1 /* 2 * Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved. 3 * 4 * The soft updates code is derived from the appendix of a University 5 * of Michigan technical report (Gregory R. Ganger and Yale N. Patt, 6 * "Soft Updates: A Solution to the Metadata Update Problem in File 7 * Systems", CSE-TR-254-95, August 1995). 8 * 9 * Further information about soft updates can be obtained from: 10 * 11 * Marshall Kirk McKusick http://www.mckusick.com/softdep/ 12 * 1614 Oxford Street mckusick@mckusick.com 13 * Berkeley, CA 94709-1608 +1-510-843-9542 14 * USA 15 * 16 * Redistribution and use in source and binary forms, with or without 17 * modification, are permitted provided that the following conditions 18 * are met: 19 * 20 * 1. Redistributions of source code must retain the above copyright 21 * notice, this list of conditions and the following disclaimer. 22 * 2. Redistributions in binary form must reproduce the above copyright 23 * notice, this list of conditions and the following disclaimer in the 24 * documentation and/or other materials provided with the distribution. 25 * 26 * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY 27 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 28 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 29 * DISCLAIMED. IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR 30 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * from: @(#)ffs_softdep.c 9.59 (McKusick) 6/21/00 39 * $FreeBSD: src/sys/ufs/ffs/ffs_softdep.c,v 1.57.2.11 2002/02/05 18:46:53 dillon Exp $ 40 * $DragonFly: src/sys/vfs/ufs/ffs_softdep.c,v 1.57 2008/06/28 17:59:51 dillon Exp $ 41 */ 42 43 /* 44 * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide. 45 */ 46 #ifndef DIAGNOSTIC 47 #define DIAGNOSTIC 48 #endif 49 #ifndef DEBUG 50 #define DEBUG 51 #endif 52 53 #include <sys/param.h> 54 #include <sys/kernel.h> 55 #include <sys/systm.h> 56 #include <sys/buf.h> 57 #include <sys/malloc.h> 58 #include <sys/mount.h> 59 #include <sys/proc.h> 60 #include <sys/syslog.h> 61 #include <sys/vnode.h> 62 #include <sys/conf.h> 63 #include <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 static int softdep_checkread(struct buf *bp); 208 static int softdep_checkwrite(struct buf *bp); 209 210 static struct bio_ops softdep_bioops = { 211 .io_start = softdep_disk_io_initiation, 212 .io_complete = softdep_disk_write_complete, 213 .io_deallocate = softdep_deallocate_dependencies, 214 .io_fsync = softdep_fsync, 215 .io_sync = softdep_process_worklist, 216 .io_movedeps = softdep_move_dependencies, 217 .io_countdeps = softdep_count_dependencies, 218 .io_checkread = softdep_checkread, 219 .io_checkwrite = softdep_checkwrite 220 }; 221 222 /* 223 * Locking primitives. 224 * 225 * For a uniprocessor, all we need to do is protect against disk 226 * interrupts. For a multiprocessor, this lock would have to be 227 * a mutex. A single mutex is used throughout this file, though 228 * finer grain locking could be used if contention warranted it. 229 * 230 * For a multiprocessor, the sleep call would accept a lock and 231 * release it after the sleep processing was complete. In a uniprocessor 232 * implementation there is no such interlock, so we simple mark 233 * the places where it needs to be done with the `interlocked' form 234 * of the lock calls. Since the uniprocessor sleep already interlocks 235 * the spl, there is nothing that really needs to be done. 236 */ 237 #ifndef /* NOT */ DEBUG 238 static struct lockit { 239 } lk = { 0 }; 240 #define ACQUIRE_LOCK(lk) crit_enter_id("softupdates"); 241 #define FREE_LOCK(lk) crit_exit_id("softupdates"); 242 243 #else /* DEBUG */ 244 #define NOHOLDER ((struct thread *)-1) 245 #define SPECIAL_FLAG ((struct thread *)-2) 246 static struct lockit { 247 int lkt_spl; 248 struct thread *lkt_held; 249 } lk = { 0, NOHOLDER }; 250 static int lockcnt; 251 252 static void acquire_lock(struct lockit *); 253 static void free_lock(struct lockit *); 254 void softdep_panic(char *); 255 256 #define ACQUIRE_LOCK(lk) acquire_lock(lk) 257 #define FREE_LOCK(lk) free_lock(lk) 258 259 static void 260 acquire_lock(struct lockit *lk) 261 { 262 thread_t holder; 263 264 if (lk->lkt_held != NOHOLDER) { 265 holder = lk->lkt_held; 266 FREE_LOCK(lk); 267 if (holder == curthread) 268 panic("softdep_lock: locking against myself"); 269 else 270 panic("softdep_lock: lock held by %p", holder); 271 } 272 crit_enter_id("softupdates"); 273 lk->lkt_held = curthread; 274 lockcnt++; 275 } 276 277 static void 278 free_lock(struct lockit *lk) 279 { 280 281 if (lk->lkt_held == NOHOLDER) 282 panic("softdep_unlock: lock not held"); 283 lk->lkt_held = NOHOLDER; 284 crit_exit_id("softupdates"); 285 } 286 287 /* 288 * Function to release soft updates lock and panic. 289 */ 290 void 291 softdep_panic(char *msg) 292 { 293 294 if (lk.lkt_held != NOHOLDER) 295 FREE_LOCK(&lk); 296 panic(msg); 297 } 298 #endif /* DEBUG */ 299 300 static int interlocked_sleep(struct lockit *, int, void *, int, 301 const char *, int); 302 303 /* 304 * When going to sleep, we must save our SPL so that it does 305 * not get lost if some other process uses the lock while we 306 * are sleeping. We restore it after we have slept. This routine 307 * wraps the interlocking with functions that sleep. The list 308 * below enumerates the available set of operations. 309 */ 310 #define UNKNOWN 0 311 #define SLEEP 1 312 #define LOCKBUF 2 313 314 static int 315 interlocked_sleep(struct lockit *lk, int op, void *ident, int flags, 316 const char *wmesg, int timo) 317 { 318 thread_t holder; 319 int s, retval; 320 321 s = lk->lkt_spl; 322 # ifdef DEBUG 323 if (lk->lkt_held == NOHOLDER) 324 panic("interlocked_sleep: lock not held"); 325 lk->lkt_held = NOHOLDER; 326 # endif /* DEBUG */ 327 switch (op) { 328 case SLEEP: 329 retval = tsleep(ident, flags, wmesg, timo); 330 break; 331 case LOCKBUF: 332 retval = BUF_LOCK((struct buf *)ident, flags); 333 break; 334 default: 335 panic("interlocked_sleep: unknown operation"); 336 } 337 # ifdef DEBUG 338 if (lk->lkt_held != NOHOLDER) { 339 holder = lk->lkt_held; 340 FREE_LOCK(lk); 341 if (holder == curthread) 342 panic("interlocked_sleep: locking against self"); 343 else 344 panic("interlocked_sleep: lock held by %p", holder); 345 } 346 lk->lkt_held = curthread; 347 lockcnt++; 348 # endif /* DEBUG */ 349 lk->lkt_spl = s; 350 return (retval); 351 } 352 353 /* 354 * Place holder for real semaphores. 355 */ 356 struct sema { 357 int value; 358 thread_t holder; 359 char *name; 360 int prio; 361 int timo; 362 }; 363 static void sema_init(struct sema *, char *, int, int); 364 static int sema_get(struct sema *, struct lockit *); 365 static void sema_release(struct sema *); 366 367 static void 368 sema_init(struct sema *semap, char *name, int prio, int timo) 369 { 370 371 semap->holder = NOHOLDER; 372 semap->value = 0; 373 semap->name = name; 374 semap->prio = prio; 375 semap->timo = timo; 376 } 377 378 static int 379 sema_get(struct sema *semap, struct lockit *interlock) 380 { 381 382 if (semap->value++ > 0) { 383 if (interlock != NULL) { 384 interlocked_sleep(interlock, SLEEP, (caddr_t)semap, 385 semap->prio, semap->name, semap->timo); 386 FREE_LOCK(interlock); 387 } else { 388 tsleep((caddr_t)semap, semap->prio, semap->name, 389 semap->timo); 390 } 391 return (0); 392 } 393 semap->holder = curthread; 394 if (interlock != NULL) 395 FREE_LOCK(interlock); 396 return (1); 397 } 398 399 static void 400 sema_release(struct sema *semap) 401 { 402 403 if (semap->value <= 0 || semap->holder != curthread) { 404 if (lk.lkt_held != NOHOLDER) 405 FREE_LOCK(&lk); 406 panic("sema_release: not held"); 407 } 408 if (--semap->value > 0) { 409 semap->value = 0; 410 wakeup(semap); 411 } 412 semap->holder = NOHOLDER; 413 } 414 415 /* 416 * Worklist queue management. 417 * These routines require that the lock be held. 418 */ 419 #ifndef /* NOT */ DEBUG 420 #define WORKLIST_INSERT(head, item) do { \ 421 (item)->wk_state |= ONWORKLIST; \ 422 LIST_INSERT_HEAD(head, item, wk_list); \ 423 } while (0) 424 425 #define WORKLIST_INSERT_BP(bp, item) do { \ 426 (item)->wk_state |= ONWORKLIST; \ 427 (bp)->b_ops = &softdep_bioops; \ 428 LIST_INSERT_HEAD(&(bp)->b_dep, item, wk_list); \ 429 } while (0) 430 431 #define WORKLIST_REMOVE(item) do { \ 432 (item)->wk_state &= ~ONWORKLIST; \ 433 LIST_REMOVE(item, wk_list); \ 434 } while (0) 435 436 #define WORKITEM_FREE(item, type) FREE(item, DtoM(type)) 437 438 #else /* DEBUG */ 439 static void worklist_insert(struct workhead *, struct worklist *); 440 static void worklist_remove(struct worklist *); 441 static void workitem_free(struct worklist *, int); 442 443 #define WORKLIST_INSERT_BP(bp, item) do { \ 444 (bp)->b_ops = &softdep_bioops; \ 445 worklist_insert(&(bp)->b_dep, item); \ 446 } while (0) 447 448 #define WORKLIST_INSERT(head, item) worklist_insert(head, item) 449 #define WORKLIST_REMOVE(item) worklist_remove(item) 450 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type) 451 452 static void 453 worklist_insert(struct workhead *head, struct worklist *item) 454 { 455 456 if (lk.lkt_held == NOHOLDER) 457 panic("worklist_insert: lock not held"); 458 if (item->wk_state & ONWORKLIST) { 459 FREE_LOCK(&lk); 460 panic("worklist_insert: already on list"); 461 } 462 item->wk_state |= ONWORKLIST; 463 LIST_INSERT_HEAD(head, item, wk_list); 464 } 465 466 static void 467 worklist_remove(struct worklist *item) 468 { 469 470 if (lk.lkt_held == NOHOLDER) 471 panic("worklist_remove: lock not held"); 472 if ((item->wk_state & ONWORKLIST) == 0) { 473 FREE_LOCK(&lk); 474 panic("worklist_remove: not on list"); 475 } 476 item->wk_state &= ~ONWORKLIST; 477 LIST_REMOVE(item, wk_list); 478 } 479 480 static void 481 workitem_free(struct worklist *item, int type) 482 { 483 484 if (item->wk_state & ONWORKLIST) { 485 if (lk.lkt_held != NOHOLDER) 486 FREE_LOCK(&lk); 487 panic("workitem_free: still on list"); 488 } 489 if (item->wk_type != type) { 490 if (lk.lkt_held != NOHOLDER) 491 FREE_LOCK(&lk); 492 panic("workitem_free: type mismatch"); 493 } 494 FREE(item, DtoM(type)); 495 } 496 #endif /* DEBUG */ 497 498 /* 499 * Workitem queue management 500 */ 501 static struct workhead softdep_workitem_pending; 502 static int num_on_worklist; /* number of worklist items to be processed */ 503 static int softdep_worklist_busy; /* 1 => trying to do unmount */ 504 static int softdep_worklist_req; /* serialized waiters */ 505 static int max_softdeps; /* maximum number of structs before slowdown */ 506 static int tickdelay = 2; /* number of ticks to pause during slowdown */ 507 static int *stat_countp; /* statistic to count in proc_waiting timeout */ 508 static int proc_waiting; /* tracks whether we have a timeout posted */ 509 static struct callout handle; /* handle on posted proc_waiting timeout */ 510 static struct thread *filesys_syncer; /* proc of filesystem syncer process */ 511 static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 512 #define FLUSH_INODES 1 513 static int req_clear_remove; /* syncer process flush some freeblks */ 514 #define FLUSH_REMOVE 2 515 /* 516 * runtime statistics 517 */ 518 static int stat_worklist_push; /* number of worklist cleanups */ 519 static int stat_blk_limit_push; /* number of times block limit neared */ 520 static int stat_ino_limit_push; /* number of times inode limit neared */ 521 static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 522 static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 523 static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 524 static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 525 static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 526 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 527 static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 528 #ifdef DEBUG 529 #include <vm/vm.h> 530 #include <sys/sysctl.h> 531 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, ""); 532 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, ""); 533 SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,""); 534 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,""); 535 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,""); 536 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, ""); 537 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, ""); 538 SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, ""); 539 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, ""); 540 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, ""); 541 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, ""); 542 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, ""); 543 #endif /* DEBUG */ 544 545 /* 546 * Add an item to the end of the work queue. 547 * This routine requires that the lock be held. 548 * This is the only routine that adds items to the list. 549 * The following routine is the only one that removes items 550 * and does so in order from first to last. 551 */ 552 static void 553 add_to_worklist(struct worklist *wk) 554 { 555 static struct worklist *worklist_tail; 556 557 if (wk->wk_state & ONWORKLIST) { 558 if (lk.lkt_held != NOHOLDER) 559 FREE_LOCK(&lk); 560 panic("add_to_worklist: already on list"); 561 } 562 wk->wk_state |= ONWORKLIST; 563 if (LIST_FIRST(&softdep_workitem_pending) == NULL) 564 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list); 565 else 566 LIST_INSERT_AFTER(worklist_tail, wk, wk_list); 567 worklist_tail = wk; 568 num_on_worklist += 1; 569 } 570 571 /* 572 * Process that runs once per second to handle items in the background queue. 573 * 574 * Note that we ensure that everything is done in the order in which they 575 * appear in the queue. The code below depends on this property to ensure 576 * that blocks of a file are freed before the inode itself is freed. This 577 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 578 * until all the old ones have been purged from the dependency lists. 579 */ 580 static int 581 softdep_process_worklist(struct mount *matchmnt) 582 { 583 thread_t td = curthread; 584 int matchcnt, loopcount; 585 long starttime; 586 587 /* 588 * Record the process identifier of our caller so that we can give 589 * this process preferential treatment in request_cleanup below. 590 */ 591 filesys_syncer = td; 592 matchcnt = 0; 593 594 /* 595 * There is no danger of having multiple processes run this 596 * code, but we have to single-thread it when softdep_flushfiles() 597 * is in operation to get an accurate count of the number of items 598 * related to its mount point that are in the list. 599 */ 600 if (matchmnt == NULL) { 601 if (softdep_worklist_busy < 0) 602 return(-1); 603 softdep_worklist_busy += 1; 604 } 605 606 /* 607 * If requested, try removing inode or removal dependencies. 608 */ 609 if (req_clear_inodedeps) { 610 clear_inodedeps(td); 611 req_clear_inodedeps -= 1; 612 wakeup_one(&proc_waiting); 613 } 614 if (req_clear_remove) { 615 clear_remove(td); 616 req_clear_remove -= 1; 617 wakeup_one(&proc_waiting); 618 } 619 loopcount = 1; 620 starttime = time_second; 621 while (num_on_worklist > 0) { 622 matchcnt += process_worklist_item(matchmnt, 0); 623 624 /* 625 * If a umount operation wants to run the worklist 626 * accurately, abort. 627 */ 628 if (softdep_worklist_req && matchmnt == NULL) { 629 matchcnt = -1; 630 break; 631 } 632 633 /* 634 * If requested, try removing inode or removal dependencies. 635 */ 636 if (req_clear_inodedeps) { 637 clear_inodedeps(td); 638 req_clear_inodedeps -= 1; 639 wakeup_one(&proc_waiting); 640 } 641 if (req_clear_remove) { 642 clear_remove(td); 643 req_clear_remove -= 1; 644 wakeup_one(&proc_waiting); 645 } 646 /* 647 * We do not generally want to stop for buffer space, but if 648 * we are really being a buffer hog, we will stop and wait. 649 */ 650 if (loopcount++ % 128 == 0) 651 bwillinode(1); 652 /* 653 * Never allow processing to run for more than one 654 * second. Otherwise the other syncer tasks may get 655 * excessively backlogged. 656 */ 657 if (starttime != time_second && matchmnt == NULL) { 658 matchcnt = -1; 659 break; 660 } 661 } 662 if (matchmnt == NULL) { 663 --softdep_worklist_busy; 664 if (softdep_worklist_req && softdep_worklist_busy == 0) 665 wakeup(&softdep_worklist_req); 666 } 667 return (matchcnt); 668 } 669 670 /* 671 * Process one item on the worklist. 672 */ 673 static int 674 process_worklist_item(struct mount *matchmnt, int flags) 675 { 676 struct worklist *wk; 677 struct dirrem *dirrem; 678 struct fs *matchfs; 679 struct vnode *vp; 680 int matchcnt = 0; 681 682 matchfs = NULL; 683 if (matchmnt != NULL) 684 matchfs = VFSTOUFS(matchmnt)->um_fs; 685 ACQUIRE_LOCK(&lk); 686 /* 687 * Normally we just process each item on the worklist in order. 688 * However, if we are in a situation where we cannot lock any 689 * inodes, we have to skip over any dirrem requests whose 690 * vnodes are resident and locked. 691 */ 692 LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) { 693 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM) 694 break; 695 dirrem = WK_DIRREM(wk); 696 vp = ufs_ihashlookup(VFSTOUFS(dirrem->dm_mnt)->um_dev, 697 dirrem->dm_oldinum); 698 if (vp == NULL || !vn_islocked(vp)) 699 break; 700 } 701 if (wk == 0) { 702 FREE_LOCK(&lk); 703 return (0); 704 } 705 WORKLIST_REMOVE(wk); 706 num_on_worklist -= 1; 707 FREE_LOCK(&lk); 708 switch (wk->wk_type) { 709 710 case D_DIRREM: 711 /* removal of a directory entry */ 712 if (WK_DIRREM(wk)->dm_mnt == matchmnt) 713 matchcnt += 1; 714 handle_workitem_remove(WK_DIRREM(wk)); 715 break; 716 717 case D_FREEBLKS: 718 /* releasing blocks and/or fragments from a file */ 719 if (WK_FREEBLKS(wk)->fb_fs == matchfs) 720 matchcnt += 1; 721 handle_workitem_freeblocks(WK_FREEBLKS(wk)); 722 break; 723 724 case D_FREEFRAG: 725 /* releasing a fragment when replaced as a file grows */ 726 if (WK_FREEFRAG(wk)->ff_fs == matchfs) 727 matchcnt += 1; 728 handle_workitem_freefrag(WK_FREEFRAG(wk)); 729 break; 730 731 case D_FREEFILE: 732 /* releasing an inode when its link count drops to 0 */ 733 if (WK_FREEFILE(wk)->fx_fs == matchfs) 734 matchcnt += 1; 735 handle_workitem_freefile(WK_FREEFILE(wk)); 736 break; 737 738 default: 739 panic("%s_process_worklist: Unknown type %s", 740 "softdep", TYPENAME(wk->wk_type)); 741 /* NOTREACHED */ 742 } 743 return (matchcnt); 744 } 745 746 /* 747 * Move dependencies from one buffer to another. 748 */ 749 static void 750 softdep_move_dependencies(struct buf *oldbp, struct buf *newbp) 751 { 752 struct worklist *wk, *wktail; 753 754 if (LIST_FIRST(&newbp->b_dep) != NULL) 755 panic("softdep_move_dependencies: need merge code"); 756 wktail = NULL; 757 ACQUIRE_LOCK(&lk); 758 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 759 LIST_REMOVE(wk, wk_list); 760 if (wktail == NULL) 761 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 762 else 763 LIST_INSERT_AFTER(wktail, wk, wk_list); 764 wktail = wk; 765 newbp->b_ops = &softdep_bioops; 766 } 767 FREE_LOCK(&lk); 768 } 769 770 /* 771 * Purge the work list of all items associated with a particular mount point. 772 */ 773 int 774 softdep_flushfiles(struct mount *oldmnt, int flags) 775 { 776 struct vnode *devvp; 777 int error, loopcnt; 778 779 /* 780 * Await our turn to clear out the queue, then serialize access. 781 */ 782 while (softdep_worklist_busy != 0) { 783 softdep_worklist_req += 1; 784 tsleep(&softdep_worklist_req, 0, "softflush", 0); 785 softdep_worklist_req -= 1; 786 } 787 softdep_worklist_busy = -1; 788 789 if ((error = ffs_flushfiles(oldmnt, flags)) != 0) { 790 softdep_worklist_busy = 0; 791 if (softdep_worklist_req) 792 wakeup(&softdep_worklist_req); 793 return (error); 794 } 795 /* 796 * Alternately flush the block device associated with the mount 797 * point and process any dependencies that the flushing 798 * creates. In theory, this loop can happen at most twice, 799 * but we give it a few extra just to be sure. 800 */ 801 devvp = VFSTOUFS(oldmnt)->um_devvp; 802 for (loopcnt = 10; loopcnt > 0; ) { 803 if (softdep_process_worklist(oldmnt) == 0) { 804 loopcnt--; 805 /* 806 * Do another flush in case any vnodes were brought in 807 * as part of the cleanup operations. 808 */ 809 if ((error = ffs_flushfiles(oldmnt, flags)) != 0) 810 break; 811 /* 812 * If we still found nothing to do, we are really done. 813 */ 814 if (softdep_process_worklist(oldmnt) == 0) 815 break; 816 } 817 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 818 error = VOP_FSYNC(devvp, MNT_WAIT); 819 vn_unlock(devvp); 820 if (error) 821 break; 822 } 823 softdep_worklist_busy = 0; 824 if (softdep_worklist_req) 825 wakeup(&softdep_worklist_req); 826 827 /* 828 * If we are unmounting then it is an error to fail. If we 829 * are simply trying to downgrade to read-only, then filesystem 830 * activity can keep us busy forever, so we just fail with EBUSY. 831 */ 832 if (loopcnt == 0) { 833 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 834 panic("softdep_flushfiles: looping"); 835 error = EBUSY; 836 } 837 return (error); 838 } 839 840 /* 841 * Structure hashing. 842 * 843 * There are three types of structures that can be looked up: 844 * 1) pagedep structures identified by mount point, inode number, 845 * and logical block. 846 * 2) inodedep structures identified by mount point and inode number. 847 * 3) newblk structures identified by mount point and 848 * physical block number. 849 * 850 * The "pagedep" and "inodedep" dependency structures are hashed 851 * separately from the file blocks and inodes to which they correspond. 852 * This separation helps when the in-memory copy of an inode or 853 * file block must be replaced. It also obviates the need to access 854 * an inode or file page when simply updating (or de-allocating) 855 * dependency structures. Lookup of newblk structures is needed to 856 * find newly allocated blocks when trying to associate them with 857 * their allocdirect or allocindir structure. 858 * 859 * The lookup routines optionally create and hash a new instance when 860 * an existing entry is not found. 861 */ 862 #define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 863 #define NODELAY 0x0002 /* cannot do background work */ 864 865 /* 866 * Structures and routines associated with pagedep caching. 867 */ 868 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl; 869 u_long pagedep_hash; /* size of hash table - 1 */ 870 #define PAGEDEP_HASH(mp, inum, lbn) \ 871 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \ 872 pagedep_hash]) 873 static struct sema pagedep_in_progress; 874 875 /* 876 * Helper routine for pagedep_lookup() 877 */ 878 static __inline 879 struct pagedep * 880 pagedep_find(struct pagedep_hashhead *pagedephd, ino_t ino, ufs_lbn_t lbn, 881 struct mount *mp) 882 { 883 struct pagedep *pagedep; 884 885 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 886 if (ino == pagedep->pd_ino && 887 lbn == pagedep->pd_lbn && 888 mp == pagedep->pd_mnt) { 889 return (pagedep); 890 } 891 } 892 return(NULL); 893 } 894 895 /* 896 * Look up a pagedep. Return 1 if found, 0 if not found. 897 * If not found, allocate if DEPALLOC flag is passed. 898 * Found or allocated entry is returned in pagedeppp. 899 * This routine must be called with splbio interrupts blocked. 900 */ 901 static int 902 pagedep_lookup(struct inode *ip, ufs_lbn_t lbn, int flags, 903 struct pagedep **pagedeppp) 904 { 905 struct pagedep *pagedep; 906 struct pagedep_hashhead *pagedephd; 907 struct mount *mp; 908 int i; 909 910 #ifdef DEBUG 911 if (lk.lkt_held == NOHOLDER) 912 panic("pagedep_lookup: lock not held"); 913 #endif 914 mp = ITOV(ip)->v_mount; 915 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn); 916 top: 917 *pagedeppp = pagedep_find(pagedephd, ip->i_number, lbn, mp); 918 if (*pagedeppp) 919 return(1); 920 if ((flags & DEPALLOC) == 0) 921 return (0); 922 if (sema_get(&pagedep_in_progress, &lk) == 0) { 923 ACQUIRE_LOCK(&lk); 924 goto top; 925 } 926 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP, 927 M_SOFTDEP_FLAGS | M_ZERO); 928 929 if (pagedep_find(pagedephd, ip->i_number, lbn, mp)) { 930 kprintf("pagedep_lookup: blocking race avoided\n"); 931 ACQUIRE_LOCK(&lk); 932 sema_release(&pagedep_in_progress); 933 kfree(pagedep, M_PAGEDEP); 934 goto top; 935 } 936 937 pagedep->pd_list.wk_type = D_PAGEDEP; 938 pagedep->pd_mnt = mp; 939 pagedep->pd_ino = ip->i_number; 940 pagedep->pd_lbn = lbn; 941 LIST_INIT(&pagedep->pd_dirremhd); 942 LIST_INIT(&pagedep->pd_pendinghd); 943 for (i = 0; i < DAHASHSZ; i++) 944 LIST_INIT(&pagedep->pd_diraddhd[i]); 945 ACQUIRE_LOCK(&lk); 946 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 947 sema_release(&pagedep_in_progress); 948 *pagedeppp = pagedep; 949 return (0); 950 } 951 952 /* 953 * Structures and routines associated with inodedep caching. 954 */ 955 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl; 956 static u_long inodedep_hash; /* size of hash table - 1 */ 957 static long num_inodedep; /* number of inodedep allocated */ 958 #define INODEDEP_HASH(fs, inum) \ 959 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash]) 960 static struct sema inodedep_in_progress; 961 962 /* 963 * Helper routine for inodedep_lookup() 964 */ 965 static __inline 966 struct inodedep * 967 inodedep_find(struct inodedep_hashhead *inodedephd, struct fs *fs, ino_t inum) 968 { 969 struct inodedep *inodedep; 970 971 LIST_FOREACH(inodedep, inodedephd, id_hash) { 972 if (inum == inodedep->id_ino && fs == inodedep->id_fs) 973 return(inodedep); 974 } 975 return (NULL); 976 } 977 978 /* 979 * Look up a inodedep. Return 1 if found, 0 if not found. 980 * If not found, allocate if DEPALLOC flag is passed. 981 * Found or allocated entry is returned in inodedeppp. 982 * This routine must be called with splbio interrupts blocked. 983 */ 984 static int 985 inodedep_lookup(struct fs *fs, ino_t inum, int flags, 986 struct inodedep **inodedeppp) 987 { 988 struct inodedep *inodedep; 989 struct inodedep_hashhead *inodedephd; 990 int firsttry; 991 992 #ifdef DEBUG 993 if (lk.lkt_held == NOHOLDER) 994 panic("inodedep_lookup: lock not held"); 995 #endif 996 firsttry = 1; 997 inodedephd = INODEDEP_HASH(fs, inum); 998 top: 999 *inodedeppp = inodedep_find(inodedephd, fs, inum); 1000 if (*inodedeppp) 1001 return (1); 1002 if ((flags & DEPALLOC) == 0) 1003 return (0); 1004 /* 1005 * If we are over our limit, try to improve the situation. 1006 */ 1007 if (num_inodedep > max_softdeps && firsttry && 1008 speedup_syncer() == 0 && (flags & NODELAY) == 0 && 1009 request_cleanup(FLUSH_INODES, 1)) { 1010 firsttry = 0; 1011 goto top; 1012 } 1013 if (sema_get(&inodedep_in_progress, &lk) == 0) { 1014 ACQUIRE_LOCK(&lk); 1015 goto top; 1016 } 1017 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep), 1018 M_INODEDEP, M_SOFTDEP_FLAGS | M_ZERO); 1019 if (inodedep_find(inodedephd, fs, inum)) { 1020 kprintf("inodedep_lookup: blocking race avoided\n"); 1021 ACQUIRE_LOCK(&lk); 1022 sema_release(&inodedep_in_progress); 1023 kfree(inodedep, M_INODEDEP); 1024 goto top; 1025 } 1026 inodedep->id_list.wk_type = D_INODEDEP; 1027 inodedep->id_fs = fs; 1028 inodedep->id_ino = inum; 1029 inodedep->id_state = ALLCOMPLETE; 1030 inodedep->id_nlinkdelta = 0; 1031 inodedep->id_savedino = NULL; 1032 inodedep->id_savedsize = -1; 1033 inodedep->id_buf = NULL; 1034 LIST_INIT(&inodedep->id_pendinghd); 1035 LIST_INIT(&inodedep->id_inowait); 1036 LIST_INIT(&inodedep->id_bufwait); 1037 TAILQ_INIT(&inodedep->id_inoupdt); 1038 TAILQ_INIT(&inodedep->id_newinoupdt); 1039 ACQUIRE_LOCK(&lk); 1040 num_inodedep += 1; 1041 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 1042 sema_release(&inodedep_in_progress); 1043 *inodedeppp = inodedep; 1044 return (0); 1045 } 1046 1047 /* 1048 * Structures and routines associated with newblk caching. 1049 */ 1050 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl; 1051 u_long newblk_hash; /* size of hash table - 1 */ 1052 #define NEWBLK_HASH(fs, inum) \ 1053 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash]) 1054 static struct sema newblk_in_progress; 1055 1056 /* 1057 * Helper routine for newblk_lookup() 1058 */ 1059 static __inline 1060 struct newblk * 1061 newblk_find(struct newblk_hashhead *newblkhd, struct fs *fs, 1062 ufs_daddr_t newblkno) 1063 { 1064 struct newblk *newblk; 1065 1066 LIST_FOREACH(newblk, newblkhd, nb_hash) { 1067 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs) 1068 return (newblk); 1069 } 1070 return(NULL); 1071 } 1072 1073 /* 1074 * Look up a newblk. Return 1 if found, 0 if not found. 1075 * If not found, allocate if DEPALLOC flag is passed. 1076 * Found or allocated entry is returned in newblkpp. 1077 */ 1078 static int 1079 newblk_lookup(struct fs *fs, ufs_daddr_t newblkno, int flags, 1080 struct newblk **newblkpp) 1081 { 1082 struct newblk *newblk; 1083 struct newblk_hashhead *newblkhd; 1084 1085 newblkhd = NEWBLK_HASH(fs, newblkno); 1086 top: 1087 *newblkpp = newblk_find(newblkhd, fs, newblkno); 1088 if (*newblkpp) 1089 return(1); 1090 if ((flags & DEPALLOC) == 0) 1091 return (0); 1092 if (sema_get(&newblk_in_progress, 0) == 0) 1093 goto top; 1094 MALLOC(newblk, struct newblk *, sizeof(struct newblk), 1095 M_NEWBLK, M_SOFTDEP_FLAGS | M_ZERO); 1096 1097 if (newblk_find(newblkhd, fs, newblkno)) { 1098 kprintf("newblk_lookup: blocking race avoided\n"); 1099 sema_release(&pagedep_in_progress); 1100 kfree(newblk, M_NEWBLK); 1101 goto top; 1102 } 1103 newblk->nb_state = 0; 1104 newblk->nb_fs = fs; 1105 newblk->nb_newblkno = newblkno; 1106 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 1107 sema_release(&newblk_in_progress); 1108 *newblkpp = newblk; 1109 return (0); 1110 } 1111 1112 /* 1113 * Executed during filesystem system initialization before 1114 * mounting any filesystems. 1115 */ 1116 void 1117 softdep_initialize(void) 1118 { 1119 callout_init(&handle); 1120 1121 LIST_INIT(&mkdirlisthd); 1122 LIST_INIT(&softdep_workitem_pending); 1123 max_softdeps = min(desiredvnodes * 8, 1124 M_INODEDEP->ks_limit / (2 * sizeof(struct inodedep))); 1125 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 1126 &pagedep_hash); 1127 sema_init(&pagedep_in_progress, "pagedep", 0, 0); 1128 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash); 1129 sema_init(&inodedep_in_progress, "inodedep", 0, 0); 1130 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash); 1131 sema_init(&newblk_in_progress, "newblk", 0, 0); 1132 add_bio_ops(&softdep_bioops); 1133 } 1134 1135 /* 1136 * Called at mount time to notify the dependency code that a 1137 * filesystem wishes to use it. 1138 */ 1139 int 1140 softdep_mount(struct vnode *devvp, struct mount *mp, struct fs *fs) 1141 { 1142 struct csum cstotal; 1143 struct cg *cgp; 1144 struct buf *bp; 1145 int error, cyl; 1146 1147 mp->mnt_flag &= ~MNT_ASYNC; 1148 mp->mnt_flag |= MNT_SOFTDEP; 1149 mp->mnt_bioops = &softdep_bioops; 1150 /* 1151 * When doing soft updates, the counters in the 1152 * superblock may have gotten out of sync, so we have 1153 * to scan the cylinder groups and recalculate them. 1154 */ 1155 if (fs->fs_clean != 0) 1156 return (0); 1157 bzero(&cstotal, sizeof cstotal); 1158 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 1159 if ((error = bread(devvp, fsbtodoff(fs, cgtod(fs, cyl)), 1160 fs->fs_cgsize, &bp)) != 0) { 1161 brelse(bp); 1162 return (error); 1163 } 1164 cgp = (struct cg *)bp->b_data; 1165 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 1166 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 1167 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 1168 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 1169 fs->fs_cs(fs, cyl) = cgp->cg_cs; 1170 brelse(bp); 1171 } 1172 #ifdef DEBUG 1173 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 1174 kprintf("ffs_mountfs: superblock updated for soft updates\n"); 1175 #endif 1176 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 1177 return (0); 1178 } 1179 1180 /* 1181 * Protecting the freemaps (or bitmaps). 1182 * 1183 * To eliminate the need to execute fsck before mounting a filesystem 1184 * after a power failure, one must (conservatively) guarantee that the 1185 * on-disk copy of the bitmaps never indicate that a live inode or block is 1186 * free. So, when a block or inode is allocated, the bitmap should be 1187 * updated (on disk) before any new pointers. When a block or inode is 1188 * freed, the bitmap should not be updated until all pointers have been 1189 * reset. The latter dependency is handled by the delayed de-allocation 1190 * approach described below for block and inode de-allocation. The former 1191 * dependency is handled by calling the following procedure when a block or 1192 * inode is allocated. When an inode is allocated an "inodedep" is created 1193 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 1194 * Each "inodedep" is also inserted into the hash indexing structure so 1195 * that any additional link additions can be made dependent on the inode 1196 * allocation. 1197 * 1198 * The ufs filesystem maintains a number of free block counts (e.g., per 1199 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 1200 * in addition to the bitmaps. These counts are used to improve efficiency 1201 * during allocation and therefore must be consistent with the bitmaps. 1202 * There is no convenient way to guarantee post-crash consistency of these 1203 * counts with simple update ordering, for two main reasons: (1) The counts 1204 * and bitmaps for a single cylinder group block are not in the same disk 1205 * sector. If a disk write is interrupted (e.g., by power failure), one may 1206 * be written and the other not. (2) Some of the counts are located in the 1207 * superblock rather than the cylinder group block. So, we focus our soft 1208 * updates implementation on protecting the bitmaps. When mounting a 1209 * filesystem, we recompute the auxiliary counts from the bitmaps. 1210 */ 1211 1212 /* 1213 * Called just after updating the cylinder group block to allocate an inode. 1214 * 1215 * Parameters: 1216 * bp: buffer for cylgroup block with inode map 1217 * ip: inode related to allocation 1218 * newinum: new inode number being allocated 1219 */ 1220 void 1221 softdep_setup_inomapdep(struct buf *bp, struct inode *ip, ino_t newinum) 1222 { 1223 struct inodedep *inodedep; 1224 struct bmsafemap *bmsafemap; 1225 1226 /* 1227 * Create a dependency for the newly allocated inode. 1228 * Panic if it already exists as something is seriously wrong. 1229 * Otherwise add it to the dependency list for the buffer holding 1230 * the cylinder group map from which it was allocated. 1231 */ 1232 ACQUIRE_LOCK(&lk); 1233 if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) { 1234 FREE_LOCK(&lk); 1235 panic("softdep_setup_inomapdep: found inode"); 1236 } 1237 inodedep->id_buf = bp; 1238 inodedep->id_state &= ~DEPCOMPLETE; 1239 bmsafemap = bmsafemap_lookup(bp); 1240 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 1241 FREE_LOCK(&lk); 1242 } 1243 1244 /* 1245 * Called just after updating the cylinder group block to 1246 * allocate block or fragment. 1247 * 1248 * Parameters: 1249 * bp: buffer for cylgroup block with block map 1250 * fs: filesystem doing allocation 1251 * newblkno: number of newly allocated block 1252 */ 1253 void 1254 softdep_setup_blkmapdep(struct buf *bp, struct fs *fs, 1255 ufs_daddr_t newblkno) 1256 { 1257 struct newblk *newblk; 1258 struct bmsafemap *bmsafemap; 1259 1260 /* 1261 * Create a dependency for the newly allocated block. 1262 * Add it to the dependency list for the buffer holding 1263 * the cylinder group map from which it was allocated. 1264 */ 1265 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0) 1266 panic("softdep_setup_blkmapdep: found block"); 1267 ACQUIRE_LOCK(&lk); 1268 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp); 1269 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 1270 FREE_LOCK(&lk); 1271 } 1272 1273 /* 1274 * Find the bmsafemap associated with a cylinder group buffer. 1275 * If none exists, create one. The buffer must be locked when 1276 * this routine is called and this routine must be called with 1277 * splbio interrupts blocked. 1278 */ 1279 static struct bmsafemap * 1280 bmsafemap_lookup(struct buf *bp) 1281 { 1282 struct bmsafemap *bmsafemap; 1283 struct worklist *wk; 1284 1285 #ifdef DEBUG 1286 if (lk.lkt_held == NOHOLDER) 1287 panic("bmsafemap_lookup: lock not held"); 1288 #endif 1289 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 1290 if (wk->wk_type == D_BMSAFEMAP) 1291 return (WK_BMSAFEMAP(wk)); 1292 } 1293 FREE_LOCK(&lk); 1294 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap), 1295 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 1296 bmsafemap->sm_list.wk_type = D_BMSAFEMAP; 1297 bmsafemap->sm_list.wk_state = 0; 1298 bmsafemap->sm_buf = bp; 1299 LIST_INIT(&bmsafemap->sm_allocdirecthd); 1300 LIST_INIT(&bmsafemap->sm_allocindirhd); 1301 LIST_INIT(&bmsafemap->sm_inodedephd); 1302 LIST_INIT(&bmsafemap->sm_newblkhd); 1303 ACQUIRE_LOCK(&lk); 1304 WORKLIST_INSERT_BP(bp, &bmsafemap->sm_list); 1305 return (bmsafemap); 1306 } 1307 1308 /* 1309 * Direct block allocation dependencies. 1310 * 1311 * When a new block is allocated, the corresponding disk locations must be 1312 * initialized (with zeros or new data) before the on-disk inode points to 1313 * them. Also, the freemap from which the block was allocated must be 1314 * updated (on disk) before the inode's pointer. These two dependencies are 1315 * independent of each other and are needed for all file blocks and indirect 1316 * blocks that are pointed to directly by the inode. Just before the 1317 * "in-core" version of the inode is updated with a newly allocated block 1318 * number, a procedure (below) is called to setup allocation dependency 1319 * structures. These structures are removed when the corresponding 1320 * dependencies are satisfied or when the block allocation becomes obsolete 1321 * (i.e., the file is deleted, the block is de-allocated, or the block is a 1322 * fragment that gets upgraded). All of these cases are handled in 1323 * procedures described later. 1324 * 1325 * When a file extension causes a fragment to be upgraded, either to a larger 1326 * fragment or to a full block, the on-disk location may change (if the 1327 * previous fragment could not simply be extended). In this case, the old 1328 * fragment must be de-allocated, but not until after the inode's pointer has 1329 * been updated. In most cases, this is handled by later procedures, which 1330 * will construct a "freefrag" structure to be added to the workitem queue 1331 * when the inode update is complete (or obsolete). The main exception to 1332 * this is when an allocation occurs while a pending allocation dependency 1333 * (for the same block pointer) remains. This case is handled in the main 1334 * allocation dependency setup procedure by immediately freeing the 1335 * unreferenced fragments. 1336 * 1337 * Parameters: 1338 * ip: inode to which block is being added 1339 * lbn: block pointer within inode 1340 * newblkno: disk block number being added 1341 * oldblkno: previous block number, 0 unless frag 1342 * newsize: size of new block 1343 * oldsize: size of new block 1344 * bp: bp for allocated block 1345 */ 1346 void 1347 softdep_setup_allocdirect(struct inode *ip, ufs_lbn_t lbn, ufs_daddr_t newblkno, 1348 ufs_daddr_t oldblkno, long newsize, long oldsize, 1349 struct buf *bp) 1350 { 1351 struct allocdirect *adp, *oldadp; 1352 struct allocdirectlst *adphead; 1353 struct bmsafemap *bmsafemap; 1354 struct inodedep *inodedep; 1355 struct pagedep *pagedep; 1356 struct newblk *newblk; 1357 1358 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect), 1359 M_ALLOCDIRECT, M_SOFTDEP_FLAGS | M_ZERO); 1360 adp->ad_list.wk_type = D_ALLOCDIRECT; 1361 adp->ad_lbn = lbn; 1362 adp->ad_newblkno = newblkno; 1363 adp->ad_oldblkno = oldblkno; 1364 adp->ad_newsize = newsize; 1365 adp->ad_oldsize = oldsize; 1366 adp->ad_state = ATTACHED; 1367 if (newblkno == oldblkno) 1368 adp->ad_freefrag = NULL; 1369 else 1370 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1371 1372 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1373 panic("softdep_setup_allocdirect: lost block"); 1374 1375 ACQUIRE_LOCK(&lk); 1376 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep); 1377 adp->ad_inodedep = inodedep; 1378 1379 if (newblk->nb_state == DEPCOMPLETE) { 1380 adp->ad_state |= DEPCOMPLETE; 1381 adp->ad_buf = NULL; 1382 } else { 1383 bmsafemap = newblk->nb_bmsafemap; 1384 adp->ad_buf = bmsafemap->sm_buf; 1385 LIST_REMOVE(newblk, nb_deps); 1386 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1387 } 1388 LIST_REMOVE(newblk, nb_hash); 1389 FREE(newblk, M_NEWBLK); 1390 1391 WORKLIST_INSERT_BP(bp, &adp->ad_list); 1392 if (lbn >= NDADDR) { 1393 /* allocating an indirect block */ 1394 if (oldblkno != 0) { 1395 FREE_LOCK(&lk); 1396 panic("softdep_setup_allocdirect: non-zero indir"); 1397 } 1398 } else { 1399 /* 1400 * Allocating a direct block. 1401 * 1402 * If we are allocating a directory block, then we must 1403 * allocate an associated pagedep to track additions and 1404 * deletions. 1405 */ 1406 if ((ip->i_mode & IFMT) == IFDIR && 1407 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) { 1408 WORKLIST_INSERT_BP(bp, &pagedep->pd_list); 1409 } 1410 } 1411 /* 1412 * The list of allocdirects must be kept in sorted and ascending 1413 * order so that the rollback routines can quickly determine the 1414 * first uncommitted block (the size of the file stored on disk 1415 * ends at the end of the lowest committed fragment, or if there 1416 * are no fragments, at the end of the highest committed block). 1417 * Since files generally grow, the typical case is that the new 1418 * block is to be added at the end of the list. We speed this 1419 * special case by checking against the last allocdirect in the 1420 * list before laboriously traversing the list looking for the 1421 * insertion point. 1422 */ 1423 adphead = &inodedep->id_newinoupdt; 1424 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1425 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1426 /* insert at end of list */ 1427 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1428 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1429 allocdirect_merge(adphead, adp, oldadp); 1430 FREE_LOCK(&lk); 1431 return; 1432 } 1433 TAILQ_FOREACH(oldadp, adphead, ad_next) { 1434 if (oldadp->ad_lbn >= lbn) 1435 break; 1436 } 1437 if (oldadp == NULL) { 1438 FREE_LOCK(&lk); 1439 panic("softdep_setup_allocdirect: lost entry"); 1440 } 1441 /* insert in middle of list */ 1442 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1443 if (oldadp->ad_lbn == lbn) 1444 allocdirect_merge(adphead, adp, oldadp); 1445 FREE_LOCK(&lk); 1446 } 1447 1448 /* 1449 * Replace an old allocdirect dependency with a newer one. 1450 * This routine must be called with splbio interrupts blocked. 1451 * 1452 * Parameters: 1453 * adphead: head of list holding allocdirects 1454 * newadp: allocdirect being added 1455 * oldadp: existing allocdirect being checked 1456 */ 1457 static void 1458 allocdirect_merge(struct allocdirectlst *adphead, 1459 struct allocdirect *newadp, 1460 struct allocdirect *oldadp) 1461 { 1462 struct freefrag *freefrag; 1463 1464 #ifdef DEBUG 1465 if (lk.lkt_held == NOHOLDER) 1466 panic("allocdirect_merge: lock not held"); 1467 #endif 1468 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 1469 newadp->ad_oldsize != oldadp->ad_newsize || 1470 newadp->ad_lbn >= NDADDR) { 1471 FREE_LOCK(&lk); 1472 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d", 1473 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn, 1474 NDADDR); 1475 } 1476 newadp->ad_oldblkno = oldadp->ad_oldblkno; 1477 newadp->ad_oldsize = oldadp->ad_oldsize; 1478 /* 1479 * If the old dependency had a fragment to free or had never 1480 * previously had a block allocated, then the new dependency 1481 * can immediately post its freefrag and adopt the old freefrag. 1482 * This action is done by swapping the freefrag dependencies. 1483 * The new dependency gains the old one's freefrag, and the 1484 * old one gets the new one and then immediately puts it on 1485 * the worklist when it is freed by free_allocdirect. It is 1486 * not possible to do this swap when the old dependency had a 1487 * non-zero size but no previous fragment to free. This condition 1488 * arises when the new block is an extension of the old block. 1489 * Here, the first part of the fragment allocated to the new 1490 * dependency is part of the block currently claimed on disk by 1491 * the old dependency, so cannot legitimately be freed until the 1492 * conditions for the new dependency are fulfilled. 1493 */ 1494 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 1495 freefrag = newadp->ad_freefrag; 1496 newadp->ad_freefrag = oldadp->ad_freefrag; 1497 oldadp->ad_freefrag = freefrag; 1498 } 1499 free_allocdirect(adphead, oldadp, 0); 1500 } 1501 1502 /* 1503 * Allocate a new freefrag structure if needed. 1504 */ 1505 static struct freefrag * 1506 newfreefrag(struct inode *ip, ufs_daddr_t blkno, long size) 1507 { 1508 struct freefrag *freefrag; 1509 struct fs *fs; 1510 1511 if (blkno == 0) 1512 return (NULL); 1513 fs = ip->i_fs; 1514 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 1515 panic("newfreefrag: frag size"); 1516 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag), 1517 M_FREEFRAG, M_SOFTDEP_FLAGS); 1518 freefrag->ff_list.wk_type = D_FREEFRAG; 1519 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */ 1520 freefrag->ff_inum = ip->i_number; 1521 freefrag->ff_fs = fs; 1522 freefrag->ff_devvp = ip->i_devvp; 1523 freefrag->ff_blkno = blkno; 1524 freefrag->ff_fragsize = size; 1525 return (freefrag); 1526 } 1527 1528 /* 1529 * This workitem de-allocates fragments that were replaced during 1530 * file block allocation. 1531 */ 1532 static void 1533 handle_workitem_freefrag(struct freefrag *freefrag) 1534 { 1535 struct inode tip; 1536 1537 tip.i_fs = freefrag->ff_fs; 1538 tip.i_devvp = freefrag->ff_devvp; 1539 tip.i_dev = freefrag->ff_devvp->v_rdev; 1540 tip.i_number = freefrag->ff_inum; 1541 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */ 1542 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize); 1543 FREE(freefrag, M_FREEFRAG); 1544 } 1545 1546 /* 1547 * Indirect block allocation dependencies. 1548 * 1549 * The same dependencies that exist for a direct block also exist when 1550 * a new block is allocated and pointed to by an entry in a block of 1551 * indirect pointers. The undo/redo states described above are also 1552 * used here. Because an indirect block contains many pointers that 1553 * may have dependencies, a second copy of the entire in-memory indirect 1554 * block is kept. The buffer cache copy is always completely up-to-date. 1555 * The second copy, which is used only as a source for disk writes, 1556 * contains only the safe pointers (i.e., those that have no remaining 1557 * update dependencies). The second copy is freed when all pointers 1558 * are safe. The cache is not allowed to replace indirect blocks with 1559 * pending update dependencies. If a buffer containing an indirect 1560 * block with dependencies is written, these routines will mark it 1561 * dirty again. It can only be successfully written once all the 1562 * dependencies are removed. The ffs_fsync routine in conjunction with 1563 * softdep_sync_metadata work together to get all the dependencies 1564 * removed so that a file can be successfully written to disk. Three 1565 * procedures are used when setting up indirect block pointer 1566 * dependencies. The division is necessary because of the organization 1567 * of the "balloc" routine and because of the distinction between file 1568 * pages and file metadata blocks. 1569 */ 1570 1571 /* 1572 * Allocate a new allocindir structure. 1573 * 1574 * Parameters: 1575 * ip: inode for file being extended 1576 * ptrno: offset of pointer in indirect block 1577 * newblkno: disk block number being added 1578 * oldblkno: previous block number, 0 if none 1579 */ 1580 static struct allocindir * 1581 newallocindir(struct inode *ip, int ptrno, ufs_daddr_t newblkno, 1582 ufs_daddr_t oldblkno) 1583 { 1584 struct allocindir *aip; 1585 1586 MALLOC(aip, struct allocindir *, sizeof(struct allocindir), 1587 M_ALLOCINDIR, M_SOFTDEP_FLAGS | M_ZERO); 1588 aip->ai_list.wk_type = D_ALLOCINDIR; 1589 aip->ai_state = ATTACHED; 1590 aip->ai_offset = ptrno; 1591 aip->ai_newblkno = newblkno; 1592 aip->ai_oldblkno = oldblkno; 1593 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize); 1594 return (aip); 1595 } 1596 1597 /* 1598 * Called just before setting an indirect block pointer 1599 * to a newly allocated file page. 1600 * 1601 * Parameters: 1602 * ip: inode for file being extended 1603 * lbn: allocated block number within file 1604 * bp: buffer with indirect blk referencing page 1605 * ptrno: offset of pointer in indirect block 1606 * newblkno: disk block number being added 1607 * oldblkno: previous block number, 0 if none 1608 * nbp: buffer holding allocated page 1609 */ 1610 void 1611 softdep_setup_allocindir_page(struct inode *ip, ufs_lbn_t lbn, 1612 struct buf *bp, int ptrno, 1613 ufs_daddr_t newblkno, ufs_daddr_t oldblkno, 1614 struct buf *nbp) 1615 { 1616 struct allocindir *aip; 1617 struct pagedep *pagedep; 1618 1619 aip = newallocindir(ip, ptrno, newblkno, oldblkno); 1620 ACQUIRE_LOCK(&lk); 1621 /* 1622 * If we are allocating a directory page, then we must 1623 * allocate an associated pagedep to track additions and 1624 * deletions. 1625 */ 1626 if ((ip->i_mode & IFMT) == IFDIR && 1627 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1628 WORKLIST_INSERT_BP(nbp, &pagedep->pd_list); 1629 WORKLIST_INSERT_BP(nbp, &aip->ai_list); 1630 FREE_LOCK(&lk); 1631 setup_allocindir_phase2(bp, ip, aip); 1632 } 1633 1634 /* 1635 * Called just before setting an indirect block pointer to a 1636 * newly allocated indirect block. 1637 * Parameters: 1638 * nbp: newly allocated indirect block 1639 * ip: inode for file being extended 1640 * bp: indirect block referencing allocated block 1641 * ptrno: offset of pointer in indirect block 1642 * newblkno: disk block number being added 1643 */ 1644 void 1645 softdep_setup_allocindir_meta(struct buf *nbp, struct inode *ip, 1646 struct buf *bp, int ptrno, 1647 ufs_daddr_t newblkno) 1648 { 1649 struct allocindir *aip; 1650 1651 aip = newallocindir(ip, ptrno, newblkno, 0); 1652 ACQUIRE_LOCK(&lk); 1653 WORKLIST_INSERT_BP(nbp, &aip->ai_list); 1654 FREE_LOCK(&lk); 1655 setup_allocindir_phase2(bp, ip, aip); 1656 } 1657 1658 /* 1659 * Called to finish the allocation of the "aip" allocated 1660 * by one of the two routines above. 1661 * 1662 * Parameters: 1663 * bp: in-memory copy of the indirect block 1664 * ip: inode for file being extended 1665 * aip: allocindir allocated by the above routines 1666 */ 1667 static void 1668 setup_allocindir_phase2(struct buf *bp, struct inode *ip, 1669 struct allocindir *aip) 1670 { 1671 struct worklist *wk; 1672 struct indirdep *indirdep, *newindirdep; 1673 struct bmsafemap *bmsafemap; 1674 struct allocindir *oldaip; 1675 struct freefrag *freefrag; 1676 struct newblk *newblk; 1677 1678 if (bp->b_loffset >= 0) 1679 panic("setup_allocindir_phase2: not indir blk"); 1680 for (indirdep = NULL, newindirdep = NULL; ; ) { 1681 ACQUIRE_LOCK(&lk); 1682 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 1683 if (wk->wk_type != D_INDIRDEP) 1684 continue; 1685 indirdep = WK_INDIRDEP(wk); 1686 break; 1687 } 1688 if (indirdep == NULL && newindirdep) { 1689 indirdep = newindirdep; 1690 WORKLIST_INSERT_BP(bp, &indirdep->ir_list); 1691 newindirdep = NULL; 1692 } 1693 FREE_LOCK(&lk); 1694 if (indirdep) { 1695 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0, 1696 &newblk) == 0) 1697 panic("setup_allocindir: lost block"); 1698 ACQUIRE_LOCK(&lk); 1699 if (newblk->nb_state == DEPCOMPLETE) { 1700 aip->ai_state |= DEPCOMPLETE; 1701 aip->ai_buf = NULL; 1702 } else { 1703 bmsafemap = newblk->nb_bmsafemap; 1704 aip->ai_buf = bmsafemap->sm_buf; 1705 LIST_REMOVE(newblk, nb_deps); 1706 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd, 1707 aip, ai_deps); 1708 } 1709 LIST_REMOVE(newblk, nb_hash); 1710 FREE(newblk, M_NEWBLK); 1711 aip->ai_indirdep = indirdep; 1712 /* 1713 * Check to see if there is an existing dependency 1714 * for this block. If there is, merge the old 1715 * dependency into the new one. 1716 */ 1717 if (aip->ai_oldblkno == 0) 1718 oldaip = NULL; 1719 else 1720 1721 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) 1722 if (oldaip->ai_offset == aip->ai_offset) 1723 break; 1724 if (oldaip != NULL) { 1725 if (oldaip->ai_newblkno != aip->ai_oldblkno) { 1726 FREE_LOCK(&lk); 1727 panic("setup_allocindir_phase2: blkno"); 1728 } 1729 aip->ai_oldblkno = oldaip->ai_oldblkno; 1730 freefrag = oldaip->ai_freefrag; 1731 oldaip->ai_freefrag = aip->ai_freefrag; 1732 aip->ai_freefrag = freefrag; 1733 free_allocindir(oldaip, NULL); 1734 } 1735 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 1736 ((ufs_daddr_t *)indirdep->ir_savebp->b_data) 1737 [aip->ai_offset] = aip->ai_oldblkno; 1738 FREE_LOCK(&lk); 1739 } 1740 if (newindirdep) { 1741 /* 1742 * Avoid any possibility of data corruption by 1743 * ensuring that our old version is thrown away. 1744 */ 1745 newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 1746 brelse(newindirdep->ir_savebp); 1747 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP); 1748 } 1749 if (indirdep) 1750 break; 1751 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep), 1752 M_INDIRDEP, M_SOFTDEP_FLAGS); 1753 newindirdep->ir_list.wk_type = D_INDIRDEP; 1754 newindirdep->ir_state = ATTACHED; 1755 LIST_INIT(&newindirdep->ir_deplisthd); 1756 LIST_INIT(&newindirdep->ir_donehd); 1757 if (bp->b_bio2.bio_offset == NOOFFSET) { 1758 VOP_BMAP(bp->b_vp, bp->b_bio1.bio_offset, 1759 &bp->b_bio2.bio_offset, NULL, NULL, 1760 BUF_CMD_WRITE); 1761 } 1762 KKASSERT(bp->b_bio2.bio_offset != NOOFFSET); 1763 newindirdep->ir_savebp = getblk(ip->i_devvp, 1764 bp->b_bio2.bio_offset, 1765 bp->b_bcount, 0, 0); 1766 BUF_KERNPROC(newindirdep->ir_savebp); 1767 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 1768 } 1769 } 1770 1771 /* 1772 * Block de-allocation dependencies. 1773 * 1774 * When blocks are de-allocated, the on-disk pointers must be nullified before 1775 * the blocks are made available for use by other files. (The true 1776 * requirement is that old pointers must be nullified before new on-disk 1777 * pointers are set. We chose this slightly more stringent requirement to 1778 * reduce complexity.) Our implementation handles this dependency by updating 1779 * the inode (or indirect block) appropriately but delaying the actual block 1780 * de-allocation (i.e., freemap and free space count manipulation) until 1781 * after the updated versions reach stable storage. After the disk is 1782 * updated, the blocks can be safely de-allocated whenever it is convenient. 1783 * This implementation handles only the common case of reducing a file's 1784 * length to zero. Other cases are handled by the conventional synchronous 1785 * write approach. 1786 * 1787 * The ffs implementation with which we worked double-checks 1788 * the state of the block pointers and file size as it reduces 1789 * a file's length. Some of this code is replicated here in our 1790 * soft updates implementation. The freeblks->fb_chkcnt field is 1791 * used to transfer a part of this information to the procedure 1792 * that eventually de-allocates the blocks. 1793 * 1794 * This routine should be called from the routine that shortens 1795 * a file's length, before the inode's size or block pointers 1796 * are modified. It will save the block pointer information for 1797 * later release and zero the inode so that the calling routine 1798 * can release it. 1799 */ 1800 struct softdep_setup_freeblocks_info { 1801 struct fs *fs; 1802 struct inode *ip; 1803 }; 1804 1805 static int softdep_setup_freeblocks_bp(struct buf *bp, void *data); 1806 1807 /* 1808 * Parameters: 1809 * ip: The inode whose length is to be reduced 1810 * length: The new length for the file 1811 */ 1812 void 1813 softdep_setup_freeblocks(struct inode *ip, off_t length) 1814 { 1815 struct softdep_setup_freeblocks_info info; 1816 struct freeblks *freeblks; 1817 struct inodedep *inodedep; 1818 struct allocdirect *adp; 1819 struct vnode *vp; 1820 struct buf *bp; 1821 struct fs *fs; 1822 int i, error, delay; 1823 int count; 1824 1825 fs = ip->i_fs; 1826 if (length != 0) 1827 panic("softde_setup_freeblocks: non-zero length"); 1828 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks), 1829 M_FREEBLKS, M_SOFTDEP_FLAGS | M_ZERO); 1830 freeblks->fb_list.wk_type = D_FREEBLKS; 1831 freeblks->fb_state = ATTACHED; 1832 freeblks->fb_uid = ip->i_uid; 1833 freeblks->fb_previousinum = ip->i_number; 1834 freeblks->fb_devvp = ip->i_devvp; 1835 freeblks->fb_fs = fs; 1836 freeblks->fb_oldsize = ip->i_size; 1837 freeblks->fb_newsize = length; 1838 freeblks->fb_chkcnt = ip->i_blocks; 1839 for (i = 0; i < NDADDR; i++) { 1840 freeblks->fb_dblks[i] = ip->i_db[i]; 1841 ip->i_db[i] = 0; 1842 } 1843 for (i = 0; i < NIADDR; i++) { 1844 freeblks->fb_iblks[i] = ip->i_ib[i]; 1845 ip->i_ib[i] = 0; 1846 } 1847 ip->i_blocks = 0; 1848 ip->i_size = 0; 1849 /* 1850 * Push the zero'ed inode to to its disk buffer so that we are free 1851 * to delete its dependencies below. Once the dependencies are gone 1852 * the buffer can be safely released. 1853 */ 1854 if ((error = bread(ip->i_devvp, 1855 fsbtodoff(fs, ino_to_fsba(fs, ip->i_number)), 1856 (int)fs->fs_bsize, &bp)) != 0) 1857 softdep_error("softdep_setup_freeblocks", error); 1858 *((struct ufs1_dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) = 1859 ip->i_din; 1860 /* 1861 * Find and eliminate any inode dependencies. 1862 */ 1863 ACQUIRE_LOCK(&lk); 1864 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep); 1865 if ((inodedep->id_state & IOSTARTED) != 0) { 1866 FREE_LOCK(&lk); 1867 panic("softdep_setup_freeblocks: inode busy"); 1868 } 1869 /* 1870 * Add the freeblks structure to the list of operations that 1871 * must await the zero'ed inode being written to disk. If we 1872 * still have a bitmap dependency (delay == 0), then the inode 1873 * has never been written to disk, so we can process the 1874 * freeblks below once we have deleted the dependencies. 1875 */ 1876 delay = (inodedep->id_state & DEPCOMPLETE); 1877 if (delay) 1878 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list); 1879 /* 1880 * Because the file length has been truncated to zero, any 1881 * pending block allocation dependency structures associated 1882 * with this inode are obsolete and can simply be de-allocated. 1883 * We must first merge the two dependency lists to get rid of 1884 * any duplicate freefrag structures, then purge the merged list. 1885 */ 1886 merge_inode_lists(inodedep); 1887 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 1888 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 1889 FREE_LOCK(&lk); 1890 bdwrite(bp); 1891 /* 1892 * We must wait for any I/O in progress to finish so that 1893 * all potential buffers on the dirty list will be visible. 1894 * Once they are all there, walk the list and get rid of 1895 * any dependencies. 1896 */ 1897 vp = ITOV(ip); 1898 ACQUIRE_LOCK(&lk); 1899 drain_output(vp, 1); 1900 1901 info.fs = fs; 1902 info.ip = ip; 1903 do { 1904 count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 1905 softdep_setup_freeblocks_bp, &info); 1906 } while (count != 0); 1907 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0) 1908 (void)free_inodedep(inodedep); 1909 1910 if (delay) { 1911 freeblks->fb_state |= DEPCOMPLETE; 1912 /* 1913 * If the inode with zeroed block pointers is now on disk 1914 * we can start freeing blocks. Add freeblks to the worklist 1915 * instead of calling handle_workitem_freeblocks directly as 1916 * it is more likely that additional IO is needed to complete 1917 * the request here than in the !delay case. 1918 */ 1919 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 1920 add_to_worklist(&freeblks->fb_list); 1921 } 1922 1923 FREE_LOCK(&lk); 1924 /* 1925 * If the inode has never been written to disk (delay == 0), 1926 * then we can process the freeblks now that we have deleted 1927 * the dependencies. 1928 */ 1929 if (!delay) 1930 handle_workitem_freeblocks(freeblks); 1931 } 1932 1933 static int 1934 softdep_setup_freeblocks_bp(struct buf *bp, void *data) 1935 { 1936 struct softdep_setup_freeblocks_info *info = data; 1937 struct inodedep *inodedep; 1938 1939 if (getdirtybuf(&bp, MNT_WAIT) == 0) { 1940 kprintf("softdep_setup_freeblocks_bp(1): caught bp %p going away\n", bp); 1941 return(-1); 1942 } 1943 if (bp->b_vp != ITOV(info->ip) || (bp->b_flags & B_DELWRI) == 0) { 1944 kprintf("softdep_setup_freeblocks_bp(2): caught bp %p going away\n", bp); 1945 BUF_UNLOCK(bp); 1946 return(-1); 1947 } 1948 (void) inodedep_lookup(info->fs, info->ip->i_number, 0, &inodedep); 1949 deallocate_dependencies(bp, inodedep); 1950 bp->b_flags |= B_INVAL | B_NOCACHE; 1951 FREE_LOCK(&lk); 1952 brelse(bp); 1953 ACQUIRE_LOCK(&lk); 1954 return(1); 1955 } 1956 1957 /* 1958 * Reclaim any dependency structures from a buffer that is about to 1959 * be reallocated to a new vnode. The buffer must be locked, thus, 1960 * no I/O completion operations can occur while we are manipulating 1961 * its associated dependencies. The mutex is held so that other I/O's 1962 * associated with related dependencies do not occur. 1963 */ 1964 static void 1965 deallocate_dependencies(struct buf *bp, struct inodedep *inodedep) 1966 { 1967 struct worklist *wk; 1968 struct indirdep *indirdep; 1969 struct allocindir *aip; 1970 struct pagedep *pagedep; 1971 struct dirrem *dirrem; 1972 struct diradd *dap; 1973 int i; 1974 1975 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 1976 switch (wk->wk_type) { 1977 1978 case D_INDIRDEP: 1979 indirdep = WK_INDIRDEP(wk); 1980 /* 1981 * None of the indirect pointers will ever be visible, 1982 * so they can simply be tossed. GOINGAWAY ensures 1983 * that allocated pointers will be saved in the buffer 1984 * cache until they are freed. Note that they will 1985 * only be able to be found by their physical address 1986 * since the inode mapping the logical address will 1987 * be gone. The save buffer used for the safe copy 1988 * was allocated in setup_allocindir_phase2 using 1989 * the physical address so it could be used for this 1990 * purpose. Hence we swap the safe copy with the real 1991 * copy, allowing the safe copy to be freed and holding 1992 * on to the real copy for later use in indir_trunc. 1993 * 1994 * NOTE: ir_savebp is relative to the block device 1995 * so b_bio1 contains the device block number. 1996 */ 1997 if (indirdep->ir_state & GOINGAWAY) { 1998 FREE_LOCK(&lk); 1999 panic("deallocate_dependencies: already gone"); 2000 } 2001 indirdep->ir_state |= GOINGAWAY; 2002 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 2003 free_allocindir(aip, inodedep); 2004 if (bp->b_bio1.bio_offset >= 0 || 2005 bp->b_bio2.bio_offset != indirdep->ir_savebp->b_bio1.bio_offset) { 2006 FREE_LOCK(&lk); 2007 panic("deallocate_dependencies: not indir"); 2008 } 2009 bcopy(bp->b_data, indirdep->ir_savebp->b_data, 2010 bp->b_bcount); 2011 WORKLIST_REMOVE(wk); 2012 WORKLIST_INSERT_BP(indirdep->ir_savebp, wk); 2013 continue; 2014 2015 case D_PAGEDEP: 2016 pagedep = WK_PAGEDEP(wk); 2017 /* 2018 * None of the directory additions will ever be 2019 * visible, so they can simply be tossed. 2020 */ 2021 for (i = 0; i < DAHASHSZ; i++) 2022 while ((dap = 2023 LIST_FIRST(&pagedep->pd_diraddhd[i]))) 2024 free_diradd(dap); 2025 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0) 2026 free_diradd(dap); 2027 /* 2028 * Copy any directory remove dependencies to the list 2029 * to be processed after the zero'ed inode is written. 2030 * If the inode has already been written, then they 2031 * can be dumped directly onto the work list. 2032 */ 2033 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 2034 LIST_REMOVE(dirrem, dm_next); 2035 dirrem->dm_dirinum = pagedep->pd_ino; 2036 if (inodedep == NULL || 2037 (inodedep->id_state & ALLCOMPLETE) == 2038 ALLCOMPLETE) 2039 add_to_worklist(&dirrem->dm_list); 2040 else 2041 WORKLIST_INSERT(&inodedep->id_bufwait, 2042 &dirrem->dm_list); 2043 } 2044 WORKLIST_REMOVE(&pagedep->pd_list); 2045 LIST_REMOVE(pagedep, pd_hash); 2046 WORKITEM_FREE(pagedep, D_PAGEDEP); 2047 continue; 2048 2049 case D_ALLOCINDIR: 2050 free_allocindir(WK_ALLOCINDIR(wk), inodedep); 2051 continue; 2052 2053 case D_ALLOCDIRECT: 2054 case D_INODEDEP: 2055 FREE_LOCK(&lk); 2056 panic("deallocate_dependencies: Unexpected type %s", 2057 TYPENAME(wk->wk_type)); 2058 /* NOTREACHED */ 2059 2060 default: 2061 FREE_LOCK(&lk); 2062 panic("deallocate_dependencies: Unknown type %s", 2063 TYPENAME(wk->wk_type)); 2064 /* NOTREACHED */ 2065 } 2066 } 2067 } 2068 2069 /* 2070 * Free an allocdirect. Generate a new freefrag work request if appropriate. 2071 * This routine must be called with splbio interrupts blocked. 2072 */ 2073 static void 2074 free_allocdirect(struct allocdirectlst *adphead, 2075 struct allocdirect *adp, int delay) 2076 { 2077 2078 #ifdef DEBUG 2079 if (lk.lkt_held == NOHOLDER) 2080 panic("free_allocdirect: lock not held"); 2081 #endif 2082 if ((adp->ad_state & DEPCOMPLETE) == 0) 2083 LIST_REMOVE(adp, ad_deps); 2084 TAILQ_REMOVE(adphead, adp, ad_next); 2085 if ((adp->ad_state & COMPLETE) == 0) 2086 WORKLIST_REMOVE(&adp->ad_list); 2087 if (adp->ad_freefrag != NULL) { 2088 if (delay) 2089 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 2090 &adp->ad_freefrag->ff_list); 2091 else 2092 add_to_worklist(&adp->ad_freefrag->ff_list); 2093 } 2094 WORKITEM_FREE(adp, D_ALLOCDIRECT); 2095 } 2096 2097 /* 2098 * Prepare an inode to be freed. The actual free operation is not 2099 * done until the zero'ed inode has been written to disk. 2100 */ 2101 void 2102 softdep_freefile(struct vnode *pvp, ino_t ino, int mode) 2103 { 2104 struct inode *ip = VTOI(pvp); 2105 struct inodedep *inodedep; 2106 struct freefile *freefile; 2107 2108 /* 2109 * This sets up the inode de-allocation dependency. 2110 */ 2111 MALLOC(freefile, struct freefile *, sizeof(struct freefile), 2112 M_FREEFILE, M_SOFTDEP_FLAGS); 2113 freefile->fx_list.wk_type = D_FREEFILE; 2114 freefile->fx_list.wk_state = 0; 2115 freefile->fx_mode = mode; 2116 freefile->fx_oldinum = ino; 2117 freefile->fx_devvp = ip->i_devvp; 2118 freefile->fx_fs = ip->i_fs; 2119 2120 /* 2121 * If the inodedep does not exist, then the zero'ed inode has 2122 * been written to disk. If the allocated inode has never been 2123 * written to disk, then the on-disk inode is zero'ed. In either 2124 * case we can free the file immediately. 2125 */ 2126 ACQUIRE_LOCK(&lk); 2127 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 || 2128 check_inode_unwritten(inodedep)) { 2129 FREE_LOCK(&lk); 2130 handle_workitem_freefile(freefile); 2131 return; 2132 } 2133 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 2134 FREE_LOCK(&lk); 2135 } 2136 2137 /* 2138 * Check to see if an inode has never been written to disk. If 2139 * so free the inodedep and return success, otherwise return failure. 2140 * This routine must be called with splbio interrupts blocked. 2141 * 2142 * If we still have a bitmap dependency, then the inode has never 2143 * been written to disk. Drop the dependency as it is no longer 2144 * necessary since the inode is being deallocated. We set the 2145 * ALLCOMPLETE flags since the bitmap now properly shows that the 2146 * inode is not allocated. Even if the inode is actively being 2147 * written, it has been rolled back to its zero'ed state, so we 2148 * are ensured that a zero inode is what is on the disk. For short 2149 * lived files, this change will usually result in removing all the 2150 * dependencies from the inode so that it can be freed immediately. 2151 */ 2152 static int 2153 check_inode_unwritten(struct inodedep *inodedep) 2154 { 2155 2156 if ((inodedep->id_state & DEPCOMPLETE) != 0 || 2157 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2158 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2159 LIST_FIRST(&inodedep->id_inowait) != NULL || 2160 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2161 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2162 inodedep->id_nlinkdelta != 0) 2163 return (0); 2164 2165 /* 2166 * Another process might be in initiate_write_inodeblock 2167 * trying to allocate memory without holding "Softdep Lock". 2168 */ 2169 if ((inodedep->id_state & IOSTARTED) != 0 && 2170 inodedep->id_savedino == NULL) 2171 return(0); 2172 2173 inodedep->id_state |= ALLCOMPLETE; 2174 LIST_REMOVE(inodedep, id_deps); 2175 inodedep->id_buf = NULL; 2176 if (inodedep->id_state & ONWORKLIST) 2177 WORKLIST_REMOVE(&inodedep->id_list); 2178 if (inodedep->id_savedino != NULL) { 2179 FREE(inodedep->id_savedino, M_INODEDEP); 2180 inodedep->id_savedino = NULL; 2181 } 2182 if (free_inodedep(inodedep) == 0) { 2183 FREE_LOCK(&lk); 2184 panic("check_inode_unwritten: busy inode"); 2185 } 2186 return (1); 2187 } 2188 2189 /* 2190 * Try to free an inodedep structure. Return 1 if it could be freed. 2191 */ 2192 static int 2193 free_inodedep(struct inodedep *inodedep) 2194 { 2195 2196 if ((inodedep->id_state & ONWORKLIST) != 0 || 2197 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 2198 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2199 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2200 LIST_FIRST(&inodedep->id_inowait) != NULL || 2201 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2202 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2203 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL) 2204 return (0); 2205 LIST_REMOVE(inodedep, id_hash); 2206 WORKITEM_FREE(inodedep, D_INODEDEP); 2207 num_inodedep -= 1; 2208 return (1); 2209 } 2210 2211 /* 2212 * This workitem routine performs the block de-allocation. 2213 * The workitem is added to the pending list after the updated 2214 * inode block has been written to disk. As mentioned above, 2215 * checks regarding the number of blocks de-allocated (compared 2216 * to the number of blocks allocated for the file) are also 2217 * performed in this function. 2218 */ 2219 static void 2220 handle_workitem_freeblocks(struct freeblks *freeblks) 2221 { 2222 struct inode tip; 2223 ufs_daddr_t bn; 2224 struct fs *fs; 2225 int i, level, bsize; 2226 long nblocks, blocksreleased = 0; 2227 int error, allerror = 0; 2228 ufs_lbn_t baselbns[NIADDR], tmpval; 2229 2230 tip.i_number = freeblks->fb_previousinum; 2231 tip.i_devvp = freeblks->fb_devvp; 2232 tip.i_dev = freeblks->fb_devvp->v_rdev; 2233 tip.i_fs = freeblks->fb_fs; 2234 tip.i_size = freeblks->fb_oldsize; 2235 tip.i_uid = freeblks->fb_uid; 2236 fs = freeblks->fb_fs; 2237 tmpval = 1; 2238 baselbns[0] = NDADDR; 2239 for (i = 1; i < NIADDR; i++) { 2240 tmpval *= NINDIR(fs); 2241 baselbns[i] = baselbns[i - 1] + tmpval; 2242 } 2243 nblocks = btodb(fs->fs_bsize); 2244 blocksreleased = 0; 2245 /* 2246 * Indirect blocks first. 2247 */ 2248 for (level = (NIADDR - 1); level >= 0; level--) { 2249 if ((bn = freeblks->fb_iblks[level]) == 0) 2250 continue; 2251 if ((error = indir_trunc(&tip, fsbtodoff(fs, bn), level, 2252 baselbns[level], &blocksreleased)) == 0) 2253 allerror = error; 2254 ffs_blkfree(&tip, bn, fs->fs_bsize); 2255 blocksreleased += nblocks; 2256 } 2257 /* 2258 * All direct blocks or frags. 2259 */ 2260 for (i = (NDADDR - 1); i >= 0; i--) { 2261 if ((bn = freeblks->fb_dblks[i]) == 0) 2262 continue; 2263 bsize = blksize(fs, &tip, i); 2264 ffs_blkfree(&tip, bn, bsize); 2265 blocksreleased += btodb(bsize); 2266 } 2267 2268 #ifdef DIAGNOSTIC 2269 if (freeblks->fb_chkcnt != blocksreleased) 2270 kprintf("handle_workitem_freeblocks: block count\n"); 2271 if (allerror) 2272 softdep_error("handle_workitem_freeblks", allerror); 2273 #endif /* DIAGNOSTIC */ 2274 WORKITEM_FREE(freeblks, D_FREEBLKS); 2275 } 2276 2277 /* 2278 * Release blocks associated with the inode ip and stored in the indirect 2279 * block at doffset. If level is greater than SINGLE, the block is an 2280 * indirect block and recursive calls to indirtrunc must be used to 2281 * cleanse other indirect blocks. 2282 */ 2283 static int 2284 indir_trunc(struct inode *ip, off_t doffset, int level, ufs_lbn_t lbn, 2285 long *countp) 2286 { 2287 struct buf *bp; 2288 ufs_daddr_t *bap; 2289 ufs_daddr_t nb; 2290 struct fs *fs; 2291 struct worklist *wk; 2292 struct indirdep *indirdep; 2293 int i, lbnadd, nblocks; 2294 int error, allerror = 0; 2295 2296 fs = ip->i_fs; 2297 lbnadd = 1; 2298 for (i = level; i > 0; i--) 2299 lbnadd *= NINDIR(fs); 2300 /* 2301 * Get buffer of block pointers to be freed. This routine is not 2302 * called until the zero'ed inode has been written, so it is safe 2303 * to free blocks as they are encountered. Because the inode has 2304 * been zero'ed, calls to bmap on these blocks will fail. So, we 2305 * have to use the on-disk address and the block device for the 2306 * filesystem to look them up. If the file was deleted before its 2307 * indirect blocks were all written to disk, the routine that set 2308 * us up (deallocate_dependencies) will have arranged to leave 2309 * a complete copy of the indirect block in memory for our use. 2310 * Otherwise we have to read the blocks in from the disk. 2311 */ 2312 ACQUIRE_LOCK(&lk); 2313 if ((bp = findblk(ip->i_devvp, doffset)) != NULL && 2314 (wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2315 /* 2316 * bp must be ir_savebp, which is held locked for our use. 2317 */ 2318 if (wk->wk_type != D_INDIRDEP || 2319 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp || 2320 (indirdep->ir_state & GOINGAWAY) == 0) { 2321 FREE_LOCK(&lk); 2322 panic("indir_trunc: lost indirdep"); 2323 } 2324 WORKLIST_REMOVE(wk); 2325 WORKITEM_FREE(indirdep, D_INDIRDEP); 2326 if (LIST_FIRST(&bp->b_dep) != NULL) { 2327 FREE_LOCK(&lk); 2328 panic("indir_trunc: dangling dep"); 2329 } 2330 FREE_LOCK(&lk); 2331 } else { 2332 FREE_LOCK(&lk); 2333 error = bread(ip->i_devvp, doffset, (int)fs->fs_bsize, &bp); 2334 if (error) 2335 return (error); 2336 } 2337 /* 2338 * Recursively free indirect blocks. 2339 */ 2340 bap = (ufs_daddr_t *)bp->b_data; 2341 nblocks = btodb(fs->fs_bsize); 2342 for (i = NINDIR(fs) - 1; i >= 0; i--) { 2343 if ((nb = bap[i]) == 0) 2344 continue; 2345 if (level != 0) { 2346 if ((error = indir_trunc(ip, fsbtodoff(fs, nb), 2347 level - 1, lbn + (i * lbnadd), countp)) != 0) 2348 allerror = error; 2349 } 2350 ffs_blkfree(ip, nb, fs->fs_bsize); 2351 *countp += nblocks; 2352 } 2353 bp->b_flags |= B_INVAL | B_NOCACHE; 2354 brelse(bp); 2355 return (allerror); 2356 } 2357 2358 /* 2359 * Free an allocindir. 2360 * This routine must be called with splbio interrupts blocked. 2361 */ 2362 static void 2363 free_allocindir(struct allocindir *aip, struct inodedep *inodedep) 2364 { 2365 struct freefrag *freefrag; 2366 2367 #ifdef DEBUG 2368 if (lk.lkt_held == NOHOLDER) 2369 panic("free_allocindir: lock not held"); 2370 #endif 2371 if ((aip->ai_state & DEPCOMPLETE) == 0) 2372 LIST_REMOVE(aip, ai_deps); 2373 if (aip->ai_state & ONWORKLIST) 2374 WORKLIST_REMOVE(&aip->ai_list); 2375 LIST_REMOVE(aip, ai_next); 2376 if ((freefrag = aip->ai_freefrag) != NULL) { 2377 if (inodedep == NULL) 2378 add_to_worklist(&freefrag->ff_list); 2379 else 2380 WORKLIST_INSERT(&inodedep->id_bufwait, 2381 &freefrag->ff_list); 2382 } 2383 WORKITEM_FREE(aip, D_ALLOCINDIR); 2384 } 2385 2386 /* 2387 * Directory entry addition dependencies. 2388 * 2389 * When adding a new directory entry, the inode (with its incremented link 2390 * count) must be written to disk before the directory entry's pointer to it. 2391 * Also, if the inode is newly allocated, the corresponding freemap must be 2392 * updated (on disk) before the directory entry's pointer. These requirements 2393 * are met via undo/redo on the directory entry's pointer, which consists 2394 * simply of the inode number. 2395 * 2396 * As directory entries are added and deleted, the free space within a 2397 * directory block can become fragmented. The ufs filesystem will compact 2398 * a fragmented directory block to make space for a new entry. When this 2399 * occurs, the offsets of previously added entries change. Any "diradd" 2400 * dependency structures corresponding to these entries must be updated with 2401 * the new offsets. 2402 */ 2403 2404 /* 2405 * This routine is called after the in-memory inode's link 2406 * count has been incremented, but before the directory entry's 2407 * pointer to the inode has been set. 2408 * 2409 * Parameters: 2410 * bp: buffer containing directory block 2411 * dp: inode for directory 2412 * diroffset: offset of new entry in directory 2413 * newinum: inode referenced by new directory entry 2414 * newdirbp: non-NULL => contents of new mkdir 2415 */ 2416 void 2417 softdep_setup_directory_add(struct buf *bp, struct inode *dp, off_t diroffset, 2418 ino_t newinum, struct buf *newdirbp) 2419 { 2420 int offset; /* offset of new entry within directory block */ 2421 ufs_lbn_t lbn; /* block in directory containing new entry */ 2422 struct fs *fs; 2423 struct diradd *dap; 2424 struct pagedep *pagedep; 2425 struct inodedep *inodedep; 2426 struct mkdir *mkdir1, *mkdir2; 2427 2428 /* 2429 * Whiteouts have no dependencies. 2430 */ 2431 if (newinum == WINO) { 2432 if (newdirbp != NULL) 2433 bdwrite(newdirbp); 2434 return; 2435 } 2436 2437 fs = dp->i_fs; 2438 lbn = lblkno(fs, diroffset); 2439 offset = blkoff(fs, diroffset); 2440 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, 2441 M_SOFTDEP_FLAGS | M_ZERO); 2442 dap->da_list.wk_type = D_DIRADD; 2443 dap->da_offset = offset; 2444 dap->da_newinum = newinum; 2445 dap->da_state = ATTACHED; 2446 if (newdirbp == NULL) { 2447 dap->da_state |= DEPCOMPLETE; 2448 ACQUIRE_LOCK(&lk); 2449 } else { 2450 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 2451 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2452 M_SOFTDEP_FLAGS); 2453 mkdir1->md_list.wk_type = D_MKDIR; 2454 mkdir1->md_state = MKDIR_BODY; 2455 mkdir1->md_diradd = dap; 2456 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2457 M_SOFTDEP_FLAGS); 2458 mkdir2->md_list.wk_type = D_MKDIR; 2459 mkdir2->md_state = MKDIR_PARENT; 2460 mkdir2->md_diradd = dap; 2461 /* 2462 * Dependency on "." and ".." being written to disk. 2463 */ 2464 mkdir1->md_buf = newdirbp; 2465 ACQUIRE_LOCK(&lk); 2466 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 2467 WORKLIST_INSERT_BP(newdirbp, &mkdir1->md_list); 2468 FREE_LOCK(&lk); 2469 bdwrite(newdirbp); 2470 /* 2471 * Dependency on link count increase for parent directory 2472 */ 2473 ACQUIRE_LOCK(&lk); 2474 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0 2475 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2476 dap->da_state &= ~MKDIR_PARENT; 2477 WORKITEM_FREE(mkdir2, D_MKDIR); 2478 } else { 2479 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 2480 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 2481 } 2482 } 2483 /* 2484 * Link into parent directory pagedep to await its being written. 2485 */ 2486 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2487 WORKLIST_INSERT_BP(bp, &pagedep->pd_list); 2488 dap->da_pagedep = pagedep; 2489 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 2490 da_pdlist); 2491 /* 2492 * Link into its inodedep. Put it on the id_bufwait list if the inode 2493 * is not yet written. If it is written, do the post-inode write 2494 * processing to put it on the id_pendinghd list. 2495 */ 2496 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep); 2497 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 2498 diradd_inode_written(dap, inodedep); 2499 else 2500 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2501 FREE_LOCK(&lk); 2502 } 2503 2504 /* 2505 * This procedure is called to change the offset of a directory 2506 * entry when compacting a directory block which must be owned 2507 * exclusively by the caller. Note that the actual entry movement 2508 * must be done in this procedure to ensure that no I/O completions 2509 * occur while the move is in progress. 2510 * 2511 * Parameters: 2512 * dp: inode for directory 2513 * base: address of dp->i_offset 2514 * oldloc: address of old directory location 2515 * newloc: address of new directory location 2516 * entrysize: size of directory entry 2517 */ 2518 void 2519 softdep_change_directoryentry_offset(struct inode *dp, caddr_t base, 2520 caddr_t oldloc, caddr_t newloc, 2521 int entrysize) 2522 { 2523 int offset, oldoffset, newoffset; 2524 struct pagedep *pagedep; 2525 struct diradd *dap; 2526 ufs_lbn_t lbn; 2527 2528 ACQUIRE_LOCK(&lk); 2529 lbn = lblkno(dp->i_fs, dp->i_offset); 2530 offset = blkoff(dp->i_fs, dp->i_offset); 2531 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) 2532 goto done; 2533 oldoffset = offset + (oldloc - base); 2534 newoffset = offset + (newloc - base); 2535 2536 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) { 2537 if (dap->da_offset != oldoffset) 2538 continue; 2539 dap->da_offset = newoffset; 2540 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) 2541 break; 2542 LIST_REMOVE(dap, da_pdlist); 2543 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], 2544 dap, da_pdlist); 2545 break; 2546 } 2547 if (dap == NULL) { 2548 2549 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) { 2550 if (dap->da_offset == oldoffset) { 2551 dap->da_offset = newoffset; 2552 break; 2553 } 2554 } 2555 } 2556 done: 2557 bcopy(oldloc, newloc, entrysize); 2558 FREE_LOCK(&lk); 2559 } 2560 2561 /* 2562 * Free a diradd dependency structure. This routine must be called 2563 * with splbio interrupts blocked. 2564 */ 2565 static void 2566 free_diradd(struct diradd *dap) 2567 { 2568 struct dirrem *dirrem; 2569 struct pagedep *pagedep; 2570 struct inodedep *inodedep; 2571 struct mkdir *mkdir, *nextmd; 2572 2573 #ifdef DEBUG 2574 if (lk.lkt_held == NOHOLDER) 2575 panic("free_diradd: lock not held"); 2576 #endif 2577 WORKLIST_REMOVE(&dap->da_list); 2578 LIST_REMOVE(dap, da_pdlist); 2579 if ((dap->da_state & DIRCHG) == 0) { 2580 pagedep = dap->da_pagedep; 2581 } else { 2582 dirrem = dap->da_previous; 2583 pagedep = dirrem->dm_pagedep; 2584 dirrem->dm_dirinum = pagedep->pd_ino; 2585 add_to_worklist(&dirrem->dm_list); 2586 } 2587 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum, 2588 0, &inodedep) != 0) 2589 (void) free_inodedep(inodedep); 2590 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2591 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 2592 nextmd = LIST_NEXT(mkdir, md_mkdirs); 2593 if (mkdir->md_diradd != dap) 2594 continue; 2595 dap->da_state &= ~mkdir->md_state; 2596 WORKLIST_REMOVE(&mkdir->md_list); 2597 LIST_REMOVE(mkdir, md_mkdirs); 2598 WORKITEM_FREE(mkdir, D_MKDIR); 2599 } 2600 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2601 FREE_LOCK(&lk); 2602 panic("free_diradd: unfound ref"); 2603 } 2604 } 2605 WORKITEM_FREE(dap, D_DIRADD); 2606 } 2607 2608 /* 2609 * Directory entry removal dependencies. 2610 * 2611 * When removing a directory entry, the entry's inode pointer must be 2612 * zero'ed on disk before the corresponding inode's link count is decremented 2613 * (possibly freeing the inode for re-use). This dependency is handled by 2614 * updating the directory entry but delaying the inode count reduction until 2615 * after the directory block has been written to disk. After this point, the 2616 * inode count can be decremented whenever it is convenient. 2617 */ 2618 2619 /* 2620 * This routine should be called immediately after removing 2621 * a directory entry. The inode's link count should not be 2622 * decremented by the calling procedure -- the soft updates 2623 * code will do this task when it is safe. 2624 * 2625 * Parameters: 2626 * bp: buffer containing directory block 2627 * dp: inode for the directory being modified 2628 * ip: inode for directory entry being removed 2629 * isrmdir: indicates if doing RMDIR 2630 */ 2631 void 2632 softdep_setup_remove(struct buf *bp, struct inode *dp, struct inode *ip, 2633 int isrmdir) 2634 { 2635 struct dirrem *dirrem, *prevdirrem; 2636 2637 /* 2638 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 2639 */ 2640 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2641 2642 /* 2643 * If the COMPLETE flag is clear, then there were no active 2644 * entries and we want to roll back to a zeroed entry until 2645 * the new inode is committed to disk. If the COMPLETE flag is 2646 * set then we have deleted an entry that never made it to 2647 * disk. If the entry we deleted resulted from a name change, 2648 * then the old name still resides on disk. We cannot delete 2649 * its inode (returned to us in prevdirrem) until the zeroed 2650 * directory entry gets to disk. The new inode has never been 2651 * referenced on the disk, so can be deleted immediately. 2652 */ 2653 if ((dirrem->dm_state & COMPLETE) == 0) { 2654 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 2655 dm_next); 2656 FREE_LOCK(&lk); 2657 } else { 2658 if (prevdirrem != NULL) 2659 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 2660 prevdirrem, dm_next); 2661 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 2662 FREE_LOCK(&lk); 2663 handle_workitem_remove(dirrem); 2664 } 2665 } 2666 2667 /* 2668 * Allocate a new dirrem if appropriate and return it along with 2669 * its associated pagedep. Called without a lock, returns with lock. 2670 */ 2671 static long num_dirrem; /* number of dirrem allocated */ 2672 2673 /* 2674 * Parameters: 2675 * bp: buffer containing directory block 2676 * dp: inode for the directory being modified 2677 * ip: inode for directory entry being removed 2678 * isrmdir: indicates if doing RMDIR 2679 * prevdirremp: previously referenced inode, if any 2680 */ 2681 static struct dirrem * 2682 newdirrem(struct buf *bp, struct inode *dp, struct inode *ip, 2683 int isrmdir, struct dirrem **prevdirremp) 2684 { 2685 int offset; 2686 ufs_lbn_t lbn; 2687 struct diradd *dap; 2688 struct dirrem *dirrem; 2689 struct pagedep *pagedep; 2690 2691 /* 2692 * Whiteouts have no deletion dependencies. 2693 */ 2694 if (ip == NULL) 2695 panic("newdirrem: whiteout"); 2696 /* 2697 * If we are over our limit, try to improve the situation. 2698 * Limiting the number of dirrem structures will also limit 2699 * the number of freefile and freeblks structures. 2700 */ 2701 if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0) 2702 (void) request_cleanup(FLUSH_REMOVE, 0); 2703 num_dirrem += 1; 2704 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem), 2705 M_DIRREM, M_SOFTDEP_FLAGS | M_ZERO); 2706 dirrem->dm_list.wk_type = D_DIRREM; 2707 dirrem->dm_state = isrmdir ? RMDIR : 0; 2708 dirrem->dm_mnt = ITOV(ip)->v_mount; 2709 dirrem->dm_oldinum = ip->i_number; 2710 *prevdirremp = NULL; 2711 2712 ACQUIRE_LOCK(&lk); 2713 lbn = lblkno(dp->i_fs, dp->i_offset); 2714 offset = blkoff(dp->i_fs, dp->i_offset); 2715 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2716 WORKLIST_INSERT_BP(bp, &pagedep->pd_list); 2717 dirrem->dm_pagedep = pagedep; 2718 /* 2719 * Check for a diradd dependency for the same directory entry. 2720 * If present, then both dependencies become obsolete and can 2721 * be de-allocated. Check for an entry on both the pd_dirraddhd 2722 * list and the pd_pendinghd list. 2723 */ 2724 2725 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 2726 if (dap->da_offset == offset) 2727 break; 2728 if (dap == NULL) { 2729 2730 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 2731 if (dap->da_offset == offset) 2732 break; 2733 if (dap == NULL) 2734 return (dirrem); 2735 } 2736 /* 2737 * Must be ATTACHED at this point. 2738 */ 2739 if ((dap->da_state & ATTACHED) == 0) { 2740 FREE_LOCK(&lk); 2741 panic("newdirrem: not ATTACHED"); 2742 } 2743 if (dap->da_newinum != ip->i_number) { 2744 FREE_LOCK(&lk); 2745 panic("newdirrem: inum %"PRId64" should be %"PRId64, 2746 ip->i_number, dap->da_newinum); 2747 } 2748 /* 2749 * If we are deleting a changed name that never made it to disk, 2750 * then return the dirrem describing the previous inode (which 2751 * represents the inode currently referenced from this entry on disk). 2752 */ 2753 if ((dap->da_state & DIRCHG) != 0) { 2754 *prevdirremp = dap->da_previous; 2755 dap->da_state &= ~DIRCHG; 2756 dap->da_pagedep = pagedep; 2757 } 2758 /* 2759 * We are deleting an entry that never made it to disk. 2760 * Mark it COMPLETE so we can delete its inode immediately. 2761 */ 2762 dirrem->dm_state |= COMPLETE; 2763 free_diradd(dap); 2764 return (dirrem); 2765 } 2766 2767 /* 2768 * Directory entry change dependencies. 2769 * 2770 * Changing an existing directory entry requires that an add operation 2771 * be completed first followed by a deletion. The semantics for the addition 2772 * are identical to the description of adding a new entry above except 2773 * that the rollback is to the old inode number rather than zero. Once 2774 * the addition dependency is completed, the removal is done as described 2775 * in the removal routine above. 2776 */ 2777 2778 /* 2779 * This routine should be called immediately after changing 2780 * a directory entry. The inode's link count should not be 2781 * decremented by the calling procedure -- the soft updates 2782 * code will perform this task when it is safe. 2783 * 2784 * Parameters: 2785 * bp: buffer containing directory block 2786 * dp: inode for the directory being modified 2787 * ip: inode for directory entry being removed 2788 * newinum: new inode number for changed entry 2789 * isrmdir: indicates if doing RMDIR 2790 */ 2791 void 2792 softdep_setup_directory_change(struct buf *bp, struct inode *dp, 2793 struct inode *ip, ino_t newinum, 2794 int isrmdir) 2795 { 2796 int offset; 2797 struct diradd *dap = NULL; 2798 struct dirrem *dirrem, *prevdirrem; 2799 struct pagedep *pagedep; 2800 struct inodedep *inodedep; 2801 2802 offset = blkoff(dp->i_fs, dp->i_offset); 2803 2804 /* 2805 * Whiteouts do not need diradd dependencies. 2806 */ 2807 if (newinum != WINO) { 2808 MALLOC(dap, struct diradd *, sizeof(struct diradd), 2809 M_DIRADD, M_SOFTDEP_FLAGS | M_ZERO); 2810 dap->da_list.wk_type = D_DIRADD; 2811 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 2812 dap->da_offset = offset; 2813 dap->da_newinum = newinum; 2814 } 2815 2816 /* 2817 * Allocate a new dirrem and ACQUIRE_LOCK. 2818 */ 2819 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2820 pagedep = dirrem->dm_pagedep; 2821 /* 2822 * The possible values for isrmdir: 2823 * 0 - non-directory file rename 2824 * 1 - directory rename within same directory 2825 * inum - directory rename to new directory of given inode number 2826 * When renaming to a new directory, we are both deleting and 2827 * creating a new directory entry, so the link count on the new 2828 * directory should not change. Thus we do not need the followup 2829 * dirrem which is usually done in handle_workitem_remove. We set 2830 * the DIRCHG flag to tell handle_workitem_remove to skip the 2831 * followup dirrem. 2832 */ 2833 if (isrmdir > 1) 2834 dirrem->dm_state |= DIRCHG; 2835 2836 /* 2837 * Whiteouts have no additional dependencies, 2838 * so just put the dirrem on the correct list. 2839 */ 2840 if (newinum == WINO) { 2841 if ((dirrem->dm_state & COMPLETE) == 0) { 2842 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 2843 dm_next); 2844 } else { 2845 dirrem->dm_dirinum = pagedep->pd_ino; 2846 add_to_worklist(&dirrem->dm_list); 2847 } 2848 FREE_LOCK(&lk); 2849 return; 2850 } 2851 2852 /* 2853 * If the COMPLETE flag is clear, then there were no active 2854 * entries and we want to roll back to the previous inode until 2855 * the new inode is committed to disk. If the COMPLETE flag is 2856 * set, then we have deleted an entry that never made it to disk. 2857 * If the entry we deleted resulted from a name change, then the old 2858 * inode reference still resides on disk. Any rollback that we do 2859 * needs to be to that old inode (returned to us in prevdirrem). If 2860 * the entry we deleted resulted from a create, then there is 2861 * no entry on the disk, so we want to roll back to zero rather 2862 * than the uncommitted inode. In either of the COMPLETE cases we 2863 * want to immediately free the unwritten and unreferenced inode. 2864 */ 2865 if ((dirrem->dm_state & COMPLETE) == 0) { 2866 dap->da_previous = dirrem; 2867 } else { 2868 if (prevdirrem != NULL) { 2869 dap->da_previous = prevdirrem; 2870 } else { 2871 dap->da_state &= ~DIRCHG; 2872 dap->da_pagedep = pagedep; 2873 } 2874 dirrem->dm_dirinum = pagedep->pd_ino; 2875 add_to_worklist(&dirrem->dm_list); 2876 } 2877 /* 2878 * Link into its inodedep. Put it on the id_bufwait list if the inode 2879 * is not yet written. If it is written, do the post-inode write 2880 * processing to put it on the id_pendinghd list. 2881 */ 2882 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 || 2883 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2884 dap->da_state |= COMPLETE; 2885 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 2886 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 2887 } else { 2888 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 2889 dap, da_pdlist); 2890 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2891 } 2892 FREE_LOCK(&lk); 2893 } 2894 2895 /* 2896 * Called whenever the link count on an inode is changed. 2897 * It creates an inode dependency so that the new reference(s) 2898 * to the inode cannot be committed to disk until the updated 2899 * inode has been written. 2900 * 2901 * Parameters: 2902 * ip: the inode with the increased link count 2903 */ 2904 void 2905 softdep_change_linkcnt(struct inode *ip) 2906 { 2907 struct inodedep *inodedep; 2908 2909 ACQUIRE_LOCK(&lk); 2910 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 2911 if (ip->i_nlink < ip->i_effnlink) { 2912 FREE_LOCK(&lk); 2913 panic("softdep_change_linkcnt: bad delta"); 2914 } 2915 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2916 FREE_LOCK(&lk); 2917 } 2918 2919 /* 2920 * This workitem decrements the inode's link count. 2921 * If the link count reaches zero, the file is removed. 2922 */ 2923 static void 2924 handle_workitem_remove(struct dirrem *dirrem) 2925 { 2926 struct inodedep *inodedep; 2927 struct vnode *vp; 2928 struct inode *ip; 2929 ino_t oldinum; 2930 int error; 2931 2932 if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) { 2933 softdep_error("handle_workitem_remove: vget", error); 2934 return; 2935 } 2936 ip = VTOI(vp); 2937 ACQUIRE_LOCK(&lk); 2938 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){ 2939 FREE_LOCK(&lk); 2940 panic("handle_workitem_remove: lost inodedep"); 2941 } 2942 /* 2943 * Normal file deletion. 2944 */ 2945 if ((dirrem->dm_state & RMDIR) == 0) { 2946 ip->i_nlink--; 2947 ip->i_flag |= IN_CHANGE; 2948 if (ip->i_nlink < ip->i_effnlink) { 2949 FREE_LOCK(&lk); 2950 panic("handle_workitem_remove: bad file delta"); 2951 } 2952 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2953 FREE_LOCK(&lk); 2954 vput(vp); 2955 num_dirrem -= 1; 2956 WORKITEM_FREE(dirrem, D_DIRREM); 2957 return; 2958 } 2959 /* 2960 * Directory deletion. Decrement reference count for both the 2961 * just deleted parent directory entry and the reference for ".". 2962 * Next truncate the directory to length zero. When the 2963 * truncation completes, arrange to have the reference count on 2964 * the parent decremented to account for the loss of "..". 2965 */ 2966 ip->i_nlink -= 2; 2967 ip->i_flag |= IN_CHANGE; 2968 if (ip->i_nlink < ip->i_effnlink) { 2969 FREE_LOCK(&lk); 2970 panic("handle_workitem_remove: bad dir delta"); 2971 } 2972 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2973 FREE_LOCK(&lk); 2974 if ((error = ffs_truncate(vp, (off_t)0, 0, proc0.p_ucred)) != 0) 2975 softdep_error("handle_workitem_remove: truncate", error); 2976 /* 2977 * Rename a directory to a new parent. Since, we are both deleting 2978 * and creating a new directory entry, the link count on the new 2979 * directory should not change. Thus we skip the followup dirrem. 2980 */ 2981 if (dirrem->dm_state & DIRCHG) { 2982 vput(vp); 2983 num_dirrem -= 1; 2984 WORKITEM_FREE(dirrem, D_DIRREM); 2985 return; 2986 } 2987 /* 2988 * If the inodedep does not exist, then the zero'ed inode has 2989 * been written to disk. If the allocated inode has never been 2990 * written to disk, then the on-disk inode is zero'ed. In either 2991 * case we can remove the file immediately. 2992 */ 2993 ACQUIRE_LOCK(&lk); 2994 dirrem->dm_state = 0; 2995 oldinum = dirrem->dm_oldinum; 2996 dirrem->dm_oldinum = dirrem->dm_dirinum; 2997 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 || 2998 check_inode_unwritten(inodedep)) { 2999 FREE_LOCK(&lk); 3000 vput(vp); 3001 handle_workitem_remove(dirrem); 3002 return; 3003 } 3004 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 3005 FREE_LOCK(&lk); 3006 ip->i_flag |= IN_CHANGE; 3007 ffs_update(vp, 0); 3008 vput(vp); 3009 } 3010 3011 /* 3012 * Inode de-allocation dependencies. 3013 * 3014 * When an inode's link count is reduced to zero, it can be de-allocated. We 3015 * found it convenient to postpone de-allocation until after the inode is 3016 * written to disk with its new link count (zero). At this point, all of the 3017 * on-disk inode's block pointers are nullified and, with careful dependency 3018 * list ordering, all dependencies related to the inode will be satisfied and 3019 * the corresponding dependency structures de-allocated. So, if/when the 3020 * inode is reused, there will be no mixing of old dependencies with new 3021 * ones. This artificial dependency is set up by the block de-allocation 3022 * procedure above (softdep_setup_freeblocks) and completed by the 3023 * following procedure. 3024 */ 3025 static void 3026 handle_workitem_freefile(struct freefile *freefile) 3027 { 3028 struct vnode vp; 3029 struct inode tip; 3030 struct inodedep *idp; 3031 int error; 3032 3033 #ifdef DEBUG 3034 ACQUIRE_LOCK(&lk); 3035 error = inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp); 3036 FREE_LOCK(&lk); 3037 if (error) 3038 panic("handle_workitem_freefile: inodedep survived"); 3039 #endif 3040 tip.i_devvp = freefile->fx_devvp; 3041 tip.i_dev = freefile->fx_devvp->v_rdev; 3042 tip.i_fs = freefile->fx_fs; 3043 vp.v_data = &tip; 3044 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0) 3045 softdep_error("handle_workitem_freefile", error); 3046 WORKITEM_FREE(freefile, D_FREEFILE); 3047 } 3048 3049 /* 3050 * Helper function which unlinks marker element from work list and returns 3051 * the next element on the list. 3052 */ 3053 static __inline struct worklist * 3054 markernext(struct worklist *marker) 3055 { 3056 struct worklist *next; 3057 3058 next = LIST_NEXT(marker, wk_list); 3059 LIST_REMOVE(marker, wk_list); 3060 return next; 3061 } 3062 3063 /* 3064 * checkread, checkwrite 3065 * 3066 */ 3067 static int 3068 softdep_checkread(struct buf *bp) 3069 { 3070 return(0); 3071 } 3072 3073 static int 3074 softdep_checkwrite(struct buf *bp) 3075 { 3076 return(0); 3077 } 3078 3079 /* 3080 * Disk writes. 3081 * 3082 * The dependency structures constructed above are most actively used when file 3083 * system blocks are written to disk. No constraints are placed on when a 3084 * block can be written, but unsatisfied update dependencies are made safe by 3085 * modifying (or replacing) the source memory for the duration of the disk 3086 * write. When the disk write completes, the memory block is again brought 3087 * up-to-date. 3088 * 3089 * In-core inode structure reclamation. 3090 * 3091 * Because there are a finite number of "in-core" inode structures, they are 3092 * reused regularly. By transferring all inode-related dependencies to the 3093 * in-memory inode block and indexing them separately (via "inodedep"s), we 3094 * can allow "in-core" inode structures to be reused at any time and avoid 3095 * any increase in contention. 3096 * 3097 * Called just before entering the device driver to initiate a new disk I/O. 3098 * The buffer must be locked, thus, no I/O completion operations can occur 3099 * while we are manipulating its associated dependencies. 3100 * 3101 * Parameters: 3102 * bp: structure describing disk write to occur 3103 */ 3104 static void 3105 softdep_disk_io_initiation(struct buf *bp) 3106 { 3107 struct worklist *wk; 3108 struct worklist marker; 3109 struct indirdep *indirdep; 3110 3111 /* 3112 * We only care about write operations. There should never 3113 * be dependencies for reads. 3114 */ 3115 if (bp->b_cmd == BUF_CMD_READ) 3116 panic("softdep_disk_io_initiation: read"); 3117 3118 marker.wk_type = D_LAST + 1; /* Not a normal workitem */ 3119 3120 /* 3121 * Do any necessary pre-I/O processing. 3122 */ 3123 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = markernext(&marker)) { 3124 LIST_INSERT_AFTER(wk, &marker, wk_list); 3125 3126 switch (wk->wk_type) { 3127 3128 case D_PAGEDEP: 3129 initiate_write_filepage(WK_PAGEDEP(wk), bp); 3130 continue; 3131 3132 case D_INODEDEP: 3133 initiate_write_inodeblock(WK_INODEDEP(wk), bp); 3134 continue; 3135 3136 case D_INDIRDEP: 3137 indirdep = WK_INDIRDEP(wk); 3138 if (indirdep->ir_state & GOINGAWAY) 3139 panic("disk_io_initiation: indirdep gone"); 3140 /* 3141 * If there are no remaining dependencies, this 3142 * will be writing the real pointers, so the 3143 * dependency can be freed. 3144 */ 3145 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) { 3146 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 3147 brelse(indirdep->ir_savebp); 3148 /* inline expand WORKLIST_REMOVE(wk); */ 3149 wk->wk_state &= ~ONWORKLIST; 3150 LIST_REMOVE(wk, wk_list); 3151 WORKITEM_FREE(indirdep, D_INDIRDEP); 3152 continue; 3153 } 3154 /* 3155 * Replace up-to-date version with safe version. 3156 */ 3157 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount, 3158 M_INDIRDEP, M_SOFTDEP_FLAGS); 3159 ACQUIRE_LOCK(&lk); 3160 indirdep->ir_state &= ~ATTACHED; 3161 indirdep->ir_state |= UNDONE; 3162 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 3163 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 3164 bp->b_bcount); 3165 FREE_LOCK(&lk); 3166 continue; 3167 3168 case D_MKDIR: 3169 case D_BMSAFEMAP: 3170 case D_ALLOCDIRECT: 3171 case D_ALLOCINDIR: 3172 continue; 3173 3174 default: 3175 panic("handle_disk_io_initiation: Unexpected type %s", 3176 TYPENAME(wk->wk_type)); 3177 /* NOTREACHED */ 3178 } 3179 } 3180 } 3181 3182 /* 3183 * Called from within the procedure above to deal with unsatisfied 3184 * allocation dependencies in a directory. The buffer must be locked, 3185 * thus, no I/O completion operations can occur while we are 3186 * manipulating its associated dependencies. 3187 */ 3188 static void 3189 initiate_write_filepage(struct pagedep *pagedep, struct buf *bp) 3190 { 3191 struct diradd *dap; 3192 struct direct *ep; 3193 int i; 3194 3195 if (pagedep->pd_state & IOSTARTED) { 3196 /* 3197 * This can only happen if there is a driver that does not 3198 * understand chaining. Here biodone will reissue the call 3199 * to strategy for the incomplete buffers. 3200 */ 3201 kprintf("initiate_write_filepage: already started\n"); 3202 return; 3203 } 3204 pagedep->pd_state |= IOSTARTED; 3205 ACQUIRE_LOCK(&lk); 3206 for (i = 0; i < DAHASHSZ; i++) { 3207 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 3208 ep = (struct direct *) 3209 ((char *)bp->b_data + dap->da_offset); 3210 if (ep->d_ino != dap->da_newinum) { 3211 FREE_LOCK(&lk); 3212 panic("%s: dir inum %d != new %"PRId64, 3213 "initiate_write_filepage", 3214 ep->d_ino, dap->da_newinum); 3215 } 3216 if (dap->da_state & DIRCHG) 3217 ep->d_ino = dap->da_previous->dm_oldinum; 3218 else 3219 ep->d_ino = 0; 3220 dap->da_state &= ~ATTACHED; 3221 dap->da_state |= UNDONE; 3222 } 3223 } 3224 FREE_LOCK(&lk); 3225 } 3226 3227 /* 3228 * Called from within the procedure above to deal with unsatisfied 3229 * allocation dependencies in an inodeblock. The buffer must be 3230 * locked, thus, no I/O completion operations can occur while we 3231 * are manipulating its associated dependencies. 3232 * 3233 * Parameters: 3234 * bp: The inode block 3235 */ 3236 static void 3237 initiate_write_inodeblock(struct inodedep *inodedep, struct buf *bp) 3238 { 3239 struct allocdirect *adp, *lastadp; 3240 struct ufs1_dinode *dp; 3241 struct ufs1_dinode *sip; 3242 struct fs *fs; 3243 ufs_lbn_t prevlbn = 0; 3244 int i, deplist; 3245 3246 if (inodedep->id_state & IOSTARTED) 3247 panic("initiate_write_inodeblock: already started"); 3248 inodedep->id_state |= IOSTARTED; 3249 fs = inodedep->id_fs; 3250 dp = (struct ufs1_dinode *)bp->b_data + 3251 ino_to_fsbo(fs, inodedep->id_ino); 3252 /* 3253 * If the bitmap is not yet written, then the allocated 3254 * inode cannot be written to disk. 3255 */ 3256 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 3257 if (inodedep->id_savedino != NULL) 3258 panic("initiate_write_inodeblock: already doing I/O"); 3259 MALLOC(sip, struct ufs1_dinode *, 3260 sizeof(struct ufs1_dinode), M_INODEDEP, M_SOFTDEP_FLAGS); 3261 inodedep->id_savedino = sip; 3262 *inodedep->id_savedino = *dp; 3263 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 3264 dp->di_gen = inodedep->id_savedino->di_gen; 3265 return; 3266 } 3267 /* 3268 * If no dependencies, then there is nothing to roll back. 3269 */ 3270 inodedep->id_savedsize = dp->di_size; 3271 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 3272 return; 3273 /* 3274 * Set the dependencies to busy. 3275 */ 3276 ACQUIRE_LOCK(&lk); 3277 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3278 adp = TAILQ_NEXT(adp, ad_next)) { 3279 #ifdef DIAGNOSTIC 3280 if (deplist != 0 && prevlbn >= adp->ad_lbn) { 3281 FREE_LOCK(&lk); 3282 panic("softdep_write_inodeblock: lbn order"); 3283 } 3284 prevlbn = adp->ad_lbn; 3285 if (adp->ad_lbn < NDADDR && 3286 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) { 3287 FREE_LOCK(&lk); 3288 panic("%s: direct pointer #%ld mismatch %d != %d", 3289 "softdep_write_inodeblock", adp->ad_lbn, 3290 dp->di_db[adp->ad_lbn], adp->ad_newblkno); 3291 } 3292 if (adp->ad_lbn >= NDADDR && 3293 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) { 3294 FREE_LOCK(&lk); 3295 panic("%s: indirect pointer #%ld mismatch %d != %d", 3296 "softdep_write_inodeblock", adp->ad_lbn - NDADDR, 3297 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno); 3298 } 3299 deplist |= 1 << adp->ad_lbn; 3300 if ((adp->ad_state & ATTACHED) == 0) { 3301 FREE_LOCK(&lk); 3302 panic("softdep_write_inodeblock: Unknown state 0x%x", 3303 adp->ad_state); 3304 } 3305 #endif /* DIAGNOSTIC */ 3306 adp->ad_state &= ~ATTACHED; 3307 adp->ad_state |= UNDONE; 3308 } 3309 /* 3310 * The on-disk inode cannot claim to be any larger than the last 3311 * fragment that has been written. Otherwise, the on-disk inode 3312 * might have fragments that were not the last block in the file 3313 * which would corrupt the filesystem. 3314 */ 3315 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3316 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 3317 if (adp->ad_lbn >= NDADDR) 3318 break; 3319 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 3320 /* keep going until hitting a rollback to a frag */ 3321 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 3322 continue; 3323 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 3324 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 3325 #ifdef DIAGNOSTIC 3326 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) { 3327 FREE_LOCK(&lk); 3328 panic("softdep_write_inodeblock: lost dep1"); 3329 } 3330 #endif /* DIAGNOSTIC */ 3331 dp->di_db[i] = 0; 3332 } 3333 for (i = 0; i < NIADDR; i++) { 3334 #ifdef DIAGNOSTIC 3335 if (dp->di_ib[i] != 0 && 3336 (deplist & ((1 << NDADDR) << i)) == 0) { 3337 FREE_LOCK(&lk); 3338 panic("softdep_write_inodeblock: lost dep2"); 3339 } 3340 #endif /* DIAGNOSTIC */ 3341 dp->di_ib[i] = 0; 3342 } 3343 FREE_LOCK(&lk); 3344 return; 3345 } 3346 /* 3347 * If we have zero'ed out the last allocated block of the file, 3348 * roll back the size to the last currently allocated block. 3349 * We know that this last allocated block is a full-sized as 3350 * we already checked for fragments in the loop above. 3351 */ 3352 if (lastadp != NULL && 3353 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 3354 for (i = lastadp->ad_lbn; i >= 0; i--) 3355 if (dp->di_db[i] != 0) 3356 break; 3357 dp->di_size = (i + 1) * fs->fs_bsize; 3358 } 3359 /* 3360 * The only dependencies are for indirect blocks. 3361 * 3362 * The file size for indirect block additions is not guaranteed. 3363 * Such a guarantee would be non-trivial to achieve. The conventional 3364 * synchronous write implementation also does not make this guarantee. 3365 * Fsck should catch and fix discrepancies. Arguably, the file size 3366 * can be over-estimated without destroying integrity when the file 3367 * moves into the indirect blocks (i.e., is large). If we want to 3368 * postpone fsck, we are stuck with this argument. 3369 */ 3370 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 3371 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 3372 FREE_LOCK(&lk); 3373 } 3374 3375 /* 3376 * This routine is called during the completion interrupt 3377 * service routine for a disk write (from the procedure called 3378 * by the device driver to inform the filesystem caches of 3379 * a request completion). It should be called early in this 3380 * procedure, before the block is made available to other 3381 * processes or other routines are called. 3382 * 3383 * Parameters: 3384 * bp: describes the completed disk write 3385 */ 3386 static void 3387 softdep_disk_write_complete(struct buf *bp) 3388 { 3389 struct worklist *wk; 3390 struct workhead reattach; 3391 struct newblk *newblk; 3392 struct allocindir *aip; 3393 struct allocdirect *adp; 3394 struct indirdep *indirdep; 3395 struct inodedep *inodedep; 3396 struct bmsafemap *bmsafemap; 3397 3398 #ifdef DEBUG 3399 if (lk.lkt_held != NOHOLDER) 3400 panic("softdep_disk_write_complete: lock is held"); 3401 lk.lkt_held = SPECIAL_FLAG; 3402 #endif 3403 LIST_INIT(&reattach); 3404 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 3405 WORKLIST_REMOVE(wk); 3406 switch (wk->wk_type) { 3407 3408 case D_PAGEDEP: 3409 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 3410 WORKLIST_INSERT(&reattach, wk); 3411 continue; 3412 3413 case D_INODEDEP: 3414 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 3415 WORKLIST_INSERT(&reattach, wk); 3416 continue; 3417 3418 case D_BMSAFEMAP: 3419 bmsafemap = WK_BMSAFEMAP(wk); 3420 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) { 3421 newblk->nb_state |= DEPCOMPLETE; 3422 newblk->nb_bmsafemap = NULL; 3423 LIST_REMOVE(newblk, nb_deps); 3424 } 3425 while ((adp = 3426 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) { 3427 adp->ad_state |= DEPCOMPLETE; 3428 adp->ad_buf = NULL; 3429 LIST_REMOVE(adp, ad_deps); 3430 handle_allocdirect_partdone(adp); 3431 } 3432 while ((aip = 3433 LIST_FIRST(&bmsafemap->sm_allocindirhd))) { 3434 aip->ai_state |= DEPCOMPLETE; 3435 aip->ai_buf = NULL; 3436 LIST_REMOVE(aip, ai_deps); 3437 handle_allocindir_partdone(aip); 3438 } 3439 while ((inodedep = 3440 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { 3441 inodedep->id_state |= DEPCOMPLETE; 3442 LIST_REMOVE(inodedep, id_deps); 3443 inodedep->id_buf = NULL; 3444 } 3445 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 3446 continue; 3447 3448 case D_MKDIR: 3449 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 3450 continue; 3451 3452 case D_ALLOCDIRECT: 3453 adp = WK_ALLOCDIRECT(wk); 3454 adp->ad_state |= COMPLETE; 3455 handle_allocdirect_partdone(adp); 3456 continue; 3457 3458 case D_ALLOCINDIR: 3459 aip = WK_ALLOCINDIR(wk); 3460 aip->ai_state |= COMPLETE; 3461 handle_allocindir_partdone(aip); 3462 continue; 3463 3464 case D_INDIRDEP: 3465 indirdep = WK_INDIRDEP(wk); 3466 if (indirdep->ir_state & GOINGAWAY) { 3467 lk.lkt_held = NOHOLDER; 3468 panic("disk_write_complete: indirdep gone"); 3469 } 3470 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 3471 FREE(indirdep->ir_saveddata, M_INDIRDEP); 3472 indirdep->ir_saveddata = 0; 3473 indirdep->ir_state &= ~UNDONE; 3474 indirdep->ir_state |= ATTACHED; 3475 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 3476 handle_allocindir_partdone(aip); 3477 if (aip == LIST_FIRST(&indirdep->ir_donehd)) { 3478 lk.lkt_held = NOHOLDER; 3479 panic("disk_write_complete: not gone"); 3480 } 3481 } 3482 WORKLIST_INSERT(&reattach, wk); 3483 if ((bp->b_flags & B_DELWRI) == 0) 3484 stat_indir_blk_ptrs++; 3485 bdirty(bp); 3486 continue; 3487 3488 default: 3489 lk.lkt_held = NOHOLDER; 3490 panic("handle_disk_write_complete: Unknown type %s", 3491 TYPENAME(wk->wk_type)); 3492 /* NOTREACHED */ 3493 } 3494 } 3495 /* 3496 * Reattach any requests that must be redone. 3497 */ 3498 while ((wk = LIST_FIRST(&reattach)) != NULL) { 3499 WORKLIST_REMOVE(wk); 3500 WORKLIST_INSERT_BP(bp, wk); 3501 } 3502 #ifdef DEBUG 3503 if (lk.lkt_held != SPECIAL_FLAG) 3504 panic("softdep_disk_write_complete: lock lost"); 3505 lk.lkt_held = NOHOLDER; 3506 #endif 3507 } 3508 3509 /* 3510 * Called from within softdep_disk_write_complete above. Note that 3511 * this routine is always called from interrupt level with further 3512 * splbio interrupts blocked. 3513 * 3514 * Parameters: 3515 * adp: the completed allocdirect 3516 */ 3517 static void 3518 handle_allocdirect_partdone(struct allocdirect *adp) 3519 { 3520 struct allocdirect *listadp; 3521 struct inodedep *inodedep; 3522 long bsize; 3523 3524 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3525 return; 3526 if (adp->ad_buf != NULL) { 3527 lk.lkt_held = NOHOLDER; 3528 panic("handle_allocdirect_partdone: dangling dep"); 3529 } 3530 /* 3531 * The on-disk inode cannot claim to be any larger than the last 3532 * fragment that has been written. Otherwise, the on-disk inode 3533 * might have fragments that were not the last block in the file 3534 * which would corrupt the filesystem. Thus, we cannot free any 3535 * allocdirects after one whose ad_oldblkno claims a fragment as 3536 * these blocks must be rolled back to zero before writing the inode. 3537 * We check the currently active set of allocdirects in id_inoupdt. 3538 */ 3539 inodedep = adp->ad_inodedep; 3540 bsize = inodedep->id_fs->fs_bsize; 3541 TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) { 3542 /* found our block */ 3543 if (listadp == adp) 3544 break; 3545 /* continue if ad_oldlbn is not a fragment */ 3546 if (listadp->ad_oldsize == 0 || 3547 listadp->ad_oldsize == bsize) 3548 continue; 3549 /* hit a fragment */ 3550 return; 3551 } 3552 /* 3553 * If we have reached the end of the current list without 3554 * finding the just finished dependency, then it must be 3555 * on the future dependency list. Future dependencies cannot 3556 * be freed until they are moved to the current list. 3557 */ 3558 if (listadp == NULL) { 3559 #ifdef DEBUG 3560 TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next) 3561 /* found our block */ 3562 if (listadp == adp) 3563 break; 3564 if (listadp == NULL) { 3565 lk.lkt_held = NOHOLDER; 3566 panic("handle_allocdirect_partdone: lost dep"); 3567 } 3568 #endif /* DEBUG */ 3569 return; 3570 } 3571 /* 3572 * If we have found the just finished dependency, then free 3573 * it along with anything that follows it that is complete. 3574 */ 3575 for (; adp; adp = listadp) { 3576 listadp = TAILQ_NEXT(adp, ad_next); 3577 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3578 return; 3579 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 3580 } 3581 } 3582 3583 /* 3584 * Called from within softdep_disk_write_complete above. Note that 3585 * this routine is always called from interrupt level with further 3586 * splbio interrupts blocked. 3587 * 3588 * Parameters: 3589 * aip: the completed allocindir 3590 */ 3591 static void 3592 handle_allocindir_partdone(struct allocindir *aip) 3593 { 3594 struct indirdep *indirdep; 3595 3596 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 3597 return; 3598 if (aip->ai_buf != NULL) { 3599 lk.lkt_held = NOHOLDER; 3600 panic("handle_allocindir_partdone: dangling dependency"); 3601 } 3602 indirdep = aip->ai_indirdep; 3603 if (indirdep->ir_state & UNDONE) { 3604 LIST_REMOVE(aip, ai_next); 3605 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 3606 return; 3607 } 3608 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 3609 aip->ai_newblkno; 3610 LIST_REMOVE(aip, ai_next); 3611 if (aip->ai_freefrag != NULL) 3612 add_to_worklist(&aip->ai_freefrag->ff_list); 3613 WORKITEM_FREE(aip, D_ALLOCINDIR); 3614 } 3615 3616 /* 3617 * Called from within softdep_disk_write_complete above to restore 3618 * in-memory inode block contents to their most up-to-date state. Note 3619 * that this routine is always called from interrupt level with further 3620 * splbio interrupts blocked. 3621 * 3622 * Parameters: 3623 * bp: buffer containing the inode block 3624 */ 3625 static int 3626 handle_written_inodeblock(struct inodedep *inodedep, struct buf *bp) 3627 { 3628 struct worklist *wk, *filefree; 3629 struct allocdirect *adp, *nextadp; 3630 struct ufs1_dinode *dp; 3631 int hadchanges; 3632 3633 if ((inodedep->id_state & IOSTARTED) == 0) { 3634 lk.lkt_held = NOHOLDER; 3635 panic("handle_written_inodeblock: not started"); 3636 } 3637 inodedep->id_state &= ~IOSTARTED; 3638 dp = (struct ufs1_dinode *)bp->b_data + 3639 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 3640 /* 3641 * If we had to rollback the inode allocation because of 3642 * bitmaps being incomplete, then simply restore it. 3643 * Keep the block dirty so that it will not be reclaimed until 3644 * all associated dependencies have been cleared and the 3645 * corresponding updates written to disk. 3646 */ 3647 if (inodedep->id_savedino != NULL) { 3648 *dp = *inodedep->id_savedino; 3649 FREE(inodedep->id_savedino, M_INODEDEP); 3650 inodedep->id_savedino = NULL; 3651 if ((bp->b_flags & B_DELWRI) == 0) 3652 stat_inode_bitmap++; 3653 bdirty(bp); 3654 return (1); 3655 } 3656 inodedep->id_state |= COMPLETE; 3657 /* 3658 * Roll forward anything that had to be rolled back before 3659 * the inode could be updated. 3660 */ 3661 hadchanges = 0; 3662 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 3663 nextadp = TAILQ_NEXT(adp, ad_next); 3664 if (adp->ad_state & ATTACHED) { 3665 lk.lkt_held = NOHOLDER; 3666 panic("handle_written_inodeblock: new entry"); 3667 } 3668 if (adp->ad_lbn < NDADDR) { 3669 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) { 3670 lk.lkt_held = NOHOLDER; 3671 panic("%s: %s #%ld mismatch %d != %d", 3672 "handle_written_inodeblock", 3673 "direct pointer", adp->ad_lbn, 3674 dp->di_db[adp->ad_lbn], adp->ad_oldblkno); 3675 } 3676 dp->di_db[adp->ad_lbn] = adp->ad_newblkno; 3677 } else { 3678 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) { 3679 lk.lkt_held = NOHOLDER; 3680 panic("%s: %s #%ld allocated as %d", 3681 "handle_written_inodeblock", 3682 "indirect pointer", adp->ad_lbn - NDADDR, 3683 dp->di_ib[adp->ad_lbn - NDADDR]); 3684 } 3685 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno; 3686 } 3687 adp->ad_state &= ~UNDONE; 3688 adp->ad_state |= ATTACHED; 3689 hadchanges = 1; 3690 } 3691 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 3692 stat_direct_blk_ptrs++; 3693 /* 3694 * Reset the file size to its most up-to-date value. 3695 */ 3696 if (inodedep->id_savedsize == -1) { 3697 lk.lkt_held = NOHOLDER; 3698 panic("handle_written_inodeblock: bad size"); 3699 } 3700 if (dp->di_size != inodedep->id_savedsize) { 3701 dp->di_size = inodedep->id_savedsize; 3702 hadchanges = 1; 3703 } 3704 inodedep->id_savedsize = -1; 3705 /* 3706 * If there were any rollbacks in the inode block, then it must be 3707 * marked dirty so that its will eventually get written back in 3708 * its correct form. 3709 */ 3710 if (hadchanges) 3711 bdirty(bp); 3712 /* 3713 * Process any allocdirects that completed during the update. 3714 */ 3715 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 3716 handle_allocdirect_partdone(adp); 3717 /* 3718 * Process deallocations that were held pending until the 3719 * inode had been written to disk. Freeing of the inode 3720 * is delayed until after all blocks have been freed to 3721 * avoid creation of new <vfsid, inum, lbn> triples 3722 * before the old ones have been deleted. 3723 */ 3724 filefree = NULL; 3725 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 3726 WORKLIST_REMOVE(wk); 3727 switch (wk->wk_type) { 3728 3729 case D_FREEFILE: 3730 /* 3731 * We defer adding filefree to the worklist until 3732 * all other additions have been made to ensure 3733 * that it will be done after all the old blocks 3734 * have been freed. 3735 */ 3736 if (filefree != NULL) { 3737 lk.lkt_held = NOHOLDER; 3738 panic("handle_written_inodeblock: filefree"); 3739 } 3740 filefree = wk; 3741 continue; 3742 3743 case D_MKDIR: 3744 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 3745 continue; 3746 3747 case D_DIRADD: 3748 diradd_inode_written(WK_DIRADD(wk), inodedep); 3749 continue; 3750 3751 case D_FREEBLKS: 3752 wk->wk_state |= COMPLETE; 3753 if ((wk->wk_state & ALLCOMPLETE) != ALLCOMPLETE) 3754 continue; 3755 /* -- fall through -- */ 3756 case D_FREEFRAG: 3757 case D_DIRREM: 3758 add_to_worklist(wk); 3759 continue; 3760 3761 default: 3762 lk.lkt_held = NOHOLDER; 3763 panic("handle_written_inodeblock: Unknown type %s", 3764 TYPENAME(wk->wk_type)); 3765 /* NOTREACHED */ 3766 } 3767 } 3768 if (filefree != NULL) { 3769 if (free_inodedep(inodedep) == 0) { 3770 lk.lkt_held = NOHOLDER; 3771 panic("handle_written_inodeblock: live inodedep"); 3772 } 3773 add_to_worklist(filefree); 3774 return (0); 3775 } 3776 3777 /* 3778 * If no outstanding dependencies, free it. 3779 */ 3780 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0) 3781 return (0); 3782 return (hadchanges); 3783 } 3784 3785 /* 3786 * Process a diradd entry after its dependent inode has been written. 3787 * This routine must be called with splbio interrupts blocked. 3788 */ 3789 static void 3790 diradd_inode_written(struct diradd *dap, struct inodedep *inodedep) 3791 { 3792 struct pagedep *pagedep; 3793 3794 dap->da_state |= COMPLETE; 3795 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3796 if (dap->da_state & DIRCHG) 3797 pagedep = dap->da_previous->dm_pagedep; 3798 else 3799 pagedep = dap->da_pagedep; 3800 LIST_REMOVE(dap, da_pdlist); 3801 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3802 } 3803 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 3804 } 3805 3806 /* 3807 * Handle the completion of a mkdir dependency. 3808 */ 3809 static void 3810 handle_written_mkdir(struct mkdir *mkdir, int type) 3811 { 3812 struct diradd *dap; 3813 struct pagedep *pagedep; 3814 3815 if (mkdir->md_state != type) { 3816 lk.lkt_held = NOHOLDER; 3817 panic("handle_written_mkdir: bad type"); 3818 } 3819 dap = mkdir->md_diradd; 3820 dap->da_state &= ~type; 3821 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 3822 dap->da_state |= DEPCOMPLETE; 3823 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3824 if (dap->da_state & DIRCHG) 3825 pagedep = dap->da_previous->dm_pagedep; 3826 else 3827 pagedep = dap->da_pagedep; 3828 LIST_REMOVE(dap, da_pdlist); 3829 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3830 } 3831 LIST_REMOVE(mkdir, md_mkdirs); 3832 WORKITEM_FREE(mkdir, D_MKDIR); 3833 } 3834 3835 /* 3836 * Called from within softdep_disk_write_complete above. 3837 * A write operation was just completed. Removed inodes can 3838 * now be freed and associated block pointers may be committed. 3839 * Note that this routine is always called from interrupt level 3840 * with further splbio interrupts blocked. 3841 * 3842 * Parameters: 3843 * bp: buffer containing the written page 3844 */ 3845 static int 3846 handle_written_filepage(struct pagedep *pagedep, struct buf *bp) 3847 { 3848 struct dirrem *dirrem; 3849 struct diradd *dap, *nextdap; 3850 struct direct *ep; 3851 int i, chgs; 3852 3853 if ((pagedep->pd_state & IOSTARTED) == 0) { 3854 lk.lkt_held = NOHOLDER; 3855 panic("handle_written_filepage: not started"); 3856 } 3857 pagedep->pd_state &= ~IOSTARTED; 3858 /* 3859 * Process any directory removals that have been committed. 3860 */ 3861 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 3862 LIST_REMOVE(dirrem, dm_next); 3863 dirrem->dm_dirinum = pagedep->pd_ino; 3864 add_to_worklist(&dirrem->dm_list); 3865 } 3866 /* 3867 * Free any directory additions that have been committed. 3868 */ 3869 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 3870 free_diradd(dap); 3871 /* 3872 * Uncommitted directory entries must be restored. 3873 */ 3874 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 3875 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 3876 dap = nextdap) { 3877 nextdap = LIST_NEXT(dap, da_pdlist); 3878 if (dap->da_state & ATTACHED) { 3879 lk.lkt_held = NOHOLDER; 3880 panic("handle_written_filepage: attached"); 3881 } 3882 ep = (struct direct *) 3883 ((char *)bp->b_data + dap->da_offset); 3884 ep->d_ino = dap->da_newinum; 3885 dap->da_state &= ~UNDONE; 3886 dap->da_state |= ATTACHED; 3887 chgs = 1; 3888 /* 3889 * If the inode referenced by the directory has 3890 * been written out, then the dependency can be 3891 * moved to the pending list. 3892 */ 3893 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3894 LIST_REMOVE(dap, da_pdlist); 3895 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 3896 da_pdlist); 3897 } 3898 } 3899 } 3900 /* 3901 * If there were any rollbacks in the directory, then it must be 3902 * marked dirty so that its will eventually get written back in 3903 * its correct form. 3904 */ 3905 if (chgs) { 3906 if ((bp->b_flags & B_DELWRI) == 0) 3907 stat_dir_entry++; 3908 bdirty(bp); 3909 } 3910 /* 3911 * If no dependencies remain, the pagedep will be freed. 3912 * Otherwise it will remain to update the page before it 3913 * is written back to disk. 3914 */ 3915 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) { 3916 for (i = 0; i < DAHASHSZ; i++) 3917 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL) 3918 break; 3919 if (i == DAHASHSZ) { 3920 LIST_REMOVE(pagedep, pd_hash); 3921 WORKITEM_FREE(pagedep, D_PAGEDEP); 3922 return (0); 3923 } 3924 } 3925 return (1); 3926 } 3927 3928 /* 3929 * Writing back in-core inode structures. 3930 * 3931 * The filesystem only accesses an inode's contents when it occupies an 3932 * "in-core" inode structure. These "in-core" structures are separate from 3933 * the page frames used to cache inode blocks. Only the latter are 3934 * transferred to/from the disk. So, when the updated contents of the 3935 * "in-core" inode structure are copied to the corresponding in-memory inode 3936 * block, the dependencies are also transferred. The following procedure is 3937 * called when copying a dirty "in-core" inode to a cached inode block. 3938 */ 3939 3940 /* 3941 * Called when an inode is loaded from disk. If the effective link count 3942 * differed from the actual link count when it was last flushed, then we 3943 * need to ensure that the correct effective link count is put back. 3944 * 3945 * Parameters: 3946 * ip: the "in_core" copy of the inode 3947 */ 3948 void 3949 softdep_load_inodeblock(struct inode *ip) 3950 { 3951 struct inodedep *inodedep; 3952 3953 /* 3954 * Check for alternate nlink count. 3955 */ 3956 ip->i_effnlink = ip->i_nlink; 3957 ACQUIRE_LOCK(&lk); 3958 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3959 FREE_LOCK(&lk); 3960 return; 3961 } 3962 ip->i_effnlink -= inodedep->id_nlinkdelta; 3963 FREE_LOCK(&lk); 3964 } 3965 3966 /* 3967 * This routine is called just before the "in-core" inode 3968 * information is to be copied to the in-memory inode block. 3969 * Recall that an inode block contains several inodes. If 3970 * the force flag is set, then the dependencies will be 3971 * cleared so that the update can always be made. Note that 3972 * the buffer is locked when this routine is called, so we 3973 * will never be in the middle of writing the inode block 3974 * to disk. 3975 * 3976 * Parameters: 3977 * ip: the "in_core" copy of the inode 3978 * bp: the buffer containing the inode block 3979 * waitfor: nonzero => update must be allowed 3980 */ 3981 void 3982 softdep_update_inodeblock(struct inode *ip, struct buf *bp, 3983 int waitfor) 3984 { 3985 struct inodedep *inodedep; 3986 struct worklist *wk; 3987 int error, gotit; 3988 3989 /* 3990 * If the effective link count is not equal to the actual link 3991 * count, then we must track the difference in an inodedep while 3992 * the inode is (potentially) tossed out of the cache. Otherwise, 3993 * if there is no existing inodedep, then there are no dependencies 3994 * to track. 3995 */ 3996 ACQUIRE_LOCK(&lk); 3997 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3998 FREE_LOCK(&lk); 3999 if (ip->i_effnlink != ip->i_nlink) 4000 panic("softdep_update_inodeblock: bad link count"); 4001 return; 4002 } 4003 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) { 4004 FREE_LOCK(&lk); 4005 panic("softdep_update_inodeblock: bad delta"); 4006 } 4007 /* 4008 * Changes have been initiated. Anything depending on these 4009 * changes cannot occur until this inode has been written. 4010 */ 4011 inodedep->id_state &= ~COMPLETE; 4012 if ((inodedep->id_state & ONWORKLIST) == 0) 4013 WORKLIST_INSERT_BP(bp, &inodedep->id_list); 4014 /* 4015 * Any new dependencies associated with the incore inode must 4016 * now be moved to the list associated with the buffer holding 4017 * the in-memory copy of the inode. Once merged process any 4018 * allocdirects that are completed by the merger. 4019 */ 4020 merge_inode_lists(inodedep); 4021 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL) 4022 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); 4023 /* 4024 * Now that the inode has been pushed into the buffer, the 4025 * operations dependent on the inode being written to disk 4026 * can be moved to the id_bufwait so that they will be 4027 * processed when the buffer I/O completes. 4028 */ 4029 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 4030 WORKLIST_REMOVE(wk); 4031 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 4032 } 4033 /* 4034 * Newly allocated inodes cannot be written until the bitmap 4035 * that allocates them have been written (indicated by 4036 * DEPCOMPLETE being set in id_state). If we are doing a 4037 * forced sync (e.g., an fsync on a file), we force the bitmap 4038 * to be written so that the update can be done. 4039 */ 4040 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) { 4041 FREE_LOCK(&lk); 4042 return; 4043 } 4044 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 4045 FREE_LOCK(&lk); 4046 if (gotit && 4047 (error = bwrite(inodedep->id_buf)) != 0) 4048 softdep_error("softdep_update_inodeblock: bwrite", error); 4049 } 4050 4051 /* 4052 * Merge the new inode dependency list (id_newinoupdt) into the old 4053 * inode dependency list (id_inoupdt). This routine must be called 4054 * with splbio interrupts blocked. 4055 */ 4056 static void 4057 merge_inode_lists(struct inodedep *inodedep) 4058 { 4059 struct allocdirect *listadp, *newadp; 4060 4061 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 4062 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) { 4063 if (listadp->ad_lbn < newadp->ad_lbn) { 4064 listadp = TAILQ_NEXT(listadp, ad_next); 4065 continue; 4066 } 4067 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 4068 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 4069 if (listadp->ad_lbn == newadp->ad_lbn) { 4070 allocdirect_merge(&inodedep->id_inoupdt, newadp, 4071 listadp); 4072 listadp = newadp; 4073 } 4074 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 4075 } 4076 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) { 4077 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 4078 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next); 4079 } 4080 } 4081 4082 /* 4083 * If we are doing an fsync, then we must ensure that any directory 4084 * entries for the inode have been written after the inode gets to disk. 4085 * 4086 * Parameters: 4087 * vp: the "in_core" copy of the inode 4088 */ 4089 static int 4090 softdep_fsync(struct vnode *vp) 4091 { 4092 struct inodedep *inodedep; 4093 struct pagedep *pagedep; 4094 struct worklist *wk; 4095 struct diradd *dap; 4096 struct mount *mnt; 4097 struct vnode *pvp; 4098 struct inode *ip; 4099 struct buf *bp; 4100 struct fs *fs; 4101 int error, flushparent; 4102 ino_t parentino; 4103 ufs_lbn_t lbn; 4104 4105 /* 4106 * Move check from original kernel code, possibly not needed any 4107 * more with the per-mount bioops. 4108 */ 4109 if ((vp->v_mount->mnt_flag & MNT_SOFTDEP) == 0) 4110 return (0); 4111 4112 ip = VTOI(vp); 4113 fs = ip->i_fs; 4114 ACQUIRE_LOCK(&lk); 4115 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) { 4116 FREE_LOCK(&lk); 4117 return (0); 4118 } 4119 if (LIST_FIRST(&inodedep->id_inowait) != NULL || 4120 LIST_FIRST(&inodedep->id_bufwait) != NULL || 4121 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 4122 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) { 4123 FREE_LOCK(&lk); 4124 panic("softdep_fsync: pending ops"); 4125 } 4126 for (error = 0, flushparent = 0; ; ) { 4127 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 4128 break; 4129 if (wk->wk_type != D_DIRADD) { 4130 FREE_LOCK(&lk); 4131 panic("softdep_fsync: Unexpected type %s", 4132 TYPENAME(wk->wk_type)); 4133 } 4134 dap = WK_DIRADD(wk); 4135 /* 4136 * Flush our parent if this directory entry 4137 * has a MKDIR_PARENT dependency. 4138 */ 4139 if (dap->da_state & DIRCHG) 4140 pagedep = dap->da_previous->dm_pagedep; 4141 else 4142 pagedep = dap->da_pagedep; 4143 mnt = pagedep->pd_mnt; 4144 parentino = pagedep->pd_ino; 4145 lbn = pagedep->pd_lbn; 4146 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) { 4147 FREE_LOCK(&lk); 4148 panic("softdep_fsync: dirty"); 4149 } 4150 flushparent = dap->da_state & MKDIR_PARENT; 4151 /* 4152 * If we are being fsync'ed as part of vgone'ing this vnode, 4153 * then we will not be able to release and recover the 4154 * vnode below, so we just have to give up on writing its 4155 * directory entry out. It will eventually be written, just 4156 * not now, but then the user was not asking to have it 4157 * written, so we are not breaking any promises. 4158 */ 4159 if (vp->v_flag & VRECLAIMED) 4160 break; 4161 /* 4162 * We prevent deadlock by always fetching inodes from the 4163 * root, moving down the directory tree. Thus, when fetching 4164 * our parent directory, we must unlock ourselves before 4165 * requesting the lock on our parent. See the comment in 4166 * ufs_lookup for details on possible races. 4167 */ 4168 FREE_LOCK(&lk); 4169 vn_unlock(vp); 4170 error = VFS_VGET(mnt, parentino, &pvp); 4171 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4172 if (error != 0) 4173 return (error); 4174 if (flushparent) { 4175 if ((error = ffs_update(pvp, 1)) != 0) { 4176 vput(pvp); 4177 return (error); 4178 } 4179 } 4180 /* 4181 * Flush directory page containing the inode's name. 4182 */ 4183 error = bread(pvp, lblktodoff(fs, lbn), blksize(fs, VTOI(pvp), lbn), &bp); 4184 if (error == 0) 4185 error = bwrite(bp); 4186 vput(pvp); 4187 if (error != 0) 4188 return (error); 4189 ACQUIRE_LOCK(&lk); 4190 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) 4191 break; 4192 } 4193 FREE_LOCK(&lk); 4194 return (0); 4195 } 4196 4197 /* 4198 * Flush all the dirty bitmaps associated with the block device 4199 * before flushing the rest of the dirty blocks so as to reduce 4200 * the number of dependencies that will have to be rolled back. 4201 */ 4202 static int softdep_fsync_mountdev_bp(struct buf *bp, void *data); 4203 4204 void 4205 softdep_fsync_mountdev(struct vnode *vp) 4206 { 4207 if (!vn_isdisk(vp, NULL)) 4208 panic("softdep_fsync_mountdev: vnode not a disk"); 4209 ACQUIRE_LOCK(&lk); 4210 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 4211 softdep_fsync_mountdev_bp, vp); 4212 drain_output(vp, 1); 4213 FREE_LOCK(&lk); 4214 } 4215 4216 static int 4217 softdep_fsync_mountdev_bp(struct buf *bp, void *data) 4218 { 4219 struct worklist *wk; 4220 struct vnode *vp = data; 4221 4222 /* 4223 * If it is already scheduled, skip to the next buffer. 4224 */ 4225 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) 4226 return(0); 4227 if (bp->b_vp != vp || (bp->b_flags & B_DELWRI) == 0) { 4228 BUF_UNLOCK(bp); 4229 kprintf("softdep_fsync_mountdev_bp: warning, buffer %p ripped out from under vnode %p\n", bp, vp); 4230 return(0); 4231 } 4232 /* 4233 * We are only interested in bitmaps with outstanding 4234 * dependencies. 4235 */ 4236 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 4237 wk->wk_type != D_BMSAFEMAP) { 4238 BUF_UNLOCK(bp); 4239 return(0); 4240 } 4241 bremfree(bp); 4242 FREE_LOCK(&lk); 4243 (void) bawrite(bp); 4244 ACQUIRE_LOCK(&lk); 4245 return(0); 4246 } 4247 4248 /* 4249 * This routine is called when we are trying to synchronously flush a 4250 * file. This routine must eliminate any filesystem metadata dependencies 4251 * so that the syncing routine can succeed by pushing the dirty blocks 4252 * associated with the file. If any I/O errors occur, they are returned. 4253 */ 4254 struct softdep_sync_metadata_info { 4255 struct vnode *vp; 4256 int waitfor; 4257 }; 4258 4259 static int softdep_sync_metadata_bp(struct buf *bp, void *data); 4260 4261 int 4262 softdep_sync_metadata(struct vnode *vp, struct thread *td) 4263 { 4264 struct softdep_sync_metadata_info info; 4265 int error, waitfor; 4266 4267 /* 4268 * Check whether this vnode is involved in a filesystem 4269 * that is doing soft dependency processing. 4270 */ 4271 if (!vn_isdisk(vp, NULL)) { 4272 if (!DOINGSOFTDEP(vp)) 4273 return (0); 4274 } else 4275 if (vp->v_rdev->si_mountpoint == NULL || 4276 (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0) 4277 return (0); 4278 /* 4279 * Ensure that any direct block dependencies have been cleared. 4280 */ 4281 ACQUIRE_LOCK(&lk); 4282 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) { 4283 FREE_LOCK(&lk); 4284 return (error); 4285 } 4286 /* 4287 * For most files, the only metadata dependencies are the 4288 * cylinder group maps that allocate their inode or blocks. 4289 * The block allocation dependencies can be found by traversing 4290 * the dependency lists for any buffers that remain on their 4291 * dirty buffer list. The inode allocation dependency will 4292 * be resolved when the inode is updated with MNT_WAIT. 4293 * This work is done in two passes. The first pass grabs most 4294 * of the buffers and begins asynchronously writing them. The 4295 * only way to wait for these asynchronous writes is to sleep 4296 * on the filesystem vnode which may stay busy for a long time 4297 * if the filesystem is active. So, instead, we make a second 4298 * pass over the dependencies blocking on each write. In the 4299 * usual case we will be blocking against a write that we 4300 * initiated, so when it is done the dependency will have been 4301 * resolved. Thus the second pass is expected to end quickly. 4302 */ 4303 waitfor = MNT_NOWAIT; 4304 top: 4305 /* 4306 * We must wait for any I/O in progress to finish so that 4307 * all potential buffers on the dirty list will be visible. 4308 */ 4309 drain_output(vp, 1); 4310 info.vp = vp; 4311 info.waitfor = waitfor; 4312 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 4313 softdep_sync_metadata_bp, &info); 4314 if (error < 0) { 4315 FREE_LOCK(&lk); 4316 return(-error); /* error code */ 4317 } 4318 4319 /* 4320 * The brief unlock is to allow any pent up dependency 4321 * processing to be done. Then proceed with the second pass. 4322 */ 4323 if (waitfor == MNT_NOWAIT) { 4324 waitfor = MNT_WAIT; 4325 FREE_LOCK(&lk); 4326 ACQUIRE_LOCK(&lk); 4327 goto top; 4328 } 4329 4330 /* 4331 * If we have managed to get rid of all the dirty buffers, 4332 * then we are done. For certain directories and block 4333 * devices, we may need to do further work. 4334 * 4335 * We must wait for any I/O in progress to finish so that 4336 * all potential buffers on the dirty list will be visible. 4337 */ 4338 drain_output(vp, 1); 4339 if (RB_EMPTY(&vp->v_rbdirty_tree)) { 4340 FREE_LOCK(&lk); 4341 return (0); 4342 } 4343 4344 FREE_LOCK(&lk); 4345 /* 4346 * If we are trying to sync a block device, some of its buffers may 4347 * contain metadata that cannot be written until the contents of some 4348 * partially written files have been written to disk. The only easy 4349 * way to accomplish this is to sync the entire filesystem (luckily 4350 * this happens rarely). 4351 */ 4352 if (vn_isdisk(vp, NULL) && 4353 vp->v_rdev && 4354 vp->v_rdev->si_mountpoint && !vn_islocked(vp) && 4355 (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT)) != 0) 4356 return (error); 4357 return (0); 4358 } 4359 4360 static int 4361 softdep_sync_metadata_bp(struct buf *bp, void *data) 4362 { 4363 struct softdep_sync_metadata_info *info = data; 4364 struct pagedep *pagedep; 4365 struct allocdirect *adp; 4366 struct allocindir *aip; 4367 struct worklist *wk; 4368 struct buf *nbp; 4369 int error; 4370 int i; 4371 4372 if (getdirtybuf(&bp, MNT_WAIT) == 0) { 4373 kprintf("softdep_sync_metadata_bp(1): caught buf %p going away\n", bp); 4374 return (1); 4375 } 4376 if (bp->b_vp != info->vp || (bp->b_flags & B_DELWRI) == 0) { 4377 kprintf("softdep_sync_metadata_bp(2): caught buf %p going away vp %p\n", bp, info->vp); 4378 BUF_UNLOCK(bp); 4379 return(1); 4380 } 4381 4382 /* 4383 * As we hold the buffer locked, none of its dependencies 4384 * will disappear. 4385 */ 4386 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 4387 switch (wk->wk_type) { 4388 4389 case D_ALLOCDIRECT: 4390 adp = WK_ALLOCDIRECT(wk); 4391 if (adp->ad_state & DEPCOMPLETE) 4392 break; 4393 nbp = adp->ad_buf; 4394 if (getdirtybuf(&nbp, info->waitfor) == 0) 4395 break; 4396 FREE_LOCK(&lk); 4397 if (info->waitfor == MNT_NOWAIT) { 4398 bawrite(nbp); 4399 } else if ((error = bwrite(nbp)) != 0) { 4400 bawrite(bp); 4401 ACQUIRE_LOCK(&lk); 4402 return (-error); 4403 } 4404 ACQUIRE_LOCK(&lk); 4405 break; 4406 4407 case D_ALLOCINDIR: 4408 aip = WK_ALLOCINDIR(wk); 4409 if (aip->ai_state & DEPCOMPLETE) 4410 break; 4411 nbp = aip->ai_buf; 4412 if (getdirtybuf(&nbp, info->waitfor) == 0) 4413 break; 4414 FREE_LOCK(&lk); 4415 if (info->waitfor == MNT_NOWAIT) { 4416 bawrite(nbp); 4417 } else if ((error = bwrite(nbp)) != 0) { 4418 bawrite(bp); 4419 ACQUIRE_LOCK(&lk); 4420 return (-error); 4421 } 4422 ACQUIRE_LOCK(&lk); 4423 break; 4424 4425 case D_INDIRDEP: 4426 restart: 4427 4428 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) { 4429 if (aip->ai_state & DEPCOMPLETE) 4430 continue; 4431 nbp = aip->ai_buf; 4432 if (getdirtybuf(&nbp, MNT_WAIT) == 0) 4433 goto restart; 4434 FREE_LOCK(&lk); 4435 if ((error = bwrite(nbp)) != 0) { 4436 bawrite(bp); 4437 ACQUIRE_LOCK(&lk); 4438 return (-error); 4439 } 4440 ACQUIRE_LOCK(&lk); 4441 goto restart; 4442 } 4443 break; 4444 4445 case D_INODEDEP: 4446 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs, 4447 WK_INODEDEP(wk)->id_ino)) != 0) { 4448 FREE_LOCK(&lk); 4449 bawrite(bp); 4450 ACQUIRE_LOCK(&lk); 4451 return (-error); 4452 } 4453 break; 4454 4455 case D_PAGEDEP: 4456 /* 4457 * We are trying to sync a directory that may 4458 * have dependencies on both its own metadata 4459 * and/or dependencies on the inodes of any 4460 * recently allocated files. We walk its diradd 4461 * lists pushing out the associated inode. 4462 */ 4463 pagedep = WK_PAGEDEP(wk); 4464 for (i = 0; i < DAHASHSZ; i++) { 4465 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 4466 continue; 4467 if ((error = 4468 flush_pagedep_deps(info->vp, 4469 pagedep->pd_mnt, 4470 &pagedep->pd_diraddhd[i]))) { 4471 FREE_LOCK(&lk); 4472 bawrite(bp); 4473 ACQUIRE_LOCK(&lk); 4474 return (-error); 4475 } 4476 } 4477 break; 4478 4479 case D_MKDIR: 4480 /* 4481 * This case should never happen if the vnode has 4482 * been properly sync'ed. However, if this function 4483 * is used at a place where the vnode has not yet 4484 * been sync'ed, this dependency can show up. So, 4485 * rather than panic, just flush it. 4486 */ 4487 nbp = WK_MKDIR(wk)->md_buf; 4488 if (getdirtybuf(&nbp, info->waitfor) == 0) 4489 break; 4490 FREE_LOCK(&lk); 4491 if (info->waitfor == MNT_NOWAIT) { 4492 bawrite(nbp); 4493 } else if ((error = bwrite(nbp)) != 0) { 4494 bawrite(bp); 4495 ACQUIRE_LOCK(&lk); 4496 return (-error); 4497 } 4498 ACQUIRE_LOCK(&lk); 4499 break; 4500 4501 case D_BMSAFEMAP: 4502 /* 4503 * This case should never happen if the vnode has 4504 * been properly sync'ed. However, if this function 4505 * is used at a place where the vnode has not yet 4506 * been sync'ed, this dependency can show up. So, 4507 * rather than panic, just flush it. 4508 * 4509 * nbp can wind up == bp if a device node for the 4510 * same filesystem is being fsynced at the same time, 4511 * leading to a panic if we don't catch the case. 4512 */ 4513 nbp = WK_BMSAFEMAP(wk)->sm_buf; 4514 if (nbp == bp) 4515 break; 4516 if (getdirtybuf(&nbp, info->waitfor) == 0) 4517 break; 4518 FREE_LOCK(&lk); 4519 if (info->waitfor == MNT_NOWAIT) { 4520 bawrite(nbp); 4521 } else if ((error = bwrite(nbp)) != 0) { 4522 bawrite(bp); 4523 ACQUIRE_LOCK(&lk); 4524 return (-error); 4525 } 4526 ACQUIRE_LOCK(&lk); 4527 break; 4528 4529 default: 4530 FREE_LOCK(&lk); 4531 panic("softdep_sync_metadata: Unknown type %s", 4532 TYPENAME(wk->wk_type)); 4533 /* NOTREACHED */ 4534 } 4535 } 4536 FREE_LOCK(&lk); 4537 bawrite(bp); 4538 ACQUIRE_LOCK(&lk); 4539 return(0); 4540 } 4541 4542 /* 4543 * Flush the dependencies associated with an inodedep. 4544 * Called with splbio blocked. 4545 */ 4546 static int 4547 flush_inodedep_deps(struct fs *fs, ino_t ino) 4548 { 4549 struct inodedep *inodedep; 4550 struct allocdirect *adp; 4551 int error, waitfor; 4552 struct buf *bp; 4553 4554 /* 4555 * This work is done in two passes. The first pass grabs most 4556 * of the buffers and begins asynchronously writing them. The 4557 * only way to wait for these asynchronous writes is to sleep 4558 * on the filesystem vnode which may stay busy for a long time 4559 * if the filesystem is active. So, instead, we make a second 4560 * pass over the dependencies blocking on each write. In the 4561 * usual case we will be blocking against a write that we 4562 * initiated, so when it is done the dependency will have been 4563 * resolved. Thus the second pass is expected to end quickly. 4564 * We give a brief window at the top of the loop to allow 4565 * any pending I/O to complete. 4566 */ 4567 for (waitfor = MNT_NOWAIT; ; ) { 4568 FREE_LOCK(&lk); 4569 ACQUIRE_LOCK(&lk); 4570 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4571 return (0); 4572 TAILQ_FOREACH(adp, &inodedep->id_inoupdt, ad_next) { 4573 if (adp->ad_state & DEPCOMPLETE) 4574 continue; 4575 bp = adp->ad_buf; 4576 if (getdirtybuf(&bp, waitfor) == 0) { 4577 if (waitfor == MNT_NOWAIT) 4578 continue; 4579 break; 4580 } 4581 FREE_LOCK(&lk); 4582 if (waitfor == MNT_NOWAIT) { 4583 bawrite(bp); 4584 } else if ((error = bwrite(bp)) != 0) { 4585 ACQUIRE_LOCK(&lk); 4586 return (error); 4587 } 4588 ACQUIRE_LOCK(&lk); 4589 break; 4590 } 4591 if (adp != NULL) 4592 continue; 4593 TAILQ_FOREACH(adp, &inodedep->id_newinoupdt, ad_next) { 4594 if (adp->ad_state & DEPCOMPLETE) 4595 continue; 4596 bp = adp->ad_buf; 4597 if (getdirtybuf(&bp, waitfor) == 0) { 4598 if (waitfor == MNT_NOWAIT) 4599 continue; 4600 break; 4601 } 4602 FREE_LOCK(&lk); 4603 if (waitfor == MNT_NOWAIT) { 4604 bawrite(bp); 4605 } else if ((error = bwrite(bp)) != 0) { 4606 ACQUIRE_LOCK(&lk); 4607 return (error); 4608 } 4609 ACQUIRE_LOCK(&lk); 4610 break; 4611 } 4612 if (adp != NULL) 4613 continue; 4614 /* 4615 * If pass2, we are done, otherwise do pass 2. 4616 */ 4617 if (waitfor == MNT_WAIT) 4618 break; 4619 waitfor = MNT_WAIT; 4620 } 4621 /* 4622 * Try freeing inodedep in case all dependencies have been removed. 4623 */ 4624 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0) 4625 (void) free_inodedep(inodedep); 4626 return (0); 4627 } 4628 4629 /* 4630 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 4631 * Called with splbio blocked. 4632 */ 4633 static int 4634 flush_pagedep_deps(struct vnode *pvp, struct mount *mp, 4635 struct diraddhd *diraddhdp) 4636 { 4637 struct inodedep *inodedep; 4638 struct ufsmount *ump; 4639 struct diradd *dap; 4640 struct vnode *vp; 4641 int gotit, error = 0; 4642 struct buf *bp; 4643 ino_t inum; 4644 4645 ump = VFSTOUFS(mp); 4646 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 4647 /* 4648 * Flush ourselves if this directory entry 4649 * has a MKDIR_PARENT dependency. 4650 */ 4651 if (dap->da_state & MKDIR_PARENT) { 4652 FREE_LOCK(&lk); 4653 if ((error = ffs_update(pvp, 1)) != 0) 4654 break; 4655 ACQUIRE_LOCK(&lk); 4656 /* 4657 * If that cleared dependencies, go on to next. 4658 */ 4659 if (dap != LIST_FIRST(diraddhdp)) 4660 continue; 4661 if (dap->da_state & MKDIR_PARENT) { 4662 FREE_LOCK(&lk); 4663 panic("flush_pagedep_deps: MKDIR_PARENT"); 4664 } 4665 } 4666 /* 4667 * A newly allocated directory must have its "." and 4668 * ".." entries written out before its name can be 4669 * committed in its parent. We do not want or need 4670 * the full semantics of a synchronous VOP_FSYNC as 4671 * that may end up here again, once for each directory 4672 * level in the filesystem. Instead, we push the blocks 4673 * and wait for them to clear. We have to fsync twice 4674 * because the first call may choose to defer blocks 4675 * that still have dependencies, but deferral will 4676 * happen at most once. 4677 */ 4678 inum = dap->da_newinum; 4679 if (dap->da_state & MKDIR_BODY) { 4680 FREE_LOCK(&lk); 4681 if ((error = VFS_VGET(mp, inum, &vp)) != 0) 4682 break; 4683 if ((error=VOP_FSYNC(vp, MNT_NOWAIT)) || 4684 (error=VOP_FSYNC(vp, MNT_NOWAIT))) { 4685 vput(vp); 4686 break; 4687 } 4688 drain_output(vp, 0); 4689 vput(vp); 4690 ACQUIRE_LOCK(&lk); 4691 /* 4692 * If that cleared dependencies, go on to next. 4693 */ 4694 if (dap != LIST_FIRST(diraddhdp)) 4695 continue; 4696 if (dap->da_state & MKDIR_BODY) { 4697 FREE_LOCK(&lk); 4698 panic("flush_pagedep_deps: MKDIR_BODY"); 4699 } 4700 } 4701 /* 4702 * Flush the inode on which the directory entry depends. 4703 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 4704 * the only remaining dependency is that the updated inode 4705 * count must get pushed to disk. The inode has already 4706 * been pushed into its inode buffer (via VOP_UPDATE) at 4707 * the time of the reference count change. So we need only 4708 * locate that buffer, ensure that there will be no rollback 4709 * caused by a bitmap dependency, then write the inode buffer. 4710 */ 4711 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) { 4712 FREE_LOCK(&lk); 4713 panic("flush_pagedep_deps: lost inode"); 4714 } 4715 /* 4716 * If the inode still has bitmap dependencies, 4717 * push them to disk. 4718 */ 4719 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4720 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 4721 FREE_LOCK(&lk); 4722 if (gotit && (error = bwrite(inodedep->id_buf)) != 0) 4723 break; 4724 ACQUIRE_LOCK(&lk); 4725 if (dap != LIST_FIRST(diraddhdp)) 4726 continue; 4727 } 4728 /* 4729 * If the inode is still sitting in a buffer waiting 4730 * to be written, push it to disk. 4731 */ 4732 FREE_LOCK(&lk); 4733 if ((error = bread(ump->um_devvp, 4734 fsbtodoff(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), 4735 (int)ump->um_fs->fs_bsize, &bp)) != 0) 4736 break; 4737 if ((error = bwrite(bp)) != 0) 4738 break; 4739 ACQUIRE_LOCK(&lk); 4740 /* 4741 * If we have failed to get rid of all the dependencies 4742 * then something is seriously wrong. 4743 */ 4744 if (dap == LIST_FIRST(diraddhdp)) { 4745 FREE_LOCK(&lk); 4746 panic("flush_pagedep_deps: flush failed"); 4747 } 4748 } 4749 if (error) 4750 ACQUIRE_LOCK(&lk); 4751 return (error); 4752 } 4753 4754 /* 4755 * A large burst of file addition or deletion activity can drive the 4756 * memory load excessively high. First attempt to slow things down 4757 * using the techniques below. If that fails, this routine requests 4758 * the offending operations to fall back to running synchronously 4759 * until the memory load returns to a reasonable level. 4760 */ 4761 int 4762 softdep_slowdown(struct vnode *vp) 4763 { 4764 int max_softdeps_hard; 4765 4766 max_softdeps_hard = max_softdeps * 11 / 10; 4767 if (num_dirrem < max_softdeps_hard / 2 && 4768 num_inodedep < max_softdeps_hard) 4769 return (0); 4770 stat_sync_limit_hit += 1; 4771 return (1); 4772 } 4773 4774 /* 4775 * If memory utilization has gotten too high, deliberately slow things 4776 * down and speed up the I/O processing. 4777 */ 4778 static int 4779 request_cleanup(int resource, int islocked) 4780 { 4781 struct thread *td = curthread; /* XXX */ 4782 4783 /* 4784 * We never hold up the filesystem syncer process. 4785 */ 4786 if (td == filesys_syncer) 4787 return (0); 4788 /* 4789 * First check to see if the work list has gotten backlogged. 4790 * If it has, co-opt this process to help clean up two entries. 4791 * Because this process may hold inodes locked, we cannot 4792 * handle any remove requests that might block on a locked 4793 * inode as that could lead to deadlock. 4794 */ 4795 if (num_on_worklist > max_softdeps / 10) { 4796 if (islocked) 4797 FREE_LOCK(&lk); 4798 process_worklist_item(NULL, LK_NOWAIT); 4799 process_worklist_item(NULL, LK_NOWAIT); 4800 stat_worklist_push += 2; 4801 if (islocked) 4802 ACQUIRE_LOCK(&lk); 4803 return(1); 4804 } 4805 4806 /* 4807 * If we are resource constrained on inode dependencies, try 4808 * flushing some dirty inodes. Otherwise, we are constrained 4809 * by file deletions, so try accelerating flushes of directories 4810 * with removal dependencies. We would like to do the cleanup 4811 * here, but we probably hold an inode locked at this point and 4812 * that might deadlock against one that we try to clean. So, 4813 * the best that we can do is request the syncer daemon to do 4814 * the cleanup for us. 4815 */ 4816 switch (resource) { 4817 4818 case FLUSH_INODES: 4819 stat_ino_limit_push += 1; 4820 req_clear_inodedeps += 1; 4821 stat_countp = &stat_ino_limit_hit; 4822 break; 4823 4824 case FLUSH_REMOVE: 4825 stat_blk_limit_push += 1; 4826 req_clear_remove += 1; 4827 stat_countp = &stat_blk_limit_hit; 4828 break; 4829 4830 default: 4831 if (islocked) 4832 FREE_LOCK(&lk); 4833 panic("request_cleanup: unknown type"); 4834 } 4835 /* 4836 * Hopefully the syncer daemon will catch up and awaken us. 4837 * We wait at most tickdelay before proceeding in any case. 4838 */ 4839 if (islocked == 0) 4840 ACQUIRE_LOCK(&lk); 4841 proc_waiting += 1; 4842 if (!callout_active(&handle)) 4843 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2, 4844 pause_timer, NULL); 4845 interlocked_sleep(&lk, SLEEP, (caddr_t)&proc_waiting, 0, 4846 "softupdate", 0); 4847 proc_waiting -= 1; 4848 if (islocked == 0) 4849 FREE_LOCK(&lk); 4850 return (1); 4851 } 4852 4853 /* 4854 * Awaken processes pausing in request_cleanup and clear proc_waiting 4855 * to indicate that there is no longer a timer running. 4856 */ 4857 void 4858 pause_timer(void *arg) 4859 { 4860 *stat_countp += 1; 4861 wakeup_one(&proc_waiting); 4862 if (proc_waiting > 0) 4863 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2, 4864 pause_timer, NULL); 4865 else 4866 callout_deactivate(&handle); 4867 } 4868 4869 /* 4870 * Flush out a directory with at least one removal dependency in an effort to 4871 * reduce the number of dirrem, freefile, and freeblks dependency structures. 4872 */ 4873 static void 4874 clear_remove(struct thread *td) 4875 { 4876 struct pagedep_hashhead *pagedephd; 4877 struct pagedep *pagedep; 4878 static int next = 0; 4879 struct mount *mp; 4880 struct vnode *vp; 4881 int error, cnt; 4882 ino_t ino; 4883 4884 ACQUIRE_LOCK(&lk); 4885 for (cnt = 0; cnt < pagedep_hash; cnt++) { 4886 pagedephd = &pagedep_hashtbl[next++]; 4887 if (next >= pagedep_hash) 4888 next = 0; 4889 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 4890 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL) 4891 continue; 4892 mp = pagedep->pd_mnt; 4893 ino = pagedep->pd_ino; 4894 FREE_LOCK(&lk); 4895 if ((error = VFS_VGET(mp, ino, &vp)) != 0) { 4896 softdep_error("clear_remove: vget", error); 4897 return; 4898 } 4899 if ((error = VOP_FSYNC(vp, MNT_NOWAIT))) 4900 softdep_error("clear_remove: fsync", error); 4901 drain_output(vp, 0); 4902 vput(vp); 4903 return; 4904 } 4905 } 4906 FREE_LOCK(&lk); 4907 } 4908 4909 /* 4910 * Clear out a block of dirty inodes in an effort to reduce 4911 * the number of inodedep dependency structures. 4912 */ 4913 struct clear_inodedeps_info { 4914 struct fs *fs; 4915 struct mount *mp; 4916 }; 4917 4918 static int 4919 clear_inodedeps_mountlist_callback(struct mount *mp, void *data) 4920 { 4921 struct clear_inodedeps_info *info = data; 4922 4923 if ((mp->mnt_flag & MNT_SOFTDEP) && info->fs == VFSTOUFS(mp)->um_fs) { 4924 info->mp = mp; 4925 return(-1); 4926 } 4927 return(0); 4928 } 4929 4930 static void 4931 clear_inodedeps(struct thread *td) 4932 { 4933 struct clear_inodedeps_info info; 4934 struct inodedep_hashhead *inodedephd; 4935 struct inodedep *inodedep; 4936 static int next = 0; 4937 struct vnode *vp; 4938 struct fs *fs; 4939 int error, cnt; 4940 ino_t firstino, lastino, ino; 4941 4942 ACQUIRE_LOCK(&lk); 4943 /* 4944 * Pick a random inode dependency to be cleared. 4945 * We will then gather up all the inodes in its block 4946 * that have dependencies and flush them out. 4947 */ 4948 for (cnt = 0; cnt < inodedep_hash; cnt++) { 4949 inodedephd = &inodedep_hashtbl[next++]; 4950 if (next >= inodedep_hash) 4951 next = 0; 4952 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 4953 break; 4954 } 4955 if (inodedep == NULL) { 4956 FREE_LOCK(&lk); 4957 return; 4958 } 4959 /* 4960 * Ugly code to find mount point given pointer to superblock. 4961 */ 4962 fs = inodedep->id_fs; 4963 info.mp = NULL; 4964 info.fs = fs; 4965 mountlist_scan(clear_inodedeps_mountlist_callback, 4966 &info, MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 4967 /* 4968 * Find the last inode in the block with dependencies. 4969 */ 4970 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 4971 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 4972 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0) 4973 break; 4974 /* 4975 * Asynchronously push all but the last inode with dependencies. 4976 * Synchronously push the last inode with dependencies to ensure 4977 * that the inode block gets written to free up the inodedeps. 4978 */ 4979 for (ino = firstino; ino <= lastino; ino++) { 4980 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4981 continue; 4982 FREE_LOCK(&lk); 4983 if ((error = VFS_VGET(info.mp, ino, &vp)) != 0) { 4984 softdep_error("clear_inodedeps: vget", error); 4985 return; 4986 } 4987 if (ino == lastino) { 4988 if ((error = VOP_FSYNC(vp, MNT_WAIT))) 4989 softdep_error("clear_inodedeps: fsync1", error); 4990 } else { 4991 if ((error = VOP_FSYNC(vp, MNT_NOWAIT))) 4992 softdep_error("clear_inodedeps: fsync2", error); 4993 drain_output(vp, 0); 4994 } 4995 vput(vp); 4996 ACQUIRE_LOCK(&lk); 4997 } 4998 FREE_LOCK(&lk); 4999 } 5000 5001 /* 5002 * Function to determine if the buffer has outstanding dependencies 5003 * that will cause a roll-back if the buffer is written. If wantcount 5004 * is set, return number of dependencies, otherwise just yes or no. 5005 */ 5006 static int 5007 softdep_count_dependencies(struct buf *bp, int wantcount) 5008 { 5009 struct worklist *wk; 5010 struct inodedep *inodedep; 5011 struct indirdep *indirdep; 5012 struct allocindir *aip; 5013 struct pagedep *pagedep; 5014 struct diradd *dap; 5015 int i, retval; 5016 5017 retval = 0; 5018 ACQUIRE_LOCK(&lk); 5019 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5020 switch (wk->wk_type) { 5021 5022 case D_INODEDEP: 5023 inodedep = WK_INODEDEP(wk); 5024 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 5025 /* bitmap allocation dependency */ 5026 retval += 1; 5027 if (!wantcount) 5028 goto out; 5029 } 5030 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 5031 /* direct block pointer dependency */ 5032 retval += 1; 5033 if (!wantcount) 5034 goto out; 5035 } 5036 continue; 5037 5038 case D_INDIRDEP: 5039 indirdep = WK_INDIRDEP(wk); 5040 5041 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 5042 /* indirect block pointer dependency */ 5043 retval += 1; 5044 if (!wantcount) 5045 goto out; 5046 } 5047 continue; 5048 5049 case D_PAGEDEP: 5050 pagedep = WK_PAGEDEP(wk); 5051 for (i = 0; i < DAHASHSZ; i++) { 5052 5053 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 5054 /* directory entry dependency */ 5055 retval += 1; 5056 if (!wantcount) 5057 goto out; 5058 } 5059 } 5060 continue; 5061 5062 case D_BMSAFEMAP: 5063 case D_ALLOCDIRECT: 5064 case D_ALLOCINDIR: 5065 case D_MKDIR: 5066 /* never a dependency on these blocks */ 5067 continue; 5068 5069 default: 5070 FREE_LOCK(&lk); 5071 panic("softdep_check_for_rollback: Unexpected type %s", 5072 TYPENAME(wk->wk_type)); 5073 /* NOTREACHED */ 5074 } 5075 } 5076 out: 5077 FREE_LOCK(&lk); 5078 return retval; 5079 } 5080 5081 /* 5082 * Acquire exclusive access to a buffer. 5083 * Must be called with splbio blocked. 5084 * Return 1 if buffer was acquired. 5085 */ 5086 static int 5087 getdirtybuf(struct buf **bpp, int waitfor) 5088 { 5089 struct buf *bp; 5090 int error; 5091 5092 for (;;) { 5093 if ((bp = *bpp) == NULL) 5094 return (0); 5095 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) 5096 break; 5097 if (waitfor != MNT_WAIT) 5098 return (0); 5099 error = interlocked_sleep(&lk, LOCKBUF, bp, 5100 LK_EXCLUSIVE | LK_SLEEPFAIL, 0, 0); 5101 if (error != ENOLCK) { 5102 FREE_LOCK(&lk); 5103 panic("getdirtybuf: inconsistent lock"); 5104 } 5105 } 5106 if ((bp->b_flags & B_DELWRI) == 0) { 5107 BUF_UNLOCK(bp); 5108 return (0); 5109 } 5110 bremfree(bp); 5111 return (1); 5112 } 5113 5114 /* 5115 * Wait for pending output on a vnode to complete. 5116 * Must be called with vnode locked. 5117 */ 5118 static void 5119 drain_output(struct vnode *vp, int islocked) 5120 { 5121 5122 if (!islocked) 5123 ACQUIRE_LOCK(&lk); 5124 while (vp->v_track_write.bk_active) { 5125 vp->v_track_write.bk_waitflag = 1; 5126 interlocked_sleep(&lk, SLEEP, &vp->v_track_write, 5127 0, "drainvp", 0); 5128 } 5129 if (!islocked) 5130 FREE_LOCK(&lk); 5131 } 5132 5133 /* 5134 * Called whenever a buffer that is being invalidated or reallocated 5135 * contains dependencies. This should only happen if an I/O error has 5136 * occurred. The routine is called with the buffer locked. 5137 */ 5138 static void 5139 softdep_deallocate_dependencies(struct buf *bp) 5140 { 5141 if ((bp->b_flags & B_ERROR) == 0) 5142 panic("softdep_deallocate_dependencies: dangling deps"); 5143 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntfromname, bp->b_error); 5144 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 5145 } 5146 5147 /* 5148 * Function to handle asynchronous write errors in the filesystem. 5149 */ 5150 void 5151 softdep_error(char *func, int error) 5152 { 5153 5154 /* XXX should do something better! */ 5155 kprintf("%s: got error %d while accessing filesystem\n", func, error); 5156 } 5157