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, 0); 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 lwkt_gettoken(&vp->v_token); 1904 do { 1905 count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 1906 softdep_setup_freeblocks_bp, &info); 1907 } while (count != 0); 1908 lwkt_reltoken(&vp->v_token); 1909 1910 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0) 1911 (void)free_inodedep(inodedep); 1912 1913 if (delay) { 1914 freeblks->fb_state |= DEPCOMPLETE; 1915 /* 1916 * If the inode with zeroed block pointers is now on disk 1917 * we can start freeing blocks. Add freeblks to the worklist 1918 * instead of calling handle_workitem_freeblocks directly as 1919 * it is more likely that additional IO is needed to complete 1920 * the request here than in the !delay case. 1921 */ 1922 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 1923 add_to_worklist(&freeblks->fb_list); 1924 } 1925 1926 FREE_LOCK(&lk); 1927 /* 1928 * If the inode has never been written to disk (delay == 0), 1929 * then we can process the freeblks now that we have deleted 1930 * the dependencies. 1931 */ 1932 if (!delay) 1933 handle_workitem_freeblocks(freeblks); 1934 } 1935 1936 static int 1937 softdep_setup_freeblocks_bp(struct buf *bp, void *data) 1938 { 1939 struct softdep_setup_freeblocks_info *info = data; 1940 struct inodedep *inodedep; 1941 1942 if (getdirtybuf(&bp, MNT_WAIT) == 0) { 1943 kprintf("softdep_setup_freeblocks_bp(1): caught bp %p going away\n", bp); 1944 return(-1); 1945 } 1946 if (bp->b_vp != ITOV(info->ip) || (bp->b_flags & B_DELWRI) == 0) { 1947 kprintf("softdep_setup_freeblocks_bp(2): caught bp %p going away\n", bp); 1948 BUF_UNLOCK(bp); 1949 return(-1); 1950 } 1951 (void) inodedep_lookup(info->fs, info->ip->i_number, 0, &inodedep); 1952 deallocate_dependencies(bp, inodedep); 1953 bp->b_flags |= B_INVAL | B_NOCACHE; 1954 FREE_LOCK(&lk); 1955 brelse(bp); 1956 ACQUIRE_LOCK(&lk); 1957 return(1); 1958 } 1959 1960 /* 1961 * Reclaim any dependency structures from a buffer that is about to 1962 * be reallocated to a new vnode. The buffer must be locked, thus, 1963 * no I/O completion operations can occur while we are manipulating 1964 * its associated dependencies. The mutex is held so that other I/O's 1965 * associated with related dependencies do not occur. 1966 */ 1967 static void 1968 deallocate_dependencies(struct buf *bp, struct inodedep *inodedep) 1969 { 1970 struct worklist *wk; 1971 struct indirdep *indirdep; 1972 struct allocindir *aip; 1973 struct pagedep *pagedep; 1974 struct dirrem *dirrem; 1975 struct diradd *dap; 1976 int i; 1977 1978 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 1979 switch (wk->wk_type) { 1980 1981 case D_INDIRDEP: 1982 indirdep = WK_INDIRDEP(wk); 1983 /* 1984 * None of the indirect pointers will ever be visible, 1985 * so they can simply be tossed. GOINGAWAY ensures 1986 * that allocated pointers will be saved in the buffer 1987 * cache until they are freed. Note that they will 1988 * only be able to be found by their physical address 1989 * since the inode mapping the logical address will 1990 * be gone. The save buffer used for the safe copy 1991 * was allocated in setup_allocindir_phase2 using 1992 * the physical address so it could be used for this 1993 * purpose. Hence we swap the safe copy with the real 1994 * copy, allowing the safe copy to be freed and holding 1995 * on to the real copy for later use in indir_trunc. 1996 * 1997 * NOTE: ir_savebp is relative to the block device 1998 * so b_bio1 contains the device block number. 1999 */ 2000 if (indirdep->ir_state & GOINGAWAY) { 2001 FREE_LOCK(&lk); 2002 panic("deallocate_dependencies: already gone"); 2003 } 2004 indirdep->ir_state |= GOINGAWAY; 2005 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 2006 free_allocindir(aip, inodedep); 2007 if (bp->b_bio1.bio_offset >= 0 || 2008 bp->b_bio2.bio_offset != indirdep->ir_savebp->b_bio1.bio_offset) { 2009 FREE_LOCK(&lk); 2010 panic("deallocate_dependencies: not indir"); 2011 } 2012 bcopy(bp->b_data, indirdep->ir_savebp->b_data, 2013 bp->b_bcount); 2014 WORKLIST_REMOVE(wk); 2015 WORKLIST_INSERT_BP(indirdep->ir_savebp, wk); 2016 continue; 2017 2018 case D_PAGEDEP: 2019 pagedep = WK_PAGEDEP(wk); 2020 /* 2021 * None of the directory additions will ever be 2022 * visible, so they can simply be tossed. 2023 */ 2024 for (i = 0; i < DAHASHSZ; i++) 2025 while ((dap = 2026 LIST_FIRST(&pagedep->pd_diraddhd[i]))) 2027 free_diradd(dap); 2028 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0) 2029 free_diradd(dap); 2030 /* 2031 * Copy any directory remove dependencies to the list 2032 * to be processed after the zero'ed inode is written. 2033 * If the inode has already been written, then they 2034 * can be dumped directly onto the work list. 2035 */ 2036 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 2037 LIST_REMOVE(dirrem, dm_next); 2038 dirrem->dm_dirinum = pagedep->pd_ino; 2039 if (inodedep == NULL || 2040 (inodedep->id_state & ALLCOMPLETE) == 2041 ALLCOMPLETE) 2042 add_to_worklist(&dirrem->dm_list); 2043 else 2044 WORKLIST_INSERT(&inodedep->id_bufwait, 2045 &dirrem->dm_list); 2046 } 2047 WORKLIST_REMOVE(&pagedep->pd_list); 2048 LIST_REMOVE(pagedep, pd_hash); 2049 WORKITEM_FREE(pagedep, D_PAGEDEP); 2050 continue; 2051 2052 case D_ALLOCINDIR: 2053 free_allocindir(WK_ALLOCINDIR(wk), inodedep); 2054 continue; 2055 2056 case D_ALLOCDIRECT: 2057 case D_INODEDEP: 2058 FREE_LOCK(&lk); 2059 panic("deallocate_dependencies: Unexpected type %s", 2060 TYPENAME(wk->wk_type)); 2061 /* NOTREACHED */ 2062 2063 default: 2064 FREE_LOCK(&lk); 2065 panic("deallocate_dependencies: Unknown type %s", 2066 TYPENAME(wk->wk_type)); 2067 /* NOTREACHED */ 2068 } 2069 } 2070 } 2071 2072 /* 2073 * Free an allocdirect. Generate a new freefrag work request if appropriate. 2074 * This routine must be called with splbio interrupts blocked. 2075 */ 2076 static void 2077 free_allocdirect(struct allocdirectlst *adphead, 2078 struct allocdirect *adp, int delay) 2079 { 2080 2081 #ifdef DEBUG 2082 if (lk.lkt_held == NOHOLDER) 2083 panic("free_allocdirect: lock not held"); 2084 #endif 2085 if ((adp->ad_state & DEPCOMPLETE) == 0) 2086 LIST_REMOVE(adp, ad_deps); 2087 TAILQ_REMOVE(adphead, adp, ad_next); 2088 if ((adp->ad_state & COMPLETE) == 0) 2089 WORKLIST_REMOVE(&adp->ad_list); 2090 if (adp->ad_freefrag != NULL) { 2091 if (delay) 2092 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 2093 &adp->ad_freefrag->ff_list); 2094 else 2095 add_to_worklist(&adp->ad_freefrag->ff_list); 2096 } 2097 WORKITEM_FREE(adp, D_ALLOCDIRECT); 2098 } 2099 2100 /* 2101 * Prepare an inode to be freed. The actual free operation is not 2102 * done until the zero'ed inode has been written to disk. 2103 */ 2104 void 2105 softdep_freefile(struct vnode *pvp, ino_t ino, int mode) 2106 { 2107 struct inode *ip = VTOI(pvp); 2108 struct inodedep *inodedep; 2109 struct freefile *freefile; 2110 2111 /* 2112 * This sets up the inode de-allocation dependency. 2113 */ 2114 MALLOC(freefile, struct freefile *, sizeof(struct freefile), 2115 M_FREEFILE, M_SOFTDEP_FLAGS); 2116 freefile->fx_list.wk_type = D_FREEFILE; 2117 freefile->fx_list.wk_state = 0; 2118 freefile->fx_mode = mode; 2119 freefile->fx_oldinum = ino; 2120 freefile->fx_devvp = ip->i_devvp; 2121 freefile->fx_fs = ip->i_fs; 2122 2123 /* 2124 * If the inodedep does not exist, then the zero'ed inode has 2125 * been written to disk. If the allocated inode has never been 2126 * written to disk, then the on-disk inode is zero'ed. In either 2127 * case we can free the file immediately. 2128 */ 2129 ACQUIRE_LOCK(&lk); 2130 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 || 2131 check_inode_unwritten(inodedep)) { 2132 FREE_LOCK(&lk); 2133 handle_workitem_freefile(freefile); 2134 return; 2135 } 2136 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 2137 FREE_LOCK(&lk); 2138 } 2139 2140 /* 2141 * Check to see if an inode has never been written to disk. If 2142 * so free the inodedep and return success, otherwise return failure. 2143 * This routine must be called with splbio interrupts blocked. 2144 * 2145 * If we still have a bitmap dependency, then the inode has never 2146 * been written to disk. Drop the dependency as it is no longer 2147 * necessary since the inode is being deallocated. We set the 2148 * ALLCOMPLETE flags since the bitmap now properly shows that the 2149 * inode is not allocated. Even if the inode is actively being 2150 * written, it has been rolled back to its zero'ed state, so we 2151 * are ensured that a zero inode is what is on the disk. For short 2152 * lived files, this change will usually result in removing all the 2153 * dependencies from the inode so that it can be freed immediately. 2154 */ 2155 static int 2156 check_inode_unwritten(struct inodedep *inodedep) 2157 { 2158 2159 if ((inodedep->id_state & DEPCOMPLETE) != 0 || 2160 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2161 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2162 LIST_FIRST(&inodedep->id_inowait) != NULL || 2163 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2164 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2165 inodedep->id_nlinkdelta != 0) 2166 return (0); 2167 2168 /* 2169 * Another process might be in initiate_write_inodeblock 2170 * trying to allocate memory without holding "Softdep Lock". 2171 */ 2172 if ((inodedep->id_state & IOSTARTED) != 0 && 2173 inodedep->id_savedino == NULL) 2174 return(0); 2175 2176 inodedep->id_state |= ALLCOMPLETE; 2177 LIST_REMOVE(inodedep, id_deps); 2178 inodedep->id_buf = NULL; 2179 if (inodedep->id_state & ONWORKLIST) 2180 WORKLIST_REMOVE(&inodedep->id_list); 2181 if (inodedep->id_savedino != NULL) { 2182 FREE(inodedep->id_savedino, M_INODEDEP); 2183 inodedep->id_savedino = NULL; 2184 } 2185 if (free_inodedep(inodedep) == 0) { 2186 FREE_LOCK(&lk); 2187 panic("check_inode_unwritten: busy inode"); 2188 } 2189 return (1); 2190 } 2191 2192 /* 2193 * Try to free an inodedep structure. Return 1 if it could be freed. 2194 */ 2195 static int 2196 free_inodedep(struct inodedep *inodedep) 2197 { 2198 2199 if ((inodedep->id_state & ONWORKLIST) != 0 || 2200 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 2201 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2202 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2203 LIST_FIRST(&inodedep->id_inowait) != NULL || 2204 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2205 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2206 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL) 2207 return (0); 2208 LIST_REMOVE(inodedep, id_hash); 2209 WORKITEM_FREE(inodedep, D_INODEDEP); 2210 num_inodedep -= 1; 2211 return (1); 2212 } 2213 2214 /* 2215 * This workitem routine performs the block de-allocation. 2216 * The workitem is added to the pending list after the updated 2217 * inode block has been written to disk. As mentioned above, 2218 * checks regarding the number of blocks de-allocated (compared 2219 * to the number of blocks allocated for the file) are also 2220 * performed in this function. 2221 */ 2222 static void 2223 handle_workitem_freeblocks(struct freeblks *freeblks) 2224 { 2225 struct inode tip; 2226 ufs_daddr_t bn; 2227 struct fs *fs; 2228 int i, level, bsize; 2229 long nblocks, blocksreleased = 0; 2230 int error, allerror = 0; 2231 ufs_lbn_t baselbns[NIADDR], tmpval; 2232 2233 tip.i_number = freeblks->fb_previousinum; 2234 tip.i_devvp = freeblks->fb_devvp; 2235 tip.i_dev = freeblks->fb_devvp->v_rdev; 2236 tip.i_fs = freeblks->fb_fs; 2237 tip.i_size = freeblks->fb_oldsize; 2238 tip.i_uid = freeblks->fb_uid; 2239 fs = freeblks->fb_fs; 2240 tmpval = 1; 2241 baselbns[0] = NDADDR; 2242 for (i = 1; i < NIADDR; i++) { 2243 tmpval *= NINDIR(fs); 2244 baselbns[i] = baselbns[i - 1] + tmpval; 2245 } 2246 nblocks = btodb(fs->fs_bsize); 2247 blocksreleased = 0; 2248 /* 2249 * Indirect blocks first. 2250 */ 2251 for (level = (NIADDR - 1); level >= 0; level--) { 2252 if ((bn = freeblks->fb_iblks[level]) == 0) 2253 continue; 2254 if ((error = indir_trunc(&tip, fsbtodoff(fs, bn), level, 2255 baselbns[level], &blocksreleased)) == 0) 2256 allerror = error; 2257 ffs_blkfree(&tip, bn, fs->fs_bsize); 2258 blocksreleased += nblocks; 2259 } 2260 /* 2261 * All direct blocks or frags. 2262 */ 2263 for (i = (NDADDR - 1); i >= 0; i--) { 2264 if ((bn = freeblks->fb_dblks[i]) == 0) 2265 continue; 2266 bsize = blksize(fs, &tip, i); 2267 ffs_blkfree(&tip, bn, bsize); 2268 blocksreleased += btodb(bsize); 2269 } 2270 2271 #ifdef DIAGNOSTIC 2272 if (freeblks->fb_chkcnt != blocksreleased) 2273 kprintf("handle_workitem_freeblocks: block count\n"); 2274 if (allerror) 2275 softdep_error("handle_workitem_freeblks", allerror); 2276 #endif /* DIAGNOSTIC */ 2277 WORKITEM_FREE(freeblks, D_FREEBLKS); 2278 } 2279 2280 /* 2281 * Release blocks associated with the inode ip and stored in the indirect 2282 * block at doffset. If level is greater than SINGLE, the block is an 2283 * indirect block and recursive calls to indirtrunc must be used to 2284 * cleanse other indirect blocks. 2285 */ 2286 static int 2287 indir_trunc(struct inode *ip, off_t doffset, int level, ufs_lbn_t lbn, 2288 long *countp) 2289 { 2290 struct buf *bp; 2291 ufs_daddr_t *bap; 2292 ufs_daddr_t nb; 2293 struct fs *fs; 2294 struct worklist *wk; 2295 struct indirdep *indirdep; 2296 int i, lbnadd, nblocks; 2297 int error, allerror = 0; 2298 2299 fs = ip->i_fs; 2300 lbnadd = 1; 2301 for (i = level; i > 0; i--) 2302 lbnadd *= NINDIR(fs); 2303 /* 2304 * Get buffer of block pointers to be freed. This routine is not 2305 * called until the zero'ed inode has been written, so it is safe 2306 * to free blocks as they are encountered. Because the inode has 2307 * been zero'ed, calls to bmap on these blocks will fail. So, we 2308 * have to use the on-disk address and the block device for the 2309 * filesystem to look them up. If the file was deleted before its 2310 * indirect blocks were all written to disk, the routine that set 2311 * us up (deallocate_dependencies) will have arranged to leave 2312 * a complete copy of the indirect block in memory for our use. 2313 * Otherwise we have to read the blocks in from the disk. 2314 */ 2315 ACQUIRE_LOCK(&lk); 2316 if ((bp = findblk(ip->i_devvp, doffset, FINDBLK_TEST)) != NULL && 2317 (wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2318 /* 2319 * bp must be ir_savebp, which is held locked for our use. 2320 */ 2321 if (wk->wk_type != D_INDIRDEP || 2322 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp || 2323 (indirdep->ir_state & GOINGAWAY) == 0) { 2324 FREE_LOCK(&lk); 2325 panic("indir_trunc: lost indirdep"); 2326 } 2327 WORKLIST_REMOVE(wk); 2328 WORKITEM_FREE(indirdep, D_INDIRDEP); 2329 if (LIST_FIRST(&bp->b_dep) != NULL) { 2330 FREE_LOCK(&lk); 2331 panic("indir_trunc: dangling dep"); 2332 } 2333 FREE_LOCK(&lk); 2334 } else { 2335 FREE_LOCK(&lk); 2336 error = bread(ip->i_devvp, doffset, (int)fs->fs_bsize, &bp); 2337 if (error) 2338 return (error); 2339 } 2340 /* 2341 * Recursively free indirect blocks. 2342 */ 2343 bap = (ufs_daddr_t *)bp->b_data; 2344 nblocks = btodb(fs->fs_bsize); 2345 for (i = NINDIR(fs) - 1; i >= 0; i--) { 2346 if ((nb = bap[i]) == 0) 2347 continue; 2348 if (level != 0) { 2349 if ((error = indir_trunc(ip, fsbtodoff(fs, nb), 2350 level - 1, lbn + (i * lbnadd), countp)) != 0) 2351 allerror = error; 2352 } 2353 ffs_blkfree(ip, nb, fs->fs_bsize); 2354 *countp += nblocks; 2355 } 2356 bp->b_flags |= B_INVAL | B_NOCACHE; 2357 brelse(bp); 2358 return (allerror); 2359 } 2360 2361 /* 2362 * Free an allocindir. 2363 * This routine must be called with splbio interrupts blocked. 2364 */ 2365 static void 2366 free_allocindir(struct allocindir *aip, struct inodedep *inodedep) 2367 { 2368 struct freefrag *freefrag; 2369 2370 #ifdef DEBUG 2371 if (lk.lkt_held == NOHOLDER) 2372 panic("free_allocindir: lock not held"); 2373 #endif 2374 if ((aip->ai_state & DEPCOMPLETE) == 0) 2375 LIST_REMOVE(aip, ai_deps); 2376 if (aip->ai_state & ONWORKLIST) 2377 WORKLIST_REMOVE(&aip->ai_list); 2378 LIST_REMOVE(aip, ai_next); 2379 if ((freefrag = aip->ai_freefrag) != NULL) { 2380 if (inodedep == NULL) 2381 add_to_worklist(&freefrag->ff_list); 2382 else 2383 WORKLIST_INSERT(&inodedep->id_bufwait, 2384 &freefrag->ff_list); 2385 } 2386 WORKITEM_FREE(aip, D_ALLOCINDIR); 2387 } 2388 2389 /* 2390 * Directory entry addition dependencies. 2391 * 2392 * When adding a new directory entry, the inode (with its incremented link 2393 * count) must be written to disk before the directory entry's pointer to it. 2394 * Also, if the inode is newly allocated, the corresponding freemap must be 2395 * updated (on disk) before the directory entry's pointer. These requirements 2396 * are met via undo/redo on the directory entry's pointer, which consists 2397 * simply of the inode number. 2398 * 2399 * As directory entries are added and deleted, the free space within a 2400 * directory block can become fragmented. The ufs filesystem will compact 2401 * a fragmented directory block to make space for a new entry. When this 2402 * occurs, the offsets of previously added entries change. Any "diradd" 2403 * dependency structures corresponding to these entries must be updated with 2404 * the new offsets. 2405 */ 2406 2407 /* 2408 * This routine is called after the in-memory inode's link 2409 * count has been incremented, but before the directory entry's 2410 * pointer to the inode has been set. 2411 * 2412 * Parameters: 2413 * bp: buffer containing directory block 2414 * dp: inode for directory 2415 * diroffset: offset of new entry in directory 2416 * newinum: inode referenced by new directory entry 2417 * newdirbp: non-NULL => contents of new mkdir 2418 */ 2419 void 2420 softdep_setup_directory_add(struct buf *bp, struct inode *dp, off_t diroffset, 2421 ino_t newinum, struct buf *newdirbp) 2422 { 2423 int offset; /* offset of new entry within directory block */ 2424 ufs_lbn_t lbn; /* block in directory containing new entry */ 2425 struct fs *fs; 2426 struct diradd *dap; 2427 struct pagedep *pagedep; 2428 struct inodedep *inodedep; 2429 struct mkdir *mkdir1, *mkdir2; 2430 2431 /* 2432 * Whiteouts have no dependencies. 2433 */ 2434 if (newinum == WINO) { 2435 if (newdirbp != NULL) 2436 bdwrite(newdirbp); 2437 return; 2438 } 2439 2440 fs = dp->i_fs; 2441 lbn = lblkno(fs, diroffset); 2442 offset = blkoff(fs, diroffset); 2443 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, 2444 M_SOFTDEP_FLAGS | M_ZERO); 2445 dap->da_list.wk_type = D_DIRADD; 2446 dap->da_offset = offset; 2447 dap->da_newinum = newinum; 2448 dap->da_state = ATTACHED; 2449 if (newdirbp == NULL) { 2450 dap->da_state |= DEPCOMPLETE; 2451 ACQUIRE_LOCK(&lk); 2452 } else { 2453 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 2454 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2455 M_SOFTDEP_FLAGS); 2456 mkdir1->md_list.wk_type = D_MKDIR; 2457 mkdir1->md_state = MKDIR_BODY; 2458 mkdir1->md_diradd = dap; 2459 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2460 M_SOFTDEP_FLAGS); 2461 mkdir2->md_list.wk_type = D_MKDIR; 2462 mkdir2->md_state = MKDIR_PARENT; 2463 mkdir2->md_diradd = dap; 2464 /* 2465 * Dependency on "." and ".." being written to disk. 2466 */ 2467 mkdir1->md_buf = newdirbp; 2468 ACQUIRE_LOCK(&lk); 2469 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 2470 WORKLIST_INSERT_BP(newdirbp, &mkdir1->md_list); 2471 FREE_LOCK(&lk); 2472 bdwrite(newdirbp); 2473 /* 2474 * Dependency on link count increase for parent directory 2475 */ 2476 ACQUIRE_LOCK(&lk); 2477 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0 2478 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2479 dap->da_state &= ~MKDIR_PARENT; 2480 WORKITEM_FREE(mkdir2, D_MKDIR); 2481 } else { 2482 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 2483 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 2484 } 2485 } 2486 /* 2487 * Link into parent directory pagedep to await its being written. 2488 */ 2489 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2490 WORKLIST_INSERT_BP(bp, &pagedep->pd_list); 2491 dap->da_pagedep = pagedep; 2492 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 2493 da_pdlist); 2494 /* 2495 * Link into its inodedep. Put it on the id_bufwait list if the inode 2496 * is not yet written. If it is written, do the post-inode write 2497 * processing to put it on the id_pendinghd list. 2498 */ 2499 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep); 2500 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 2501 diradd_inode_written(dap, inodedep); 2502 else 2503 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2504 FREE_LOCK(&lk); 2505 } 2506 2507 /* 2508 * This procedure is called to change the offset of a directory 2509 * entry when compacting a directory block which must be owned 2510 * exclusively by the caller. Note that the actual entry movement 2511 * must be done in this procedure to ensure that no I/O completions 2512 * occur while the move is in progress. 2513 * 2514 * Parameters: 2515 * dp: inode for directory 2516 * base: address of dp->i_offset 2517 * oldloc: address of old directory location 2518 * newloc: address of new directory location 2519 * entrysize: size of directory entry 2520 */ 2521 void 2522 softdep_change_directoryentry_offset(struct inode *dp, caddr_t base, 2523 caddr_t oldloc, caddr_t newloc, 2524 int entrysize) 2525 { 2526 int offset, oldoffset, newoffset; 2527 struct pagedep *pagedep; 2528 struct diradd *dap; 2529 ufs_lbn_t lbn; 2530 2531 ACQUIRE_LOCK(&lk); 2532 lbn = lblkno(dp->i_fs, dp->i_offset); 2533 offset = blkoff(dp->i_fs, dp->i_offset); 2534 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) 2535 goto done; 2536 oldoffset = offset + (oldloc - base); 2537 newoffset = offset + (newloc - base); 2538 2539 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) { 2540 if (dap->da_offset != oldoffset) 2541 continue; 2542 dap->da_offset = newoffset; 2543 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) 2544 break; 2545 LIST_REMOVE(dap, da_pdlist); 2546 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], 2547 dap, da_pdlist); 2548 break; 2549 } 2550 if (dap == NULL) { 2551 2552 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) { 2553 if (dap->da_offset == oldoffset) { 2554 dap->da_offset = newoffset; 2555 break; 2556 } 2557 } 2558 } 2559 done: 2560 bcopy(oldloc, newloc, entrysize); 2561 FREE_LOCK(&lk); 2562 } 2563 2564 /* 2565 * Free a diradd dependency structure. This routine must be called 2566 * with splbio interrupts blocked. 2567 */ 2568 static void 2569 free_diradd(struct diradd *dap) 2570 { 2571 struct dirrem *dirrem; 2572 struct pagedep *pagedep; 2573 struct inodedep *inodedep; 2574 struct mkdir *mkdir, *nextmd; 2575 2576 #ifdef DEBUG 2577 if (lk.lkt_held == NOHOLDER) 2578 panic("free_diradd: lock not held"); 2579 #endif 2580 WORKLIST_REMOVE(&dap->da_list); 2581 LIST_REMOVE(dap, da_pdlist); 2582 if ((dap->da_state & DIRCHG) == 0) { 2583 pagedep = dap->da_pagedep; 2584 } else { 2585 dirrem = dap->da_previous; 2586 pagedep = dirrem->dm_pagedep; 2587 dirrem->dm_dirinum = pagedep->pd_ino; 2588 add_to_worklist(&dirrem->dm_list); 2589 } 2590 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum, 2591 0, &inodedep) != 0) 2592 (void) free_inodedep(inodedep); 2593 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2594 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 2595 nextmd = LIST_NEXT(mkdir, md_mkdirs); 2596 if (mkdir->md_diradd != dap) 2597 continue; 2598 dap->da_state &= ~mkdir->md_state; 2599 WORKLIST_REMOVE(&mkdir->md_list); 2600 LIST_REMOVE(mkdir, md_mkdirs); 2601 WORKITEM_FREE(mkdir, D_MKDIR); 2602 } 2603 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2604 FREE_LOCK(&lk); 2605 panic("free_diradd: unfound ref"); 2606 } 2607 } 2608 WORKITEM_FREE(dap, D_DIRADD); 2609 } 2610 2611 /* 2612 * Directory entry removal dependencies. 2613 * 2614 * When removing a directory entry, the entry's inode pointer must be 2615 * zero'ed on disk before the corresponding inode's link count is decremented 2616 * (possibly freeing the inode for re-use). This dependency is handled by 2617 * updating the directory entry but delaying the inode count reduction until 2618 * after the directory block has been written to disk. After this point, the 2619 * inode count can be decremented whenever it is convenient. 2620 */ 2621 2622 /* 2623 * This routine should be called immediately after removing 2624 * a directory entry. The inode's link count should not be 2625 * decremented by the calling procedure -- the soft updates 2626 * code will do this task when it is safe. 2627 * 2628 * Parameters: 2629 * bp: buffer containing directory block 2630 * dp: inode for the directory being modified 2631 * ip: inode for directory entry being removed 2632 * isrmdir: indicates if doing RMDIR 2633 */ 2634 void 2635 softdep_setup_remove(struct buf *bp, struct inode *dp, struct inode *ip, 2636 int isrmdir) 2637 { 2638 struct dirrem *dirrem, *prevdirrem; 2639 2640 /* 2641 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 2642 */ 2643 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2644 2645 /* 2646 * If the COMPLETE flag is clear, then there were no active 2647 * entries and we want to roll back to a zeroed entry until 2648 * the new inode is committed to disk. If the COMPLETE flag is 2649 * set then we have deleted an entry that never made it to 2650 * disk. If the entry we deleted resulted from a name change, 2651 * then the old name still resides on disk. We cannot delete 2652 * its inode (returned to us in prevdirrem) until the zeroed 2653 * directory entry gets to disk. The new inode has never been 2654 * referenced on the disk, so can be deleted immediately. 2655 */ 2656 if ((dirrem->dm_state & COMPLETE) == 0) { 2657 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 2658 dm_next); 2659 FREE_LOCK(&lk); 2660 } else { 2661 if (prevdirrem != NULL) 2662 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 2663 prevdirrem, dm_next); 2664 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 2665 FREE_LOCK(&lk); 2666 handle_workitem_remove(dirrem); 2667 } 2668 } 2669 2670 /* 2671 * Allocate a new dirrem if appropriate and return it along with 2672 * its associated pagedep. Called without a lock, returns with lock. 2673 */ 2674 static long num_dirrem; /* number of dirrem allocated */ 2675 2676 /* 2677 * Parameters: 2678 * bp: buffer containing directory block 2679 * dp: inode for the directory being modified 2680 * ip: inode for directory entry being removed 2681 * isrmdir: indicates if doing RMDIR 2682 * prevdirremp: previously referenced inode, if any 2683 */ 2684 static struct dirrem * 2685 newdirrem(struct buf *bp, struct inode *dp, struct inode *ip, 2686 int isrmdir, struct dirrem **prevdirremp) 2687 { 2688 int offset; 2689 ufs_lbn_t lbn; 2690 struct diradd *dap; 2691 struct dirrem *dirrem; 2692 struct pagedep *pagedep; 2693 2694 /* 2695 * Whiteouts have no deletion dependencies. 2696 */ 2697 if (ip == NULL) 2698 panic("newdirrem: whiteout"); 2699 /* 2700 * If we are over our limit, try to improve the situation. 2701 * Limiting the number of dirrem structures will also limit 2702 * the number of freefile and freeblks structures. 2703 */ 2704 if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0) 2705 (void) request_cleanup(FLUSH_REMOVE, 0); 2706 num_dirrem += 1; 2707 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem), 2708 M_DIRREM, M_SOFTDEP_FLAGS | M_ZERO); 2709 dirrem->dm_list.wk_type = D_DIRREM; 2710 dirrem->dm_state = isrmdir ? RMDIR : 0; 2711 dirrem->dm_mnt = ITOV(ip)->v_mount; 2712 dirrem->dm_oldinum = ip->i_number; 2713 *prevdirremp = NULL; 2714 2715 ACQUIRE_LOCK(&lk); 2716 lbn = lblkno(dp->i_fs, dp->i_offset); 2717 offset = blkoff(dp->i_fs, dp->i_offset); 2718 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2719 WORKLIST_INSERT_BP(bp, &pagedep->pd_list); 2720 dirrem->dm_pagedep = pagedep; 2721 /* 2722 * Check for a diradd dependency for the same directory entry. 2723 * If present, then both dependencies become obsolete and can 2724 * be de-allocated. Check for an entry on both the pd_dirraddhd 2725 * list and the pd_pendinghd list. 2726 */ 2727 2728 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 2729 if (dap->da_offset == offset) 2730 break; 2731 if (dap == NULL) { 2732 2733 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 2734 if (dap->da_offset == offset) 2735 break; 2736 if (dap == NULL) 2737 return (dirrem); 2738 } 2739 /* 2740 * Must be ATTACHED at this point. 2741 */ 2742 if ((dap->da_state & ATTACHED) == 0) { 2743 FREE_LOCK(&lk); 2744 panic("newdirrem: not ATTACHED"); 2745 } 2746 if (dap->da_newinum != ip->i_number) { 2747 FREE_LOCK(&lk); 2748 panic("newdirrem: inum %"PRId64" should be %"PRId64, 2749 ip->i_number, dap->da_newinum); 2750 } 2751 /* 2752 * If we are deleting a changed name that never made it to disk, 2753 * then return the dirrem describing the previous inode (which 2754 * represents the inode currently referenced from this entry on disk). 2755 */ 2756 if ((dap->da_state & DIRCHG) != 0) { 2757 *prevdirremp = dap->da_previous; 2758 dap->da_state &= ~DIRCHG; 2759 dap->da_pagedep = pagedep; 2760 } 2761 /* 2762 * We are deleting an entry that never made it to disk. 2763 * Mark it COMPLETE so we can delete its inode immediately. 2764 */ 2765 dirrem->dm_state |= COMPLETE; 2766 free_diradd(dap); 2767 return (dirrem); 2768 } 2769 2770 /* 2771 * Directory entry change dependencies. 2772 * 2773 * Changing an existing directory entry requires that an add operation 2774 * be completed first followed by a deletion. The semantics for the addition 2775 * are identical to the description of adding a new entry above except 2776 * that the rollback is to the old inode number rather than zero. Once 2777 * the addition dependency is completed, the removal is done as described 2778 * in the removal routine above. 2779 */ 2780 2781 /* 2782 * This routine should be called immediately after changing 2783 * a directory entry. The inode's link count should not be 2784 * decremented by the calling procedure -- the soft updates 2785 * code will perform this task when it is safe. 2786 * 2787 * Parameters: 2788 * bp: buffer containing directory block 2789 * dp: inode for the directory being modified 2790 * ip: inode for directory entry being removed 2791 * newinum: new inode number for changed entry 2792 * isrmdir: indicates if doing RMDIR 2793 */ 2794 void 2795 softdep_setup_directory_change(struct buf *bp, struct inode *dp, 2796 struct inode *ip, ino_t newinum, 2797 int isrmdir) 2798 { 2799 int offset; 2800 struct diradd *dap = NULL; 2801 struct dirrem *dirrem, *prevdirrem; 2802 struct pagedep *pagedep; 2803 struct inodedep *inodedep; 2804 2805 offset = blkoff(dp->i_fs, dp->i_offset); 2806 2807 /* 2808 * Whiteouts do not need diradd dependencies. 2809 */ 2810 if (newinum != WINO) { 2811 MALLOC(dap, struct diradd *, sizeof(struct diradd), 2812 M_DIRADD, M_SOFTDEP_FLAGS | M_ZERO); 2813 dap->da_list.wk_type = D_DIRADD; 2814 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 2815 dap->da_offset = offset; 2816 dap->da_newinum = newinum; 2817 } 2818 2819 /* 2820 * Allocate a new dirrem and ACQUIRE_LOCK. 2821 */ 2822 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2823 pagedep = dirrem->dm_pagedep; 2824 /* 2825 * The possible values for isrmdir: 2826 * 0 - non-directory file rename 2827 * 1 - directory rename within same directory 2828 * inum - directory rename to new directory of given inode number 2829 * When renaming to a new directory, we are both deleting and 2830 * creating a new directory entry, so the link count on the new 2831 * directory should not change. Thus we do not need the followup 2832 * dirrem which is usually done in handle_workitem_remove. We set 2833 * the DIRCHG flag to tell handle_workitem_remove to skip the 2834 * followup dirrem. 2835 */ 2836 if (isrmdir > 1) 2837 dirrem->dm_state |= DIRCHG; 2838 2839 /* 2840 * Whiteouts have no additional dependencies, 2841 * so just put the dirrem on the correct list. 2842 */ 2843 if (newinum == WINO) { 2844 if ((dirrem->dm_state & COMPLETE) == 0) { 2845 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 2846 dm_next); 2847 } else { 2848 dirrem->dm_dirinum = pagedep->pd_ino; 2849 add_to_worklist(&dirrem->dm_list); 2850 } 2851 FREE_LOCK(&lk); 2852 return; 2853 } 2854 2855 /* 2856 * If the COMPLETE flag is clear, then there were no active 2857 * entries and we want to roll back to the previous inode until 2858 * the new inode is committed to disk. If the COMPLETE flag is 2859 * set, then we have deleted an entry that never made it to disk. 2860 * If the entry we deleted resulted from a name change, then the old 2861 * inode reference still resides on disk. Any rollback that we do 2862 * needs to be to that old inode (returned to us in prevdirrem). If 2863 * the entry we deleted resulted from a create, then there is 2864 * no entry on the disk, so we want to roll back to zero rather 2865 * than the uncommitted inode. In either of the COMPLETE cases we 2866 * want to immediately free the unwritten and unreferenced inode. 2867 */ 2868 if ((dirrem->dm_state & COMPLETE) == 0) { 2869 dap->da_previous = dirrem; 2870 } else { 2871 if (prevdirrem != NULL) { 2872 dap->da_previous = prevdirrem; 2873 } else { 2874 dap->da_state &= ~DIRCHG; 2875 dap->da_pagedep = pagedep; 2876 } 2877 dirrem->dm_dirinum = pagedep->pd_ino; 2878 add_to_worklist(&dirrem->dm_list); 2879 } 2880 /* 2881 * Link into its inodedep. Put it on the id_bufwait list if the inode 2882 * is not yet written. If it is written, do the post-inode write 2883 * processing to put it on the id_pendinghd list. 2884 */ 2885 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 || 2886 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2887 dap->da_state |= COMPLETE; 2888 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 2889 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 2890 } else { 2891 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 2892 dap, da_pdlist); 2893 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2894 } 2895 FREE_LOCK(&lk); 2896 } 2897 2898 /* 2899 * Called whenever the link count on an inode is changed. 2900 * It creates an inode dependency so that the new reference(s) 2901 * to the inode cannot be committed to disk until the updated 2902 * inode has been written. 2903 * 2904 * Parameters: 2905 * ip: the inode with the increased link count 2906 */ 2907 void 2908 softdep_change_linkcnt(struct inode *ip) 2909 { 2910 struct inodedep *inodedep; 2911 2912 ACQUIRE_LOCK(&lk); 2913 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 2914 if (ip->i_nlink < ip->i_effnlink) { 2915 FREE_LOCK(&lk); 2916 panic("softdep_change_linkcnt: bad delta"); 2917 } 2918 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2919 FREE_LOCK(&lk); 2920 } 2921 2922 /* 2923 * This workitem decrements the inode's link count. 2924 * If the link count reaches zero, the file is removed. 2925 */ 2926 static void 2927 handle_workitem_remove(struct dirrem *dirrem) 2928 { 2929 struct inodedep *inodedep; 2930 struct vnode *vp; 2931 struct inode *ip; 2932 ino_t oldinum; 2933 int error; 2934 2935 error = VFS_VGET(dirrem->dm_mnt, NULL, dirrem->dm_oldinum, &vp); 2936 if (error) { 2937 softdep_error("handle_workitem_remove: vget", error); 2938 return; 2939 } 2940 ip = VTOI(vp); 2941 ACQUIRE_LOCK(&lk); 2942 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){ 2943 FREE_LOCK(&lk); 2944 panic("handle_workitem_remove: lost inodedep"); 2945 } 2946 /* 2947 * Normal file deletion. 2948 */ 2949 if ((dirrem->dm_state & RMDIR) == 0) { 2950 ip->i_nlink--; 2951 ip->i_flag |= IN_CHANGE; 2952 if (ip->i_nlink < ip->i_effnlink) { 2953 FREE_LOCK(&lk); 2954 panic("handle_workitem_remove: bad file delta"); 2955 } 2956 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2957 FREE_LOCK(&lk); 2958 vput(vp); 2959 num_dirrem -= 1; 2960 WORKITEM_FREE(dirrem, D_DIRREM); 2961 return; 2962 } 2963 /* 2964 * Directory deletion. Decrement reference count for both the 2965 * just deleted parent directory entry and the reference for ".". 2966 * Next truncate the directory to length zero. When the 2967 * truncation completes, arrange to have the reference count on 2968 * the parent decremented to account for the loss of "..". 2969 */ 2970 ip->i_nlink -= 2; 2971 ip->i_flag |= IN_CHANGE; 2972 if (ip->i_nlink < ip->i_effnlink) { 2973 FREE_LOCK(&lk); 2974 panic("handle_workitem_remove: bad dir delta"); 2975 } 2976 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2977 FREE_LOCK(&lk); 2978 if ((error = ffs_truncate(vp, (off_t)0, 0, proc0.p_ucred)) != 0) 2979 softdep_error("handle_workitem_remove: truncate", error); 2980 /* 2981 * Rename a directory to a new parent. Since, we are both deleting 2982 * and creating a new directory entry, the link count on the new 2983 * directory should not change. Thus we skip the followup dirrem. 2984 */ 2985 if (dirrem->dm_state & DIRCHG) { 2986 vput(vp); 2987 num_dirrem -= 1; 2988 WORKITEM_FREE(dirrem, D_DIRREM); 2989 return; 2990 } 2991 /* 2992 * If the inodedep does not exist, then the zero'ed inode has 2993 * been written to disk. If the allocated inode has never been 2994 * written to disk, then the on-disk inode is zero'ed. In either 2995 * case we can remove the file immediately. 2996 */ 2997 ACQUIRE_LOCK(&lk); 2998 dirrem->dm_state = 0; 2999 oldinum = dirrem->dm_oldinum; 3000 dirrem->dm_oldinum = dirrem->dm_dirinum; 3001 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 || 3002 check_inode_unwritten(inodedep)) { 3003 FREE_LOCK(&lk); 3004 vput(vp); 3005 handle_workitem_remove(dirrem); 3006 return; 3007 } 3008 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 3009 FREE_LOCK(&lk); 3010 ip->i_flag |= IN_CHANGE; 3011 ffs_update(vp, 0); 3012 vput(vp); 3013 } 3014 3015 /* 3016 * Inode de-allocation dependencies. 3017 * 3018 * When an inode's link count is reduced to zero, it can be de-allocated. We 3019 * found it convenient to postpone de-allocation until after the inode is 3020 * written to disk with its new link count (zero). At this point, all of the 3021 * on-disk inode's block pointers are nullified and, with careful dependency 3022 * list ordering, all dependencies related to the inode will be satisfied and 3023 * the corresponding dependency structures de-allocated. So, if/when the 3024 * inode is reused, there will be no mixing of old dependencies with new 3025 * ones. This artificial dependency is set up by the block de-allocation 3026 * procedure above (softdep_setup_freeblocks) and completed by the 3027 * following procedure. 3028 */ 3029 static void 3030 handle_workitem_freefile(struct freefile *freefile) 3031 { 3032 struct vnode vp; 3033 struct inode tip; 3034 struct inodedep *idp; 3035 int error; 3036 3037 #ifdef DEBUG 3038 ACQUIRE_LOCK(&lk); 3039 error = inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp); 3040 FREE_LOCK(&lk); 3041 if (error) 3042 panic("handle_workitem_freefile: inodedep survived"); 3043 #endif 3044 tip.i_devvp = freefile->fx_devvp; 3045 tip.i_dev = freefile->fx_devvp->v_rdev; 3046 tip.i_fs = freefile->fx_fs; 3047 vp.v_data = &tip; 3048 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0) 3049 softdep_error("handle_workitem_freefile", error); 3050 WORKITEM_FREE(freefile, D_FREEFILE); 3051 } 3052 3053 /* 3054 * Helper function which unlinks marker element from work list and returns 3055 * the next element on the list. 3056 */ 3057 static __inline struct worklist * 3058 markernext(struct worklist *marker) 3059 { 3060 struct worklist *next; 3061 3062 next = LIST_NEXT(marker, wk_list); 3063 LIST_REMOVE(marker, wk_list); 3064 return next; 3065 } 3066 3067 /* 3068 * checkread, checkwrite 3069 * 3070 */ 3071 static int 3072 softdep_checkread(struct buf *bp) 3073 { 3074 return(0); 3075 } 3076 3077 static int 3078 softdep_checkwrite(struct buf *bp) 3079 { 3080 return(0); 3081 } 3082 3083 /* 3084 * Disk writes. 3085 * 3086 * The dependency structures constructed above are most actively used when file 3087 * system blocks are written to disk. No constraints are placed on when a 3088 * block can be written, but unsatisfied update dependencies are made safe by 3089 * modifying (or replacing) the source memory for the duration of the disk 3090 * write. When the disk write completes, the memory block is again brought 3091 * up-to-date. 3092 * 3093 * In-core inode structure reclamation. 3094 * 3095 * Because there are a finite number of "in-core" inode structures, they are 3096 * reused regularly. By transferring all inode-related dependencies to the 3097 * in-memory inode block and indexing them separately (via "inodedep"s), we 3098 * can allow "in-core" inode structures to be reused at any time and avoid 3099 * any increase in contention. 3100 * 3101 * Called just before entering the device driver to initiate a new disk I/O. 3102 * The buffer must be locked, thus, no I/O completion operations can occur 3103 * while we are manipulating its associated dependencies. 3104 * 3105 * Parameters: 3106 * bp: structure describing disk write to occur 3107 */ 3108 static void 3109 softdep_disk_io_initiation(struct buf *bp) 3110 { 3111 struct worklist *wk; 3112 struct worklist marker; 3113 struct indirdep *indirdep; 3114 3115 /* 3116 * We only care about write operations. There should never 3117 * be dependencies for reads. 3118 */ 3119 if (bp->b_cmd == BUF_CMD_READ) 3120 panic("softdep_disk_io_initiation: read"); 3121 3122 marker.wk_type = D_LAST + 1; /* Not a normal workitem */ 3123 3124 /* 3125 * Do any necessary pre-I/O processing. 3126 */ 3127 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = markernext(&marker)) { 3128 LIST_INSERT_AFTER(wk, &marker, wk_list); 3129 3130 switch (wk->wk_type) { 3131 3132 case D_PAGEDEP: 3133 initiate_write_filepage(WK_PAGEDEP(wk), bp); 3134 continue; 3135 3136 case D_INODEDEP: 3137 initiate_write_inodeblock(WK_INODEDEP(wk), bp); 3138 continue; 3139 3140 case D_INDIRDEP: 3141 indirdep = WK_INDIRDEP(wk); 3142 if (indirdep->ir_state & GOINGAWAY) 3143 panic("disk_io_initiation: indirdep gone"); 3144 /* 3145 * If there are no remaining dependencies, this 3146 * will be writing the real pointers, so the 3147 * dependency can be freed. 3148 */ 3149 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) { 3150 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 3151 brelse(indirdep->ir_savebp); 3152 /* inline expand WORKLIST_REMOVE(wk); */ 3153 wk->wk_state &= ~ONWORKLIST; 3154 LIST_REMOVE(wk, wk_list); 3155 WORKITEM_FREE(indirdep, D_INDIRDEP); 3156 continue; 3157 } 3158 /* 3159 * Replace up-to-date version with safe version. 3160 */ 3161 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount, 3162 M_INDIRDEP, M_SOFTDEP_FLAGS); 3163 ACQUIRE_LOCK(&lk); 3164 indirdep->ir_state &= ~ATTACHED; 3165 indirdep->ir_state |= UNDONE; 3166 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 3167 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 3168 bp->b_bcount); 3169 FREE_LOCK(&lk); 3170 continue; 3171 3172 case D_MKDIR: 3173 case D_BMSAFEMAP: 3174 case D_ALLOCDIRECT: 3175 case D_ALLOCINDIR: 3176 continue; 3177 3178 default: 3179 panic("handle_disk_io_initiation: Unexpected type %s", 3180 TYPENAME(wk->wk_type)); 3181 /* NOTREACHED */ 3182 } 3183 } 3184 } 3185 3186 /* 3187 * Called from within the procedure above to deal with unsatisfied 3188 * allocation dependencies in a directory. The buffer must be locked, 3189 * thus, no I/O completion operations can occur while we are 3190 * manipulating its associated dependencies. 3191 */ 3192 static void 3193 initiate_write_filepage(struct pagedep *pagedep, struct buf *bp) 3194 { 3195 struct diradd *dap; 3196 struct direct *ep; 3197 int i; 3198 3199 if (pagedep->pd_state & IOSTARTED) { 3200 /* 3201 * This can only happen if there is a driver that does not 3202 * understand chaining. Here biodone will reissue the call 3203 * to strategy for the incomplete buffers. 3204 */ 3205 kprintf("initiate_write_filepage: already started\n"); 3206 return; 3207 } 3208 pagedep->pd_state |= IOSTARTED; 3209 ACQUIRE_LOCK(&lk); 3210 for (i = 0; i < DAHASHSZ; i++) { 3211 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 3212 ep = (struct direct *) 3213 ((char *)bp->b_data + dap->da_offset); 3214 if (ep->d_ino != dap->da_newinum) { 3215 FREE_LOCK(&lk); 3216 panic("%s: dir inum %d != new %"PRId64, 3217 "initiate_write_filepage", 3218 ep->d_ino, dap->da_newinum); 3219 } 3220 if (dap->da_state & DIRCHG) 3221 ep->d_ino = dap->da_previous->dm_oldinum; 3222 else 3223 ep->d_ino = 0; 3224 dap->da_state &= ~ATTACHED; 3225 dap->da_state |= UNDONE; 3226 } 3227 } 3228 FREE_LOCK(&lk); 3229 } 3230 3231 /* 3232 * Called from within the procedure above to deal with unsatisfied 3233 * allocation dependencies in an inodeblock. The buffer must be 3234 * locked, thus, no I/O completion operations can occur while we 3235 * are manipulating its associated dependencies. 3236 * 3237 * Parameters: 3238 * bp: The inode block 3239 */ 3240 static void 3241 initiate_write_inodeblock(struct inodedep *inodedep, struct buf *bp) 3242 { 3243 struct allocdirect *adp, *lastadp; 3244 struct ufs1_dinode *dp; 3245 struct ufs1_dinode *sip; 3246 struct fs *fs; 3247 ufs_lbn_t prevlbn = 0; 3248 int i, deplist; 3249 3250 if (inodedep->id_state & IOSTARTED) 3251 panic("initiate_write_inodeblock: already started"); 3252 inodedep->id_state |= IOSTARTED; 3253 fs = inodedep->id_fs; 3254 dp = (struct ufs1_dinode *)bp->b_data + 3255 ino_to_fsbo(fs, inodedep->id_ino); 3256 /* 3257 * If the bitmap is not yet written, then the allocated 3258 * inode cannot be written to disk. 3259 */ 3260 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 3261 if (inodedep->id_savedino != NULL) 3262 panic("initiate_write_inodeblock: already doing I/O"); 3263 MALLOC(sip, struct ufs1_dinode *, 3264 sizeof(struct ufs1_dinode), M_INODEDEP, M_SOFTDEP_FLAGS); 3265 inodedep->id_savedino = sip; 3266 *inodedep->id_savedino = *dp; 3267 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 3268 dp->di_gen = inodedep->id_savedino->di_gen; 3269 return; 3270 } 3271 /* 3272 * If no dependencies, then there is nothing to roll back. 3273 */ 3274 inodedep->id_savedsize = dp->di_size; 3275 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 3276 return; 3277 /* 3278 * Set the dependencies to busy. 3279 */ 3280 ACQUIRE_LOCK(&lk); 3281 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3282 adp = TAILQ_NEXT(adp, ad_next)) { 3283 #ifdef DIAGNOSTIC 3284 if (deplist != 0 && prevlbn >= adp->ad_lbn) { 3285 FREE_LOCK(&lk); 3286 panic("softdep_write_inodeblock: lbn order"); 3287 } 3288 prevlbn = adp->ad_lbn; 3289 if (adp->ad_lbn < NDADDR && 3290 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) { 3291 FREE_LOCK(&lk); 3292 panic("%s: direct pointer #%ld mismatch %d != %d", 3293 "softdep_write_inodeblock", adp->ad_lbn, 3294 dp->di_db[adp->ad_lbn], adp->ad_newblkno); 3295 } 3296 if (adp->ad_lbn >= NDADDR && 3297 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) { 3298 FREE_LOCK(&lk); 3299 panic("%s: indirect pointer #%ld mismatch %d != %d", 3300 "softdep_write_inodeblock", adp->ad_lbn - NDADDR, 3301 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno); 3302 } 3303 deplist |= 1 << adp->ad_lbn; 3304 if ((adp->ad_state & ATTACHED) == 0) { 3305 FREE_LOCK(&lk); 3306 panic("softdep_write_inodeblock: Unknown state 0x%x", 3307 adp->ad_state); 3308 } 3309 #endif /* DIAGNOSTIC */ 3310 adp->ad_state &= ~ATTACHED; 3311 adp->ad_state |= UNDONE; 3312 } 3313 /* 3314 * The on-disk inode cannot claim to be any larger than the last 3315 * fragment that has been written. Otherwise, the on-disk inode 3316 * might have fragments that were not the last block in the file 3317 * which would corrupt the filesystem. 3318 */ 3319 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3320 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 3321 if (adp->ad_lbn >= NDADDR) 3322 break; 3323 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 3324 /* keep going until hitting a rollback to a frag */ 3325 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 3326 continue; 3327 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 3328 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 3329 #ifdef DIAGNOSTIC 3330 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) { 3331 FREE_LOCK(&lk); 3332 panic("softdep_write_inodeblock: lost dep1"); 3333 } 3334 #endif /* DIAGNOSTIC */ 3335 dp->di_db[i] = 0; 3336 } 3337 for (i = 0; i < NIADDR; i++) { 3338 #ifdef DIAGNOSTIC 3339 if (dp->di_ib[i] != 0 && 3340 (deplist & ((1 << NDADDR) << i)) == 0) { 3341 FREE_LOCK(&lk); 3342 panic("softdep_write_inodeblock: lost dep2"); 3343 } 3344 #endif /* DIAGNOSTIC */ 3345 dp->di_ib[i] = 0; 3346 } 3347 FREE_LOCK(&lk); 3348 return; 3349 } 3350 /* 3351 * If we have zero'ed out the last allocated block of the file, 3352 * roll back the size to the last currently allocated block. 3353 * We know that this last allocated block is a full-sized as 3354 * we already checked for fragments in the loop above. 3355 */ 3356 if (lastadp != NULL && 3357 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 3358 for (i = lastadp->ad_lbn; i >= 0; i--) 3359 if (dp->di_db[i] != 0) 3360 break; 3361 dp->di_size = (i + 1) * fs->fs_bsize; 3362 } 3363 /* 3364 * The only dependencies are for indirect blocks. 3365 * 3366 * The file size for indirect block additions is not guaranteed. 3367 * Such a guarantee would be non-trivial to achieve. The conventional 3368 * synchronous write implementation also does not make this guarantee. 3369 * Fsck should catch and fix discrepancies. Arguably, the file size 3370 * can be over-estimated without destroying integrity when the file 3371 * moves into the indirect blocks (i.e., is large). If we want to 3372 * postpone fsck, we are stuck with this argument. 3373 */ 3374 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 3375 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 3376 FREE_LOCK(&lk); 3377 } 3378 3379 /* 3380 * This routine is called during the completion interrupt 3381 * service routine for a disk write (from the procedure called 3382 * by the device driver to inform the filesystem caches of 3383 * a request completion). It should be called early in this 3384 * procedure, before the block is made available to other 3385 * processes or other routines are called. 3386 * 3387 * Parameters: 3388 * bp: describes the completed disk write 3389 */ 3390 static void 3391 softdep_disk_write_complete(struct buf *bp) 3392 { 3393 struct worklist *wk; 3394 struct workhead reattach; 3395 struct newblk *newblk; 3396 struct allocindir *aip; 3397 struct allocdirect *adp; 3398 struct indirdep *indirdep; 3399 struct inodedep *inodedep; 3400 struct bmsafemap *bmsafemap; 3401 3402 #ifdef DEBUG 3403 if (lk.lkt_held != NOHOLDER) 3404 panic("softdep_disk_write_complete: lock is held"); 3405 lk.lkt_held = SPECIAL_FLAG; 3406 #endif 3407 LIST_INIT(&reattach); 3408 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 3409 WORKLIST_REMOVE(wk); 3410 switch (wk->wk_type) { 3411 3412 case D_PAGEDEP: 3413 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 3414 WORKLIST_INSERT(&reattach, wk); 3415 continue; 3416 3417 case D_INODEDEP: 3418 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 3419 WORKLIST_INSERT(&reattach, wk); 3420 continue; 3421 3422 case D_BMSAFEMAP: 3423 bmsafemap = WK_BMSAFEMAP(wk); 3424 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) { 3425 newblk->nb_state |= DEPCOMPLETE; 3426 newblk->nb_bmsafemap = NULL; 3427 LIST_REMOVE(newblk, nb_deps); 3428 } 3429 while ((adp = 3430 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) { 3431 adp->ad_state |= DEPCOMPLETE; 3432 adp->ad_buf = NULL; 3433 LIST_REMOVE(adp, ad_deps); 3434 handle_allocdirect_partdone(adp); 3435 } 3436 while ((aip = 3437 LIST_FIRST(&bmsafemap->sm_allocindirhd))) { 3438 aip->ai_state |= DEPCOMPLETE; 3439 aip->ai_buf = NULL; 3440 LIST_REMOVE(aip, ai_deps); 3441 handle_allocindir_partdone(aip); 3442 } 3443 while ((inodedep = 3444 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { 3445 inodedep->id_state |= DEPCOMPLETE; 3446 LIST_REMOVE(inodedep, id_deps); 3447 inodedep->id_buf = NULL; 3448 } 3449 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 3450 continue; 3451 3452 case D_MKDIR: 3453 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 3454 continue; 3455 3456 case D_ALLOCDIRECT: 3457 adp = WK_ALLOCDIRECT(wk); 3458 adp->ad_state |= COMPLETE; 3459 handle_allocdirect_partdone(adp); 3460 continue; 3461 3462 case D_ALLOCINDIR: 3463 aip = WK_ALLOCINDIR(wk); 3464 aip->ai_state |= COMPLETE; 3465 handle_allocindir_partdone(aip); 3466 continue; 3467 3468 case D_INDIRDEP: 3469 indirdep = WK_INDIRDEP(wk); 3470 if (indirdep->ir_state & GOINGAWAY) { 3471 lk.lkt_held = NOHOLDER; 3472 panic("disk_write_complete: indirdep gone"); 3473 } 3474 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 3475 FREE(indirdep->ir_saveddata, M_INDIRDEP); 3476 indirdep->ir_saveddata = 0; 3477 indirdep->ir_state &= ~UNDONE; 3478 indirdep->ir_state |= ATTACHED; 3479 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 3480 handle_allocindir_partdone(aip); 3481 if (aip == LIST_FIRST(&indirdep->ir_donehd)) { 3482 lk.lkt_held = NOHOLDER; 3483 panic("disk_write_complete: not gone"); 3484 } 3485 } 3486 WORKLIST_INSERT(&reattach, wk); 3487 if ((bp->b_flags & B_DELWRI) == 0) 3488 stat_indir_blk_ptrs++; 3489 bdirty(bp); 3490 continue; 3491 3492 default: 3493 lk.lkt_held = NOHOLDER; 3494 panic("handle_disk_write_complete: Unknown type %s", 3495 TYPENAME(wk->wk_type)); 3496 /* NOTREACHED */ 3497 } 3498 } 3499 /* 3500 * Reattach any requests that must be redone. 3501 */ 3502 while ((wk = LIST_FIRST(&reattach)) != NULL) { 3503 WORKLIST_REMOVE(wk); 3504 WORKLIST_INSERT_BP(bp, wk); 3505 } 3506 #ifdef DEBUG 3507 if (lk.lkt_held != SPECIAL_FLAG) 3508 panic("softdep_disk_write_complete: lock lost"); 3509 lk.lkt_held = NOHOLDER; 3510 #endif 3511 } 3512 3513 /* 3514 * Called from within softdep_disk_write_complete above. Note that 3515 * this routine is always called from interrupt level with further 3516 * splbio interrupts blocked. 3517 * 3518 * Parameters: 3519 * adp: the completed allocdirect 3520 */ 3521 static void 3522 handle_allocdirect_partdone(struct allocdirect *adp) 3523 { 3524 struct allocdirect *listadp; 3525 struct inodedep *inodedep; 3526 long bsize; 3527 3528 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3529 return; 3530 if (adp->ad_buf != NULL) { 3531 lk.lkt_held = NOHOLDER; 3532 panic("handle_allocdirect_partdone: dangling dep"); 3533 } 3534 /* 3535 * The on-disk inode cannot claim to be any larger than the last 3536 * fragment that has been written. Otherwise, the on-disk inode 3537 * might have fragments that were not the last block in the file 3538 * which would corrupt the filesystem. Thus, we cannot free any 3539 * allocdirects after one whose ad_oldblkno claims a fragment as 3540 * these blocks must be rolled back to zero before writing the inode. 3541 * We check the currently active set of allocdirects in id_inoupdt. 3542 */ 3543 inodedep = adp->ad_inodedep; 3544 bsize = inodedep->id_fs->fs_bsize; 3545 TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) { 3546 /* found our block */ 3547 if (listadp == adp) 3548 break; 3549 /* continue if ad_oldlbn is not a fragment */ 3550 if (listadp->ad_oldsize == 0 || 3551 listadp->ad_oldsize == bsize) 3552 continue; 3553 /* hit a fragment */ 3554 return; 3555 } 3556 /* 3557 * If we have reached the end of the current list without 3558 * finding the just finished dependency, then it must be 3559 * on the future dependency list. Future dependencies cannot 3560 * be freed until they are moved to the current list. 3561 */ 3562 if (listadp == NULL) { 3563 #ifdef DEBUG 3564 TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next) 3565 /* found our block */ 3566 if (listadp == adp) 3567 break; 3568 if (listadp == NULL) { 3569 lk.lkt_held = NOHOLDER; 3570 panic("handle_allocdirect_partdone: lost dep"); 3571 } 3572 #endif /* DEBUG */ 3573 return; 3574 } 3575 /* 3576 * If we have found the just finished dependency, then free 3577 * it along with anything that follows it that is complete. 3578 */ 3579 for (; adp; adp = listadp) { 3580 listadp = TAILQ_NEXT(adp, ad_next); 3581 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3582 return; 3583 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 3584 } 3585 } 3586 3587 /* 3588 * Called from within softdep_disk_write_complete above. Note that 3589 * this routine is always called from interrupt level with further 3590 * splbio interrupts blocked. 3591 * 3592 * Parameters: 3593 * aip: the completed allocindir 3594 */ 3595 static void 3596 handle_allocindir_partdone(struct allocindir *aip) 3597 { 3598 struct indirdep *indirdep; 3599 3600 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 3601 return; 3602 if (aip->ai_buf != NULL) { 3603 lk.lkt_held = NOHOLDER; 3604 panic("handle_allocindir_partdone: dangling dependency"); 3605 } 3606 indirdep = aip->ai_indirdep; 3607 if (indirdep->ir_state & UNDONE) { 3608 LIST_REMOVE(aip, ai_next); 3609 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 3610 return; 3611 } 3612 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 3613 aip->ai_newblkno; 3614 LIST_REMOVE(aip, ai_next); 3615 if (aip->ai_freefrag != NULL) 3616 add_to_worklist(&aip->ai_freefrag->ff_list); 3617 WORKITEM_FREE(aip, D_ALLOCINDIR); 3618 } 3619 3620 /* 3621 * Called from within softdep_disk_write_complete above to restore 3622 * in-memory inode block contents to their most up-to-date state. Note 3623 * that this routine is always called from interrupt level with further 3624 * splbio interrupts blocked. 3625 * 3626 * Parameters: 3627 * bp: buffer containing the inode block 3628 */ 3629 static int 3630 handle_written_inodeblock(struct inodedep *inodedep, struct buf *bp) 3631 { 3632 struct worklist *wk, *filefree; 3633 struct allocdirect *adp, *nextadp; 3634 struct ufs1_dinode *dp; 3635 int hadchanges; 3636 3637 if ((inodedep->id_state & IOSTARTED) == 0) { 3638 lk.lkt_held = NOHOLDER; 3639 panic("handle_written_inodeblock: not started"); 3640 } 3641 inodedep->id_state &= ~IOSTARTED; 3642 dp = (struct ufs1_dinode *)bp->b_data + 3643 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 3644 /* 3645 * If we had to rollback the inode allocation because of 3646 * bitmaps being incomplete, then simply restore it. 3647 * Keep the block dirty so that it will not be reclaimed until 3648 * all associated dependencies have been cleared and the 3649 * corresponding updates written to disk. 3650 */ 3651 if (inodedep->id_savedino != NULL) { 3652 *dp = *inodedep->id_savedino; 3653 FREE(inodedep->id_savedino, M_INODEDEP); 3654 inodedep->id_savedino = NULL; 3655 if ((bp->b_flags & B_DELWRI) == 0) 3656 stat_inode_bitmap++; 3657 bdirty(bp); 3658 return (1); 3659 } 3660 inodedep->id_state |= COMPLETE; 3661 /* 3662 * Roll forward anything that had to be rolled back before 3663 * the inode could be updated. 3664 */ 3665 hadchanges = 0; 3666 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 3667 nextadp = TAILQ_NEXT(adp, ad_next); 3668 if (adp->ad_state & ATTACHED) { 3669 lk.lkt_held = NOHOLDER; 3670 panic("handle_written_inodeblock: new entry"); 3671 } 3672 if (adp->ad_lbn < NDADDR) { 3673 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) { 3674 lk.lkt_held = NOHOLDER; 3675 panic("%s: %s #%ld mismatch %d != %d", 3676 "handle_written_inodeblock", 3677 "direct pointer", adp->ad_lbn, 3678 dp->di_db[adp->ad_lbn], adp->ad_oldblkno); 3679 } 3680 dp->di_db[adp->ad_lbn] = adp->ad_newblkno; 3681 } else { 3682 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) { 3683 lk.lkt_held = NOHOLDER; 3684 panic("%s: %s #%ld allocated as %d", 3685 "handle_written_inodeblock", 3686 "indirect pointer", adp->ad_lbn - NDADDR, 3687 dp->di_ib[adp->ad_lbn - NDADDR]); 3688 } 3689 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno; 3690 } 3691 adp->ad_state &= ~UNDONE; 3692 adp->ad_state |= ATTACHED; 3693 hadchanges = 1; 3694 } 3695 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 3696 stat_direct_blk_ptrs++; 3697 /* 3698 * Reset the file size to its most up-to-date value. 3699 */ 3700 if (inodedep->id_savedsize == -1) { 3701 lk.lkt_held = NOHOLDER; 3702 panic("handle_written_inodeblock: bad size"); 3703 } 3704 if (dp->di_size != inodedep->id_savedsize) { 3705 dp->di_size = inodedep->id_savedsize; 3706 hadchanges = 1; 3707 } 3708 inodedep->id_savedsize = -1; 3709 /* 3710 * If there were any rollbacks in the inode block, then it must be 3711 * marked dirty so that its will eventually get written back in 3712 * its correct form. 3713 */ 3714 if (hadchanges) 3715 bdirty(bp); 3716 /* 3717 * Process any allocdirects that completed during the update. 3718 */ 3719 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 3720 handle_allocdirect_partdone(adp); 3721 /* 3722 * Process deallocations that were held pending until the 3723 * inode had been written to disk. Freeing of the inode 3724 * is delayed until after all blocks have been freed to 3725 * avoid creation of new <vfsid, inum, lbn> triples 3726 * before the old ones have been deleted. 3727 */ 3728 filefree = NULL; 3729 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 3730 WORKLIST_REMOVE(wk); 3731 switch (wk->wk_type) { 3732 3733 case D_FREEFILE: 3734 /* 3735 * We defer adding filefree to the worklist until 3736 * all other additions have been made to ensure 3737 * that it will be done after all the old blocks 3738 * have been freed. 3739 */ 3740 if (filefree != NULL) { 3741 lk.lkt_held = NOHOLDER; 3742 panic("handle_written_inodeblock: filefree"); 3743 } 3744 filefree = wk; 3745 continue; 3746 3747 case D_MKDIR: 3748 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 3749 continue; 3750 3751 case D_DIRADD: 3752 diradd_inode_written(WK_DIRADD(wk), inodedep); 3753 continue; 3754 3755 case D_FREEBLKS: 3756 wk->wk_state |= COMPLETE; 3757 if ((wk->wk_state & ALLCOMPLETE) != ALLCOMPLETE) 3758 continue; 3759 /* -- fall through -- */ 3760 case D_FREEFRAG: 3761 case D_DIRREM: 3762 add_to_worklist(wk); 3763 continue; 3764 3765 default: 3766 lk.lkt_held = NOHOLDER; 3767 panic("handle_written_inodeblock: Unknown type %s", 3768 TYPENAME(wk->wk_type)); 3769 /* NOTREACHED */ 3770 } 3771 } 3772 if (filefree != NULL) { 3773 if (free_inodedep(inodedep) == 0) { 3774 lk.lkt_held = NOHOLDER; 3775 panic("handle_written_inodeblock: live inodedep"); 3776 } 3777 add_to_worklist(filefree); 3778 return (0); 3779 } 3780 3781 /* 3782 * If no outstanding dependencies, free it. 3783 */ 3784 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0) 3785 return (0); 3786 return (hadchanges); 3787 } 3788 3789 /* 3790 * Process a diradd entry after its dependent inode has been written. 3791 * This routine must be called with splbio interrupts blocked. 3792 */ 3793 static void 3794 diradd_inode_written(struct diradd *dap, struct inodedep *inodedep) 3795 { 3796 struct pagedep *pagedep; 3797 3798 dap->da_state |= COMPLETE; 3799 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3800 if (dap->da_state & DIRCHG) 3801 pagedep = dap->da_previous->dm_pagedep; 3802 else 3803 pagedep = dap->da_pagedep; 3804 LIST_REMOVE(dap, da_pdlist); 3805 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3806 } 3807 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 3808 } 3809 3810 /* 3811 * Handle the completion of a mkdir dependency. 3812 */ 3813 static void 3814 handle_written_mkdir(struct mkdir *mkdir, int type) 3815 { 3816 struct diradd *dap; 3817 struct pagedep *pagedep; 3818 3819 if (mkdir->md_state != type) { 3820 lk.lkt_held = NOHOLDER; 3821 panic("handle_written_mkdir: bad type"); 3822 } 3823 dap = mkdir->md_diradd; 3824 dap->da_state &= ~type; 3825 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 3826 dap->da_state |= DEPCOMPLETE; 3827 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3828 if (dap->da_state & DIRCHG) 3829 pagedep = dap->da_previous->dm_pagedep; 3830 else 3831 pagedep = dap->da_pagedep; 3832 LIST_REMOVE(dap, da_pdlist); 3833 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3834 } 3835 LIST_REMOVE(mkdir, md_mkdirs); 3836 WORKITEM_FREE(mkdir, D_MKDIR); 3837 } 3838 3839 /* 3840 * Called from within softdep_disk_write_complete above. 3841 * A write operation was just completed. Removed inodes can 3842 * now be freed and associated block pointers may be committed. 3843 * Note that this routine is always called from interrupt level 3844 * with further splbio interrupts blocked. 3845 * 3846 * Parameters: 3847 * bp: buffer containing the written page 3848 */ 3849 static int 3850 handle_written_filepage(struct pagedep *pagedep, struct buf *bp) 3851 { 3852 struct dirrem *dirrem; 3853 struct diradd *dap, *nextdap; 3854 struct direct *ep; 3855 int i, chgs; 3856 3857 if ((pagedep->pd_state & IOSTARTED) == 0) { 3858 lk.lkt_held = NOHOLDER; 3859 panic("handle_written_filepage: not started"); 3860 } 3861 pagedep->pd_state &= ~IOSTARTED; 3862 /* 3863 * Process any directory removals that have been committed. 3864 */ 3865 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 3866 LIST_REMOVE(dirrem, dm_next); 3867 dirrem->dm_dirinum = pagedep->pd_ino; 3868 add_to_worklist(&dirrem->dm_list); 3869 } 3870 /* 3871 * Free any directory additions that have been committed. 3872 */ 3873 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 3874 free_diradd(dap); 3875 /* 3876 * Uncommitted directory entries must be restored. 3877 */ 3878 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 3879 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 3880 dap = nextdap) { 3881 nextdap = LIST_NEXT(dap, da_pdlist); 3882 if (dap->da_state & ATTACHED) { 3883 lk.lkt_held = NOHOLDER; 3884 panic("handle_written_filepage: attached"); 3885 } 3886 ep = (struct direct *) 3887 ((char *)bp->b_data + dap->da_offset); 3888 ep->d_ino = dap->da_newinum; 3889 dap->da_state &= ~UNDONE; 3890 dap->da_state |= ATTACHED; 3891 chgs = 1; 3892 /* 3893 * If the inode referenced by the directory has 3894 * been written out, then the dependency can be 3895 * moved to the pending list. 3896 */ 3897 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3898 LIST_REMOVE(dap, da_pdlist); 3899 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 3900 da_pdlist); 3901 } 3902 } 3903 } 3904 /* 3905 * If there were any rollbacks in the directory, then it must be 3906 * marked dirty so that its will eventually get written back in 3907 * its correct form. 3908 */ 3909 if (chgs) { 3910 if ((bp->b_flags & B_DELWRI) == 0) 3911 stat_dir_entry++; 3912 bdirty(bp); 3913 } 3914 /* 3915 * If no dependencies remain, the pagedep will be freed. 3916 * Otherwise it will remain to update the page before it 3917 * is written back to disk. 3918 */ 3919 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) { 3920 for (i = 0; i < DAHASHSZ; i++) 3921 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL) 3922 break; 3923 if (i == DAHASHSZ) { 3924 LIST_REMOVE(pagedep, pd_hash); 3925 WORKITEM_FREE(pagedep, D_PAGEDEP); 3926 return (0); 3927 } 3928 } 3929 return (1); 3930 } 3931 3932 /* 3933 * Writing back in-core inode structures. 3934 * 3935 * The filesystem only accesses an inode's contents when it occupies an 3936 * "in-core" inode structure. These "in-core" structures are separate from 3937 * the page frames used to cache inode blocks. Only the latter are 3938 * transferred to/from the disk. So, when the updated contents of the 3939 * "in-core" inode structure are copied to the corresponding in-memory inode 3940 * block, the dependencies are also transferred. The following procedure is 3941 * called when copying a dirty "in-core" inode to a cached inode block. 3942 */ 3943 3944 /* 3945 * Called when an inode is loaded from disk. If the effective link count 3946 * differed from the actual link count when it was last flushed, then we 3947 * need to ensure that the correct effective link count is put back. 3948 * 3949 * Parameters: 3950 * ip: the "in_core" copy of the inode 3951 */ 3952 void 3953 softdep_load_inodeblock(struct inode *ip) 3954 { 3955 struct inodedep *inodedep; 3956 3957 /* 3958 * Check for alternate nlink count. 3959 */ 3960 ip->i_effnlink = ip->i_nlink; 3961 ACQUIRE_LOCK(&lk); 3962 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3963 FREE_LOCK(&lk); 3964 return; 3965 } 3966 ip->i_effnlink -= inodedep->id_nlinkdelta; 3967 FREE_LOCK(&lk); 3968 } 3969 3970 /* 3971 * This routine is called just before the "in-core" inode 3972 * information is to be copied to the in-memory inode block. 3973 * Recall that an inode block contains several inodes. If 3974 * the force flag is set, then the dependencies will be 3975 * cleared so that the update can always be made. Note that 3976 * the buffer is locked when this routine is called, so we 3977 * will never be in the middle of writing the inode block 3978 * to disk. 3979 * 3980 * Parameters: 3981 * ip: the "in_core" copy of the inode 3982 * bp: the buffer containing the inode block 3983 * waitfor: nonzero => update must be allowed 3984 */ 3985 void 3986 softdep_update_inodeblock(struct inode *ip, struct buf *bp, 3987 int waitfor) 3988 { 3989 struct inodedep *inodedep; 3990 struct worklist *wk; 3991 int error, gotit; 3992 3993 /* 3994 * If the effective link count is not equal to the actual link 3995 * count, then we must track the difference in an inodedep while 3996 * the inode is (potentially) tossed out of the cache. Otherwise, 3997 * if there is no existing inodedep, then there are no dependencies 3998 * to track. 3999 */ 4000 ACQUIRE_LOCK(&lk); 4001 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 4002 FREE_LOCK(&lk); 4003 if (ip->i_effnlink != ip->i_nlink) 4004 panic("softdep_update_inodeblock: bad link count"); 4005 return; 4006 } 4007 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) { 4008 FREE_LOCK(&lk); 4009 panic("softdep_update_inodeblock: bad delta"); 4010 } 4011 /* 4012 * Changes have been initiated. Anything depending on these 4013 * changes cannot occur until this inode has been written. 4014 */ 4015 inodedep->id_state &= ~COMPLETE; 4016 if ((inodedep->id_state & ONWORKLIST) == 0) 4017 WORKLIST_INSERT_BP(bp, &inodedep->id_list); 4018 /* 4019 * Any new dependencies associated with the incore inode must 4020 * now be moved to the list associated with the buffer holding 4021 * the in-memory copy of the inode. Once merged process any 4022 * allocdirects that are completed by the merger. 4023 */ 4024 merge_inode_lists(inodedep); 4025 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL) 4026 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); 4027 /* 4028 * Now that the inode has been pushed into the buffer, the 4029 * operations dependent on the inode being written to disk 4030 * can be moved to the id_bufwait so that they will be 4031 * processed when the buffer I/O completes. 4032 */ 4033 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 4034 WORKLIST_REMOVE(wk); 4035 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 4036 } 4037 /* 4038 * Newly allocated inodes cannot be written until the bitmap 4039 * that allocates them have been written (indicated by 4040 * DEPCOMPLETE being set in id_state). If we are doing a 4041 * forced sync (e.g., an fsync on a file), we force the bitmap 4042 * to be written so that the update can be done. 4043 */ 4044 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) { 4045 FREE_LOCK(&lk); 4046 return; 4047 } 4048 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 4049 FREE_LOCK(&lk); 4050 if (gotit && 4051 (error = bwrite(inodedep->id_buf)) != 0) 4052 softdep_error("softdep_update_inodeblock: bwrite", error); 4053 } 4054 4055 /* 4056 * Merge the new inode dependency list (id_newinoupdt) into the old 4057 * inode dependency list (id_inoupdt). This routine must be called 4058 * with splbio interrupts blocked. 4059 */ 4060 static void 4061 merge_inode_lists(struct inodedep *inodedep) 4062 { 4063 struct allocdirect *listadp, *newadp; 4064 4065 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 4066 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) { 4067 if (listadp->ad_lbn < newadp->ad_lbn) { 4068 listadp = TAILQ_NEXT(listadp, ad_next); 4069 continue; 4070 } 4071 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 4072 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 4073 if (listadp->ad_lbn == newadp->ad_lbn) { 4074 allocdirect_merge(&inodedep->id_inoupdt, newadp, 4075 listadp); 4076 listadp = newadp; 4077 } 4078 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 4079 } 4080 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) { 4081 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 4082 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next); 4083 } 4084 } 4085 4086 /* 4087 * If we are doing an fsync, then we must ensure that any directory 4088 * entries for the inode have been written after the inode gets to disk. 4089 * 4090 * Parameters: 4091 * vp: the "in_core" copy of the inode 4092 */ 4093 static int 4094 softdep_fsync(struct vnode *vp) 4095 { 4096 struct inodedep *inodedep; 4097 struct pagedep *pagedep; 4098 struct worklist *wk; 4099 struct diradd *dap; 4100 struct mount *mnt; 4101 struct vnode *pvp; 4102 struct inode *ip; 4103 struct buf *bp; 4104 struct fs *fs; 4105 int error, flushparent; 4106 ino_t parentino; 4107 ufs_lbn_t lbn; 4108 4109 /* 4110 * Move check from original kernel code, possibly not needed any 4111 * more with the per-mount bioops. 4112 */ 4113 if ((vp->v_mount->mnt_flag & MNT_SOFTDEP) == 0) 4114 return (0); 4115 4116 ip = VTOI(vp); 4117 fs = ip->i_fs; 4118 ACQUIRE_LOCK(&lk); 4119 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) { 4120 FREE_LOCK(&lk); 4121 return (0); 4122 } 4123 if (LIST_FIRST(&inodedep->id_inowait) != NULL || 4124 LIST_FIRST(&inodedep->id_bufwait) != NULL || 4125 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 4126 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) { 4127 FREE_LOCK(&lk); 4128 panic("softdep_fsync: pending ops"); 4129 } 4130 for (error = 0, flushparent = 0; ; ) { 4131 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 4132 break; 4133 if (wk->wk_type != D_DIRADD) { 4134 FREE_LOCK(&lk); 4135 panic("softdep_fsync: Unexpected type %s", 4136 TYPENAME(wk->wk_type)); 4137 } 4138 dap = WK_DIRADD(wk); 4139 /* 4140 * Flush our parent if this directory entry 4141 * has a MKDIR_PARENT dependency. 4142 */ 4143 if (dap->da_state & DIRCHG) 4144 pagedep = dap->da_previous->dm_pagedep; 4145 else 4146 pagedep = dap->da_pagedep; 4147 mnt = pagedep->pd_mnt; 4148 parentino = pagedep->pd_ino; 4149 lbn = pagedep->pd_lbn; 4150 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) { 4151 FREE_LOCK(&lk); 4152 panic("softdep_fsync: dirty"); 4153 } 4154 flushparent = dap->da_state & MKDIR_PARENT; 4155 /* 4156 * If we are being fsync'ed as part of vgone'ing this vnode, 4157 * then we will not be able to release and recover the 4158 * vnode below, so we just have to give up on writing its 4159 * directory entry out. It will eventually be written, just 4160 * not now, but then the user was not asking to have it 4161 * written, so we are not breaking any promises. 4162 */ 4163 if (vp->v_flag & VRECLAIMED) 4164 break; 4165 /* 4166 * We prevent deadlock by always fetching inodes from the 4167 * root, moving down the directory tree. Thus, when fetching 4168 * our parent directory, we must unlock ourselves before 4169 * requesting the lock on our parent. See the comment in 4170 * ufs_lookup for details on possible races. 4171 */ 4172 FREE_LOCK(&lk); 4173 vn_unlock(vp); 4174 error = VFS_VGET(mnt, NULL, parentino, &pvp); 4175 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4176 if (error != 0) 4177 return (error); 4178 if (flushparent) { 4179 if ((error = ffs_update(pvp, 1)) != 0) { 4180 vput(pvp); 4181 return (error); 4182 } 4183 } 4184 /* 4185 * Flush directory page containing the inode's name. 4186 */ 4187 error = bread(pvp, lblktodoff(fs, lbn), blksize(fs, VTOI(pvp), lbn), &bp); 4188 if (error == 0) 4189 error = bwrite(bp); 4190 vput(pvp); 4191 if (error != 0) 4192 return (error); 4193 ACQUIRE_LOCK(&lk); 4194 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) 4195 break; 4196 } 4197 FREE_LOCK(&lk); 4198 return (0); 4199 } 4200 4201 /* 4202 * Flush all the dirty bitmaps associated with the block device 4203 * before flushing the rest of the dirty blocks so as to reduce 4204 * the number of dependencies that will have to be rolled back. 4205 */ 4206 static int softdep_fsync_mountdev_bp(struct buf *bp, void *data); 4207 4208 void 4209 softdep_fsync_mountdev(struct vnode *vp) 4210 { 4211 if (!vn_isdisk(vp, NULL)) 4212 panic("softdep_fsync_mountdev: vnode not a disk"); 4213 ACQUIRE_LOCK(&lk); 4214 lwkt_gettoken(&vp->v_token); 4215 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 4216 softdep_fsync_mountdev_bp, vp); 4217 lwkt_reltoken(&vp->v_token); 4218 drain_output(vp, 1); 4219 FREE_LOCK(&lk); 4220 } 4221 4222 static int 4223 softdep_fsync_mountdev_bp(struct buf *bp, void *data) 4224 { 4225 struct worklist *wk; 4226 struct vnode *vp = data; 4227 4228 /* 4229 * If it is already scheduled, skip to the next buffer. 4230 */ 4231 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) 4232 return(0); 4233 if (bp->b_vp != vp || (bp->b_flags & B_DELWRI) == 0) { 4234 BUF_UNLOCK(bp); 4235 kprintf("softdep_fsync_mountdev_bp: warning, buffer %p ripped out from under vnode %p\n", bp, vp); 4236 return(0); 4237 } 4238 /* 4239 * We are only interested in bitmaps with outstanding 4240 * dependencies. 4241 */ 4242 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 4243 wk->wk_type != D_BMSAFEMAP) { 4244 BUF_UNLOCK(bp); 4245 return(0); 4246 } 4247 bremfree(bp); 4248 FREE_LOCK(&lk); 4249 (void) bawrite(bp); 4250 ACQUIRE_LOCK(&lk); 4251 return(0); 4252 } 4253 4254 /* 4255 * This routine is called when we are trying to synchronously flush a 4256 * file. This routine must eliminate any filesystem metadata dependencies 4257 * so that the syncing routine can succeed by pushing the dirty blocks 4258 * associated with the file. If any I/O errors occur, they are returned. 4259 */ 4260 struct softdep_sync_metadata_info { 4261 struct vnode *vp; 4262 int waitfor; 4263 }; 4264 4265 static int softdep_sync_metadata_bp(struct buf *bp, void *data); 4266 4267 int 4268 softdep_sync_metadata(struct vnode *vp, struct thread *td) 4269 { 4270 struct softdep_sync_metadata_info info; 4271 int error, waitfor; 4272 4273 /* 4274 * Check whether this vnode is involved in a filesystem 4275 * that is doing soft dependency processing. 4276 */ 4277 if (!vn_isdisk(vp, NULL)) { 4278 if (!DOINGSOFTDEP(vp)) 4279 return (0); 4280 } else 4281 if (vp->v_rdev->si_mountpoint == NULL || 4282 (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0) 4283 return (0); 4284 /* 4285 * Ensure that any direct block dependencies have been cleared. 4286 */ 4287 ACQUIRE_LOCK(&lk); 4288 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) { 4289 FREE_LOCK(&lk); 4290 return (error); 4291 } 4292 /* 4293 * For most files, the only metadata dependencies are the 4294 * cylinder group maps that allocate their inode or blocks. 4295 * The block allocation dependencies can be found by traversing 4296 * the dependency lists for any buffers that remain on their 4297 * dirty buffer list. The inode allocation dependency will 4298 * be resolved when the inode is updated with MNT_WAIT. 4299 * This work is done in two passes. The first pass grabs most 4300 * of the buffers and begins asynchronously writing them. The 4301 * only way to wait for these asynchronous writes is to sleep 4302 * on the filesystem vnode which may stay busy for a long time 4303 * if the filesystem is active. So, instead, we make a second 4304 * pass over the dependencies blocking on each write. In the 4305 * usual case we will be blocking against a write that we 4306 * initiated, so when it is done the dependency will have been 4307 * resolved. Thus the second pass is expected to end quickly. 4308 */ 4309 waitfor = MNT_NOWAIT; 4310 top: 4311 /* 4312 * We must wait for any I/O in progress to finish so that 4313 * all potential buffers on the dirty list will be visible. 4314 */ 4315 drain_output(vp, 1); 4316 4317 info.vp = vp; 4318 info.waitfor = waitfor; 4319 lwkt_gettoken(&vp->v_token); 4320 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 4321 softdep_sync_metadata_bp, &info); 4322 lwkt_reltoken(&vp->v_token); 4323 if (error < 0) { 4324 FREE_LOCK(&lk); 4325 return(-error); /* error code */ 4326 } 4327 4328 /* 4329 * The brief unlock is to allow any pent up dependency 4330 * processing to be done. Then proceed with the second pass. 4331 */ 4332 if (waitfor == MNT_NOWAIT) { 4333 waitfor = MNT_WAIT; 4334 FREE_LOCK(&lk); 4335 ACQUIRE_LOCK(&lk); 4336 goto top; 4337 } 4338 4339 /* 4340 * If we have managed to get rid of all the dirty buffers, 4341 * then we are done. For certain directories and block 4342 * devices, we may need to do further work. 4343 * 4344 * We must wait for any I/O in progress to finish so that 4345 * all potential buffers on the dirty list will be visible. 4346 */ 4347 drain_output(vp, 1); 4348 if (RB_EMPTY(&vp->v_rbdirty_tree)) { 4349 FREE_LOCK(&lk); 4350 return (0); 4351 } 4352 4353 FREE_LOCK(&lk); 4354 /* 4355 * If we are trying to sync a block device, some of its buffers may 4356 * contain metadata that cannot be written until the contents of some 4357 * partially written files have been written to disk. The only easy 4358 * way to accomplish this is to sync the entire filesystem (luckily 4359 * this happens rarely). 4360 */ 4361 if (vn_isdisk(vp, NULL) && 4362 vp->v_rdev && 4363 vp->v_rdev->si_mountpoint && !vn_islocked(vp) && 4364 (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT)) != 0) 4365 return (error); 4366 return (0); 4367 } 4368 4369 static int 4370 softdep_sync_metadata_bp(struct buf *bp, void *data) 4371 { 4372 struct softdep_sync_metadata_info *info = data; 4373 struct pagedep *pagedep; 4374 struct allocdirect *adp; 4375 struct allocindir *aip; 4376 struct worklist *wk; 4377 struct buf *nbp; 4378 int error; 4379 int i; 4380 4381 if (getdirtybuf(&bp, MNT_WAIT) == 0) { 4382 kprintf("softdep_sync_metadata_bp(1): caught buf %p going away\n", bp); 4383 return (1); 4384 } 4385 if (bp->b_vp != info->vp || (bp->b_flags & B_DELWRI) == 0) { 4386 kprintf("softdep_sync_metadata_bp(2): caught buf %p going away vp %p\n", bp, info->vp); 4387 BUF_UNLOCK(bp); 4388 return(1); 4389 } 4390 4391 /* 4392 * As we hold the buffer locked, none of its dependencies 4393 * will disappear. 4394 */ 4395 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 4396 switch (wk->wk_type) { 4397 4398 case D_ALLOCDIRECT: 4399 adp = WK_ALLOCDIRECT(wk); 4400 if (adp->ad_state & DEPCOMPLETE) 4401 break; 4402 nbp = adp->ad_buf; 4403 if (getdirtybuf(&nbp, info->waitfor) == 0) 4404 break; 4405 FREE_LOCK(&lk); 4406 if (info->waitfor == MNT_NOWAIT) { 4407 bawrite(nbp); 4408 } else if ((error = bwrite(nbp)) != 0) { 4409 bawrite(bp); 4410 ACQUIRE_LOCK(&lk); 4411 return (-error); 4412 } 4413 ACQUIRE_LOCK(&lk); 4414 break; 4415 4416 case D_ALLOCINDIR: 4417 aip = WK_ALLOCINDIR(wk); 4418 if (aip->ai_state & DEPCOMPLETE) 4419 break; 4420 nbp = aip->ai_buf; 4421 if (getdirtybuf(&nbp, info->waitfor) == 0) 4422 break; 4423 FREE_LOCK(&lk); 4424 if (info->waitfor == MNT_NOWAIT) { 4425 bawrite(nbp); 4426 } else if ((error = bwrite(nbp)) != 0) { 4427 bawrite(bp); 4428 ACQUIRE_LOCK(&lk); 4429 return (-error); 4430 } 4431 ACQUIRE_LOCK(&lk); 4432 break; 4433 4434 case D_INDIRDEP: 4435 restart: 4436 4437 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) { 4438 if (aip->ai_state & DEPCOMPLETE) 4439 continue; 4440 nbp = aip->ai_buf; 4441 if (getdirtybuf(&nbp, MNT_WAIT) == 0) 4442 goto restart; 4443 FREE_LOCK(&lk); 4444 if ((error = bwrite(nbp)) != 0) { 4445 bawrite(bp); 4446 ACQUIRE_LOCK(&lk); 4447 return (-error); 4448 } 4449 ACQUIRE_LOCK(&lk); 4450 goto restart; 4451 } 4452 break; 4453 4454 case D_INODEDEP: 4455 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs, 4456 WK_INODEDEP(wk)->id_ino)) != 0) { 4457 FREE_LOCK(&lk); 4458 bawrite(bp); 4459 ACQUIRE_LOCK(&lk); 4460 return (-error); 4461 } 4462 break; 4463 4464 case D_PAGEDEP: 4465 /* 4466 * We are trying to sync a directory that may 4467 * have dependencies on both its own metadata 4468 * and/or dependencies on the inodes of any 4469 * recently allocated files. We walk its diradd 4470 * lists pushing out the associated inode. 4471 */ 4472 pagedep = WK_PAGEDEP(wk); 4473 for (i = 0; i < DAHASHSZ; i++) { 4474 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 4475 continue; 4476 if ((error = 4477 flush_pagedep_deps(info->vp, 4478 pagedep->pd_mnt, 4479 &pagedep->pd_diraddhd[i]))) { 4480 FREE_LOCK(&lk); 4481 bawrite(bp); 4482 ACQUIRE_LOCK(&lk); 4483 return (-error); 4484 } 4485 } 4486 break; 4487 4488 case D_MKDIR: 4489 /* 4490 * This case should never happen if the vnode has 4491 * been properly sync'ed. However, if this function 4492 * is used at a place where the vnode has not yet 4493 * been sync'ed, this dependency can show up. So, 4494 * rather than panic, just flush it. 4495 */ 4496 nbp = WK_MKDIR(wk)->md_buf; 4497 if (getdirtybuf(&nbp, info->waitfor) == 0) 4498 break; 4499 FREE_LOCK(&lk); 4500 if (info->waitfor == MNT_NOWAIT) { 4501 bawrite(nbp); 4502 } else if ((error = bwrite(nbp)) != 0) { 4503 bawrite(bp); 4504 ACQUIRE_LOCK(&lk); 4505 return (-error); 4506 } 4507 ACQUIRE_LOCK(&lk); 4508 break; 4509 4510 case D_BMSAFEMAP: 4511 /* 4512 * This case should never happen if the vnode has 4513 * been properly sync'ed. However, if this function 4514 * is used at a place where the vnode has not yet 4515 * been sync'ed, this dependency can show up. So, 4516 * rather than panic, just flush it. 4517 * 4518 * nbp can wind up == bp if a device node for the 4519 * same filesystem is being fsynced at the same time, 4520 * leading to a panic if we don't catch the case. 4521 */ 4522 nbp = WK_BMSAFEMAP(wk)->sm_buf; 4523 if (nbp == bp) 4524 break; 4525 if (getdirtybuf(&nbp, info->waitfor) == 0) 4526 break; 4527 FREE_LOCK(&lk); 4528 if (info->waitfor == MNT_NOWAIT) { 4529 bawrite(nbp); 4530 } else if ((error = bwrite(nbp)) != 0) { 4531 bawrite(bp); 4532 ACQUIRE_LOCK(&lk); 4533 return (-error); 4534 } 4535 ACQUIRE_LOCK(&lk); 4536 break; 4537 4538 default: 4539 FREE_LOCK(&lk); 4540 panic("softdep_sync_metadata: Unknown type %s", 4541 TYPENAME(wk->wk_type)); 4542 /* NOTREACHED */ 4543 } 4544 } 4545 FREE_LOCK(&lk); 4546 bawrite(bp); 4547 ACQUIRE_LOCK(&lk); 4548 return(0); 4549 } 4550 4551 /* 4552 * Flush the dependencies associated with an inodedep. 4553 * Called with splbio blocked. 4554 */ 4555 static int 4556 flush_inodedep_deps(struct fs *fs, ino_t ino) 4557 { 4558 struct inodedep *inodedep; 4559 struct allocdirect *adp; 4560 int error, waitfor; 4561 struct buf *bp; 4562 4563 /* 4564 * This work is done in two passes. The first pass grabs most 4565 * of the buffers and begins asynchronously writing them. The 4566 * only way to wait for these asynchronous writes is to sleep 4567 * on the filesystem vnode which may stay busy for a long time 4568 * if the filesystem is active. So, instead, we make a second 4569 * pass over the dependencies blocking on each write. In the 4570 * usual case we will be blocking against a write that we 4571 * initiated, so when it is done the dependency will have been 4572 * resolved. Thus the second pass is expected to end quickly. 4573 * We give a brief window at the top of the loop to allow 4574 * any pending I/O to complete. 4575 */ 4576 for (waitfor = MNT_NOWAIT; ; ) { 4577 FREE_LOCK(&lk); 4578 ACQUIRE_LOCK(&lk); 4579 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4580 return (0); 4581 TAILQ_FOREACH(adp, &inodedep->id_inoupdt, ad_next) { 4582 if (adp->ad_state & DEPCOMPLETE) 4583 continue; 4584 bp = adp->ad_buf; 4585 if (getdirtybuf(&bp, waitfor) == 0) { 4586 if (waitfor == MNT_NOWAIT) 4587 continue; 4588 break; 4589 } 4590 FREE_LOCK(&lk); 4591 if (waitfor == MNT_NOWAIT) { 4592 bawrite(bp); 4593 } else if ((error = bwrite(bp)) != 0) { 4594 ACQUIRE_LOCK(&lk); 4595 return (error); 4596 } 4597 ACQUIRE_LOCK(&lk); 4598 break; 4599 } 4600 if (adp != NULL) 4601 continue; 4602 TAILQ_FOREACH(adp, &inodedep->id_newinoupdt, ad_next) { 4603 if (adp->ad_state & DEPCOMPLETE) 4604 continue; 4605 bp = adp->ad_buf; 4606 if (getdirtybuf(&bp, waitfor) == 0) { 4607 if (waitfor == MNT_NOWAIT) 4608 continue; 4609 break; 4610 } 4611 FREE_LOCK(&lk); 4612 if (waitfor == MNT_NOWAIT) { 4613 bawrite(bp); 4614 } else if ((error = bwrite(bp)) != 0) { 4615 ACQUIRE_LOCK(&lk); 4616 return (error); 4617 } 4618 ACQUIRE_LOCK(&lk); 4619 break; 4620 } 4621 if (adp != NULL) 4622 continue; 4623 /* 4624 * If pass2, we are done, otherwise do pass 2. 4625 */ 4626 if (waitfor == MNT_WAIT) 4627 break; 4628 waitfor = MNT_WAIT; 4629 } 4630 /* 4631 * Try freeing inodedep in case all dependencies have been removed. 4632 */ 4633 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0) 4634 (void) free_inodedep(inodedep); 4635 return (0); 4636 } 4637 4638 /* 4639 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 4640 * Called with splbio blocked. 4641 */ 4642 static int 4643 flush_pagedep_deps(struct vnode *pvp, struct mount *mp, 4644 struct diraddhd *diraddhdp) 4645 { 4646 struct inodedep *inodedep; 4647 struct ufsmount *ump; 4648 struct diradd *dap; 4649 struct vnode *vp; 4650 int gotit, error = 0; 4651 struct buf *bp; 4652 ino_t inum; 4653 4654 ump = VFSTOUFS(mp); 4655 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 4656 /* 4657 * Flush ourselves if this directory entry 4658 * has a MKDIR_PARENT dependency. 4659 */ 4660 if (dap->da_state & MKDIR_PARENT) { 4661 FREE_LOCK(&lk); 4662 if ((error = ffs_update(pvp, 1)) != 0) 4663 break; 4664 ACQUIRE_LOCK(&lk); 4665 /* 4666 * If that cleared dependencies, go on to next. 4667 */ 4668 if (dap != LIST_FIRST(diraddhdp)) 4669 continue; 4670 if (dap->da_state & MKDIR_PARENT) { 4671 FREE_LOCK(&lk); 4672 panic("flush_pagedep_deps: MKDIR_PARENT"); 4673 } 4674 } 4675 /* 4676 * A newly allocated directory must have its "." and 4677 * ".." entries written out before its name can be 4678 * committed in its parent. We do not want or need 4679 * the full semantics of a synchronous VOP_FSYNC as 4680 * that may end up here again, once for each directory 4681 * level in the filesystem. Instead, we push the blocks 4682 * and wait for them to clear. We have to fsync twice 4683 * because the first call may choose to defer blocks 4684 * that still have dependencies, but deferral will 4685 * happen at most once. 4686 */ 4687 inum = dap->da_newinum; 4688 if (dap->da_state & MKDIR_BODY) { 4689 FREE_LOCK(&lk); 4690 if ((error = VFS_VGET(mp, NULL, inum, &vp)) != 0) 4691 break; 4692 if ((error=VOP_FSYNC(vp, MNT_NOWAIT, 0)) || 4693 (error=VOP_FSYNC(vp, MNT_NOWAIT, 0))) { 4694 vput(vp); 4695 break; 4696 } 4697 drain_output(vp, 0); 4698 vput(vp); 4699 ACQUIRE_LOCK(&lk); 4700 /* 4701 * If that cleared dependencies, go on to next. 4702 */ 4703 if (dap != LIST_FIRST(diraddhdp)) 4704 continue; 4705 if (dap->da_state & MKDIR_BODY) { 4706 FREE_LOCK(&lk); 4707 panic("flush_pagedep_deps: MKDIR_BODY"); 4708 } 4709 } 4710 /* 4711 * Flush the inode on which the directory entry depends. 4712 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 4713 * the only remaining dependency is that the updated inode 4714 * count must get pushed to disk. The inode has already 4715 * been pushed into its inode buffer (via VOP_UPDATE) at 4716 * the time of the reference count change. So we need only 4717 * locate that buffer, ensure that there will be no rollback 4718 * caused by a bitmap dependency, then write the inode buffer. 4719 */ 4720 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) { 4721 FREE_LOCK(&lk); 4722 panic("flush_pagedep_deps: lost inode"); 4723 } 4724 /* 4725 * If the inode still has bitmap dependencies, 4726 * push them to disk. 4727 */ 4728 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4729 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 4730 FREE_LOCK(&lk); 4731 if (gotit && (error = bwrite(inodedep->id_buf)) != 0) 4732 break; 4733 ACQUIRE_LOCK(&lk); 4734 if (dap != LIST_FIRST(diraddhdp)) 4735 continue; 4736 } 4737 /* 4738 * If the inode is still sitting in a buffer waiting 4739 * to be written, push it to disk. 4740 */ 4741 FREE_LOCK(&lk); 4742 if ((error = bread(ump->um_devvp, 4743 fsbtodoff(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), 4744 (int)ump->um_fs->fs_bsize, &bp)) != 0) 4745 break; 4746 if ((error = bwrite(bp)) != 0) 4747 break; 4748 ACQUIRE_LOCK(&lk); 4749 /* 4750 * If we have failed to get rid of all the dependencies 4751 * then something is seriously wrong. 4752 */ 4753 if (dap == LIST_FIRST(diraddhdp)) { 4754 FREE_LOCK(&lk); 4755 panic("flush_pagedep_deps: flush failed"); 4756 } 4757 } 4758 if (error) 4759 ACQUIRE_LOCK(&lk); 4760 return (error); 4761 } 4762 4763 /* 4764 * A large burst of file addition or deletion activity can drive the 4765 * memory load excessively high. First attempt to slow things down 4766 * using the techniques below. If that fails, this routine requests 4767 * the offending operations to fall back to running synchronously 4768 * until the memory load returns to a reasonable level. 4769 */ 4770 int 4771 softdep_slowdown(struct vnode *vp) 4772 { 4773 int max_softdeps_hard; 4774 4775 max_softdeps_hard = max_softdeps * 11 / 10; 4776 if (num_dirrem < max_softdeps_hard / 2 && 4777 num_inodedep < max_softdeps_hard) 4778 return (0); 4779 stat_sync_limit_hit += 1; 4780 return (1); 4781 } 4782 4783 /* 4784 * If memory utilization has gotten too high, deliberately slow things 4785 * down and speed up the I/O processing. 4786 */ 4787 static int 4788 request_cleanup(int resource, int islocked) 4789 { 4790 struct thread *td = curthread; /* XXX */ 4791 4792 /* 4793 * We never hold up the filesystem syncer process. 4794 */ 4795 if (td == filesys_syncer) 4796 return (0); 4797 /* 4798 * First check to see if the work list has gotten backlogged. 4799 * If it has, co-opt this process to help clean up two entries. 4800 * Because this process may hold inodes locked, we cannot 4801 * handle any remove requests that might block on a locked 4802 * inode as that could lead to deadlock. 4803 */ 4804 if (num_on_worklist > max_softdeps / 10) { 4805 if (islocked) 4806 FREE_LOCK(&lk); 4807 process_worklist_item(NULL, LK_NOWAIT); 4808 process_worklist_item(NULL, LK_NOWAIT); 4809 stat_worklist_push += 2; 4810 if (islocked) 4811 ACQUIRE_LOCK(&lk); 4812 return(1); 4813 } 4814 4815 /* 4816 * If we are resource constrained on inode dependencies, try 4817 * flushing some dirty inodes. Otherwise, we are constrained 4818 * by file deletions, so try accelerating flushes of directories 4819 * with removal dependencies. We would like to do the cleanup 4820 * here, but we probably hold an inode locked at this point and 4821 * that might deadlock against one that we try to clean. So, 4822 * the best that we can do is request the syncer daemon to do 4823 * the cleanup for us. 4824 */ 4825 switch (resource) { 4826 4827 case FLUSH_INODES: 4828 stat_ino_limit_push += 1; 4829 req_clear_inodedeps += 1; 4830 stat_countp = &stat_ino_limit_hit; 4831 break; 4832 4833 case FLUSH_REMOVE: 4834 stat_blk_limit_push += 1; 4835 req_clear_remove += 1; 4836 stat_countp = &stat_blk_limit_hit; 4837 break; 4838 4839 default: 4840 if (islocked) 4841 FREE_LOCK(&lk); 4842 panic("request_cleanup: unknown type"); 4843 } 4844 /* 4845 * Hopefully the syncer daemon will catch up and awaken us. 4846 * We wait at most tickdelay before proceeding in any case. 4847 */ 4848 if (islocked == 0) 4849 ACQUIRE_LOCK(&lk); 4850 proc_waiting += 1; 4851 if (!callout_active(&handle)) 4852 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2, 4853 pause_timer, NULL); 4854 interlocked_sleep(&lk, SLEEP, (caddr_t)&proc_waiting, 0, 4855 "softupdate", 0); 4856 proc_waiting -= 1; 4857 if (islocked == 0) 4858 FREE_LOCK(&lk); 4859 return (1); 4860 } 4861 4862 /* 4863 * Awaken processes pausing in request_cleanup and clear proc_waiting 4864 * to indicate that there is no longer a timer running. 4865 */ 4866 void 4867 pause_timer(void *arg) 4868 { 4869 *stat_countp += 1; 4870 wakeup_one(&proc_waiting); 4871 if (proc_waiting > 0) 4872 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2, 4873 pause_timer, NULL); 4874 else 4875 callout_deactivate(&handle); 4876 } 4877 4878 /* 4879 * Flush out a directory with at least one removal dependency in an effort to 4880 * reduce the number of dirrem, freefile, and freeblks dependency structures. 4881 */ 4882 static void 4883 clear_remove(struct thread *td) 4884 { 4885 struct pagedep_hashhead *pagedephd; 4886 struct pagedep *pagedep; 4887 static int next = 0; 4888 struct mount *mp; 4889 struct vnode *vp; 4890 int error, cnt; 4891 ino_t ino; 4892 4893 ACQUIRE_LOCK(&lk); 4894 for (cnt = 0; cnt < pagedep_hash; cnt++) { 4895 pagedephd = &pagedep_hashtbl[next++]; 4896 if (next >= pagedep_hash) 4897 next = 0; 4898 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 4899 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL) 4900 continue; 4901 mp = pagedep->pd_mnt; 4902 ino = pagedep->pd_ino; 4903 FREE_LOCK(&lk); 4904 if ((error = VFS_VGET(mp, NULL, ino, &vp)) != 0) { 4905 softdep_error("clear_remove: vget", error); 4906 return; 4907 } 4908 if ((error = VOP_FSYNC(vp, MNT_NOWAIT, 0))) 4909 softdep_error("clear_remove: fsync", error); 4910 drain_output(vp, 0); 4911 vput(vp); 4912 return; 4913 } 4914 } 4915 FREE_LOCK(&lk); 4916 } 4917 4918 /* 4919 * Clear out a block of dirty inodes in an effort to reduce 4920 * the number of inodedep dependency structures. 4921 */ 4922 struct clear_inodedeps_info { 4923 struct fs *fs; 4924 struct mount *mp; 4925 }; 4926 4927 static int 4928 clear_inodedeps_mountlist_callback(struct mount *mp, void *data) 4929 { 4930 struct clear_inodedeps_info *info = data; 4931 4932 if ((mp->mnt_flag & MNT_SOFTDEP) && info->fs == VFSTOUFS(mp)->um_fs) { 4933 info->mp = mp; 4934 return(-1); 4935 } 4936 return(0); 4937 } 4938 4939 static void 4940 clear_inodedeps(struct thread *td) 4941 { 4942 struct clear_inodedeps_info info; 4943 struct inodedep_hashhead *inodedephd; 4944 struct inodedep *inodedep; 4945 static int next = 0; 4946 struct vnode *vp; 4947 struct fs *fs; 4948 int error, cnt; 4949 ino_t firstino, lastino, ino; 4950 4951 ACQUIRE_LOCK(&lk); 4952 /* 4953 * Pick a random inode dependency to be cleared. 4954 * We will then gather up all the inodes in its block 4955 * that have dependencies and flush them out. 4956 */ 4957 for (cnt = 0; cnt < inodedep_hash; cnt++) { 4958 inodedephd = &inodedep_hashtbl[next++]; 4959 if (next >= inodedep_hash) 4960 next = 0; 4961 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 4962 break; 4963 } 4964 if (inodedep == NULL) { 4965 FREE_LOCK(&lk); 4966 return; 4967 } 4968 /* 4969 * Ugly code to find mount point given pointer to superblock. 4970 */ 4971 fs = inodedep->id_fs; 4972 info.mp = NULL; 4973 info.fs = fs; 4974 mountlist_scan(clear_inodedeps_mountlist_callback, 4975 &info, MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 4976 /* 4977 * Find the last inode in the block with dependencies. 4978 */ 4979 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 4980 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 4981 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0) 4982 break; 4983 /* 4984 * Asynchronously push all but the last inode with dependencies. 4985 * Synchronously push the last inode with dependencies to ensure 4986 * that the inode block gets written to free up the inodedeps. 4987 */ 4988 for (ino = firstino; ino <= lastino; ino++) { 4989 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4990 continue; 4991 FREE_LOCK(&lk); 4992 if ((error = VFS_VGET(info.mp, NULL, ino, &vp)) != 0) { 4993 softdep_error("clear_inodedeps: vget", error); 4994 return; 4995 } 4996 if (ino == lastino) { 4997 if ((error = VOP_FSYNC(vp, MNT_WAIT, 0))) 4998 softdep_error("clear_inodedeps: fsync1", error); 4999 } else { 5000 if ((error = VOP_FSYNC(vp, MNT_NOWAIT, 0))) 5001 softdep_error("clear_inodedeps: fsync2", error); 5002 drain_output(vp, 0); 5003 } 5004 vput(vp); 5005 ACQUIRE_LOCK(&lk); 5006 } 5007 FREE_LOCK(&lk); 5008 } 5009 5010 /* 5011 * Function to determine if the buffer has outstanding dependencies 5012 * that will cause a roll-back if the buffer is written. If wantcount 5013 * is set, return number of dependencies, otherwise just yes or no. 5014 */ 5015 static int 5016 softdep_count_dependencies(struct buf *bp, int wantcount) 5017 { 5018 struct worklist *wk; 5019 struct inodedep *inodedep; 5020 struct indirdep *indirdep; 5021 struct allocindir *aip; 5022 struct pagedep *pagedep; 5023 struct diradd *dap; 5024 int i, retval; 5025 5026 retval = 0; 5027 ACQUIRE_LOCK(&lk); 5028 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5029 switch (wk->wk_type) { 5030 5031 case D_INODEDEP: 5032 inodedep = WK_INODEDEP(wk); 5033 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 5034 /* bitmap allocation dependency */ 5035 retval += 1; 5036 if (!wantcount) 5037 goto out; 5038 } 5039 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 5040 /* direct block pointer dependency */ 5041 retval += 1; 5042 if (!wantcount) 5043 goto out; 5044 } 5045 continue; 5046 5047 case D_INDIRDEP: 5048 indirdep = WK_INDIRDEP(wk); 5049 5050 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 5051 /* indirect block pointer dependency */ 5052 retval += 1; 5053 if (!wantcount) 5054 goto out; 5055 } 5056 continue; 5057 5058 case D_PAGEDEP: 5059 pagedep = WK_PAGEDEP(wk); 5060 for (i = 0; i < DAHASHSZ; i++) { 5061 5062 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 5063 /* directory entry dependency */ 5064 retval += 1; 5065 if (!wantcount) 5066 goto out; 5067 } 5068 } 5069 continue; 5070 5071 case D_BMSAFEMAP: 5072 case D_ALLOCDIRECT: 5073 case D_ALLOCINDIR: 5074 case D_MKDIR: 5075 /* never a dependency on these blocks */ 5076 continue; 5077 5078 default: 5079 FREE_LOCK(&lk); 5080 panic("softdep_check_for_rollback: Unexpected type %s", 5081 TYPENAME(wk->wk_type)); 5082 /* NOTREACHED */ 5083 } 5084 } 5085 out: 5086 FREE_LOCK(&lk); 5087 return retval; 5088 } 5089 5090 /* 5091 * Acquire exclusive access to a buffer. 5092 * Must be called with splbio blocked. 5093 * Return 1 if buffer was acquired. 5094 */ 5095 static int 5096 getdirtybuf(struct buf **bpp, int waitfor) 5097 { 5098 struct buf *bp; 5099 int error; 5100 5101 for (;;) { 5102 if ((bp = *bpp) == NULL) 5103 return (0); 5104 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) 5105 break; 5106 if (waitfor != MNT_WAIT) 5107 return (0); 5108 error = interlocked_sleep(&lk, LOCKBUF, bp, 5109 LK_EXCLUSIVE | LK_SLEEPFAIL, 0, 0); 5110 if (error != ENOLCK) { 5111 FREE_LOCK(&lk); 5112 panic("getdirtybuf: inconsistent lock"); 5113 } 5114 } 5115 if ((bp->b_flags & B_DELWRI) == 0) { 5116 BUF_UNLOCK(bp); 5117 return (0); 5118 } 5119 bremfree(bp); 5120 return (1); 5121 } 5122 5123 /* 5124 * Wait for pending output on a vnode to complete. 5125 * Must be called with vnode locked. 5126 */ 5127 static void 5128 drain_output(struct vnode *vp, int islocked) 5129 { 5130 5131 if (!islocked) 5132 ACQUIRE_LOCK(&lk); 5133 while (bio_track_active(&vp->v_track_write)) { 5134 FREE_LOCK(&lk); 5135 bio_track_wait(&vp->v_track_write, 0, 0); 5136 ACQUIRE_LOCK(&lk); 5137 } 5138 if (!islocked) 5139 FREE_LOCK(&lk); 5140 } 5141 5142 /* 5143 * Called whenever a buffer that is being invalidated or reallocated 5144 * contains dependencies. This should only happen if an I/O error has 5145 * occurred. The routine is called with the buffer locked. 5146 */ 5147 static void 5148 softdep_deallocate_dependencies(struct buf *bp) 5149 { 5150 if ((bp->b_flags & B_ERROR) == 0) 5151 panic("softdep_deallocate_dependencies: dangling deps"); 5152 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntfromname, bp->b_error); 5153 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 5154 } 5155 5156 /* 5157 * Function to handle asynchronous write errors in the filesystem. 5158 */ 5159 void 5160 softdep_error(char *func, int error) 5161 { 5162 5163 /* XXX should do something better! */ 5164 kprintf("%s: got error %d while accessing filesystem\n", func, error); 5165 } 5166