * on the future dependency list. Future dependencies cannot 3525 * be freed until they are moved to the current list. 3526 */ 3527 if (listadp == NULL) { 3528 #ifdef DEBUG 3529 TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next) 3530 /* found our block */ 3531 if (listadp == adp) 3532 break; 3533 if (listadp == NULL) { 3534 lk.lkt_held = NOHOLDER; 3535 panic("handle_allocdirect_partdone: lost dep"); 3536 } 3537 #endif /* DEBUG */ 3538 return; 3539 } 3540 /* 3541 * If we have found the just finished dependency, then free 3542 * it along with anything that follows it that is complete. 3543 */ 3544 for (; adp; adp = listadp) { 3545 listadp = TAILQ_NEXT(adp, ad_next); 3546 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3547 return; 3548 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 3549 } 3550 } 3551 3552 /* 3553 * Called from within softdep_disk_write_complete above. Note that 3554 * this routine is always called from interrupt level with further 3555 * splbio interrupts blocked. 3556 * 3557 * Parameters: 3558 * aip: the completed allocindir 3559 */ 3560 static void 3561 handle_allocindir_partdone(struct allocindir *aip) 3562 { 3563 struct indirdep *indirdep; 3564 3565 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 3566 return; 3567 if (aip->ai_buf != NULL) { 3568 lk.lkt_held = NOHOLDER; 3569 panic("handle_allocindir_partdone: dangling dependency"); 3570 } 3571 indirdep = aip->ai_indirdep; 3572 if (indirdep->ir_state & UNDONE) { 3573 LIST_REMOVE(aip, ai_next); 3574 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 3575 return; 3576 } 3577 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 3578 aip->ai_newblkno; 3579 LIST_REMOVE(aip, ai_next); 3580 if (aip->ai_freefrag != NULL) 3581 add_to_worklist(&aip->ai_freefrag->ff_list); 3582 WORKITEM_FREE(aip, D_ALLOCINDIR); 3583 } 3584 3585 /* 3586 * Called from within softdep_disk_write_complete above to restore 3587 * in-memory inode block contents to their most up-to-date state. Note 3588 * that this routine is always called from interrupt level with further 3589 * splbio interrupts blocked. 3590 * 3591 * Parameters: 3592 * bp: buffer containing the inode block 3593 */ 3594 static int 3595 handle_written_inodeblock(struct inodedep *inodedep, struct buf *bp) 3596 { 3597 struct worklist *wk, *filefree; 3598 struct allocdirect *adp, *nextadp; 3599 struct ufs1_dinode *dp; 3600 int hadchanges; 3601 3602 if ((inodedep->id_state & IOSTARTED) == 0) { 3603 lk.lkt_held = NOHOLDER; 3604 panic("handle_written_inodeblock: not started"); 3605 } 3606 inodedep->id_state &= ~IOSTARTED; 3607 dp = (struct ufs1_dinode *)bp->b_data + 3608 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 3609 /* 3610 * If we had to rollback the inode allocation because of 3611 * bitmaps being incomplete, then simply restore it. 3612 * Keep the block dirty so that it will not be reclaimed until 3613 * all associated dependencies have been cleared and the 3614 * corresponding updates written to disk. 3615 */ 3616 if (inodedep->id_savedino != NULL) { 3617 *dp = *inodedep->id_savedino; 3618 FREE(inodedep->id_savedino, M_INODEDEP); 3619 inodedep->id_savedino = NULL; 3620 if ((bp->b_flags & B_DELWRI) == 0) 3621 stat_inode_bitmap++; 3622 bdirty(bp); 3623 return (1); 3624 } 3625 inodedep->id_state |= COMPLETE; 3626 /* 3627 * Roll forward anything that had to be rolled back before 3628 * the inode could be updated. 3629 */ 3630 hadchanges = 0; 3631 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 3632 nextadp = TAILQ_NEXT(adp, ad_next); 3633 if (adp->ad_state & ATTACHED) { 3634 lk.lkt_held = NOHOLDER; 3635 panic("handle_written_inodeblock: new entry"); 3636 } 3637 if (adp->ad_lbn < NDADDR) { 3638 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) { 3639 lk.lkt_held = NOHOLDER; 3640 panic("%s: %s #%ld mismatch %d != %d", 3641 "handle_written_inodeblock", 3642 "direct pointer", adp->ad_lbn, 3643 dp->di_db[adp->ad_lbn], adp->ad_oldblkno); 3644 } 3645 dp->di_db[adp->ad_lbn] = adp->ad_newblkno; 3646 } else { 3647 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) { 3648 lk.lkt_held = NOHOLDER; 3649 panic("%s: %s #%ld allocated as %d", 3650 "handle_written_inodeblock", 3651 "indirect pointer", adp->ad_lbn - NDADDR, 3652 dp->di_ib[adp->ad_lbn - NDADDR]); 3653 } 3654 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno; 3655 } 3656 adp->ad_state &= ~UNDONE; 3657 adp->ad_state |= ATTACHED; 3658 hadchanges = 1; 3659 } 3660 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 3661 stat_direct_blk_ptrs++; 3662 /* 3663 * Reset the file size to its most up-to-date value. 3664 */ 3665 if (inodedep->id_savedsize == -1) { 3666 lk.lkt_held = NOHOLDER; 3667 panic("handle_written_inodeblock: bad size"); 3668 } 3669 if (dp->di_size != inodedep->id_savedsize) { 3670 dp->di_size = inodedep->id_savedsize; 3671 hadchanges = 1; 3672 } 3673 inodedep->id_savedsize = -1; 3674 /* 3675 * If there were any rollbacks in the inode block, then it must be 3676 * marked dirty so that its will eventually get written back in 3677 * its correct form. 3678 */ 3679 if (hadchanges) 3680 bdirty(bp); 3681 /* 3682 * Process any allocdirects that completed during the update. 3683 */ 3684 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 3685 handle_allocdirect_partdone(adp); 3686 /* 3687 * Process deallocations that were held pending until the 3688 * inode had been written to disk. Freeing of the inode 3689 * is delayed until after all blocks have been freed to 3690 * avoid creation of new <vfsid, inum, lbn> triples 3691 * before the old ones have been deleted. 3692 */ 3693 filefree = NULL; 3694 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 3695 WORKLIST_REMOVE(wk); 3696 switch (wk->wk_type) { 3697 3698 case D_FREEFILE: 3699 /* 3700 * We defer adding filefree to the worklist until 3701 * all other additions have been made to ensure 3702 * that it will be done after all the old blocks 3703 * have been freed. 3704 */ 3705 if (filefree != NULL) { 3706 lk.lkt_held = NOHOLDER; 3707 panic("handle_written_inodeblock: filefree"); 3708 } 3709 filefree = wk; 3710 continue; 3711 3712 case D_MKDIR: 3713 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 3714 continue; 3715 3716 case D_DIRADD: 3717 diradd_inode_written(WK_DIRADD(wk), inodedep); 3718 continue; 3719 3720 case D_FREEBLKS: 3721 wk->wk_state |= COMPLETE; 3722 if ((wk->wk_state & ALLCOMPLETE) != ALLCOMPLETE) 3723 continue; 3724 /* -- fall through -- */ 3725 case D_FREEFRAG: 3726 case D_DIRREM: 3727 add_to_worklist(wk); 3728 continue; 3729 3730 default: 3731 lk.lkt_held = NOHOLDER; 3732 panic("handle_written_inodeblock: Unknown type %s", 3733 TYPENAME(wk->wk_type)); 3734 /* NOTREACHED */ 3735 } 3736 } 3737 if (filefree != NULL) { 3738 if (free_inodedep(inodedep) == 0) { 3739 lk.lkt_held = NOHOLDER; 3740 panic("handle_written_inodeblock: live inodedep"); 3741 } 3742 add_to_worklist(filefree); 3743 return (0); 3744 } 3745 3746 /* 3747 * If no outstanding dependencies, free it. 3748 */ 3749 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0) 3750 return (0); 3751 return (hadchanges); 3752 } 3753 3754 /* 3755 * Process a diradd entry after its dependent inode has been written. 3756 * This routine must be called with splbio interrupts blocked. 3757 */ 3758 static void 3759 diradd_inode_written(struct diradd *dap, struct inodedep *inodedep) 3760 { 3761 struct pagedep *pagedep; 3762 3763 dap->da_state |= COMPLETE; 3764 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3765 if (dap->da_state & DIRCHG) 3766 pagedep = dap->da_previous->dm_pagedep; 3767 else 3768 pagedep = dap->da_pagedep; 3769 LIST_REMOVE(dap, da_pdlist); 3770 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3771 } 3772 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 3773 } 3774 3775 /* 3776 * Handle the completion of a mkdir dependency. 3777 */ 3778 static void 3779 handle_written_mkdir(struct mkdir *mkdir, int type) 3780 { 3781 struct diradd *dap; 3782 struct pagedep *pagedep; 3783 3784 if (mkdir->md_state != type) { 3785 lk.lkt_held = NOHOLDER; 3786 panic("handle_written_mkdir: bad type"); 3787 } 3788 dap = mkdir->md_diradd; 3789 dap->da_state &= ~type; 3790 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 3791 dap->da_state |= DEPCOMPLETE; 3792 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3793 if (dap->da_state & DIRCHG) 3794 pagedep = dap->da_previous->dm_pagedep; 3795 else 3796 pagedep = dap->da_pagedep; 3797 LIST_REMOVE(dap, da_pdlist); 3798 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3799 } 3800 LIST_REMOVE(mkdir, md_mkdirs); 3801 WORKITEM_FREE(mkdir, D_MKDIR); 3802 } 3803 3804 /* 3805 * Called from within softdep_disk_write_complete above. 3806 * A write operation was just completed. Removed inodes can 3807 * now be freed and associated block pointers may be committed. 3808 * Note that this routine is always called from interrupt level 3809 * with further splbio interrupts blocked. 3810 * 3811 * Parameters: 3812 * bp: buffer containing the written page 3813 */ 3814 static int 3815 handle_written_filepage(struct pagedep *pagedep, struct buf *bp) 3816 { 3817 struct dirrem *dirrem; 3818 struct diradd *dap, *nextdap; 3819 struct direct *ep; 3820 int i, chgs; 3821 3822 if ((pagedep->pd_state & IOSTARTED) == 0) { 3823 lk.lkt_held = NOHOLDER; 3824 panic("handle_written_filepage: not started"); 3825 } 3826 pagedep->pd_state &= ~IOSTARTED; 3827 /* 3828 * Process any directory removals that have been committed. 3829 */ 3830 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 3831 LIST_REMOVE(dirrem, dm_next); 3832 dirrem->dm_dirinum = pagedep->pd_ino; 3833 add_to_worklist(&dirrem->dm_list); 3834 } 3835 /* 3836 * Free any directory additions that have been committed. 3837 */ 3838 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 3839 free_diradd(dap); 3840 /* 3841 * Uncommitted directory entries must be restored. 3842 */ 3843 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 3844 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 3845 dap = nextdap) { 3846 nextdap = LIST_NEXT(dap, da_pdlist); 3847 if (dap->da_state & ATTACHED) { 3848 lk.lkt_held = NOHOLDER; 3849 panic("handle_written_filepage: attached"); 3850 } 3851 ep = (struct direct *) 3852 ((char *)bp->b_data + dap->da_offset); 3853 ep->d_ino = dap->da_newinum; 3854 dap->da_state &= ~UNDONE; 3855 dap->da_state |= ATTACHED; 3856 chgs = 1; 3857 /* 3858 * If the inode referenced by the directory has 3859 * been written out, then the dependency can be 3860 * moved to the pending list. 3861 */ 3862 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3863 LIST_REMOVE(dap, da_pdlist); 3864 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 3865 da_pdlist); 3866 } 3867 } 3868 } 3869 /* 3870 * If there were any rollbacks in the directory, then it must be 3871 * marked dirty so that its will eventually get written back in 3872 * its correct form. 3873 */ 3874 if (chgs) { 3875 if ((bp->b_flags & B_DELWRI) == 0) 3876 stat_dir_entry++; 3877 bdirty(bp); 3878 } 3879 /* 3880 * If no dependencies remain, the pagedep will be freed. 3881 * Otherwise it will remain to update the page before it 3882 * is written back to disk. 3883 */ 3884 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) { 3885 for (i = 0; i < DAHASHSZ; i++) 3886 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL) 3887 break; 3888 if (i == DAHASHSZ) { 3889 LIST_REMOVE(pagedep, pd_hash); 3890 WORKITEM_FREE(pagedep, D_PAGEDEP); 3891 return (0); 3892 } 3893 } 3894 return (1); 3895 } 3896 3897 /* 3898 * Writing back in-core inode structures. 3899 * 3900 * The filesystem only accesses an inode's contents when it occupies an 3901 * "in-core" inode structure. These "in-core" structures are separate from 3902 * the page frames used to cache inode blocks. Only the latter are 3903 * transferred to/from the disk. So, when the updated contents of the 3904 * "in-core" inode structure are copied to the corresponding in-memory inode 3905 * block, the dependencies are also transferred. The following procedure is 3906 * called when copying a dirty "in-core" inode to a cached inode block. 3907 */ 3908 3909 /* 3910 * Called when an inode is loaded from disk. If the effective link count 3911 * differed from the actual link count when it was last flushed, then we 3912 * need to ensure that the correct effective link count is put back. 3913 * 3914 * Parameters: 3915 * ip: the "in_core" copy of the inode 3916 */ 3917 void 3918 softdep_load_inodeblock(struct inode *ip) 3919 { 3920 struct inodedep *inodedep; 3921 3922 /* 3923 * Check for alternate nlink count. 3924 */ 3925 ip->i_effnlink = ip->i_nlink; 3926 ACQUIRE_LOCK(&lk); 3927 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3928 FREE_LOCK(&lk); 3929 return; 3930 } 3931 ip->i_effnlink -= inodedep->id_nlinkdelta; 3932 FREE_LOCK(&lk); 3933 } 3934 3935 /* 3936 * This routine is called just before the "in-core" inode 3937 * information is to be copied to the in-memory inode block. 3938 * Recall that an inode block contains several inodes. If 3939 * the force flag is set, then the dependencies will be 3940 * cleared so that the update can always be made. Note that 3941 * the buffer is locked when this routine is called, so we 3942 * will never be in the middle of writing the inode block 3943 * to disk. 3944 * 3945 * Parameters: 3946 * ip: the "in_core" copy of the inode 3947 * bp: the buffer containing the inode block 3948 * waitfor: nonzero => update must be allowed 3949 */ 3950 void 3951 softdep_update_inodeblock(struct inode *ip, struct buf *bp, 3952 int waitfor) 3953 { 3954 struct inodedep *inodedep; 3955 struct worklist *wk; 3956 int error, gotit; 3957 3958 /* 3959 * If the effective link count is not equal to the actual link 3960 * count, then we must track the difference in an inodedep while 3961 * the inode is (potentially) tossed out of the cache. Otherwise, 3962 * if there is no existing inodedep, then there are no dependencies 3963 * to track. 3964 */ 3965 ACQUIRE_LOCK(&lk); 3966 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3967 FREE_LOCK(&lk); 3968 if (ip->i_effnlink != ip->i_nlink) 3969 panic("softdep_update_inodeblock: bad link count"); 3970 return; 3971 } 3972 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) { 3973 FREE_LOCK(&lk); 3974 panic("softdep_update_inodeblock: bad delta"); 3975 } 3976 /* 3977 * Changes have been initiated. Anything depending on these 3978 * changes cannot occur until this inode has been written. 3979 */ 3980 inodedep->id_state &= ~COMPLETE; 3981 if ((inodedep->id_state & ONWORKLIST) == 0) 3982 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 3983 /* 3984 * Any new dependencies associated with the incore inode must 3985 * now be moved to the list associated with the buffer holding 3986 * the in-memory copy of the inode. Once merged process any 3987 * allocdirects that are completed by the merger. 3988 */ 3989 merge_inode_lists(inodedep); 3990 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL) 3991 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); 3992 /* 3993 * Now that the inode has been pushed into the buffer, the 3994 * operations dependent on the inode being written to disk 3995 * can be moved to the id_bufwait so that they will be 3996 * processed when the buffer I/O completes. 3997 */ 3998 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 3999 WORKLIST_REMOVE(wk); 4000 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 4001 } 4002 /* 4003 * Newly allocated inodes cannot be written until the bitmap 4004 * that allocates them have been written (indicated by 4005 * DEPCOMPLETE being set in id_state). If we are doing a 4006 * forced sync (e.g., an fsync on a file), we force the bitmap 4007 * to be written so that the update can be done. 4008 */ 4009 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) { 4010 FREE_LOCK(&lk); 4011 return; 4012 } 4013 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 4014 FREE_LOCK(&lk); 4015 if (gotit && 4016 (error = bwrite(inodedep->id_buf)) != 0) 4017 softdep_error("softdep_update_inodeblock: bwrite", error); 4018 } 4019 4020 /* 4021 * Merge the new inode dependency list (id_newinoupdt) into the old 4022 * inode dependency list (id_inoupdt). This routine must be called 4023 * with splbio interrupts blocked. 4024 */ 4025 static void 4026 merge_inode_lists(struct inodedep *inodedep) 4027 { 4028 struct allocdirect *listadp, *newadp; 4029 4030 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 4031 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) { 4032 if (listadp->ad_lbn < newadp->ad_lbn) { 4033 listadp = TAILQ_NEXT(listadp, ad_next); 4034 continue; 4035 } 4036 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 4037 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 4038 if (listadp->ad_lbn == newadp->ad_lbn) { 4039 allocdirect_merge(&inodedep->id_inoupdt, newadp, 4040 listadp); 4041 listadp = newadp; 4042 } 4043 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 4044 } 4045 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) { 4046 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 4047 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next); 4048 } 4049 } 4050 4051 /* 4052 * If we are doing an fsync, then we must ensure that any directory 4053 * entries for the inode have been written after the inode gets to disk. 4054 * 4055 * Parameters: 4056 * vp: the "in_core" copy of the inode 4057 */ 4058 static int 4059 softdep_fsync(struct vnode *vp) 4060 { 4061 struct inodedep *inodedep; 4062 struct pagedep *pagedep; 4063 struct worklist *wk; 4064 struct diradd *dap; 4065 struct mount *mnt; 4066 struct vnode *pvp; 4067 struct inode *ip; 4068 struct buf *bp; 4069 struct fs *fs; 4070 int error, flushparent; 4071 ino_t parentino; 4072 ufs_lbn_t lbn; 4073 4074 ip = VTOI(vp); 4075 fs = ip->i_fs; 4076 ACQUIRE_LOCK(&lk); 4077 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) { 4078 FREE_LOCK(&lk); 4079 return (0); 4080 } 4081 if (LIST_FIRST(&inodedep->id_inowait) != NULL || 4082 LIST_FIRST(&inodedep->id_bufwait) != NULL || 4083 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 4084 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) { 4085 FREE_LOCK(&lk); 4086 panic("softdep_fsync: pending ops"); 4087 } 4088 for (error = 0, flushparent = 0; ; ) { 4089 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 4090 break; 4091 if (wk->wk_type != D_DIRADD) { 4092 FREE_LOCK(&lk); 4093 panic("softdep_fsync: Unexpected type %s", 4094 TYPENAME(wk->wk_type)); 4095 } 4096 dap = WK_DIRADD(wk); 4097 /* 4098 * Flush our parent if this directory entry 4099 * has a MKDIR_PARENT dependency. 4100 */ 4101 if (dap->da_state & DIRCHG) 4102 pagedep = dap->da_previous->dm_pagedep; 4103 else 4104 pagedep = dap->da_pagedep; 4105 mnt = pagedep->pd_mnt; 4106 parentino = pagedep->pd_ino; 4107 lbn = pagedep->pd_lbn; 4108 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) { 4109 FREE_LOCK(&lk); 4110 panic("softdep_fsync: dirty"); 4111 } 4112 flushparent = dap->da_state & MKDIR_PARENT; 4113 /* 4114 * If we are being fsync'ed as part of vgone'ing this vnode, 4115 * then we will not be able to release and recover the 4116 * vnode below, so we just have to give up on writing its 4117 * directory entry out. It will eventually be written, just 4118 * not now, but then the user was not asking to have it 4119 * written, so we are not breaking any promises. 4120 */ 4121 if (vp->v_flag & VRECLAIMED) 4122 break; 4123 /* 4124 * We prevent deadlock by always fetching inodes from the 4125 * root, moving down the directory tree. Thus, when fetching 4126 * our parent directory, we must unlock ourselves before 4127 * requesting the lock on our parent. See the comment in 4128 * ufs_lookup for details on possible races. 4129 */ 4130 FREE_LOCK(&lk); 4131 vn_unlock(vp); 4132 error = VFS_VGET(mnt, parentino, &pvp); 4133 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4134 if (error != 0) 4135 return (error); 4136 if (flushparent) { 4137 if ((error = ffs_update(pvp, 1)) != 0) { 4138 vput(pvp); 4139 return (error); 4140 } 4141 } 4142 /* 4143 * Flush directory page containing the inode's name. 4144 */ 4145 error = bread(pvp, lblktodoff(fs, lbn), blksize(fs, VTOI(pvp), lbn), &bp); 4146 if (error == 0) 4147 error = bwrite(bp); 4148 vput(pvp); 4149 if (error != 0) 4150 return (error); 4151 ACQUIRE_LOCK(&lk); 4152 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) 4153 break; 4154 } 4155 FREE_LOCK(&lk); 4156 return (0); 4157 } 4158 4159 /* 4160 * Flush all the dirty bitmaps associated with the block device 4161 * before flushing the rest of the dirty blocks so as to reduce 4162 * the number of dependencies that will have to be rolled back. 4163 */ 4164 static int softdep_fsync_mountdev_bp(struct buf *bp, void *data); 4165 4166 void 4167 softdep_fsync_mountdev(struct vnode *vp) 4168 { 4169 if (!vn_isdisk(vp, NULL)) 4170 panic("softdep_fsync_mountdev: vnode not a disk"); 4171 ACQUIRE_LOCK(&lk); 4172 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 4173 softdep_fsync_mountdev_bp, vp); 4174 drain_output(vp, 1); 4175 FREE_LOCK(&lk); 4176 } 4177 4178 static int 4179 softdep_fsync_mountdev_bp(struct buf *bp, void *data) 4180 { 4181 struct worklist *wk; 4182 struct vnode *vp = data; 4183 4184 /* 4185 * If it is already scheduled, skip to the next buffer. 4186 */ 4187 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) 4188 return(0); 4189 if (bp->b_vp != vp || (bp->b_flags & B_DELWRI) == 0) { 4190 BUF_UNLOCK(bp); 4191 kprintf("softdep_fsync_mountdev_bp: warning, buffer %p ripped out from under vnode %p\n", bp, vp); 4192 return(0); 4193 } 4194 /* 4195 * We are only interested in bitmaps with outstanding 4196 * dependencies. 4197 */ 4198 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 4199 wk->wk_type != D_BMSAFEMAP) { 4200 BUF_UNLOCK(bp); 4201 return(0); 4202 } 4203 bremfree(bp); 4204 FREE_LOCK(&lk); 4205 (void) bawrite(bp); 4206 ACQUIRE_LOCK(&lk); 4207 return(0); 4208 } 4209 4210 /* 4211 * This routine is called when we are trying to synchronously flush a 4212 * file. This routine must eliminate any filesystem metadata dependencies 4213 * so that the syncing routine can succeed by pushing the dirty blocks 4214 * associated with the file. If any I/O errors occur, they are returned. 4215 */ 4216 struct softdep_sync_metadata_info { 4217 struct vnode *vp; 4218 int waitfor; 4219 }; 4220 4221 static int softdep_sync_metadata_bp(struct buf *bp, void *data); 4222 4223 int 4224 softdep_sync_metadata(struct vnode *vp, struct thread *td) 4225 { 4226 struct softdep_sync_metadata_info info; 4227 int error, waitfor; 4228 4229 /* 4230 * Check whether this vnode is involved in a filesystem 4231 * that is doing soft dependency processing. 4232 */ 4233 if (!vn_isdisk(vp, NULL)) { 4234 if (!DOINGSOFTDEP(vp)) 4235 return (0); 4236 } else 4237 if (vp->v_rdev->si_mountpoint == NULL || 4238 (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0) 4239 return (0); 4240 /* 4241 * Ensure that any direct block dependencies have been cleared. 4242 */ 4243 ACQUIRE_LOCK(&lk); 4244 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) { 4245 FREE_LOCK(&lk); 4246 return (error); 4247 } 4248 /* 4249 * For most files, the only metadata dependencies are the 4250 * cylinder group maps that allocate their inode or blocks. 4251 * The block allocation dependencies can be found by traversing 4252 * the dependency lists for any buffers that remain on their 4253 * dirty buffer list. The inode allocation dependency will 4254 * be resolved when the inode is updated with MNT_WAIT. 4255 * This work is done in two passes. The first pass grabs most 4256 * of the buffers and begins asynchronously writing them. The 4257 * only way to wait for these asynchronous writes is to sleep 4258 * on the filesystem vnode which may stay busy for a long time 4259 * if the filesystem is active. So, instead, we make a second 4260 * pass over the dependencies blocking on each write. In the 4261 * usual case we will be blocking against a write that we 4262 * initiated, so when it is done the dependency will have been 4263 * resolved. Thus the second pass is expected to end quickly. 4264 */ 4265 waitfor = MNT_NOWAIT; 4266 top: 4267 /* 4268 * We must wait for any I/O in progress to finish so that 4269 * all potential buffers on the dirty list will be visible. 4270 */ 4271 drain_output(vp, 1); 4272 info.vp = vp; 4273 info.waitfor = waitfor; 4274 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 4275 softdep_sync_metadata_bp, &info); 4276 if (error < 0) { 4277 FREE_LOCK(&lk); 4278 return(-error); /* error code */ 4279 } 4280 4281 /* 4282 * The brief unlock is to allow any pent up dependency 4283 * processing to be done. Then proceed with the second pass. 4284 */ 4285 if (waitfor == MNT_NOWAIT) { 4286 waitfor = MNT_WAIT; 4287 FREE_LOCK(&lk); 4288 ACQUIRE_LOCK(&lk); 4289 goto top; 4290 } 4291 4292 /* 4293 * If we have managed to get rid of all the dirty buffers, 4294 * then we are done. For certain directories and block 4295 * devices, we may need to do further work. 4296 * 4297 * We must wait for any I/O in progress to finish so that 4298 * all potential buffers on the dirty list will be visible. 4299 */ 4300 drain_output(vp, 1); 4301 if (RB_EMPTY(&vp->v_rbdirty_tree)) { 4302 FREE_LOCK(&lk); 4303 return (0); 4304 } 4305 4306 FREE_LOCK(&lk); 4307 /* 4308 * If we are trying to sync a block device, some of its buffers may 4309 * contain metadata that cannot be written until the contents of some 4310 * partially written files have been written to disk. The only easy 4311 * way to accomplish this is to sync the entire filesystem (luckily 4312 * this happens rarely). 4313 */ 4314 if (vn_isdisk(vp, NULL) && 4315 vp->v_rdev && 4316 vp->v_rdev->si_mountpoint && !vn_islocked(vp) && 4317 (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT)) != 0) 4318 return (error); 4319 return (0); 4320 } 4321 4322 static int 4323 softdep_sync_metadata_bp(struct buf *bp, void *data) 4324 { 4325 struct softdep_sync_metadata_info *info = data; 4326 struct pagedep *pagedep; 4327 struct allocdirect *adp; 4328 struct allocindir *aip; 4329 struct worklist *wk; 4330 struct buf *nbp; 4331 int error; 4332 int i; 4333 4334 if (getdirtybuf(&bp, MNT_WAIT) == 0) { 4335 kprintf("softdep_sync_metadata_bp(1): caught buf %p going away\n", bp); 4336 return (1); 4337 } 4338 if (bp->b_vp != info->vp || (bp->b_flags & B_DELWRI) == 0) { 4339 kprintf("softdep_sync_metadata_bp(2): caught buf %p going away vp %p\n", bp, info->vp); 4340 BUF_UNLOCK(bp); 4341 return(1); 4342 } 4343 4344 /* 4345 * As we hold the buffer locked, none of its dependencies 4346 * will disappear. 4347 */ 4348 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 4349 switch (wk->wk_type) { 4350 4351 case D_ALLOCDIRECT: 4352 adp = WK_ALLOCDIRECT(wk); 4353 if (adp->ad_state & DEPCOMPLETE) 4354 break; 4355 nbp = adp->ad_buf; 4356 if (getdirtybuf(&nbp, info->waitfor) == 0) 4357 break; 4358 FREE_LOCK(&lk); 4359 if (info->waitfor == MNT_NOWAIT) { 4360 bawrite(nbp); 4361 } else if ((error = bwrite(nbp)) != 0) { 4362 bawrite(bp); 4363 ACQUIRE_LOCK(&lk); 4364 return (-error); 4365 } 4366 ACQUIRE_LOCK(&lk); 4367 break; 4368 4369 case D_ALLOCINDIR: 4370 aip = WK_ALLOCINDIR(wk); 4371 if (aip->ai_state & DEPCOMPLETE) 4372 break; 4373 nbp = aip->ai_buf; 4374 if (getdirtybuf(&nbp, info->waitfor) == 0) 4375 break; 4376 FREE_LOCK(&lk); 4377 if (info->waitfor == MNT_NOWAIT) { 4378 bawrite(nbp); 4379 } else if ((error = bwrite(nbp)) != 0) { 4380 bawrite(bp); 4381 ACQUIRE_LOCK(&lk); 4382 return (-error); 4383 } 4384 ACQUIRE_LOCK(&lk); 4385 break; 4386 4387 case D_INDIRDEP: 4388 restart: 4389 4390 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) { 4391 if (aip->ai_state & DEPCOMPLETE) 4392 continue; 4393 nbp = aip->ai_buf; 4394 if (getdirtybuf(&nbp, MNT_WAIT) == 0) 4395 goto restart; 4396 FREE_LOCK(&lk); 4397 if ((error = bwrite(nbp)) != 0) { 4398 bawrite(bp); 4399 ACQUIRE_LOCK(&lk); 4400 return (-error); 4401 } 4402 ACQUIRE_LOCK(&lk); 4403 goto restart; 4404 } 4405 break; 4406 4407 case D_INODEDEP: 4408 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs, 4409 WK_INODEDEP(wk)->id_ino)) != 0) { 4410 FREE_LOCK(&lk); 4411 bawrite(bp); 4412 ACQUIRE_LOCK(&lk); 4413 return (-error); 4414 } 4415 break; 4416 4417 case D_PAGEDEP: 4418 /* 4419 * We are trying to sync a directory that may 4420 * have dependencies on both its own metadata 4421 * and/or dependencies on the inodes of any 4422 * recently allocated files. We walk its diradd 4423 * lists pushing out the associated inode. 4424 */ 4425 pagedep = WK_PAGEDEP(wk); 4426 for (i = 0; i < DAHASHSZ; i++) { 4427 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 4428 continue; 4429 if ((error = 4430 flush_pagedep_deps(info->vp, 4431 pagedep->pd_mnt, 4432 &pagedep->pd_diraddhd[i]))) { 4433 FREE_LOCK(&lk); 4434 bawrite(bp); 4435 ACQUIRE_LOCK(&lk); 4436 return (-error); 4437 } 4438 } 4439 break; 4440 4441 case D_MKDIR: 4442 /* 4443 * This case should never happen if the vnode has 4444 * been properly sync'ed. However, if this function 4445 * is used at a place where the vnode has not yet 4446 * been sync'ed, this dependency can show up. So, 4447 * rather than panic, just flush it. 4448 */ 4449 nbp = WK_MKDIR(wk)->md_buf; 4450 if (getdirtybuf(&nbp, info->waitfor) == 0) 4451 break; 4452 FREE_LOCK(&lk); 4453 if (info->waitfor == MNT_NOWAIT) { 4454 bawrite(nbp); 4455 } else if ((error = bwrite(nbp)) != 0) { 4456 bawrite(bp); 4457 ACQUIRE_LOCK(&lk); 4458 return (-error); 4459 } 4460 ACQUIRE_LOCK(&lk); 4461 break; 4462 4463 case D_BMSAFEMAP: 4464 /* 4465 * This case should never happen if the vnode has 4466 * been properly sync'ed. However, if this function 4467 * is used at a place where the vnode has not yet 4468 * been sync'ed, this dependency can show up. So, 4469 * rather than panic, just flush it. 4470 * 4471 * nbp can wind up == bp if a device node for the 4472 * same filesystem is being fsynced at the same time, 4473 * leading to a panic if we don't catch the case. 4474 */ 4475 nbp = WK_BMSAFEMAP(wk)->sm_buf; 4476 if (nbp == bp) 4477 break; 4478 if (getdirtybuf(&nbp, info->waitfor) == 0) 4479 break; 4480 FREE_LOCK(&lk); 4481 if (info->waitfor == MNT_NOWAIT) { 4482 bawrite(nbp); 4483 } else if ((error = bwrite(nbp)) != 0) { 4484 bawrite(bp); 4485 ACQUIRE_LOCK(&lk); 4486 return (-error); 4487 } 4488 ACQUIRE_LOCK(&lk); 4489 break; 4490 4491 default: 4492 FREE_LOCK(&lk); 4493 panic("softdep_sync_metadata: Unknown type %s", 4494 TYPENAME(wk->wk_type)); 4495 /* NOTREACHED */ 4496 } 4497 } 4498 FREE_LOCK(&lk); 4499 bawrite(bp); 4500 ACQUIRE_LOCK(&lk); 4501 return(0); 4502 } 4503 4504 /* 4505 * Flush the dependencies associated with an inodedep. 4506 * Called with splbio blocked. 4507 */ 4508 static int 4509 flush_inodedep_deps(struct fs *fs, ino_t ino) 4510 { 4511 struct inodedep *inodedep; 4512 struct allocdirect *adp; 4513 int error, waitfor; 4514 struct buf *bp; 4515 4516 /* 4517 * This work is done in two passes. The first pass grabs most 4518 * of the buffers and begins asynchronously writing them. The 4519 * only way to wait for these asynchronous writes is to sleep 4520 * on the filesystem vnode which may stay busy for a long time 4521 * if the filesystem is active. So, instead, we make a second 4522 * pass over the dependencies blocking on each write. In the 4523 * usual case we will be blocking against a write that we 4524 * initiated, so when it is done the dependency will have been 4525 * resolved. Thus the second pass is expected to end quickly. 4526 * We give a brief window at the top of the loop to allow 4527 * any pending I/O to complete. 4528 */ 4529 for (waitfor = MNT_NOWAIT; ; ) { 4530 FREE_LOCK(&lk); 4531 ACQUIRE_LOCK(&lk); 4532 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4533 return (0); 4534 TAILQ_FOREACH(adp, &inodedep->id_inoupdt, ad_next) { 4535 if (adp->ad_state & DEPCOMPLETE) 4536 continue; 4537 bp = adp->ad_buf; 4538 if (getdirtybuf(&bp, waitfor) == 0) { 4539 if (waitfor == MNT_NOWAIT) 4540 continue; 4541 break; 4542 } 4543 FREE_LOCK(&lk); 4544 if (waitfor == MNT_NOWAIT) { 4545 bawrite(bp); 4546 } else if ((error = bwrite(bp)) != 0) { 4547 ACQUIRE_LOCK(&lk); 4548 return (error); 4549 } 4550 ACQUIRE_LOCK(&lk); 4551 break; 4552 } 4553 if (adp != NULL) 4554 continue; 4555 TAILQ_FOREACH(adp, &inodedep->id_newinoupdt, ad_next) { 4556 if (adp->ad_state & DEPCOMPLETE) 4557 continue; 4558 bp = adp->ad_buf; 4559 if (getdirtybuf(&bp, waitfor) == 0) { 4560 if (waitfor == MNT_NOWAIT) 4561 continue; 4562 break; 4563 } 4564 FREE_LOCK(&lk); 4565 if (waitfor == MNT_NOWAIT) { 4566 bawrite(bp); 4567 } else if ((error = bwrite(bp)) != 0) { 4568 ACQUIRE_LOCK(&lk); 4569 return (error); 4570 } 4571 ACQUIRE_LOCK(&lk); 4572 break; 4573 } 4574 if (adp != NULL) 4575 continue; 4576 /* 4577 * If pass2, we are done, otherwise do pass 2. 4578 */ 4579 if (waitfor == MNT_WAIT) 4580 break; 4581 waitfor = MNT_WAIT; 4582 } 4583 /* 4584 * Try freeing inodedep in case all dependencies have been removed. 4585 */ 4586 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0) 4587 (void) free_inodedep(inodedep); 4588 return (0); 4589 } 4590 4591 /* 4592 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 4593 * Called with splbio blocked. 4594 */ 4595 static int 4596 flush_pagedep_deps(struct vnode *pvp, struct mount *mp, 4597 struct diraddhd *diraddhdp) 4598 { 4599 struct inodedep *inodedep; 4600 struct ufsmount *ump; 4601 struct diradd *dap; 4602 struct vnode *vp; 4603 int gotit, error = 0; 4604 struct buf *bp; 4605 ino_t inum; 4606 4607 ump = VFSTOUFS(mp); 4608 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 4609 /* 4610 * Flush ourselves if this directory entry 4611 * has a MKDIR_PARENT dependency. 4612 */ 4613 if (dap->da_state & MKDIR_PARENT) { 4614 FREE_LOCK(&lk); 4615 if ((error = ffs_update(pvp, 1)) != 0) 4616 break; 4617 ACQUIRE_LOCK(&lk); 4618 /* 4619 * If that cleared dependencies, go on to next. 4620 */ 4621 if (dap != LIST_FIRST(diraddhdp)) 4622 continue; 4623 if (dap->da_state & MKDIR_PARENT) { 4624 FREE_LOCK(&lk); 4625 panic("flush_pagedep_deps: MKDIR_PARENT"); 4626 } 4627 } 4628 /* 4629 * A newly allocated directory must have its "." and 4630 * ".." entries written out before its name can be 4631 * committed in its parent. We do not want or need 4632 * the full semantics of a synchronous VOP_FSYNC as 4633 * that may end up here again, once for each directory 4634 * level in the filesystem. Instead, we push the blocks 4635 * and wait for them to clear. We have to fsync twice 4636 * because the first call may choose to defer blocks 4637 * that still have dependencies, but deferral will 4638 * happen at most once. 4639 */ 4640 inum = dap->da_newinum; 4641 if (dap->da_state & MKDIR_BODY) { 4642 FREE_LOCK(&lk); 4643 if ((error = VFS_VGET(mp, inum, &vp)) != 0) 4644 break; 4645 if ((error=VOP_FSYNC(vp, MNT_NOWAIT)) || 4646 (error=VOP_FSYNC(vp, MNT_NOWAIT))) { 4647 vput(vp); 4648 break; 4649 } 4650 drain_output(vp, 0); 4651 vput(vp); 4652 ACQUIRE_LOCK(&lk); 4653 /* 4654 * If that cleared dependencies, go on to next. 4655 */ 4656 if (dap != LIST_FIRST(diraddhdp)) 4657 continue; 4658 if (dap->da_state & MKDIR_BODY) { 4659 FREE_LOCK(&lk); 4660 panic("flush_pagedep_deps: MKDIR_BODY"); 4661 } 4662 } 4663 /* 4664 * Flush the inode on which the directory entry depends. 4665 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 4666 * the only remaining dependency is that the updated inode 4667 * count must get pushed to disk. The inode has already 4668 * been pushed into its inode buffer (via VOP_UPDATE) at 4669 * the time of the reference count change. So we need only 4670 * locate that buffer, ensure that there will be no rollback 4671 * caused by a bitmap dependency, then write the inode buffer. 4672 */ 4673 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) { 4674 FREE_LOCK(&lk); 4675 panic("flush_pagedep_deps: lost inode"); 4676 } 4677 /* 4678 * If the inode still has bitmap dependencies, 4679 * push them to disk. 4680 */ 4681 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4682 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 4683 FREE_LOCK(&lk); 4684 if (gotit && (error = bwrite(inodedep->id_buf)) != 0) 4685 break; 4686 ACQUIRE_LOCK(&lk); 4687 if (dap != LIST_FIRST(diraddhdp)) 4688 continue; 4689 } 4690 /* 4691 * If the inode is still sitting in a buffer waiting 4692 * to be written, push it to disk. 4693 */ 4694 FREE_LOCK(&lk); 4695 if ((error = bread(ump->um_devvp, 4696 fsbtodoff(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), 4697 (int)ump->um_fs->fs_bsize, &bp)) != 0) 4698 break; 4699 if ((error = bwrite(bp)) != 0) 4700 break; 4701 ACQUIRE_LOCK(&lk); 4702 /* 4703 * If we have failed to get rid of all the dependencies 4704 * then something is seriously wrong. 4705 */ 4706 if (dap == LIST_FIRST(diraddhdp)) { 4707 FREE_LOCK(&lk); 4708 panic("flush_pagedep_deps: flush failed"); 4709 } 4710 } 4711 if (error) 4712 ACQUIRE_LOCK(&lk); 4713 return (error); 4714 } 4715 4716 /* 4717 * A large burst of file addition or deletion activity can drive the 4718 * memory load excessively high. First attempt to slow things down 4719 * using the techniques below. If that fails, this routine requests 4720 * the offending operations to fall back to running synchronously 4721 * until the memory load returns to a reasonable level. 4722 */ 4723 int 4724 softdep_slowdown(struct vnode *vp) 4725 { 4726 int max_softdeps_hard; 4727 4728 max_softdeps_hard = max_softdeps * 11 / 10; 4729 if (num_dirrem < max_softdeps_hard / 2 && 4730 num_inodedep < max_softdeps_hard) 4731 return (0); 4732 stat_sync_limit_hit += 1; 4733 return (1); 4734 } 4735 4736 /* 4737 * If memory utilization has gotten too high, deliberately slow things 4738 * down and speed up the I/O processing. 4739 */ 4740 static int 4741 request_cleanup(int resource, int islocked) 4742 { 4743 struct thread *td = curthread; /* XXX */ 4744 4745 /* 4746 * We never hold up the filesystem syncer process. 4747 */ 4748 if (td == filesys_syncer) 4749 return (0); 4750 /* 4751 * First check to see if the work list has gotten backlogged. 4752 * If it has, co-opt this process to help clean up two entries. 4753 * Because this process may hold inodes locked, we cannot 4754 * handle any remove requests that might block on a locked 4755 * inode as that could lead to deadlock. 4756 */ 4757 if (num_on_worklist > max_softdeps / 10) { 4758 if (islocked) 4759 FREE_LOCK(&lk); 4760 process_worklist_item(NULL, LK_NOWAIT); 4761 process_worklist_item(NULL, LK_NOWAIT); 4762 stat_worklist_push += 2; 4763 if (islocked) 4764 ACQUIRE_LOCK(&lk); 4765 return(1); 4766 } 4767 4768 /* 4769 * If we are resource constrained on inode dependencies, try 4770 * flushing some dirty inodes. Otherwise, we are constrained 4771 * by file deletions, so try accelerating flushes of directories 4772 * with removal dependencies. We would like to do the cleanup 4773 * here, but we probably hold an inode locked at this point and 4774 * that might deadlock against one that we try to clean. So, 4775 * the best that we can do is request the syncer daemon to do 4776 * the cleanup for us. 4777 */ 4778 switch (resource) { 4779 4780 case FLUSH_INODES: 4781 stat_ino_limit_push += 1; 4782 req_clear_inodedeps += 1; 4783 stat_countp = &stat_ino_limit_hit; 4784 break; 4785 4786 case FLUSH_REMOVE: 4787 stat_blk_limit_push += 1; 4788 req_clear_remove += 1; 4789 stat_countp = &stat_blk_limit_hit; 4790 break; 4791 4792 default: 4793 if (islocked) 4794 FREE_LOCK(&lk); 4795 panic("request_cleanup: unknown type"); 4796 } 4797 /* 4798 * Hopefully the syncer daemon will catch up and awaken us. 4799 * We wait at most tickdelay before proceeding in any case. 4800 */ 4801 if (islocked == 0) 4802 ACQUIRE_LOCK(&lk); 4803 proc_waiting += 1; 4804 if (!callout_active(&handle)) 4805 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2, 4806 pause_timer, NULL); 4807 interlocked_sleep(&lk, SLEEP, (caddr_t)&proc_waiting, 0, 4808 "softupdate", 0); 4809 proc_waiting -= 1; 4810 if (islocked == 0) 4811 FREE_LOCK(&lk); 4812 return (1); 4813 } 4814 4815 /* 4816 * Awaken processes pausing in request_cleanup and clear proc_waiting 4817 * to indicate that there is no longer a timer running. 4818 */ 4819 void 4820 pause_timer(void *arg) 4821 { 4822 *stat_countp += 1; 4823 wakeup_one(&proc_waiting); 4824 if (proc_waiting > 0) 4825 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2, 4826 pause_timer, NULL); 4827 else 4828 callout_deactivate(&handle); 4829 } 4830 4831 /* 4832 * Flush out a directory with at least one removal dependency in an effort to 4833 * reduce the number of dirrem, freefile, and freeblks dependency structures. 4834 */ 4835 static void 4836 clear_remove(struct thread *td) 4837 { 4838 struct pagedep_hashhead *pagedephd; 4839 struct pagedep *pagedep; 4840 static int next = 0; 4841 struct mount *mp; 4842 struct vnode *vp; 4843 int error, cnt; 4844 ino_t ino; 4845 4846 ACQUIRE_LOCK(&lk); 4847 for (cnt = 0; cnt < pagedep_hash; cnt++) { 4848 pagedephd = &pagedep_hashtbl[next++]; 4849 if (next >= pagedep_hash) 4850 next = 0; 4851 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 4852 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL) 4853 continue; 4854 mp = pagedep->pd_mnt; 4855 ino = pagedep->pd_ino; 4856 FREE_LOCK(&lk); 4857 if ((error = VFS_VGET(mp, ino, &vp)) != 0) { 4858 softdep_error("clear_remove: vget", error); 4859 return; 4860 } 4861 if ((error = VOP_FSYNC(vp, MNT_NOWAIT))) 4862 softdep_error("clear_remove: fsync", error); 4863 drain_output(vp, 0); 4864 vput(vp); 4865 return; 4866 } 4867 } 4868 FREE_LOCK(&lk); 4869 } 4870 4871 /* 4872 * Clear out a block of dirty inodes in an effort to reduce 4873 * the number of inodedep dependency structures. 4874 */ 4875 struct clear_inodedeps_info { 4876 struct fs *fs; 4877 struct mount *mp; 4878 }; 4879 4880 static int 4881 clear_inodedeps_mountlist_callback(struct mount *mp, void *data) 4882 { 4883 struct clear_inodedeps_info *info = data; 4884 4885 if ((mp->mnt_flag & MNT_SOFTDEP) && info->fs == VFSTOUFS(mp)->um_fs) { 4886 info->mp = mp; 4887 return(-1); 4888 } 4889 return(0); 4890 } 4891 4892 static void 4893 clear_inodedeps(struct thread *td) 4894 { 4895 struct clear_inodedeps_info info; 4896 struct inodedep_hashhead *inodedephd; 4897 struct inodedep *inodedep; 4898 static int next = 0; 4899 struct vnode *vp; 4900 struct fs *fs; 4901 int error, cnt; 4902 ino_t firstino, lastino, ino; 4903 4904 ACQUIRE_LOCK(&lk); 4905 /* 4906 * Pick a random inode dependency to be cleared. 4907 * We will then gather up all the inodes in its block 4908 * that have dependencies and flush them out. 4909 */ 4910 for (cnt = 0; cnt < inodedep_hash; cnt++) { 4911 inodedephd = &inodedep_hashtbl[next++]; 4912 if (next >= inodedep_hash) 4913 next = 0; 4914 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 4915 break; 4916 } 4917 if (inodedep == NULL) { 4918 FREE_LOCK(&lk); 4919 return; 4920 } 4921 /* 4922 * Ugly code to find mount point given pointer to superblock. 4923 */ 4924 fs = inodedep->id_fs; 4925 info.mp = NULL; 4926 info.fs = fs; 4927 mountlist_scan(clear_inodedeps_mountlist_callback, 4928 &info, MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 4929 /* 4930 * Find the last inode in the block with dependencies. 4931 */ 4932 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 4933 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 4934 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0) 4935 break; 4936 /* 4937 * Asynchronously push all but the last inode with dependencies. 4938 * Synchronously push the last inode with dependencies to ensure 4939 * that the inode block gets written to free up the inodedeps. 4940 */ 4941 for (ino = firstino; ino <= lastino; ino++) { 4942 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4943 continue; 4944 FREE_LOCK(&lk); 4945 if ((error = VFS_VGET(info.mp, ino, &vp)) != 0) { 4946 softdep_error("clear_inodedeps: vget", error); 4947 return; 4948 } 4949 if (ino == lastino) { 4950 if ((error = VOP_FSYNC(vp, MNT_WAIT))) 4951 softdep_error("clear_inodedeps: fsync1", error); 4952 } else { 4953 if ((error = VOP_FSYNC(vp, MNT_NOWAIT))) 4954 softdep_error("clear_inodedeps: fsync2", error); 4955 drain_output(vp, 0); 4956 } 4957 vput(vp); 4958 ACQUIRE_LOCK(&lk); 4959 } 4960 FREE_LOCK(&lk); 4961 } 4962 4963 /* 4964 * Function to determine if the buffer has outstanding dependencies 4965 * that will cause a roll-back if the buffer is written. If wantcount 4966 * is set, return number of dependencies, otherwise just yes or no. 4967 */ 4968 static int 4969 softdep_count_dependencies(struct buf *bp, int wantcount) 4970 { 4971 struct worklist *wk; 4972 struct inodedep *inodedep; 4973 struct indirdep *indirdep; 4974 struct allocindir *aip; 4975 struct pagedep *pagedep; 4976 struct diradd *dap; 4977 int i, retval; 4978 4979 retval = 0; 4980 ACQUIRE_LOCK(&lk); 4981 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 4982 switch (wk->wk_type) { 4983 4984 case D_INODEDEP: 4985 inodedep = WK_INODEDEP(wk); 4986 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4987 /* bitmap allocation dependency */ 4988 retval += 1; 4989 if (!wantcount) 4990 goto out; 4991 } 4992 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 4993 /* direct block pointer dependency */ 4994 retval += 1; 4995 if (!wantcount) 4996 goto out; 4997 } 4998 continue; 4999 5000 case D_INDIRDEP: 5001 indirdep = WK_INDIRDEP(wk); 5002 5003 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 5004 /* indirect block pointer dependency */ 5005 retval += 1; 5006 if (!wantcount) 5007 goto out; 5008 } 5009 continue; 5010 5011 case D_PAGEDEP: 5012 pagedep = WK_PAGEDEP(wk); 5013 for (i = 0; i < DAHASHSZ; i++) { 5014 5015 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 5016 /* directory entry dependency */ 5017 retval += 1; 5018 if (!wantcount) 5019 goto out; 5020 } 5021 } 5022 continue; 5023 5024 case D_BMSAFEMAP: 5025 case D_ALLOCDIRECT: 5026 case D_ALLOCINDIR: 5027 case D_MKDIR: 5028 /* never a dependency on these blocks */ 5029 continue; 5030 5031 default: 5032 FREE_LOCK(&lk); 5033 panic("softdep_check_for_rollback: Unexpected type %s", 5034 TYPENAME(wk->wk_type)); 5035 /* NOTREACHED */ 5036 } 5037 } 5038 out: 5039 FREE_LOCK(&lk); 5040 return retval; 5041 } 5042 5043 /* 5044 * Acquire exclusive access to a buffer. 5045 * Must be called with splbio blocked. 5046 * Return 1 if buffer was acquired. 5047 */ 5048 static int 5049 getdirtybuf(struct buf **bpp, int waitfor) 5050 { 5051 struct buf *bp; 5052 int error; 5053 5054 for (;;) { 5055 if ((bp = *bpp) == NULL) 5056 return (0); 5057 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) 5058 break; 5059 if (waitfor != MNT_WAIT) 5060 return (0); 5061 error = interlocked_sleep(&lk, LOCKBUF, bp, 5062 LK_EXCLUSIVE | LK_SLEEPFAIL, 0, 0); 5063 if (error != ENOLCK) { 5064 FREE_LOCK(&lk); 5065 panic("getdirtybuf: inconsistent lock"); 5066 } 5067 } 5068 if ((bp->b_flags & B_DELWRI) == 0) { 5069 BUF_UNLOCK(bp); 5070 return (0); 5071 } 5072 bremfree(bp); 5073 return (1); 5074 } 5075 5076 /* 5077 * Wait for pending output on a vnode to complete. 5078 * Must be called with vnode locked. 5079 */ 5080 static void 5081 drain_output(struct vnode *vp, int islocked) 5082 { 5083 5084 if (!islocked) 5085 ACQUIRE_LOCK(&lk); 5086 while (vp->v_track_write.bk_active) { 5087 vp->v_track_write.bk_waitflag = 1; 5088 interlocked_sleep(&lk, SLEEP, &vp->v_track_write, 5089 0, "drainvp", 0); 5090 } 5091 if (!islocked) 5092 FREE_LOCK(&lk); 5093 } 5094 5095 /* 5096 * Called whenever a buffer that is being invalidated or reallocated 5097 * contains dependencies. This should only happen if an I/O error has 5098 * occurred. The routine is called with the buffer locked. 5099 */ 5100 static void 5101 softdep_deallocate_dependencies(struct buf *bp) 5102 { 5103 5104 if ((bp->b_flags & B_ERROR) == 0) 5105 panic("softdep_deallocate_dependencies: dangling deps"); 5106 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntfromname, bp->b_error); 5107 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 5108 } 5109 5110 /* 5111 * Function to handle asynchronous write errors in the filesystem. 5112 */ 5113 void 5114 softdep_error(char *func, int error) 5115 { 5116 5117 /* XXX should do something better! */ 5118 kprintf("%s: got error %d while accessing filesystem\n", func, error); 5119 } 5120