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 1.55 2008/01/05 14:02:41 swildner Exp $ 41 */ 42 43 /* 44 * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide. 45 */ 46 #ifndef DIAGNOSTIC 47 #define DIAGNOSTIC 48 #endif 49 #ifndef DEBUG 50 #define DEBUG 51 #endif 52 53 #include <sys/param.h> 54 #include <sys/kernel.h> 55 #include <sys/systm.h> 56 #include <sys/buf.h> 57 #include <sys/malloc.h> 58 #include <sys/mount.h> 59 #include <sys/proc.h> 60 #include <sys/syslog.h> 61 #include <sys/vnode.h> 62 #include <sys/conf.h> 63 #include <sys/buf2.h> 64 #include <machine/inttypes.h> 65 #include "dir.h" 66 #include "quota.h" 67 #include "inode.h" 68 #include "ufsmount.h" 69 #include "fs.h" 70 #include "softdep.h" 71 #include "ffs_extern.h" 72 #include "ufs_extern.h" 73 74 #include <sys/thread2.h> 75 76 /* 77 * These definitions need to be adapted to the system to which 78 * this file is being ported. 79 */ 80 /* 81 * malloc types defined for the softdep system. 82 */ 83 MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies"); 84 MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies"); 85 MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation"); 86 MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map"); 87 MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode"); 88 MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies"); 89 MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block"); 90 MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode"); 91 MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode"); 92 MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated"); 93 MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry"); 94 MALLOC_DEFINE(M_MKDIR, "mkdir","New directory"); 95 MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted"); 96 97 #define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE) 98 99 #define D_PAGEDEP 0 100 #define D_INODEDEP 1 101 #define D_NEWBLK 2 102 #define D_BMSAFEMAP 3 103 #define D_ALLOCDIRECT 4 104 #define D_INDIRDEP 5 105 #define D_ALLOCINDIR 6 106 #define D_FREEFRAG 7 107 #define D_FREEBLKS 8 108 #define D_FREEFILE 9 109 #define D_DIRADD 10 110 #define D_MKDIR 11 111 #define D_DIRREM 12 112 #define D_LAST D_DIRREM 113 114 /* 115 * translate from workitem type to memory type 116 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 117 */ 118 static struct malloc_type *memtype[] = { 119 M_PAGEDEP, 120 M_INODEDEP, 121 M_NEWBLK, 122 M_BMSAFEMAP, 123 M_ALLOCDIRECT, 124 M_INDIRDEP, 125 M_ALLOCINDIR, 126 M_FREEFRAG, 127 M_FREEBLKS, 128 M_FREEFILE, 129 M_DIRADD, 130 M_MKDIR, 131 M_DIRREM 132 }; 133 134 #define DtoM(type) (memtype[type]) 135 136 /* 137 * Names of malloc types. 138 */ 139 #define TYPENAME(type) \ 140 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???") 141 /* 142 * End system adaptaion definitions. 143 */ 144 145 /* 146 * Internal function prototypes. 147 */ 148 static void softdep_error(char *, int); 149 static void drain_output(struct vnode *, int); 150 static int getdirtybuf(struct buf **, int); 151 static void clear_remove(struct thread *); 152 static void clear_inodedeps(struct thread *); 153 static int flush_pagedep_deps(struct vnode *, struct mount *, 154 struct diraddhd *); 155 static int flush_inodedep_deps(struct fs *, ino_t); 156 static int handle_written_filepage(struct pagedep *, struct buf *); 157 static void diradd_inode_written(struct diradd *, struct inodedep *); 158 static int handle_written_inodeblock(struct inodedep *, struct buf *); 159 static void handle_allocdirect_partdone(struct allocdirect *); 160 static void handle_allocindir_partdone(struct allocindir *); 161 static void initiate_write_filepage(struct pagedep *, struct buf *); 162 static void handle_written_mkdir(struct mkdir *, int); 163 static void initiate_write_inodeblock(struct inodedep *, struct buf *); 164 static void handle_workitem_freefile(struct freefile *); 165 static void handle_workitem_remove(struct dirrem *); 166 static struct dirrem *newdirrem(struct buf *, struct inode *, 167 struct inode *, int, struct dirrem **); 168 static void free_diradd(struct diradd *); 169 static void free_allocindir(struct allocindir *, struct inodedep *); 170 static int indir_trunc (struct inode *, off_t, int, ufs_lbn_t, long *); 171 static void deallocate_dependencies(struct buf *, struct inodedep *); 172 static void free_allocdirect(struct allocdirectlst *, 173 struct allocdirect *, int); 174 static int check_inode_unwritten(struct inodedep *); 175 static int free_inodedep(struct inodedep *); 176 static void handle_workitem_freeblocks(struct freeblks *); 177 static void merge_inode_lists(struct inodedep *); 178 static void setup_allocindir_phase2(struct buf *, struct inode *, 179 struct allocindir *); 180 static struct allocindir *newallocindir(struct inode *, int, ufs_daddr_t, 181 ufs_daddr_t); 182 static void handle_workitem_freefrag(struct freefrag *); 183 static struct freefrag *newfreefrag(struct inode *, ufs_daddr_t, long); 184 static void allocdirect_merge(struct allocdirectlst *, 185 struct allocdirect *, struct allocdirect *); 186 static struct bmsafemap *bmsafemap_lookup(struct buf *); 187 static int newblk_lookup(struct fs *, ufs_daddr_t, int, 188 struct newblk **); 189 static int inodedep_lookup(struct fs *, ino_t, int, struct inodedep **); 190 static int pagedep_lookup(struct inode *, ufs_lbn_t, int, 191 struct pagedep **); 192 static void pause_timer(void *); 193 static int request_cleanup(int, int); 194 static int process_worklist_item(struct mount *, int); 195 static void add_to_worklist(struct worklist *); 196 197 /* 198 * Exported softdep operations. 199 */ 200 static void softdep_disk_io_initiation(struct buf *); 201 static void softdep_disk_write_complete(struct buf *); 202 static void softdep_deallocate_dependencies(struct buf *); 203 static int softdep_fsync(struct vnode *); 204 static int softdep_process_worklist(struct mount *); 205 static void softdep_move_dependencies(struct buf *, struct buf *); 206 static int softdep_count_dependencies(struct buf *bp, int); 207 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 bwillwrite(); 652 /* 653 * Never allow processing to run for more than one 654 * second. Otherwise the other syncer tasks may get 655 * excessively backlogged. 656 */ 657 if (starttime != time_second && matchmnt == NULL) { 658 matchcnt = -1; 659 break; 660 } 661 } 662 if (matchmnt == NULL) { 663 --softdep_worklist_busy; 664 if (softdep_worklist_req && softdep_worklist_busy == 0) 665 wakeup(&softdep_worklist_req); 666 } 667 return (matchcnt); 668 } 669 670 /* 671 * Process one item on the worklist. 672 */ 673 static int 674 process_worklist_item(struct mount *matchmnt, int flags) 675 { 676 struct worklist *wk; 677 struct dirrem *dirrem; 678 struct fs *matchfs; 679 struct vnode *vp; 680 int matchcnt = 0; 681 682 matchfs = NULL; 683 if (matchmnt != NULL) 684 matchfs = VFSTOUFS(matchmnt)->um_fs; 685 ACQUIRE_LOCK(&lk); 686 /* 687 * Normally we just process each item on the worklist in order. 688 * However, if we are in a situation where we cannot lock any 689 * inodes, we have to skip over any dirrem requests whose 690 * vnodes are resident and locked. 691 */ 692 LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) { 693 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM) 694 break; 695 dirrem = WK_DIRREM(wk); 696 vp = ufs_ihashlookup(VFSTOUFS(dirrem->dm_mnt)->um_dev, 697 dirrem->dm_oldinum); 698 if (vp == NULL || !vn_islocked(vp)) 699 break; 700 } 701 if (wk == 0) { 702 FREE_LOCK(&lk); 703 return (0); 704 } 705 WORKLIST_REMOVE(wk); 706 num_on_worklist -= 1; 707 FREE_LOCK(&lk); 708 switch (wk->wk_type) { 709 710 case D_DIRREM: 711 /* removal of a directory entry */ 712 if (WK_DIRREM(wk)->dm_mnt == matchmnt) 713 matchcnt += 1; 714 handle_workitem_remove(WK_DIRREM(wk)); 715 break; 716 717 case D_FREEBLKS: 718 /* releasing blocks and/or fragments from a file */ 719 if (WK_FREEBLKS(wk)->fb_fs == matchfs) 720 matchcnt += 1; 721 handle_workitem_freeblocks(WK_FREEBLKS(wk)); 722 break; 723 724 case D_FREEFRAG: 725 /* releasing a fragment when replaced as a file grows */ 726 if (WK_FREEFRAG(wk)->ff_fs == matchfs) 727 matchcnt += 1; 728 handle_workitem_freefrag(WK_FREEFRAG(wk)); 729 break; 730 731 case D_FREEFILE: 732 /* releasing an inode when its link count drops to 0 */ 733 if (WK_FREEFILE(wk)->fx_fs == matchfs) 734 matchcnt += 1; 735 handle_workitem_freefile(WK_FREEFILE(wk)); 736 break; 737 738 default: 739 panic("%s_process_worklist: Unknown type %s", 740 "softdep", TYPENAME(wk->wk_type)); 741 /* NOTREACHED */ 742 } 743 return (matchcnt); 744 } 745 746 /* 747 * Move dependencies from one buffer to another. 748 */ 749 static void 750 softdep_move_dependencies(struct buf *oldbp, struct buf *newbp) 751 { 752 struct worklist *wk, *wktail; 753 754 if (LIST_FIRST(&newbp->b_dep) != NULL) 755 panic("softdep_move_dependencies: need merge code"); 756 wktail = NULL; 757 ACQUIRE_LOCK(&lk); 758 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 759 LIST_REMOVE(wk, wk_list); 760 if (wktail == NULL) 761 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 762 else 763 LIST_INSERT_AFTER(wktail, wk, wk_list); 764 wktail = wk; 765 newbp->b_ops = &softdep_bioops; 766 } 767 FREE_LOCK(&lk); 768 } 769 770 /* 771 * Purge the work list of all items associated with a particular mount point. 772 */ 773 int 774 softdep_flushfiles(struct mount *oldmnt, int flags) 775 { 776 struct vnode *devvp; 777 int error, loopcnt; 778 779 /* 780 * Await our turn to clear out the queue, then serialize access. 781 */ 782 while (softdep_worklist_busy != 0) { 783 softdep_worklist_req += 1; 784 tsleep(&softdep_worklist_req, 0, "softflush", 0); 785 softdep_worklist_req -= 1; 786 } 787 softdep_worklist_busy = -1; 788 789 if ((error = ffs_flushfiles(oldmnt, flags)) != 0) { 790 softdep_worklist_busy = 0; 791 if (softdep_worklist_req) 792 wakeup(&softdep_worklist_req); 793 return (error); 794 } 795 /* 796 * Alternately flush the block device associated with the mount 797 * point and process any dependencies that the flushing 798 * creates. In theory, this loop can happen at most twice, 799 * but we give it a few extra just to be sure. 800 */ 801 devvp = VFSTOUFS(oldmnt)->um_devvp; 802 for (loopcnt = 10; loopcnt > 0; ) { 803 if (softdep_process_worklist(oldmnt) == 0) { 804 loopcnt--; 805 /* 806 * Do another flush in case any vnodes were brought in 807 * as part of the cleanup operations. 808 */ 809 if ((error = ffs_flushfiles(oldmnt, flags)) != 0) 810 break; 811 /* 812 * If we still found nothing to do, we are really done. 813 */ 814 if (softdep_process_worklist(oldmnt) == 0) 815 break; 816 } 817 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 818 error = VOP_FSYNC(devvp, MNT_WAIT); 819 vn_unlock(devvp); 820 if (error) 821 break; 822 } 823 softdep_worklist_busy = 0; 824 if (softdep_worklist_req) 825 wakeup(&softdep_worklist_req); 826 827 /* 828 * If we are unmounting then it is an error to fail. If we 829 * are simply trying to downgrade to read-only, then filesystem 830 * activity can keep us busy forever, so we just fail with EBUSY. 831 */ 832 if (loopcnt == 0) { 833 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 834 panic("softdep_flushfiles: looping"); 835 error = EBUSY; 836 } 837 return (error); 838 } 839 840 /* 841 * Structure hashing. 842 * 843 * There are three types of structures that can be looked up: 844 * 1) pagedep structures identified by mount point, inode number, 845 * and logical block. 846 * 2) inodedep structures identified by mount point and inode number. 847 * 3) newblk structures identified by mount point and 848 * physical block number. 849 * 850 * The "pagedep" and "inodedep" dependency structures are hashed 851 * separately from the file blocks and inodes to which they correspond. 852 * This separation helps when the in-memory copy of an inode or 853 * file block must be replaced. It also obviates the need to access 854 * an inode or file page when simply updating (or de-allocating) 855 * dependency structures. Lookup of newblk structures is needed to 856 * find newly allocated blocks when trying to associate them with 857 * their allocdirect or allocindir structure. 858 * 859 * The lookup routines optionally create and hash a new instance when 860 * an existing entry is not found. 861 */ 862 #define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 863 #define NODELAY 0x0002 /* cannot do background work */ 864 865 /* 866 * Structures and routines associated with pagedep caching. 867 */ 868 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl; 869 u_long pagedep_hash; /* size of hash table - 1 */ 870 #define PAGEDEP_HASH(mp, inum, lbn) \ 871 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \ 872 pagedep_hash]) 873 static struct sema pagedep_in_progress; 874 875 /* 876 * Helper routine for pagedep_lookup() 877 */ 878 static __inline 879 struct pagedep * 880 pagedep_find(struct pagedep_hashhead *pagedephd, ino_t ino, ufs_lbn_t lbn, 881 struct mount *mp) 882 { 883 struct pagedep *pagedep; 884 885 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 886 if (ino == pagedep->pd_ino && 887 lbn == pagedep->pd_lbn && 888 mp == pagedep->pd_mnt) { 889 return (pagedep); 890 } 891 } 892 return(NULL); 893 } 894 895 /* 896 * Look up a pagedep. Return 1 if found, 0 if not found. 897 * If not found, allocate if DEPALLOC flag is passed. 898 * Found or allocated entry is returned in pagedeppp. 899 * This routine must be called with splbio interrupts blocked. 900 */ 901 static int 902 pagedep_lookup(struct inode *ip, ufs_lbn_t lbn, int flags, 903 struct pagedep **pagedeppp) 904 { 905 struct pagedep *pagedep; 906 struct pagedep_hashhead *pagedephd; 907 struct mount *mp; 908 int i; 909 910 #ifdef DEBUG 911 if (lk.lkt_held == NOHOLDER) 912 panic("pagedep_lookup: lock not held"); 913 #endif 914 mp = ITOV(ip)->v_mount; 915 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn); 916 top: 917 *pagedeppp = pagedep_find(pagedephd, ip->i_number, lbn, mp); 918 if (*pagedeppp) 919 return(1); 920 if ((flags & DEPALLOC) == 0) 921 return (0); 922 if (sema_get(&pagedep_in_progress, &lk) == 0) { 923 ACQUIRE_LOCK(&lk); 924 goto top; 925 } 926 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP, 927 M_SOFTDEP_FLAGS | M_ZERO); 928 929 if (pagedep_find(pagedephd, ip->i_number, lbn, mp)) { 930 kprintf("pagedep_lookup: blocking race avoided\n"); 931 ACQUIRE_LOCK(&lk); 932 sema_release(&pagedep_in_progress); 933 kfree(pagedep, M_PAGEDEP); 934 goto top; 935 } 936 937 pagedep->pd_list.wk_type = D_PAGEDEP; 938 pagedep->pd_mnt = mp; 939 pagedep->pd_ino = ip->i_number; 940 pagedep->pd_lbn = lbn; 941 LIST_INIT(&pagedep->pd_dirremhd); 942 LIST_INIT(&pagedep->pd_pendinghd); 943 for (i = 0; i < DAHASHSZ; i++) 944 LIST_INIT(&pagedep->pd_diraddhd[i]); 945 ACQUIRE_LOCK(&lk); 946 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 947 sema_release(&pagedep_in_progress); 948 *pagedeppp = pagedep; 949 return (0); 950 } 951 952 /* 953 * Structures and routines associated with inodedep caching. 954 */ 955 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl; 956 static u_long inodedep_hash; /* size of hash table - 1 */ 957 static long num_inodedep; /* number of inodedep allocated */ 958 #define INODEDEP_HASH(fs, inum) \ 959 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash]) 960 static struct sema inodedep_in_progress; 961 962 /* 963 * Helper routine for inodedep_lookup() 964 */ 965 static __inline 966 struct inodedep * 967 inodedep_find(struct inodedep_hashhead *inodedephd, struct fs *fs, ino_t inum) 968 { 969 struct inodedep *inodedep; 970 971 LIST_FOREACH(inodedep, inodedephd, id_hash) { 972 if (inum == inodedep->id_ino && fs == inodedep->id_fs) 973 return(inodedep); 974 } 975 return (NULL); 976 } 977 978 /* 979 * Look up a inodedep. Return 1 if found, 0 if not found. 980 * If not found, allocate if DEPALLOC flag is passed. 981 * Found or allocated entry is returned in inodedeppp. 982 * This routine must be called with splbio interrupts blocked. 983 */ 984 static int 985 inodedep_lookup(struct fs *fs, ino_t inum, int flags, 986 struct inodedep **inodedeppp) 987 { 988 struct inodedep *inodedep; 989 struct inodedep_hashhead *inodedephd; 990 int firsttry; 991 992 #ifdef DEBUG 993 if (lk.lkt_held == NOHOLDER) 994 panic("inodedep_lookup: lock not held"); 995 #endif 996 firsttry = 1; 997 inodedephd = INODEDEP_HASH(fs, inum); 998 top: 999 *inodedeppp = inodedep_find(inodedephd, fs, inum); 1000 if (*inodedeppp) 1001 return (1); 1002 if ((flags & DEPALLOC) == 0) 1003 return (0); 1004 /* 1005 * If we are over our limit, try to improve the situation. 1006 */ 1007 if (num_inodedep > max_softdeps && firsttry && 1008 speedup_syncer() == 0 && (flags & NODELAY) == 0 && 1009 request_cleanup(FLUSH_INODES, 1)) { 1010 firsttry = 0; 1011 goto top; 1012 } 1013 if (sema_get(&inodedep_in_progress, &lk) == 0) { 1014 ACQUIRE_LOCK(&lk); 1015 goto top; 1016 } 1017 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep), 1018 M_INODEDEP, M_SOFTDEP_FLAGS | M_ZERO); 1019 if (inodedep_find(inodedephd, fs, inum)) { 1020 kprintf("inodedep_lookup: blocking race avoided\n"); 1021 ACQUIRE_LOCK(&lk); 1022 sema_release(&inodedep_in_progress); 1023 kfree(inodedep, M_INODEDEP); 1024 goto top; 1025 } 1026 inodedep->id_list.wk_type = D_INODEDEP; 1027 inodedep->id_fs = fs; 1028 inodedep->id_ino = inum; 1029 inodedep->id_state = ALLCOMPLETE; 1030 inodedep->id_nlinkdelta = 0; 1031 inodedep->id_savedino = NULL; 1032 inodedep->id_savedsize = -1; 1033 inodedep->id_buf = NULL; 1034 LIST_INIT(&inodedep->id_pendinghd); 1035 LIST_INIT(&inodedep->id_inowait); 1036 LIST_INIT(&inodedep->id_bufwait); 1037 TAILQ_INIT(&inodedep->id_inoupdt); 1038 TAILQ_INIT(&inodedep->id_newinoupdt); 1039 ACQUIRE_LOCK(&lk); 1040 num_inodedep += 1; 1041 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 1042 sema_release(&inodedep_in_progress); 1043 *inodedeppp = inodedep; 1044 return (0); 1045 } 1046 1047 /* 1048 * Structures and routines associated with newblk caching. 1049 */ 1050 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl; 1051 u_long newblk_hash; /* size of hash table - 1 */ 1052 #define NEWBLK_HASH(fs, inum) \ 1053 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash]) 1054 static struct sema newblk_in_progress; 1055 1056 /* 1057 * Helper routine for newblk_lookup() 1058 */ 1059 static __inline 1060 struct newblk * 1061 newblk_find(struct newblk_hashhead *newblkhd, struct fs *fs, 1062 ufs_daddr_t newblkno) 1063 { 1064 struct newblk *newblk; 1065 1066 LIST_FOREACH(newblk, newblkhd, nb_hash) { 1067 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs) 1068 return (newblk); 1069 } 1070 return(NULL); 1071 } 1072 1073 /* 1074 * Look up a newblk. Return 1 if found, 0 if not found. 1075 * If not found, allocate if DEPALLOC flag is passed. 1076 * Found or allocated entry is returned in newblkpp. 1077 */ 1078 static int 1079 newblk_lookup(struct fs *fs, ufs_daddr_t newblkno, int flags, 1080 struct newblk **newblkpp) 1081 { 1082 struct newblk *newblk; 1083 struct newblk_hashhead *newblkhd; 1084 1085 newblkhd = NEWBLK_HASH(fs, newblkno); 1086 top: 1087 *newblkpp = newblk_find(newblkhd, fs, newblkno); 1088 if (*newblkpp) 1089 return(1); 1090 if ((flags & DEPALLOC) == 0) 1091 return (0); 1092 if (sema_get(&newblk_in_progress, 0) == 0) 1093 goto top; 1094 MALLOC(newblk, struct newblk *, sizeof(struct newblk), 1095 M_NEWBLK, M_SOFTDEP_FLAGS | M_ZERO); 1096 1097 if (newblk_find(newblkhd, fs, newblkno)) { 1098 kprintf("newblk_lookup: blocking race avoided\n"); 1099 sema_release(&pagedep_in_progress); 1100 kfree(newblk, M_NEWBLK); 1101 goto top; 1102 } 1103 newblk->nb_state = 0; 1104 newblk->nb_fs = fs; 1105 newblk->nb_newblkno = newblkno; 1106 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 1107 sema_release(&newblk_in_progress); 1108 *newblkpp = newblk; 1109 return (0); 1110 } 1111 1112 /* 1113 * Executed during filesystem system initialization before 1114 * mounting any filesystems. 1115 */ 1116 void 1117 softdep_initialize(void) 1118 { 1119 callout_init(&handle); 1120 1121 LIST_INIT(&mkdirlisthd); 1122 LIST_INIT(&softdep_workitem_pending); 1123 max_softdeps = min(desiredvnodes * 8, 1124 M_INODEDEP->ks_limit / (2 * sizeof(struct inodedep))); 1125 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 1126 &pagedep_hash); 1127 sema_init(&pagedep_in_progress, "pagedep", 0, 0); 1128 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash); 1129 sema_init(&inodedep_in_progress, "inodedep", 0, 0); 1130 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash); 1131 sema_init(&newblk_in_progress, "newblk", 0, 0); 1132 add_bio_ops(&softdep_bioops); 1133 } 1134 1135 /* 1136 * Called at mount time to notify the dependency code that a 1137 * filesystem wishes to use it. 1138 */ 1139 int 1140 softdep_mount(struct vnode *devvp, struct mount *mp, struct fs *fs) 1141 { 1142 struct csum cstotal; 1143 struct cg *cgp; 1144 struct buf *bp; 1145 int error, cyl; 1146 1147 mp->mnt_flag &= ~MNT_ASYNC; 1148 mp->mnt_flag |= MNT_SOFTDEP; 1149 mp->mnt_bioops = &softdep_bioops; 1150 /* 1151 * When doing soft updates, the counters in the 1152 * superblock may have gotten out of sync, so we have 1153 * to scan the cylinder groups and recalculate them. 1154 */ 1155 if (fs->fs_clean != 0) 1156 return (0); 1157 bzero(&cstotal, sizeof cstotal); 1158 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 1159 if ((error = bread(devvp, fsbtodoff(fs, cgtod(fs, cyl)), 1160 fs->fs_cgsize, &bp)) != 0) { 1161 brelse(bp); 1162 return (error); 1163 } 1164 cgp = (struct cg *)bp->b_data; 1165 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 1166 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 1167 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 1168 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 1169 fs->fs_cs(fs, cyl) = cgp->cg_cs; 1170 brelse(bp); 1171 } 1172 #ifdef DEBUG 1173 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 1174 kprintf("ffs_mountfs: superblock updated for soft updates\n"); 1175 #endif 1176 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 1177 return (0); 1178 } 1179 1180 /* 1181 * Protecting the freemaps (or bitmaps). 1182 * 1183 * To eliminate the need to execute fsck before mounting a filesystem 1184 * after a power failure, one must (conservatively) guarantee that the 1185 * on-disk copy of the bitmaps never indicate that a live inode or block is 1186 * free. So, when a block or inode is allocated, the bitmap should be 1187 * updated (on disk) before any new pointers. When a block or inode is 1188 * freed, the bitmap should not be updated until all pointers have been 1189 * reset. The latter dependency is handled by the delayed de-allocation 1190 * approach described below for block and inode de-allocation. The former 1191 * dependency is handled by calling the following procedure when a block or 1192 * inode is allocated. When an inode is allocated an "inodedep" is created 1193 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 1194 * Each "inodedep" is also inserted into the hash indexing structure so 1195 * that any additional link additions can be made dependent on the inode 1196 * allocation. 1197 * 1198 * The ufs filesystem maintains a number of free block counts (e.g., per 1199 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 1200 * in addition to the bitmaps. These counts are used to improve efficiency 1201 * during allocation and therefore must be consistent with the bitmaps. 1202 * There is no convenient way to guarantee post-crash consistency of these 1203 * counts with simple update ordering, for two main reasons: (1) The counts 1204 * and bitmaps for a single cylinder group block are not in the same disk 1205 * sector. If a disk write is interrupted (e.g., by power failure), one may 1206 * be written and the other not. (2) Some of the counts are located in the 1207 * superblock rather than the cylinder group block. So, we focus our soft 1208 * updates implementation on protecting the bitmaps. When mounting a 1209 * filesystem, we recompute the auxiliary counts from the bitmaps. 1210 */ 1211 1212 /* 1213 * Called just after updating the cylinder group block to allocate an inode. 1214 * 1215 * Parameters: 1216 * bp: buffer for cylgroup block with inode map 1217 * ip: inode related to allocation 1218 * newinum: new inode number being allocated 1219 */ 1220 void 1221 softdep_setup_inomapdep(struct buf *bp, struct inode *ip, ino_t newinum) 1222 { 1223 struct inodedep *inodedep; 1224 struct bmsafemap *bmsafemap; 1225 1226 /* 1227 * Create a dependency for the newly allocated inode. 1228 * Panic if it already exists as something is seriously wrong. 1229 * Otherwise add it to the dependency list for the buffer holding 1230 * the cylinder group map from which it was allocated. 1231 */ 1232 ACQUIRE_LOCK(&lk); 1233 if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) { 1234 FREE_LOCK(&lk); 1235 panic("softdep_setup_inomapdep: found inode"); 1236 } 1237 inodedep->id_buf = bp; 1238 inodedep->id_state &= ~DEPCOMPLETE; 1239 bmsafemap = bmsafemap_lookup(bp); 1240 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 1241 FREE_LOCK(&lk); 1242 } 1243 1244 /* 1245 * Called just after updating the cylinder group block to 1246 * allocate block or fragment. 1247 * 1248 * Parameters: 1249 * bp: buffer for cylgroup block with block map 1250 * fs: filesystem doing allocation 1251 * newblkno: number of newly allocated block 1252 */ 1253 void 1254 softdep_setup_blkmapdep(struct buf *bp, struct fs *fs, 1255 ufs_daddr_t newblkno) 1256 { 1257 struct newblk *newblk; 1258 struct bmsafemap *bmsafemap; 1259 1260 /* 1261 * Create a dependency for the newly allocated block. 1262 * Add it to the dependency list for the buffer holding 1263 * the cylinder group map from which it was allocated. 1264 */ 1265 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0) 1266 panic("softdep_setup_blkmapdep: found block"); 1267 ACQUIRE_LOCK(&lk); 1268 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp); 1269 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 1270 FREE_LOCK(&lk); 1271 } 1272 1273 /* 1274 * Find the bmsafemap associated with a cylinder group buffer. 1275 * If none exists, create one. The buffer must be locked when 1276 * this routine is called and this routine must be called with 1277 * splbio interrupts blocked. 1278 */ 1279 static struct bmsafemap * 1280 bmsafemap_lookup(struct buf *bp) 1281 { 1282 struct bmsafemap *bmsafemap; 1283 struct worklist *wk; 1284 1285 #ifdef DEBUG 1286 if (lk.lkt_held == NOHOLDER) 1287 panic("bmsafemap_lookup: lock not held"); 1288 #endif 1289 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 1290 if (wk->wk_type == D_BMSAFEMAP) 1291 return (WK_BMSAFEMAP(wk)); 1292 } 1293 FREE_LOCK(&lk); 1294 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap), 1295 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 1296 bmsafemap->sm_list.wk_type = D_BMSAFEMAP; 1297 bmsafemap->sm_list.wk_state = 0; 1298 bmsafemap->sm_buf = bp; 1299 LIST_INIT(&bmsafemap->sm_allocdirecthd); 1300 LIST_INIT(&bmsafemap->sm_allocindirhd); 1301 LIST_INIT(&bmsafemap->sm_inodedephd); 1302 LIST_INIT(&bmsafemap->sm_newblkhd); 1303 ACQUIRE_LOCK(&lk); 1304 WORKLIST_INSERT_BP(bp, &bmsafemap->sm_list); 1305 return (bmsafemap); 1306 } 1307 1308 /* 1309 * Direct block allocation dependencies. 1310 * 1311 * When a new block is allocated, the corresponding disk locations must be 1312 * initialized (with zeros or new data) before the on-disk inode points to 1313 * them. Also, the freemap from which the block was allocated must be 1314 * updated (on disk) before the inode's pointer. These two dependencies are 1315 * independent of each other and are needed for all file blocks and indirect 1316 * blocks that are pointed to directly by the inode. Just before the 1317 * "in-core" version of the inode is updated with a newly allocated block 1318 * number, a procedure (below) is called to setup allocation dependency 1319 * structures. These structures are removed when the corresponding 1320 * dependencies are satisfied or when the block allocation becomes obsolete 1321 * (i.e., the file is deleted, the block is de-allocated, or the block is a 1322 * fragment that gets upgraded). All of these cases are handled in 1323 * procedures described later. 1324 * 1325 * When a file extension causes a fragment to be upgraded, either to a larger 1326 * fragment or to a full block, the on-disk location may change (if the 1327 * previous fragment could not simply be extended). In this case, the old 1328 * fragment must be de-allocated, but not until after the inode's pointer has 1329 * been updated. In most cases, this is handled by later procedures, which 1330 * will construct a "freefrag" structure to be added to the workitem queue 1331 * when the inode update is complete (or obsolete). The main exception to 1332 * this is when an allocation occurs while a pending allocation dependency 1333 * (for the same block pointer) remains. This case is handled in the main 1334 * allocation dependency setup procedure by immediately freeing the 1335 * unreferenced fragments. 1336 * 1337 * Parameters: 1338 * ip: inode to which block is being added 1339 * lbn: block pointer within inode 1340 * newblkno: disk block number being added 1341 * oldblkno: previous block number, 0 unless frag 1342 * newsize: size of new block 1343 * oldsize: size of new block 1344 * bp: bp for allocated block 1345 */ 1346 void 1347 softdep_setup_allocdirect(struct inode *ip, ufs_lbn_t lbn, ufs_daddr_t newblkno, 1348 ufs_daddr_t oldblkno, long newsize, long oldsize, 1349 struct buf *bp) 1350 { 1351 struct allocdirect *adp, *oldadp; 1352 struct allocdirectlst *adphead; 1353 struct bmsafemap *bmsafemap; 1354 struct inodedep *inodedep; 1355 struct pagedep *pagedep; 1356 struct newblk *newblk; 1357 1358 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect), 1359 M_ALLOCDIRECT, M_SOFTDEP_FLAGS | M_ZERO); 1360 adp->ad_list.wk_type = D_ALLOCDIRECT; 1361 adp->ad_lbn = lbn; 1362 adp->ad_newblkno = newblkno; 1363 adp->ad_oldblkno = oldblkno; 1364 adp->ad_newsize = newsize; 1365 adp->ad_oldsize = oldsize; 1366 adp->ad_state = ATTACHED; 1367 if (newblkno == oldblkno) 1368 adp->ad_freefrag = NULL; 1369 else 1370 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1371 1372 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1373 panic("softdep_setup_allocdirect: lost block"); 1374 1375 ACQUIRE_LOCK(&lk); 1376 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep); 1377 adp->ad_inodedep = inodedep; 1378 1379 if (newblk->nb_state == DEPCOMPLETE) { 1380 adp->ad_state |= DEPCOMPLETE; 1381 adp->ad_buf = NULL; 1382 } else { 1383 bmsafemap = newblk->nb_bmsafemap; 1384 adp->ad_buf = bmsafemap->sm_buf; 1385 LIST_REMOVE(newblk, nb_deps); 1386 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1387 } 1388 LIST_REMOVE(newblk, nb_hash); 1389 FREE(newblk, M_NEWBLK); 1390 1391 WORKLIST_INSERT_BP(bp, &adp->ad_list); 1392 if (lbn >= NDADDR) { 1393 /* allocating an indirect block */ 1394 if (oldblkno != 0) { 1395 FREE_LOCK(&lk); 1396 panic("softdep_setup_allocdirect: non-zero indir"); 1397 } 1398 } else { 1399 /* 1400 * Allocating a direct block. 1401 * 1402 * If we are allocating a directory block, then we must 1403 * allocate an associated pagedep to track additions and 1404 * deletions. 1405 */ 1406 if ((ip->i_mode & IFMT) == IFDIR && 1407 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) { 1408 WORKLIST_INSERT_BP(bp, &pagedep->pd_list); 1409 } 1410 } 1411 /* 1412 * The list of allocdirects must be kept in sorted and ascending 1413 * order so that the rollback routines can quickly determine the 1414 * first uncommitted block (the size of the file stored on disk 1415 * ends at the end of the lowest committed fragment, or if there 1416 * are no fragments, at the end of the highest committed block). 1417 * Since files generally grow, the typical case is that the new 1418 * block is to be added at the end of the list. We speed this 1419 * special case by checking against the last allocdirect in the 1420 * list before laboriously traversing the list looking for the 1421 * insertion point. 1422 */ 1423 adphead = &inodedep->id_newinoupdt; 1424 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1425 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1426 /* insert at end of list */ 1427 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1428 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1429 allocdirect_merge(adphead, adp, oldadp); 1430 FREE_LOCK(&lk); 1431 return; 1432 } 1433 TAILQ_FOREACH(oldadp, adphead, ad_next) { 1434 if (oldadp->ad_lbn >= lbn) 1435 break; 1436 } 1437 if (oldadp == NULL) { 1438 FREE_LOCK(&lk); 1439 panic("softdep_setup_allocdirect: lost entry"); 1440 } 1441 /* insert in middle of list */ 1442 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1443 if (oldadp->ad_lbn == lbn) 1444 allocdirect_merge(adphead, adp, oldadp); 1445 FREE_LOCK(&lk); 1446 } 1447 1448 /* 1449 * Replace an old allocdirect dependency with a newer one. 1450 * This routine must be called with splbio interrupts blocked. 1451 * 1452 * Parameters: 1453 * adphead: head of list holding allocdirects 1454 * newadp: allocdirect being added 1455 * oldadp: existing allocdirect being checked 1456 */ 1457 static void 1458 allocdirect_merge(struct allocdirectlst *adphead, 1459 struct allocdirect *newadp, 1460 struct allocdirect *oldadp) 1461 { 1462 struct freefrag *freefrag; 1463 1464 #ifdef DEBUG 1465 if (lk.lkt_held == NOHOLDER) 1466 panic("allocdirect_merge: lock not held"); 1467 #endif 1468 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 1469 newadp->ad_oldsize != oldadp->ad_newsize || 1470 newadp->ad_lbn >= NDADDR) { 1471 FREE_LOCK(&lk); 1472 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d", 1473 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn, 1474 NDADDR); 1475 } 1476 newadp->ad_oldblkno = oldadp->ad_oldblkno; 1477 newadp->ad_oldsize = oldadp->ad_oldsize; 1478 /* 1479 * If the old dependency had a fragment to free or had never 1480 * previously had a block allocated, then the new dependency 1481 * can immediately post its freefrag and adopt the old freefrag. 1482 * This action is done by swapping the freefrag dependencies. 1483 * The new dependency gains the old one's freefrag, and the 1484 * old one gets the new one and then immediately puts it on 1485 * the worklist when it is freed by free_allocdirect. It is 1486 * not possible to do this swap when the old dependency had a 1487 * non-zero size but no previous fragment to free. This condition 1488 * arises when the new block is an extension of the old block. 1489 * Here, the first part of the fragment allocated to the new 1490 * dependency is part of the block currently claimed on disk by 1491 * the old dependency, so cannot legitimately be freed until the 1492 * conditions for the new dependency are fulfilled. 1493 */ 1494 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 1495 freefrag = newadp->ad_freefrag; 1496 newadp->ad_freefrag = oldadp->ad_freefrag; 1497 oldadp->ad_freefrag = freefrag; 1498 } 1499 free_allocdirect(adphead, oldadp, 0); 1500 } 1501 1502 /* 1503 * Allocate a new freefrag structure if needed. 1504 */ 1505 static struct freefrag * 1506 newfreefrag(struct inode *ip, ufs_daddr_t blkno, long size) 1507 { 1508 struct freefrag *freefrag; 1509 struct fs *fs; 1510 1511 if (blkno == 0) 1512 return (NULL); 1513 fs = ip->i_fs; 1514 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 1515 panic("newfreefrag: frag size"); 1516 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag), 1517 M_FREEFRAG, M_SOFTDEP_FLAGS); 1518 freefrag->ff_list.wk_type = D_FREEFRAG; 1519 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */ 1520 freefrag->ff_inum = ip->i_number; 1521 freefrag->ff_fs = fs; 1522 freefrag->ff_devvp = ip->i_devvp; 1523 freefrag->ff_blkno = blkno; 1524 freefrag->ff_fragsize = size; 1525 return (freefrag); 1526 } 1527 1528 /* 1529 * This workitem de-allocates fragments that were replaced during 1530 * file block allocation. 1531 */ 1532 static void 1533 handle_workitem_freefrag(struct freefrag *freefrag) 1534 { 1535 struct inode tip; 1536 1537 tip.i_fs = freefrag->ff_fs; 1538 tip.i_devvp = freefrag->ff_devvp; 1539 tip.i_dev = freefrag->ff_devvp->v_rdev; 1540 tip.i_number = freefrag->ff_inum; 1541 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */ 1542 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize); 1543 FREE(freefrag, M_FREEFRAG); 1544 } 1545 1546 /* 1547 * Indirect block allocation dependencies. 1548 * 1549 * The same dependencies that exist for a direct block also exist when 1550 * a new block is allocated and pointed to by an entry in a block of 1551 * indirect pointers. The undo/redo states described above are also 1552 * used here. Because an indirect block contains many pointers that 1553 * may have dependencies, a second copy of the entire in-memory indirect 1554 * block is kept. The buffer cache copy is always completely up-to-date. 1555 * The second copy, which is used only as a source for disk writes, 1556 * contains only the safe pointers (i.e., those that have no remaining 1557 * update dependencies). The second copy is freed when all pointers 1558 * are safe. The cache is not allowed to replace indirect blocks with 1559 * pending update dependencies. If a buffer containing an indirect 1560 * block with dependencies is written, these routines will mark it 1561 * dirty again. It can only be successfully written once all the 1562 * dependencies are removed. The ffs_fsync routine in conjunction with 1563 * softdep_sync_metadata work together to get all the dependencies 1564 * removed so that a file can be successfully written to disk. Three 1565 * procedures are used when setting up indirect block pointer 1566 * dependencies. The division is necessary because of the organization 1567 * of the "balloc" routine and because of the distinction between file 1568 * pages and file metadata blocks. 1569 */ 1570 1571 /* 1572 * Allocate a new allocindir structure. 1573 * 1574 * Parameters: 1575 * ip: inode for file being extended 1576 * ptrno: offset of pointer in indirect block 1577 * newblkno: disk block number being added 1578 * oldblkno: previous block number, 0 if none 1579 */ 1580 static struct allocindir * 1581 newallocindir(struct inode *ip, int ptrno, ufs_daddr_t newblkno, 1582 ufs_daddr_t oldblkno) 1583 { 1584 struct allocindir *aip; 1585 1586 MALLOC(aip, struct allocindir *, sizeof(struct allocindir), 1587 M_ALLOCINDIR, M_SOFTDEP_FLAGS | M_ZERO); 1588 aip->ai_list.wk_type = D_ALLOCINDIR; 1589 aip->ai_state = ATTACHED; 1590 aip->ai_offset = ptrno; 1591 aip->ai_newblkno = newblkno; 1592 aip->ai_oldblkno = oldblkno; 1593 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize); 1594 return (aip); 1595 } 1596 1597 /* 1598 * Called just before setting an indirect block pointer 1599 * to a newly allocated file page. 1600 * 1601 * Parameters: 1602 * ip: inode for file being extended 1603 * lbn: allocated block number within file 1604 * bp: buffer with indirect blk referencing page 1605 * ptrno: offset of pointer in indirect block 1606 * newblkno: disk block number being added 1607 * oldblkno: previous block number, 0 if none 1608 * nbp: buffer holding allocated page 1609 */ 1610 void 1611 softdep_setup_allocindir_page(struct inode *ip, ufs_lbn_t lbn, 1612 struct buf *bp, int ptrno, 1613 ufs_daddr_t newblkno, ufs_daddr_t oldblkno, 1614 struct buf *nbp) 1615 { 1616 struct allocindir *aip; 1617 struct pagedep *pagedep; 1618 1619 aip = newallocindir(ip, ptrno, newblkno, oldblkno); 1620 ACQUIRE_LOCK(&lk); 1621 /* 1622 * If we are allocating a directory page, then we must 1623 * allocate an associated pagedep to track additions and 1624 * deletions. 1625 */ 1626 if ((ip->i_mode & IFMT) == IFDIR && 1627 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1628 WORKLIST_INSERT_BP(nbp, &pagedep->pd_list); 1629 WORKLIST_INSERT_BP(nbp, &aip->ai_list); 1630 FREE_LOCK(&lk); 1631 setup_allocindir_phase2(bp, ip, aip); 1632 } 1633 1634 /* 1635 * Called just before setting an indirect block pointer to a 1636 * newly allocated indirect block. 1637 * Parameters: 1638 * nbp: newly allocated indirect block 1639 * ip: inode for file being extended 1640 * bp: indirect block referencing allocated block 1641 * ptrno: offset of pointer in indirect block 1642 * newblkno: disk block number being added 1643 */ 1644 void 1645 softdep_setup_allocindir_meta(struct buf *nbp, struct inode *ip, 1646 struct buf *bp, int ptrno, 1647 ufs_daddr_t newblkno) 1648 { 1649 struct allocindir *aip; 1650 1651 aip = newallocindir(ip, ptrno, newblkno, 0); 1652 ACQUIRE_LOCK(&lk); 1653 WORKLIST_INSERT_BP(nbp, &aip->ai_list); 1654 FREE_LOCK(&lk); 1655 setup_allocindir_phase2(bp, ip, aip); 1656 } 1657 1658 /* 1659 * Called to finish the allocation of the "aip" allocated 1660 * by one of the two routines above. 1661 * 1662 * Parameters: 1663 * bp: in-memory copy of the indirect block 1664 * ip: inode for file being extended 1665 * aip: allocindir allocated by the above routines 1666 */ 1667 static void 1668 setup_allocindir_phase2(struct buf *bp, struct inode *ip, 1669 struct allocindir *aip) 1670 { 1671 struct worklist *wk; 1672 struct indirdep *indirdep, *newindirdep; 1673 struct bmsafemap *bmsafemap; 1674 struct allocindir *oldaip; 1675 struct freefrag *freefrag; 1676 struct newblk *newblk; 1677 1678 if (bp->b_loffset >= 0) 1679 panic("setup_allocindir_phase2: not indir blk"); 1680 for (indirdep = NULL, newindirdep = NULL; ; ) { 1681 ACQUIRE_LOCK(&lk); 1682 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 1683 if (wk->wk_type != D_INDIRDEP) 1684 continue; 1685 indirdep = WK_INDIRDEP(wk); 1686 break; 1687 } 1688 if (indirdep == NULL && newindirdep) { 1689 indirdep = newindirdep; 1690 WORKLIST_INSERT_BP(bp, &indirdep->ir_list); 1691 newindirdep = NULL; 1692 } 1693 FREE_LOCK(&lk); 1694 if (indirdep) { 1695 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0, 1696 &newblk) == 0) 1697 panic("setup_allocindir: lost block"); 1698 ACQUIRE_LOCK(&lk); 1699 if (newblk->nb_state == DEPCOMPLETE) { 1700 aip->ai_state |= DEPCOMPLETE; 1701 aip->ai_buf = NULL; 1702 } else { 1703 bmsafemap = newblk->nb_bmsafemap; 1704 aip->ai_buf = bmsafemap->sm_buf; 1705 LIST_REMOVE(newblk, nb_deps); 1706 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd, 1707 aip, ai_deps); 1708 } 1709 LIST_REMOVE(newblk, nb_hash); 1710 FREE(newblk, M_NEWBLK); 1711 aip->ai_indirdep = indirdep; 1712 /* 1713 * Check to see if there is an existing dependency 1714 * for this block. If there is, merge the old 1715 * dependency into the new one. 1716 */ 1717 if (aip->ai_oldblkno == 0) 1718 oldaip = NULL; 1719 else 1720 1721 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) 1722 if (oldaip->ai_offset == aip->ai_offset) 1723 break; 1724 if (oldaip != NULL) { 1725 if (oldaip->ai_newblkno != aip->ai_oldblkno) { 1726 FREE_LOCK(&lk); 1727 panic("setup_allocindir_phase2: blkno"); 1728 } 1729 aip->ai_oldblkno = oldaip->ai_oldblkno; 1730 freefrag = oldaip->ai_freefrag; 1731 oldaip->ai_freefrag = aip->ai_freefrag; 1732 aip->ai_freefrag = freefrag; 1733 free_allocindir(oldaip, NULL); 1734 } 1735 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 1736 ((ufs_daddr_t *)indirdep->ir_savebp->b_data) 1737 [aip->ai_offset] = aip->ai_oldblkno; 1738 FREE_LOCK(&lk); 1739 } 1740 if (newindirdep) { 1741 /* 1742 * Avoid any possibility of data corruption by 1743 * ensuring that our old version is thrown away. 1744 */ 1745 newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 1746 brelse(newindirdep->ir_savebp); 1747 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP); 1748 } 1749 if (indirdep) 1750 break; 1751 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep), 1752 M_INDIRDEP, M_SOFTDEP_FLAGS); 1753 newindirdep->ir_list.wk_type = D_INDIRDEP; 1754 newindirdep->ir_state = ATTACHED; 1755 LIST_INIT(&newindirdep->ir_deplisthd); 1756 LIST_INIT(&newindirdep->ir_donehd); 1757 if (bp->b_bio2.bio_offset == NOOFFSET) { 1758 VOP_BMAP(bp->b_vp, bp->b_bio1.bio_offset, 1759 &bp->b_bio2.bio_offset, NULL, NULL); 1760 } 1761 KKASSERT(bp->b_bio2.bio_offset != NOOFFSET); 1762 newindirdep->ir_savebp = getblk(ip->i_devvp, 1763 bp->b_bio2.bio_offset, 1764 bp->b_bcount, 0, 0); 1765 BUF_KERNPROC(newindirdep->ir_savebp); 1766 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 1767 } 1768 } 1769 1770 /* 1771 * Block de-allocation dependencies. 1772 * 1773 * When blocks are de-allocated, the on-disk pointers must be nullified before 1774 * the blocks are made available for use by other files. (The true 1775 * requirement is that old pointers must be nullified before new on-disk 1776 * pointers are set. We chose this slightly more stringent requirement to 1777 * reduce complexity.) Our implementation handles this dependency by updating 1778 * the inode (or indirect block) appropriately but delaying the actual block 1779 * de-allocation (i.e., freemap and free space count manipulation) until 1780 * after the updated versions reach stable storage. After the disk is 1781 * updated, the blocks can be safely de-allocated whenever it is convenient. 1782 * This implementation handles only the common case of reducing a file's 1783 * length to zero. Other cases are handled by the conventional synchronous 1784 * write approach. 1785 * 1786 * The ffs implementation with which we worked double-checks 1787 * the state of the block pointers and file size as it reduces 1788 * a file's length. Some of this code is replicated here in our 1789 * soft updates implementation. The freeblks->fb_chkcnt field is 1790 * used to transfer a part of this information to the procedure 1791 * that eventually de-allocates the blocks. 1792 * 1793 * This routine should be called from the routine that shortens 1794 * a file's length, before the inode's size or block pointers 1795 * are modified. It will save the block pointer information for 1796 * later release and zero the inode so that the calling routine 1797 * can release it. 1798 */ 1799 struct softdep_setup_freeblocks_info { 1800 struct fs *fs; 1801 struct inode *ip; 1802 }; 1803 1804 static int softdep_setup_freeblocks_bp(struct buf *bp, void *data); 1805 1806 /* 1807 * Parameters: 1808 * ip: The inode whose length is to be reduced 1809 * length: The new length for the file 1810 */ 1811 void 1812 softdep_setup_freeblocks(struct inode *ip, off_t length) 1813 { 1814 struct softdep_setup_freeblocks_info info; 1815 struct freeblks *freeblks; 1816 struct inodedep *inodedep; 1817 struct allocdirect *adp; 1818 struct vnode *vp; 1819 struct buf *bp; 1820 struct fs *fs; 1821 int i, error, delay; 1822 int count; 1823 1824 fs = ip->i_fs; 1825 if (length != 0) 1826 panic("softde_setup_freeblocks: non-zero length"); 1827 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks), 1828 M_FREEBLKS, M_SOFTDEP_FLAGS | M_ZERO); 1829 freeblks->fb_list.wk_type = D_FREEBLKS; 1830 freeblks->fb_state = ATTACHED; 1831 freeblks->fb_uid = ip->i_uid; 1832 freeblks->fb_previousinum = ip->i_number; 1833 freeblks->fb_devvp = ip->i_devvp; 1834 freeblks->fb_fs = fs; 1835 freeblks->fb_oldsize = ip->i_size; 1836 freeblks->fb_newsize = length; 1837 freeblks->fb_chkcnt = ip->i_blocks; 1838 for (i = 0; i < NDADDR; i++) { 1839 freeblks->fb_dblks[i] = ip->i_db[i]; 1840 ip->i_db[i] = 0; 1841 } 1842 for (i = 0; i < NIADDR; i++) { 1843 freeblks->fb_iblks[i] = ip->i_ib[i]; 1844 ip->i_ib[i] = 0; 1845 } 1846 ip->i_blocks = 0; 1847 ip->i_size = 0; 1848 /* 1849 * Push the zero'ed inode to to its disk buffer so that we are free 1850 * to delete its dependencies below. Once the dependencies are gone 1851 * the buffer can be safely released. 1852 */ 1853 if ((error = bread(ip->i_devvp, 1854 fsbtodoff(fs, ino_to_fsba(fs, ip->i_number)), 1855 (int)fs->fs_bsize, &bp)) != 0) 1856 softdep_error("softdep_setup_freeblocks", error); 1857 *((struct ufs1_dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) = 1858 ip->i_din; 1859 /* 1860 * Find and eliminate any inode dependencies. 1861 */ 1862 ACQUIRE_LOCK(&lk); 1863 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep); 1864 if ((inodedep->id_state & IOSTARTED) != 0) { 1865 FREE_LOCK(&lk); 1866 panic("softdep_setup_freeblocks: inode busy"); 1867 } 1868 /* 1869 * Add the freeblks structure to the list of operations that 1870 * must await the zero'ed inode being written to disk. If we 1871 * still have a bitmap dependency (delay == 0), then the inode 1872 * has never been written to disk, so we can process the 1873 * freeblks below once we have deleted the dependencies. 1874 */ 1875 delay = (inodedep->id_state & DEPCOMPLETE); 1876 if (delay) 1877 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list); 1878 /* 1879 * Because the file length has been truncated to zero, any 1880 * pending block allocation dependency structures associated 1881 * with this inode are obsolete and can simply be de-allocated. 1882 * We must first merge the two dependency lists to get rid of 1883 * any duplicate freefrag structures, then purge the merged list. 1884 */ 1885 merge_inode_lists(inodedep); 1886 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 1887 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 1888 FREE_LOCK(&lk); 1889 bdwrite(bp); 1890 /* 1891 * We must wait for any I/O in progress to finish so that 1892 * all potential buffers on the dirty list will be visible. 1893 * Once they are all there, walk the list and get rid of 1894 * any dependencies. 1895 */ 1896 vp = ITOV(ip); 1897 ACQUIRE_LOCK(&lk); 1898 drain_output(vp, 1); 1899 1900 info.fs = fs; 1901 info.ip = ip; 1902 do { 1903 count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 1904 softdep_setup_freeblocks_bp, &info); 1905 } while (count != 0); 1906 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0) 1907 (void)free_inodedep(inodedep); 1908 1909 if (delay) { 1910 freeblks->fb_state |= DEPCOMPLETE; 1911 /* 1912 * If the inode with zeroed block pointers is now on disk 1913 * we can start freeing blocks. Add freeblks to the worklist 1914 * instead of calling handle_workitem_freeblocks directly as 1915 * it is more likely that additional IO is needed to complete 1916 * the request here than in the !delay case. 1917 */ 1918 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 1919 add_to_worklist(&freeblks->fb_list); 1920 } 1921 1922 FREE_LOCK(&lk); 1923 /* 1924 * If the inode has never been written to disk (delay == 0), 1925 * then we can process the freeblks now that we have deleted 1926 * the dependencies. 1927 */ 1928 if (!delay) 1929 handle_workitem_freeblocks(freeblks); 1930 } 1931 1932 static int 1933 softdep_setup_freeblocks_bp(struct buf *bp, void *data) 1934 { 1935 struct softdep_setup_freeblocks_info *info = data; 1936 struct inodedep *inodedep; 1937 1938 if (getdirtybuf(&bp, MNT_WAIT) == 0) { 1939 kprintf("softdep_setup_freeblocks_bp(1): caught bp %p going away\n", bp); 1940 return(-1); 1941 } 1942 if (bp->b_vp != ITOV(info->ip) || (bp->b_flags & B_DELWRI) == 0) { 1943 kprintf("softdep_setup_freeblocks_bp(2): caught bp %p going away\n", bp); 1944 BUF_UNLOCK(bp); 1945 return(-1); 1946 } 1947 (void) inodedep_lookup(info->fs, info->ip->i_number, 0, &inodedep); 1948 deallocate_dependencies(bp, inodedep); 1949 bp->b_flags |= B_INVAL | B_NOCACHE; 1950 FREE_LOCK(&lk); 1951 brelse(bp); 1952 ACQUIRE_LOCK(&lk); 1953 return(1); 1954 } 1955 1956 /* 1957 * Reclaim any dependency structures from a buffer that is about to 1958 * be reallocated to a new vnode. The buffer must be locked, thus, 1959 * no I/O completion operations can occur while we are manipulating 1960 * its associated dependencies. The mutex is held so that other I/O's 1961 * associated with related dependencies do not occur. 1962 */ 1963 static void 1964 deallocate_dependencies(struct buf *bp, struct inodedep *inodedep) 1965 { 1966 struct worklist *wk; 1967 struct indirdep *indirdep; 1968 struct allocindir *aip; 1969 struct pagedep *pagedep; 1970 struct dirrem *dirrem; 1971 struct diradd *dap; 1972 int i; 1973 1974 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 1975 switch (wk->wk_type) { 1976 1977 case D_INDIRDEP: 1978 indirdep = WK_INDIRDEP(wk); 1979 /* 1980 * None of the indirect pointers will ever be visible, 1981 * so they can simply be tossed. GOINGAWAY ensures 1982 * that allocated pointers will be saved in the buffer 1983 * cache until they are freed. Note that they will 1984 * only be able to be found by their physical address 1985 * since the inode mapping the logical address will 1986 * be gone. The save buffer used for the safe copy 1987 * was allocated in setup_allocindir_phase2 using 1988 * the physical address so it could be used for this 1989 * purpose. Hence we swap the safe copy with the real 1990 * copy, allowing the safe copy to be freed and holding 1991 * on to the real copy for later use in indir_trunc. 1992 * 1993 * NOTE: ir_savebp is relative to the block device 1994 * so b_bio1 contains the device block number. 1995 */ 1996 if (indirdep->ir_state & GOINGAWAY) { 1997 FREE_LOCK(&lk); 1998 panic("deallocate_dependencies: already gone"); 1999 } 2000 indirdep->ir_state |= GOINGAWAY; 2001 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 2002 free_allocindir(aip, inodedep); 2003 if (bp->b_bio1.bio_offset >= 0 || 2004 bp->b_bio2.bio_offset != indirdep->ir_savebp->b_bio1.bio_offset) { 2005 FREE_LOCK(&lk); 2006 panic("deallocate_dependencies: not indir"); 2007 } 2008 bcopy(bp->b_data, indirdep->ir_savebp->b_data, 2009 bp->b_bcount); 2010 WORKLIST_REMOVE(wk); 2011 WORKLIST_INSERT_BP(indirdep->ir_savebp, wk); 2012 continue; 2013 2014 case D_PAGEDEP: 2015 pagedep = WK_PAGEDEP(wk); 2016 /* 2017 * None of the directory additions will ever be 2018 * visible, so they can simply be tossed. 2019 */ 2020 for (i = 0; i < DAHASHSZ; i++) 2021 while ((dap = 2022 LIST_FIRST(&pagedep->pd_diraddhd[i]))) 2023 free_diradd(dap); 2024 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0) 2025 free_diradd(dap); 2026 /* 2027 * Copy any directory remove dependencies to the list 2028 * to be processed after the zero'ed inode is written. 2029 * If the inode has already been written, then they 2030 * can be dumped directly onto the work list. 2031 */ 2032 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 2033 LIST_REMOVE(dirrem, dm_next); 2034 dirrem->dm_dirinum = pagedep->pd_ino; 2035 if (inodedep == NULL || 2036 (inodedep->id_state & ALLCOMPLETE) == 2037 ALLCOMPLETE) 2038 add_to_worklist(&dirrem->dm_list); 2039 else 2040 WORKLIST_INSERT(&inodedep->id_bufwait, 2041 &dirrem->dm_list); 2042 } 2043 WORKLIST_REMOVE(&pagedep->pd_list); 2044 LIST_REMOVE(pagedep, pd_hash); 2045 WORKITEM_FREE(pagedep, D_PAGEDEP); 2046 continue; 2047 2048 case D_ALLOCINDIR: 2049 free_allocindir(WK_ALLOCINDIR(wk), inodedep); 2050 continue; 2051 2052 case D_ALLOCDIRECT: 2053 case D_INODEDEP: 2054 FREE_LOCK(&lk); 2055 panic("deallocate_dependencies: Unexpected type %s", 2056 TYPENAME(wk->wk_type)); 2057 /* NOTREACHED */ 2058 2059 default: 2060 FREE_LOCK(&lk); 2061 panic("deallocate_dependencies: Unknown type %s", 2062 TYPENAME(wk->wk_type)); 2063 /* NOTREACHED */ 2064 } 2065 } 2066 } 2067 2068 /* 2069 * Free an allocdirect. Generate a new freefrag work request if appropriate. 2070 * This routine must be called with splbio interrupts blocked. 2071 */ 2072 static void 2073 free_allocdirect(struct allocdirectlst *adphead, 2074 struct allocdirect *adp, int delay) 2075 { 2076 2077 #ifdef DEBUG 2078 if (lk.lkt_held == NOHOLDER) 2079 panic("free_allocdirect: lock not held"); 2080 #endif 2081 if ((adp->ad_state & DEPCOMPLETE) == 0) 2082 LIST_REMOVE(adp, ad_deps); 2083 TAILQ_REMOVE(adphead, adp, ad_next); 2084 if ((adp->ad_state & COMPLETE) == 0) 2085 WORKLIST_REMOVE(&adp->ad_list); 2086 if (adp->ad_freefrag != NULL) { 2087 if (delay) 2088 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 2089 &adp->ad_freefrag->ff_list); 2090 else 2091 add_to_worklist(&adp->ad_freefrag->ff_list); 2092 } 2093 WORKITEM_FREE(adp, D_ALLOCDIRECT); 2094 } 2095 2096 /* 2097 * Prepare an inode to be freed. The actual free operation is not 2098 * done until the zero'ed inode has been written to disk. 2099 */ 2100 void 2101 softdep_freefile(struct vnode *pvp, ino_t ino, int mode) 2102 { 2103 struct inode *ip = VTOI(pvp); 2104 struct inodedep *inodedep; 2105 struct freefile *freefile; 2106 2107 /* 2108 * This sets up the inode de-allocation dependency. 2109 */ 2110 MALLOC(freefile, struct freefile *, sizeof(struct freefile), 2111 M_FREEFILE, M_SOFTDEP_FLAGS); 2112 freefile->fx_list.wk_type = D_FREEFILE; 2113 freefile->fx_list.wk_state = 0; 2114 freefile->fx_mode = mode; 2115 freefile->fx_oldinum = ino; 2116 freefile->fx_devvp = ip->i_devvp; 2117 freefile->fx_fs = ip->i_fs; 2118 2119 /* 2120 * If the inodedep does not exist, then the zero'ed inode has 2121 * been written to disk. If the allocated inode has never been 2122 * written to disk, then the on-disk inode is zero'ed. In either 2123 * case we can free the file immediately. 2124 */ 2125 ACQUIRE_LOCK(&lk); 2126 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 || 2127 check_inode_unwritten(inodedep)) { 2128 FREE_LOCK(&lk); 2129 handle_workitem_freefile(freefile); 2130 return; 2131 } 2132 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 2133 FREE_LOCK(&lk); 2134 } 2135 2136 /* 2137 * Check to see if an inode has never been written to disk. If 2138 * so free the inodedep and return success, otherwise return failure. 2139 * This routine must be called with splbio interrupts blocked. 2140 * 2141 * If we still have a bitmap dependency, then the inode has never 2142 * been written to disk. Drop the dependency as it is no longer 2143 * necessary since the inode is being deallocated. We set the 2144 * ALLCOMPLETE flags since the bitmap now properly shows that the 2145 * inode is not allocated. Even if the inode is actively being 2146 * written, it has been rolled back to its zero'ed state, so we 2147 * are ensured that a zero inode is what is on the disk. For short 2148 * lived files, this change will usually result in removing all the 2149 * dependencies from the inode so that it can be freed immediately. 2150 */ 2151 static int 2152 check_inode_unwritten(struct inodedep *inodedep) 2153 { 2154 2155 if ((inodedep->id_state & DEPCOMPLETE) != 0 || 2156 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2157 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2158 LIST_FIRST(&inodedep->id_inowait) != NULL || 2159 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2160 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2161 inodedep->id_nlinkdelta != 0) 2162 return (0); 2163 2164 /* 2165 * Another process might be in initiate_write_inodeblock 2166 * trying to allocate memory without holding "Softdep Lock". 2167 */ 2168 if ((inodedep->id_state & IOSTARTED) != 0 && 2169 inodedep->id_savedino == NULL) 2170 return(0); 2171 2172 inodedep->id_state |= ALLCOMPLETE; 2173 LIST_REMOVE(inodedep, id_deps); 2174 inodedep->id_buf = NULL; 2175 if (inodedep->id_state & ONWORKLIST) 2176 WORKLIST_REMOVE(&inodedep->id_list); 2177 if (inodedep->id_savedino != NULL) { 2178 FREE(inodedep->id_savedino, M_INODEDEP); 2179 inodedep->id_savedino = NULL; 2180 } 2181 if (free_inodedep(inodedep) == 0) { 2182 FREE_LOCK(&lk); 2183 panic("check_inode_unwritten: busy inode"); 2184 } 2185 return (1); 2186 } 2187 2188 /* 2189 * Try to free an inodedep structure. Return 1 if it could be freed. 2190 */ 2191 static int 2192 free_inodedep(struct inodedep *inodedep) 2193 { 2194 2195 if ((inodedep->id_state & ONWORKLIST) != 0 || 2196 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 2197 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2198 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2199 LIST_FIRST(&inodedep->id_inowait) != NULL || 2200 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2201 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2202 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL) 2203 return (0); 2204 LIST_REMOVE(inodedep, id_hash); 2205 WORKITEM_FREE(inodedep, D_INODEDEP); 2206 num_inodedep -= 1; 2207 return (1); 2208 } 2209 2210 /* 2211 * This workitem routine performs the block de-allocation. 2212 * The workitem is added to the pending list after the updated 2213 * inode block has been written to disk. As mentioned above, 2214 * checks regarding the number of blocks de-allocated (compared 2215 * to the number of blocks allocated for the file) are also 2216 * performed in this function. 2217 */ 2218 static void 2219 handle_workitem_freeblocks(struct freeblks *freeblks) 2220 { 2221 struct inode tip; 2222 ufs_daddr_t bn; 2223 struct fs *fs; 2224 int i, level, bsize; 2225 long nblocks, blocksreleased = 0; 2226 int error, allerror = 0; 2227 ufs_lbn_t baselbns[NIADDR], tmpval; 2228 2229 tip.i_number = freeblks->fb_previousinum; 2230 tip.i_devvp = freeblks->fb_devvp; 2231 tip.i_dev = freeblks->fb_devvp->v_rdev; 2232 tip.i_fs = freeblks->fb_fs; 2233 tip.i_size = freeblks->fb_oldsize; 2234 tip.i_uid = freeblks->fb_uid; 2235 fs = freeblks->fb_fs; 2236 tmpval = 1; 2237 baselbns[0] = NDADDR; 2238 for (i = 1; i < NIADDR; i++) { 2239 tmpval *= NINDIR(fs); 2240 baselbns[i] = baselbns[i - 1] + tmpval; 2241 } 2242 nblocks = btodb(fs->fs_bsize); 2243 blocksreleased = 0; 2244 /* 2245 * Indirect blocks first. 2246 */ 2247 for (level = (NIADDR - 1); level >= 0; level--) { 2248 if ((bn = freeblks->fb_iblks[level]) == 0) 2249 continue; 2250 if ((error = indir_trunc(&tip, fsbtodoff(fs, bn), level, 2251 baselbns[level], &blocksreleased)) == 0) 2252 allerror = error; 2253 ffs_blkfree(&tip, bn, fs->fs_bsize); 2254 blocksreleased += nblocks; 2255 } 2256 /* 2257 * All direct blocks or frags. 2258 */ 2259 for (i = (NDADDR - 1); i >= 0; i--) { 2260 if ((bn = freeblks->fb_dblks[i]) == 0) 2261 continue; 2262 bsize = blksize(fs, &tip, i); 2263 ffs_blkfree(&tip, bn, bsize); 2264 blocksreleased += btodb(bsize); 2265 } 2266 2267 #ifdef DIAGNOSTIC 2268 if (freeblks->fb_chkcnt != blocksreleased) 2269 kprintf("handle_workitem_freeblocks: block count\n"); 2270 if (allerror) 2271 softdep_error("handle_workitem_freeblks", allerror); 2272 #endif /* DIAGNOSTIC */ 2273 WORKITEM_FREE(freeblks, D_FREEBLKS); 2274 } 2275 2276 /* 2277 * Release blocks associated with the inode ip and stored in the indirect 2278 * block at doffset. If level is greater than SINGLE, the block is an 2279 * indirect block and recursive calls to indirtrunc must be used to 2280 * cleanse other indirect blocks. 2281 */ 2282 static int 2283 indir_trunc(struct inode *ip, off_t doffset, int level, ufs_lbn_t lbn, 2284 long *countp) 2285 { 2286 struct buf *bp; 2287 ufs_daddr_t *bap; 2288 ufs_daddr_t nb; 2289 struct fs *fs; 2290 struct worklist *wk; 2291 struct indirdep *indirdep; 2292 int i, lbnadd, nblocks; 2293 int error, allerror = 0; 2294 2295 fs = ip->i_fs; 2296 lbnadd = 1; 2297 for (i = level; i > 0; i--) 2298 lbnadd *= NINDIR(fs); 2299 /* 2300 * Get buffer of block pointers to be freed. This routine is not 2301 * called until the zero'ed inode has been written, so it is safe 2302 * to free blocks as they are encountered. Because the inode has 2303 * been zero'ed, calls to bmap on these blocks will fail. So, we 2304 * have to use the on-disk address and the block device for the 2305 * filesystem to look them up. If the file was deleted before its 2306 * indirect blocks were all written to disk, the routine that set 2307 * us up (deallocate_dependencies) will have arranged to leave 2308 * a complete copy of the indirect block in memory for our use. 2309 * Otherwise we have to read the blocks in from the disk. 2310 */ 2311 ACQUIRE_LOCK(&lk); 2312 if ((bp = findblk(ip->i_devvp, doffset)) != NULL && 2313 (wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2314 /* 2315 * bp must be ir_savebp, which is held locked for our use. 2316 */ 2317 if (wk->wk_type != D_INDIRDEP || 2318 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp || 2319 (indirdep->ir_state & GOINGAWAY) == 0) { 2320 FREE_LOCK(&lk); 2321 panic("indir_trunc: lost indirdep"); 2322 } 2323 WORKLIST_REMOVE(wk); 2324 WORKITEM_FREE(indirdep, D_INDIRDEP); 2325 if (LIST_FIRST(&bp->b_dep) != NULL) { 2326 FREE_LOCK(&lk); 2327 panic("indir_trunc: dangling dep"); 2328 } 2329 FREE_LOCK(&lk); 2330 } else { 2331 FREE_LOCK(&lk); 2332 error = bread(ip->i_devvp, doffset, (int)fs->fs_bsize, &bp); 2333 if (error) 2334 return (error); 2335 } 2336 /* 2337 * Recursively free indirect blocks. 2338 */ 2339 bap = (ufs_daddr_t *)bp->b_data; 2340 nblocks = btodb(fs->fs_bsize); 2341 for (i = NINDIR(fs) - 1; i >= 0; i--) { 2342 if ((nb = bap[i]) == 0) 2343 continue; 2344 if (level != 0) { 2345 if ((error = indir_trunc(ip, fsbtodoff(fs, nb), 2346 level - 1, lbn + (i * lbnadd), countp)) != 0) 2347 allerror = error; 2348 } 2349 ffs_blkfree(ip, nb, fs->fs_bsize); 2350 *countp += nblocks; 2351 } 2352 bp->b_flags |= B_INVAL | B_NOCACHE; 2353 brelse(bp); 2354 return (allerror); 2355 } 2356 2357 /* 2358 * Free an allocindir. 2359 * This routine must be called with splbio interrupts blocked. 2360 */ 2361 static void 2362 free_allocindir(struct allocindir *aip, struct inodedep *inodedep) 2363 { 2364 struct freefrag *freefrag; 2365 2366 #ifdef DEBUG 2367 if (lk.lkt_held == NOHOLDER) 2368 panic("free_allocindir: lock not held"); 2369 #endif 2370 if ((aip->ai_state & DEPCOMPLETE) == 0) 2371 LIST_REMOVE(aip, ai_deps); 2372 if (aip->ai_state & ONWORKLIST) 2373 WORKLIST_REMOVE(&aip->ai_list); 2374 LIST_REMOVE(aip, ai_next); 2375 if ((freefrag = aip->ai_freefrag) != NULL) { 2376 if (inodedep == NULL) 2377 add_to_worklist(&freefrag->ff_list); 2378 else 2379 WORKLIST_INSERT(&inodedep->id_bufwait, 2380 &freefrag->ff_list); 2381 } 2382 WORKITEM_FREE(aip, D_ALLOCINDIR); 2383 } 2384 2385 /* 2386 * Directory entry addition dependencies. 2387 * 2388 * When adding a new directory entry, the inode (with its incremented link 2389 * count) must be written to disk before the directory entry's pointer to it. 2390 * Also, if the inode is newly allocated, the corresponding freemap must be 2391 * updated (on disk) before the directory entry's pointer. These requirements 2392 * are met via undo/redo on the directory entry's pointer, which consists 2393 * simply of the inode number. 2394 * 2395 * As directory entries are added and deleted, the free space within a 2396 * directory block can become fragmented. The ufs filesystem will compact 2397 * a fragmented directory block to make space for a new entry. When this 2398 * occurs, the offsets of previously added entries change. Any "diradd" 2399 * dependency structures corresponding to these entries must be updated with 2400 * the new offsets. 2401 */ 2402 2403 /* 2404 * This routine is called after the in-memory inode's link 2405 * count has been incremented, but before the directory entry's 2406 * pointer to the inode has been set. 2407 * 2408 * Parameters: 2409 * bp: buffer containing directory block 2410 * dp: inode for directory 2411 * diroffset: offset of new entry in directory 2412 * newinum: inode referenced by new directory entry 2413 * newdirbp: non-NULL => contents of new mkdir 2414 */ 2415 void 2416 softdep_setup_directory_add(struct buf *bp, struct inode *dp, off_t diroffset, 2417 ino_t newinum, struct buf *newdirbp) 2418 { 2419 int offset; /* offset of new entry within directory block */ 2420 ufs_lbn_t lbn; /* block in directory containing new entry */ 2421 struct fs *fs; 2422 struct diradd *dap; 2423 struct pagedep *pagedep; 2424 struct inodedep *inodedep; 2425 struct mkdir *mkdir1, *mkdir2; 2426 2427 /* 2428 * Whiteouts have no dependencies. 2429 */ 2430 if (newinum == WINO) { 2431 if (newdirbp != NULL) 2432 bdwrite(newdirbp); 2433 return; 2434 } 2435 2436 fs = dp->i_fs; 2437 lbn = lblkno(fs, diroffset); 2438 offset = blkoff(fs, diroffset); 2439 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, 2440 M_SOFTDEP_FLAGS | M_ZERO); 2441 dap->da_list.wk_type = D_DIRADD; 2442 dap->da_offset = offset; 2443 dap->da_newinum = newinum; 2444 dap->da_state = ATTACHED; 2445 if (newdirbp == NULL) { 2446 dap->da_state |= DEPCOMPLETE; 2447 ACQUIRE_LOCK(&lk); 2448 } else { 2449 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 2450 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2451 M_SOFTDEP_FLAGS); 2452 mkdir1->md_list.wk_type = D_MKDIR; 2453 mkdir1->md_state = MKDIR_BODY; 2454 mkdir1->md_diradd = dap; 2455 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2456 M_SOFTDEP_FLAGS); 2457 mkdir2->md_list.wk_type = D_MKDIR; 2458 mkdir2->md_state = MKDIR_PARENT; 2459 mkdir2->md_diradd = dap; 2460 /* 2461 * Dependency on "." and ".." being written to disk. 2462 */ 2463 mkdir1->md_buf = newdirbp; 2464 ACQUIRE_LOCK(&lk); 2465 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 2466 WORKLIST_INSERT_BP(newdirbp, &mkdir1->md_list); 2467 FREE_LOCK(&lk); 2468 bdwrite(newdirbp); 2469 /* 2470 * Dependency on link count increase for parent directory 2471 */ 2472 ACQUIRE_LOCK(&lk); 2473 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0 2474 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2475 dap->da_state &= ~MKDIR_PARENT; 2476 WORKITEM_FREE(mkdir2, D_MKDIR); 2477 } else { 2478 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 2479 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 2480 } 2481 } 2482 /* 2483 * Link into parent directory pagedep to await its being written. 2484 */ 2485 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2486 WORKLIST_INSERT_BP(bp, &pagedep->pd_list); 2487 dap->da_pagedep = pagedep; 2488 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 2489 da_pdlist); 2490 /* 2491 * Link into its inodedep. Put it on the id_bufwait list if the inode 2492 * is not yet written. If it is written, do the post-inode write 2493 * processing to put it on the id_pendinghd list. 2494 */ 2495 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep); 2496 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 2497 diradd_inode_written(dap, inodedep); 2498 else 2499 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2500 FREE_LOCK(&lk); 2501 } 2502 2503 /* 2504 * This procedure is called to change the offset of a directory 2505 * entry when compacting a directory block which must be owned 2506 * exclusively by the caller. Note that the actual entry movement 2507 * must be done in this procedure to ensure that no I/O completions 2508 * occur while the move is in progress. 2509 * 2510 * Parameters: 2511 * dp: inode for directory 2512 * base: address of dp->i_offset 2513 * oldloc: address of old directory location 2514 * newloc: address of new directory location 2515 * entrysize: size of directory entry 2516 */ 2517 void 2518 softdep_change_directoryentry_offset(struct inode *dp, caddr_t base, 2519 caddr_t oldloc, caddr_t newloc, 2520 int entrysize) 2521 { 2522 int offset, oldoffset, newoffset; 2523 struct pagedep *pagedep; 2524 struct diradd *dap; 2525 ufs_lbn_t lbn; 2526 2527 ACQUIRE_LOCK(&lk); 2528 lbn = lblkno(dp->i_fs, dp->i_offset); 2529 offset = blkoff(dp->i_fs, dp->i_offset); 2530 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) 2531 goto done; 2532 oldoffset = offset + (oldloc - base); 2533 newoffset = offset + (newloc - base); 2534 2535 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) { 2536 if (dap->da_offset != oldoffset) 2537 continue; 2538 dap->da_offset = newoffset; 2539 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) 2540 break; 2541 LIST_REMOVE(dap, da_pdlist); 2542 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], 2543 dap, da_pdlist); 2544 break; 2545 } 2546 if (dap == NULL) { 2547 2548 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) { 2549 if (dap->da_offset == oldoffset) { 2550 dap->da_offset = newoffset; 2551 break; 2552 } 2553 } 2554 } 2555 done: 2556 bcopy(oldloc, newloc, entrysize); 2557 FREE_LOCK(&lk); 2558 } 2559 2560 /* 2561 * Free a diradd dependency structure. This routine must be called 2562 * with splbio interrupts blocked. 2563 */ 2564 static void 2565 free_diradd(struct diradd *dap) 2566 { 2567 struct dirrem *dirrem; 2568 struct pagedep *pagedep; 2569 struct inodedep *inodedep; 2570 struct mkdir *mkdir, *nextmd; 2571 2572 #ifdef DEBUG 2573 if (lk.lkt_held == NOHOLDER) 2574 panic("free_diradd: lock not held"); 2575 #endif 2576 WORKLIST_REMOVE(&dap->da_list); 2577 LIST_REMOVE(dap, da_pdlist); 2578 if ((dap->da_state & DIRCHG) == 0) { 2579 pagedep = dap->da_pagedep; 2580 } else { 2581 dirrem = dap->da_previous; 2582 pagedep = dirrem->dm_pagedep; 2583 dirrem->dm_dirinum = pagedep->pd_ino; 2584 add_to_worklist(&dirrem->dm_list); 2585 } 2586 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum, 2587 0, &inodedep) != 0) 2588 (void) free_inodedep(inodedep); 2589 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2590 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 2591 nextmd = LIST_NEXT(mkdir, md_mkdirs); 2592 if (mkdir->md_diradd != dap) 2593 continue; 2594 dap->da_state &= ~mkdir->md_state; 2595 WORKLIST_REMOVE(&mkdir->md_list); 2596 LIST_REMOVE(mkdir, md_mkdirs); 2597 WORKITEM_FREE(mkdir, D_MKDIR); 2598 } 2599 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2600 FREE_LOCK(&lk); 2601 panic("free_diradd: unfound ref"); 2602 } 2603 } 2604 WORKITEM_FREE(dap, D_DIRADD); 2605 } 2606 2607 /* 2608 * Directory entry removal dependencies. 2609 * 2610 * When removing a directory entry, the entry's inode pointer must be 2611 * zero'ed on disk before the corresponding inode's link count is decremented 2612 * (possibly freeing the inode for re-use). This dependency is handled by 2613 * updating the directory entry but delaying the inode count reduction until 2614 * after the directory block has been written to disk. After this point, the 2615 * inode count can be decremented whenever it is convenient. 2616 */ 2617 2618 /* 2619 * This routine should be called immediately after removing 2620 * a directory entry. The inode's link count should not be 2621 * decremented by the calling procedure -- the soft updates 2622 * code will do this task when it is safe. 2623 * 2624 * Parameters: 2625 * bp: buffer containing directory block 2626 * dp: inode for the directory being modified 2627 * ip: inode for directory entry being removed 2628 * isrmdir: indicates if doing RMDIR 2629 */ 2630 void 2631 softdep_setup_remove(struct buf *bp, struct inode *dp, struct inode *ip, 2632 int isrmdir) 2633 { 2634 struct dirrem *dirrem, *prevdirrem; 2635 2636 /* 2637 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 2638 */ 2639 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2640 2641 /* 2642 * If the COMPLETE flag is clear, then there were no active 2643 * entries and we want to roll back to a zeroed entry until 2644 * the new inode is committed to disk. If the COMPLETE flag is 2645 * set then we have deleted an entry that never made it to 2646 * disk. If the entry we deleted resulted from a name change, 2647 * then the old name still resides on disk. We cannot delete 2648 * its inode (returned to us in prevdirrem) until the zeroed 2649 * directory entry gets to disk. The new inode has never been 2650 * referenced on the disk, so can be deleted immediately. 2651 */ 2652 if ((dirrem->dm_state & COMPLETE) == 0) { 2653 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 2654 dm_next); 2655 FREE_LOCK(&lk); 2656 } else { 2657 if (prevdirrem != NULL) 2658 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 2659 prevdirrem, dm_next); 2660 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 2661 FREE_LOCK(&lk); 2662 handle_workitem_remove(dirrem); 2663 } 2664 } 2665 2666 /* 2667 * Allocate a new dirrem if appropriate and return it along with 2668 * its associated pagedep. Called without a lock, returns with lock. 2669 */ 2670 static long num_dirrem; /* number of dirrem allocated */ 2671 2672 /* 2673 * Parameters: 2674 * bp: buffer containing directory block 2675 * dp: inode for the directory being modified 2676 * ip: inode for directory entry being removed 2677 * isrmdir: indicates if doing RMDIR 2678 * prevdirremp: previously referenced inode, if any 2679 */ 2680 static struct dirrem * 2681 newdirrem(struct buf *bp, struct inode *dp, struct inode *ip, 2682 int isrmdir, struct dirrem **prevdirremp) 2683 { 2684 int offset; 2685 ufs_lbn_t lbn; 2686 struct diradd *dap; 2687 struct dirrem *dirrem; 2688 struct pagedep *pagedep; 2689 2690 /* 2691 * Whiteouts have no deletion dependencies. 2692 */ 2693 if (ip == NULL) 2694 panic("newdirrem: whiteout"); 2695 /* 2696 * If we are over our limit, try to improve the situation. 2697 * Limiting the number of dirrem structures will also limit 2698 * the number of freefile and freeblks structures. 2699 */ 2700 if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0) 2701 (void) request_cleanup(FLUSH_REMOVE, 0); 2702 num_dirrem += 1; 2703 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem), 2704 M_DIRREM, M_SOFTDEP_FLAGS | M_ZERO); 2705 dirrem->dm_list.wk_type = D_DIRREM; 2706 dirrem->dm_state = isrmdir ? RMDIR : 0; 2707 dirrem->dm_mnt = ITOV(ip)->v_mount; 2708 dirrem->dm_oldinum = ip->i_number; 2709 *prevdirremp = NULL; 2710 2711 ACQUIRE_LOCK(&lk); 2712 lbn = lblkno(dp->i_fs, dp->i_offset); 2713 offset = blkoff(dp->i_fs, dp->i_offset); 2714 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2715 WORKLIST_INSERT_BP(bp, &pagedep->pd_list); 2716 dirrem->dm_pagedep = pagedep; 2717 /* 2718 * Check for a diradd dependency for the same directory entry. 2719 * If present, then both dependencies become obsolete and can 2720 * be de-allocated. Check for an entry on both the pd_dirraddhd 2721 * list and the pd_pendinghd list. 2722 */ 2723 2724 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 2725 if (dap->da_offset == offset) 2726 break; 2727 if (dap == NULL) { 2728 2729 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 2730 if (dap->da_offset == offset) 2731 break; 2732 if (dap == NULL) 2733 return (dirrem); 2734 } 2735 /* 2736 * Must be ATTACHED at this point. 2737 */ 2738 if ((dap->da_state & ATTACHED) == 0) { 2739 FREE_LOCK(&lk); 2740 panic("newdirrem: not ATTACHED"); 2741 } 2742 if (dap->da_newinum != ip->i_number) { 2743 FREE_LOCK(&lk); 2744 panic("newdirrem: inum %"PRId64" should be %"PRId64, 2745 ip->i_number, dap->da_newinum); 2746 } 2747 /* 2748 * If we are deleting a changed name that never made it to disk, 2749 * then return the dirrem describing the previous inode (which 2750 * represents the inode currently referenced from this entry on disk). 2751 */ 2752 if ((dap->da_state & DIRCHG) != 0) { 2753 *prevdirremp = dap->da_previous; 2754 dap->da_state &= ~DIRCHG; 2755 dap->da_pagedep = pagedep; 2756 } 2757 /* 2758 * We are deleting an entry that never made it to disk. 2759 * Mark it COMPLETE so we can delete its inode immediately. 2760 */ 2761 dirrem->dm_state |= COMPLETE; 2762 free_diradd(dap); 2763 return (dirrem); 2764 } 2765 2766 /* 2767 * Directory entry change dependencies. 2768 * 2769 * Changing an existing directory entry requires that an add operation 2770 * be completed first followed by a deletion. The semantics for the addition 2771 * are identical to the description of adding a new entry above except 2772 * that the rollback is to the old inode number rather than zero. Once 2773 * the addition dependency is completed, the removal is done as described 2774 * in the removal routine above. 2775 */ 2776 2777 /* 2778 * This routine should be called immediately after changing 2779 * a directory entry. The inode's link count should not be 2780 * decremented by the calling procedure -- the soft updates 2781 * code will perform this task when it is safe. 2782 * 2783 * Parameters: 2784 * bp: buffer containing directory block 2785 * dp: inode for the directory being modified 2786 * ip: inode for directory entry being removed 2787 * newinum: new inode number for changed entry 2788 * isrmdir: indicates if doing RMDIR 2789 */ 2790 void 2791 softdep_setup_directory_change(struct buf *bp, struct inode *dp, 2792 struct inode *ip, ino_t newinum, 2793 int isrmdir) 2794 { 2795 int offset; 2796 struct diradd *dap = NULL; 2797 struct dirrem *dirrem, *prevdirrem; 2798 struct pagedep *pagedep; 2799 struct inodedep *inodedep; 2800 2801 offset = blkoff(dp->i_fs, dp->i_offset); 2802 2803 /* 2804 * Whiteouts do not need diradd dependencies. 2805 */ 2806 if (newinum != WINO) { 2807 MALLOC(dap, struct diradd *, sizeof(struct diradd), 2808 M_DIRADD, M_SOFTDEP_FLAGS | M_ZERO); 2809 dap->da_list.wk_type = D_DIRADD; 2810 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 2811 dap->da_offset = offset; 2812 dap->da_newinum = newinum; 2813 } 2814 2815 /* 2816 * Allocate a new dirrem and ACQUIRE_LOCK. 2817 */ 2818 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2819 pagedep = dirrem->dm_pagedep; 2820 /* 2821 * The possible values for isrmdir: 2822 * 0 - non-directory file rename 2823 * 1 - directory rename within same directory 2824 * inum - directory rename to new directory of given inode number 2825 * When renaming to a new directory, we are both deleting and 2826 * creating a new directory entry, so the link count on the new 2827 * directory should not change. Thus we do not need the followup 2828 * dirrem which is usually done in handle_workitem_remove. We set 2829 * the DIRCHG flag to tell handle_workitem_remove to skip the 2830 * followup dirrem. 2831 */ 2832 if (isrmdir > 1) 2833 dirrem->dm_state |= DIRCHG; 2834 2835 /* 2836 * Whiteouts have no additional dependencies, 2837 * so just put the dirrem on the correct list. 2838 */ 2839 if (newinum == WINO) { 2840 if ((dirrem->dm_state & COMPLETE) == 0) { 2841 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 2842 dm_next); 2843 } else { 2844 dirrem->dm_dirinum = pagedep->pd_ino; 2845 add_to_worklist(&dirrem->dm_list); 2846 } 2847 FREE_LOCK(&lk); 2848 return; 2849 } 2850 2851 /* 2852 * If the COMPLETE flag is clear, then there were no active 2853 * entries and we want to roll back to the previous inode until 2854 * the new inode is committed to disk. If the COMPLETE flag is 2855 * set, then we have deleted an entry that never made it to disk. 2856 * If the entry we deleted resulted from a name change, then the old 2857 * inode reference still resides on disk. Any rollback that we do 2858 * needs to be to that old inode (returned to us in prevdirrem). If 2859 * the entry we deleted resulted from a create, then there is 2860 * no entry on the disk, so we want to roll back to zero rather 2861 * than the uncommitted inode. In either of the COMPLETE cases we 2862 * want to immediately free the unwritten and unreferenced inode. 2863 */ 2864 if ((dirrem->dm_state & COMPLETE) == 0) { 2865 dap->da_previous = dirrem; 2866 } else { 2867 if (prevdirrem != NULL) { 2868 dap->da_previous = prevdirrem; 2869 } else { 2870 dap->da_state &= ~DIRCHG; 2871 dap->da_pagedep = pagedep; 2872 } 2873 dirrem->dm_dirinum = pagedep->pd_ino; 2874 add_to_worklist(&dirrem->dm_list); 2875 } 2876 /* 2877 * Link into its inodedep. Put it on the id_bufwait list if the inode 2878 * is not yet written. If it is written, do the post-inode write 2879 * processing to put it on the id_pendinghd list. 2880 */ 2881 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 || 2882 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2883 dap->da_state |= COMPLETE; 2884 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 2885 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 2886 } else { 2887 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 2888 dap, da_pdlist); 2889 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2890 } 2891 FREE_LOCK(&lk); 2892 } 2893 2894 /* 2895 * Called whenever the link count on an inode is changed. 2896 * It creates an inode dependency so that the new reference(s) 2897 * to the inode cannot be committed to disk until the updated 2898 * inode has been written. 2899 * 2900 * Parameters: 2901 * ip: the inode with the increased link count 2902 */ 2903 void 2904 softdep_change_linkcnt(struct inode *ip) 2905 { 2906 struct inodedep *inodedep; 2907 2908 ACQUIRE_LOCK(&lk); 2909 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 2910 if (ip->i_nlink < ip->i_effnlink) { 2911 FREE_LOCK(&lk); 2912 panic("softdep_change_linkcnt: bad delta"); 2913 } 2914 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2915 FREE_LOCK(&lk); 2916 } 2917 2918 /* 2919 * This workitem decrements the inode's link count. 2920 * If the link count reaches zero, the file is removed. 2921 */ 2922 static void 2923 handle_workitem_remove(struct dirrem *dirrem) 2924 { 2925 struct inodedep *inodedep; 2926 struct vnode *vp; 2927 struct inode *ip; 2928 ino_t oldinum; 2929 int error; 2930 2931 if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) { 2932 softdep_error("handle_workitem_remove: vget", error); 2933 return; 2934 } 2935 ip = VTOI(vp); 2936 ACQUIRE_LOCK(&lk); 2937 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){ 2938 FREE_LOCK(&lk); 2939 panic("handle_workitem_remove: lost inodedep"); 2940 } 2941 /* 2942 * Normal file deletion. 2943 */ 2944 if ((dirrem->dm_state & RMDIR) == 0) { 2945 ip->i_nlink--; 2946 ip->i_flag |= IN_CHANGE; 2947 if (ip->i_nlink < ip->i_effnlink) { 2948 FREE_LOCK(&lk); 2949 panic("handle_workitem_remove: bad file delta"); 2950 } 2951 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2952 FREE_LOCK(&lk); 2953 vput(vp); 2954 num_dirrem -= 1; 2955 WORKITEM_FREE(dirrem, D_DIRREM); 2956 return; 2957 } 2958 /* 2959 * Directory deletion. Decrement reference count for both the 2960 * just deleted parent directory entry and the reference for ".". 2961 * Next truncate the directory to length zero. When the 2962 * truncation completes, arrange to have the reference count on 2963 * the parent decremented to account for the loss of "..". 2964 */ 2965 ip->i_nlink -= 2; 2966 ip->i_flag |= IN_CHANGE; 2967 if (ip->i_nlink < ip->i_effnlink) { 2968 FREE_LOCK(&lk); 2969 panic("handle_workitem_remove: bad dir delta"); 2970 } 2971 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2972 FREE_LOCK(&lk); 2973 if ((error = ffs_truncate(vp, (off_t)0, 0, proc0.p_ucred)) != 0) 2974 softdep_error("handle_workitem_remove: truncate", error); 2975 /* 2976 * Rename a directory to a new parent. Since, we are both deleting 2977 * and creating a new directory entry, the link count on the new 2978 * directory should not change. Thus we skip the followup dirrem. 2979 */ 2980 if (dirrem->dm_state & DIRCHG) { 2981 vput(vp); 2982 num_dirrem -= 1; 2983 WORKITEM_FREE(dirrem, D_DIRREM); 2984 return; 2985 } 2986 /* 2987 * If the inodedep does not exist, then the zero'ed inode has 2988 * been written to disk. If the allocated inode has never been 2989 * written to disk, then the on-disk inode is zero'ed. In either 2990 * case we can remove the file immediately. 2991 */ 2992 ACQUIRE_LOCK(&lk); 2993 dirrem->dm_state = 0; 2994 oldinum = dirrem->dm_oldinum; 2995 dirrem->dm_oldinum = dirrem->dm_dirinum; 2996 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 || 2997 check_inode_unwritten(inodedep)) { 2998 FREE_LOCK(&lk); 2999 vput(vp); 3000 handle_workitem_remove(dirrem); 3001 return; 3002 } 3003 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 3004 FREE_LOCK(&lk); 3005 ip->i_flag |= IN_CHANGE; 3006 ffs_update(vp, 0); 3007 vput(vp); 3008 } 3009 3010 /* 3011 * Inode de-allocation dependencies. 3012 * 3013 * When an inode's link count is reduced to zero, it can be de-allocated. We 3014 * found it convenient to postpone de-allocation until after the inode is 3015 * written to disk with its new link count (zero). At this point, all of the 3016 * on-disk inode's block pointers are nullified and, with careful dependency 3017 * list ordering, all dependencies related to the inode will be satisfied and 3018 * the corresponding dependency structures de-allocated. So, if/when the 3019 * inode is reused, there will be no mixing of old dependencies with new 3020 * ones. This artificial dependency is set up by the block de-allocation 3021 * procedure above (softdep_setup_freeblocks) and completed by the 3022 * following procedure. 3023 */ 3024 static void 3025 handle_workitem_freefile(struct freefile *freefile) 3026 { 3027 struct vnode vp; 3028 struct inode tip; 3029 struct inodedep *idp; 3030 int error; 3031 3032 #ifdef DEBUG 3033 ACQUIRE_LOCK(&lk); 3034 error = inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp); 3035 FREE_LOCK(&lk); 3036 if (error) 3037 panic("handle_workitem_freefile: inodedep survived"); 3038 #endif 3039 tip.i_devvp = freefile->fx_devvp; 3040 tip.i_dev = freefile->fx_devvp->v_rdev; 3041 tip.i_fs = freefile->fx_fs; 3042 vp.v_data = &tip; 3043 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0) 3044 softdep_error("handle_workitem_freefile", error); 3045 WORKITEM_FREE(freefile, D_FREEFILE); 3046 } 3047 3048 /* 3049 * Helper function which unlinks marker element from work list and returns 3050 * the next element on the list. 3051 */ 3052 static __inline struct worklist * 3053 markernext(struct worklist *marker) 3054 { 3055 struct worklist *next; 3056 3057 next = LIST_NEXT(marker, wk_list); 3058 LIST_REMOVE(marker, wk_list); 3059 return next; 3060 } 3061 3062 /* 3063 * checkread, checkwrite 3064 * 3065 */ 3066 static int 3067 softdep_checkread(struct buf *bp) 3068 { 3069 return(0); 3070 } 3071 3072 static int 3073 softdep_checkwrite(struct buf *bp) 3074 { 3075 return(0); 3076 } 3077 3078 /* 3079 * Disk writes. 3080 * 3081 * The dependency structures constructed above are most actively used when file 3082 * system blocks are written to disk. No constraints are placed on when a 3083 * block can be written, but unsatisfied update dependencies are made safe by 3084 * modifying (or replacing) the source memory for the duration of the disk 3085 * write. When the disk write completes, the memory block is again brought 3086 * up-to-date. 3087 * 3088 * In-core inode structure reclamation. 3089 * 3090 * Because there are a finite number of "in-core" inode structures, they are 3091 * reused regularly. By transferring all inode-related dependencies to the 3092 * in-memory inode block and indexing them separately (via "inodedep"s), we 3093 * can allow "in-core" inode structures to be reused at any time and avoid 3094 * any increase in contention. 3095 * 3096 * Called just before entering the device driver to initiate a new disk I/O. 3097 * The buffer must be locked, thus, no I/O completion operations can occur 3098 * while we are manipulating its associated dependencies. 3099 * 3100 * Parameters: 3101 * bp: structure describing disk write to occur 3102 */ 3103 static void 3104 softdep_disk_io_initiation(struct buf *bp) 3105 { 3106 struct worklist *wk; 3107 struct worklist marker; 3108 struct indirdep *indirdep; 3109 3110 /* 3111 * We only care about write operations. There should never 3112 * be dependencies for reads. 3113 */ 3114 if (bp->b_cmd == BUF_CMD_READ) 3115 panic("softdep_disk_io_initiation: read"); 3116 3117 marker.wk_type = D_LAST + 1; /* Not a normal workitem */ 3118 3119 /* 3120 * Do any necessary pre-I/O processing. 3121 */ 3122 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = markernext(&marker)) { 3123 LIST_INSERT_AFTER(wk, &marker, wk_list); 3124 3125 switch (wk->wk_type) { 3126 3127 case D_PAGEDEP: 3128 initiate_write_filepage(WK_PAGEDEP(wk), bp); 3129 continue; 3130 3131 case D_INODEDEP: 3132 initiate_write_inodeblock(WK_INODEDEP(wk), bp); 3133 continue; 3134 3135 case D_INDIRDEP: 3136 indirdep = WK_INDIRDEP(wk); 3137 if (indirdep->ir_state & GOINGAWAY) 3138 panic("disk_io_initiation: indirdep gone"); 3139 /* 3140 * If there are no remaining dependencies, this 3141 * will be writing the real pointers, so the 3142 * dependency can be freed. 3143 */ 3144 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) { 3145 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 3146 brelse(indirdep->ir_savebp); 3147 /* inline expand WORKLIST_REMOVE(wk); */ 3148 wk->wk_state &= ~ONWORKLIST; 3149 LIST_REMOVE(wk, wk_list); 3150 WORKITEM_FREE(indirdep, D_INDIRDEP); 3151 continue; 3152 } 3153 /* 3154 * Replace up-to-date version with safe version. 3155 */ 3156 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount, 3157 M_INDIRDEP, M_SOFTDEP_FLAGS); 3158 ACQUIRE_LOCK(&lk); 3159 indirdep->ir_state &= ~ATTACHED; 3160 indirdep->ir_state |= UNDONE; 3161 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 3162 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 3163 bp->b_bcount); 3164 FREE_LOCK(&lk); 3165 continue; 3166 3167 case D_MKDIR: 3168 case D_BMSAFEMAP: 3169 case D_ALLOCDIRECT: 3170 case D_ALLOCINDIR: 3171 continue; 3172 3173 default: 3174 panic("handle_disk_io_initiation: Unexpected type %s", 3175 TYPENAME(wk->wk_type)); 3176 /* NOTREACHED */ 3177 } 3178 } 3179 } 3180 3181 /* 3182 * Called from within the procedure above to deal with unsatisfied 3183 * allocation dependencies in a directory. The buffer must be locked, 3184 * thus, no I/O completion operations can occur while we are 3185 * manipulating its associated dependencies. 3186 */ 3187 static void 3188 initiate_write_filepage(struct pagedep *pagedep, struct buf *bp) 3189 { 3190 struct diradd *dap; 3191 struct direct *ep; 3192 int i; 3193 3194 if (pagedep->pd_state & IOSTARTED) { 3195 /* 3196 * This can only happen if there is a driver that does not 3197 * understand chaining. Here biodone will reissue the call 3198 * to strategy for the incomplete buffers. 3199 */ 3200 kprintf("initiate_write_filepage: already started\n"); 3201 return; 3202 } 3203 pagedep->pd_state |= IOSTARTED; 3204 ACQUIRE_LOCK(&lk); 3205 for (i = 0; i < DAHASHSZ; i++) { 3206 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 3207 ep = (struct direct *) 3208 ((char *)bp->b_data + dap->da_offset); 3209 if (ep->d_ino != dap->da_newinum) { 3210 FREE_LOCK(&lk); 3211 panic("%s: dir inum %d != new %"PRId64, 3212 "initiate_write_filepage", 3213 ep->d_ino, dap->da_newinum); 3214 } 3215 if (dap->da_state & DIRCHG) 3216 ep->d_ino = dap->da_previous->dm_oldinum; 3217 else 3218 ep->d_ino = 0; 3219 dap->da_state &= ~ATTACHED; 3220 dap->da_state |= UNDONE; 3221 } 3222 } 3223 FREE_LOCK(&lk); 3224 } 3225 3226 /* 3227 * Called from within the procedure above to deal with unsatisfied 3228 * allocation dependencies in an inodeblock. The buffer must be 3229 * locked, thus, no I/O completion operations can occur while we 3230 * are manipulating its associated dependencies. 3231 * 3232 * Parameters: 3233 * bp: The inode block 3234 */ 3235 static void 3236 initiate_write_inodeblock(struct inodedep *inodedep, struct buf *bp) 3237 { 3238 struct allocdirect *adp, *lastadp; 3239 struct ufs1_dinode *dp; 3240 struct ufs1_dinode *sip; 3241 struct fs *fs; 3242 ufs_lbn_t prevlbn = 0; 3243 int i, deplist; 3244 3245 if (inodedep->id_state & IOSTARTED) 3246 panic("initiate_write_inodeblock: already started"); 3247 inodedep->id_state |= IOSTARTED; 3248 fs = inodedep->id_fs; 3249 dp = (struct ufs1_dinode *)bp->b_data + 3250 ino_to_fsbo(fs, inodedep->id_ino); 3251 /* 3252 * If the bitmap is not yet written, then the allocated 3253 * inode cannot be written to disk. 3254 */ 3255 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 3256 if (inodedep->id_savedino != NULL) 3257 panic("initiate_write_inodeblock: already doing I/O"); 3258 MALLOC(sip, struct ufs1_dinode *, 3259 sizeof(struct ufs1_dinode), M_INODEDEP, M_SOFTDEP_FLAGS); 3260 inodedep->id_savedino = sip; 3261 *inodedep->id_savedino = *dp; 3262 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 3263 dp->di_gen = inodedep->id_savedino->di_gen; 3264 return; 3265 } 3266 /* 3267 * If no dependencies, then there is nothing to roll back. 3268 */ 3269 inodedep->id_savedsize = dp->di_size; 3270 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 3271 return; 3272 /* 3273 * Set the dependencies to busy. 3274 */ 3275 ACQUIRE_LOCK(&lk); 3276 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3277 adp = TAILQ_NEXT(adp, ad_next)) { 3278 #ifdef DIAGNOSTIC 3279 if (deplist != 0 && prevlbn >= adp->ad_lbn) { 3280 FREE_LOCK(&lk); 3281 panic("softdep_write_inodeblock: lbn order"); 3282 } 3283 prevlbn = adp->ad_lbn; 3284 if (adp->ad_lbn < NDADDR && 3285 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) { 3286 FREE_LOCK(&lk); 3287 panic("%s: direct pointer #%ld mismatch %d != %d", 3288 "softdep_write_inodeblock", adp->ad_lbn, 3289 dp->di_db[adp->ad_lbn], adp->ad_newblkno); 3290 } 3291 if (adp->ad_lbn >= NDADDR && 3292 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) { 3293 FREE_LOCK(&lk); 3294 panic("%s: indirect pointer #%ld mismatch %d != %d", 3295 "softdep_write_inodeblock", adp->ad_lbn - NDADDR, 3296 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno); 3297 } 3298 deplist |= 1 << adp->ad_lbn; 3299 if ((adp->ad_state & ATTACHED) == 0) { 3300 FREE_LOCK(&lk); 3301 panic("softdep_write_inodeblock: Unknown state 0x%x", 3302 adp->ad_state); 3303 } 3304 #endif /* DIAGNOSTIC */ 3305 adp->ad_state &= ~ATTACHED; 3306 adp->ad_state |= UNDONE; 3307 } 3308 /* 3309 * The on-disk inode cannot claim to be any larger than the last 3310 * fragment that has been written. Otherwise, the on-disk inode 3311 * might have fragments that were not the last block in the file 3312 * which would corrupt the filesystem. 3313 */ 3314 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3315 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 3316 if (adp->ad_lbn >= NDADDR) 3317 break; 3318 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 3319 /* keep going until hitting a rollback to a frag */ 3320 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 3321 continue; 3322 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 3323 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 3324 #ifdef DIAGNOSTIC 3325 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) { 3326 FREE_LOCK(&lk); 3327 panic("softdep_write_inodeblock: lost dep1"); 3328 } 3329 #endif /* DIAGNOSTIC */ 3330 dp->di_db[i] = 0; 3331 } 3332 for (i = 0; i < NIADDR; i++) { 3333 #ifdef DIAGNOSTIC 3334 if (dp->di_ib[i] != 0 && 3335 (deplist & ((1 << NDADDR) << i)) == 0) { 3336 FREE_LOCK(&lk); 3337 panic("softdep_write_inodeblock: lost dep2"); 3338 } 3339 #endif /* DIAGNOSTIC */ 3340 dp->di_ib[i] = 0; 3341 } 3342 FREE_LOCK(&lk); 3343 return; 3344 } 3345 /* 3346 * If we have zero'ed out the last allocated block of the file, 3347 * roll back the size to the last currently allocated block. 3348 * We know that this last allocated block is a full-sized as 3349 * we already checked for fragments in the loop above. 3350 */ 3351 if (lastadp != NULL && 3352 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 3353 for (i = lastadp->ad_lbn; i >= 0; i--) 3354 if (dp->di_db[i] != 0) 3355 break; 3356 dp->di_size = (i + 1) * fs->fs_bsize; 3357 } 3358 /* 3359 * The only dependencies are for indirect blocks. 3360 * 3361 * The file size for indirect block additions is not guaranteed. 3362 * Such a guarantee would be non-trivial to achieve. The conventional 3363 * synchronous write implementation also does not make this guarantee. 3364 * Fsck should catch and fix discrepancies. Arguably, the file size 3365 * can be over-estimated without destroying integrity when the file 3366 * moves into the indirect blocks (i.e., is large). If we want to 3367 * postpone fsck, we are stuck with this argument. 3368 */ 3369 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 3370 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 3371 FREE_LOCK(&lk); 3372 } 3373 3374 /* 3375 * This routine is called during the completion interrupt 3376 * service routine for a disk write (from the procedure called 3377 * by the device driver to inform the filesystem caches of 3378 * a request completion). It should be called early in this 3379 * procedure, before the block is made available to other 3380 * processes or other routines are called. 3381 * 3382 * Parameters: 3383 * bp: describes the completed disk write 3384 */ 3385 static void 3386 softdep_disk_write_complete(struct buf *bp) 3387 { 3388 struct worklist *wk; 3389 struct workhead reattach; 3390 struct newblk *newblk; 3391 struct allocindir *aip; 3392 struct allocdirect *adp; 3393 struct indirdep *indirdep; 3394 struct inodedep *inodedep; 3395 struct bmsafemap *bmsafemap; 3396 3397 #ifdef DEBUG 3398 if (lk.lkt_held != NOHOLDER) 3399 panic("softdep_disk_write_complete: lock is held"); 3400 lk.lkt_held = SPECIAL_FLAG; 3401 #endif 3402 LIST_INIT(&reattach); 3403 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 3404 WORKLIST_REMOVE(wk); 3405 switch (wk->wk_type) { 3406 3407 case D_PAGEDEP: 3408 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 3409 WORKLIST_INSERT(&reattach, wk); 3410 continue; 3411 3412 case D_INODEDEP: 3413 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 3414 WORKLIST_INSERT(&reattach, wk); 3415 continue; 3416 3417 case D_BMSAFEMAP: 3418 bmsafemap = WK_BMSAFEMAP(wk); 3419 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) { 3420 newblk->nb_state |= DEPCOMPLETE; 3421 newblk->nb_bmsafemap = NULL; 3422 LIST_REMOVE(newblk, nb_deps); 3423 } 3424 while ((adp = 3425 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) { 3426 adp->ad_state |= DEPCOMPLETE; 3427 adp->ad_buf = NULL; 3428 LIST_REMOVE(adp, ad_deps); 3429 handle_allocdirect_partdone(adp); 3430 } 3431 while ((aip = 3432 LIST_FIRST(&bmsafemap->sm_allocindirhd))) { 3433 aip->ai_state |= DEPCOMPLETE; 3434 aip->ai_buf = NULL; 3435 LIST_REMOVE(aip, ai_deps); 3436 handle_allocindir_partdone(aip); 3437 } 3438 while ((inodedep = 3439 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { 3440 inodedep->id_state |= DEPCOMPLETE; 3441 LIST_REMOVE(inodedep, id_deps); 3442 inodedep->id_buf = NULL; 3443 } 3444 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 3445 continue; 3446 3447 case D_MKDIR: 3448 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 3449 continue; 3450 3451 case D_ALLOCDIRECT: 3452 adp = WK_ALLOCDIRECT(wk); 3453 adp->ad_state |= COMPLETE; 3454 handle_allocdirect_partdone(adp); 3455 continue; 3456 3457 case D_ALLOCINDIR: 3458 aip = WK_ALLOCINDIR(wk); 3459 aip->ai_state |= COMPLETE; 3460 handle_allocindir_partdone(aip); 3461 continue; 3462 3463 case D_INDIRDEP: 3464 indirdep = WK_INDIRDEP(wk); 3465 if (indirdep->ir_state & GOINGAWAY) { 3466 lk.lkt_held = NOHOLDER; 3467 panic("disk_write_complete: indirdep gone"); 3468 } 3469 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 3470 FREE(indirdep->ir_saveddata, M_INDIRDEP); 3471 indirdep->ir_saveddata = 0; 3472 indirdep->ir_state &= ~UNDONE; 3473 indirdep->ir_state |= ATTACHED; 3474 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 3475 handle_allocindir_partdone(aip); 3476 if (aip == LIST_FIRST(&indirdep->ir_donehd)) { 3477 lk.lkt_held = NOHOLDER; 3478 panic("disk_write_complete: not gone"); 3479 } 3480 } 3481 WORKLIST_INSERT(&reattach, wk); 3482 if ((bp->b_flags & B_DELWRI) == 0) 3483 stat_indir_blk_ptrs++; 3484 bdirty(bp); 3485 continue; 3486 3487 default: 3488 lk.lkt_held = NOHOLDER; 3489 panic("handle_disk_write_complete: Unknown type %s", 3490 TYPENAME(wk->wk_type)); 3491 /* NOTREACHED */ 3492 } 3493 } 3494 /* 3495 * Reattach any requests that must be redone. 3496 */ 3497 while ((wk = LIST_FIRST(&reattach)) != NULL) { 3498 WORKLIST_REMOVE(wk); 3499 WORKLIST_INSERT_BP(bp, wk); 3500 } 3501 #ifdef DEBUG 3502 if (lk.lkt_held != SPECIAL_FLAG) 3503 panic("softdep_disk_write_complete: lock lost"); 3504 lk.lkt_held = NOHOLDER; 3505 #endif 3506 } 3507 3508 /* 3509 * Called from within softdep_disk_write_complete above. Note that 3510 * this routine is always called from interrupt level with further 3511 * splbio interrupts blocked. 3512 * 3513 * Parameters: 3514 * adp: the completed allocdirect 3515 */ 3516 static void 3517 handle_allocdirect_partdone(struct allocdirect *adp) 3518 { 3519 struct allocdirect *listadp; 3520 struct inodedep *inodedep; 3521 long bsize; 3522 3523 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3524 return; 3525 if (adp->ad_buf != NULL) { 3526 lk.lkt_held = NOHOLDER; 3527 panic("handle_allocdirect_partdone: dangling dep"); 3528 } 3529 /* 3530 * The on-disk inode cannot claim to be any larger than the last 3531 * fragment that has been written. Otherwise, the on-disk inode 3532 * might have fragments that were not the last block in the file 3533 * which would corrupt the filesystem. Thus, we cannot free any 3534 * allocdirects after one whose ad_oldblkno claims a fragment as 3535 * these blocks must be rolled back to zero before writing the inode. 3536 * We check the currently active set of allocdirects in id_inoupdt. 3537 */ 3538 inodedep = adp->ad_inodedep; 3539 bsize = inodedep->id_fs->fs_bsize; 3540 TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) { 3541 /* found our block */ 3542 if (listadp == adp) 3543 break; 3544 /* continue if ad_oldlbn is not a fragment */ 3545 if (listadp->ad_oldsize == 0 || 3546 listadp->ad_oldsize == bsize) 3547 continue; 3548 /* hit a fragment */ 3549 return; 3550 } 3551 /* 3552 * If we have reached the end of the current list without 3553 * finding the just finished dependency, then it must be 3554 * on the future dependency list. Future dependencies cannot 3555 * be freed until they are moved to the current list. 3556 */ 3557 if (listadp == NULL) { 3558 #ifdef DEBUG 3559 TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next) 3560 /* found our block */ 3561 if (listadp == adp) 3562 break; 3563 if (listadp == NULL) { 3564 lk.lkt_held = NOHOLDER; 3565 panic("handle_allocdirect_partdone: lost dep"); 3566 } 3567 #endif /* DEBUG */ 3568 return; 3569 } 3570 /* 3571 * If we have found the just finished dependency, then free 3572 * it along with anything that follows it that is complete. 3573 */ 3574 for (; adp; adp = listadp) { 3575 listadp = TAILQ_NEXT(adp, ad_next); 3576 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3577 return; 3578 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 3579 } 3580 } 3581 3582 /* 3583 * Called from within softdep_disk_write_complete above. Note that 3584 * this routine is always called from interrupt level with further 3585 * splbio interrupts blocked. 3586 * 3587 * Parameters: 3588 * aip: the completed allocindir 3589 */ 3590 static void 3591 handle_allocindir_partdone(struct allocindir *aip) 3592 { 3593 struct indirdep *indirdep; 3594 3595 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 3596 return; 3597 if (aip->ai_buf != NULL) { 3598 lk.lkt_held = NOHOLDER; 3599 panic("handle_allocindir_partdone: dangling dependency"); 3600 } 3601 indirdep = aip->ai_indirdep; 3602 if (indirdep->ir_state & UNDONE) { 3603 LIST_REMOVE(aip, ai_next); 3604 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 3605 return; 3606 } 3607 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 3608 aip->ai_newblkno; 3609 LIST_REMOVE(aip, ai_next); 3610 if (aip->ai_freefrag != NULL) 3611 add_to_worklist(&aip->ai_freefrag->ff_list); 3612 WORKITEM_FREE(aip, D_ALLOCINDIR); 3613 } 3614 3615 /* 3616 * Called from within softdep_disk_write_complete above to restore 3617 * in-memory inode block contents to their most up-to-date state. Note 3618 * that this routine is always called from interrupt level with further 3619 * splbio interrupts blocked. 3620 * 3621 * Parameters: 3622 * bp: buffer containing the inode block 3623 */ 3624 static int 3625 handle_written_inodeblock(struct inodedep *inodedep, struct buf *bp) 3626 { 3627 struct worklist *wk, *filefree; 3628 struct allocdirect *adp, *nextadp; 3629 struct ufs1_dinode *dp; 3630 int hadchanges; 3631 3632 if ((inodedep->id_state & IOSTARTED) == 0) { 3633 lk.lkt_held = NOHOLDER; 3634 panic("handle_written_inodeblock: not started"); 3635 } 3636 inodedep->id_state &= ~IOSTARTED; 3637 dp = (struct ufs1_dinode *)bp->b_data + 3638 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 3639 /* 3640 * If we had to rollback the inode allocation because of 3641 * bitmaps being incomplete, then simply restore it. 3642 * Keep the block dirty so that it will not be reclaimed until 3643 * all associated dependencies have been cleared and the 3644 * corresponding updates written to disk. 3645 */ 3646 if (inodedep->id_savedino != NULL) { 3647 *dp = *inodedep->id_savedino; 3648 FREE(inodedep->id_savedino, M_INODEDEP); 3649 inodedep->id_savedino = NULL; 3650 if ((bp->b_flags & B_DELWRI) == 0) 3651 stat_inode_bitmap++; 3652 bdirty(bp); 3653 return (1); 3654 } 3655 inodedep->id_state |= COMPLETE; 3656 /* 3657 * Roll forward anything that had to be rolled back before 3658 * the inode could be updated. 3659 */ 3660 hadchanges = 0; 3661 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 3662 nextadp = TAILQ_NEXT(adp, ad_next); 3663 if (adp->ad_state & ATTACHED) { 3664 lk.lkt_held = NOHOLDER; 3665 panic("handle_written_inodeblock: new entry"); 3666 } 3667 if (adp->ad_lbn < NDADDR) { 3668 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) { 3669 lk.lkt_held = NOHOLDER; 3670 panic("%s: %s #%ld mismatch %d != %d", 3671 "handle_written_inodeblock", 3672 "direct pointer", adp->ad_lbn, 3673 dp->di_db[adp->ad_lbn], adp->ad_oldblkno); 3674 } 3675 dp->di_db[adp->ad_lbn] = adp->ad_newblkno; 3676 } else { 3677 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) { 3678 lk.lkt_held = NOHOLDER; 3679 panic("%s: %s #%ld allocated as %d", 3680 "handle_written_inodeblock", 3681 "indirect pointer", adp->ad_lbn - NDADDR, 3682 dp->di_ib[adp->ad_lbn - NDADDR]); 3683 } 3684 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno; 3685 } 3686 adp->ad_state &= ~UNDONE; 3687 adp->ad_state |= ATTACHED; 3688 hadchanges = 1; 3689 } 3690 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 3691 stat_direct_blk_ptrs++; 3692 /* 3693 * Reset the file size to its most up-to-date value. 3694 */ 3695 if (inodedep->id_savedsize == -1) { 3696 lk.lkt_held = NOHOLDER; 3697 panic("handle_written_inodeblock: bad size"); 3698 } 3699 if (dp->di_size != inodedep->id_savedsize) { 3700 dp->di_size = inodedep->id_savedsize; 3701 hadchanges = 1; 3702 } 3703 inodedep->id_savedsize = -1; 3704 /* 3705 * If there were any rollbacks in the inode block, then it must be 3706 * marked dirty so that its will eventually get written back in 3707 * its correct form. 3708 */ 3709 if (hadchanges) 3710 bdirty(bp); 3711 /* 3712 * Process any allocdirects that completed during the update. 3713 */ 3714 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 3715 handle_allocdirect_partdone(adp); 3716 /* 3717 * Process deallocations that were held pending until the 3718 * inode had been written to disk. Freeing of the inode 3719 * is delayed until after all blocks have been freed to 3720 * avoid creation of new <vfsid, inum, lbn> triples 3721 * before the old ones have been deleted. 3722 */ 3723 filefree = NULL; 3724 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 3725 WORKLIST_REMOVE(wk); 3726 switch (wk->wk_type) { 3727 3728 case D_FREEFILE: 3729 /* 3730 * We defer adding filefree to the worklist until 3731 * all other additions have been made to ensure 3732 * that it will be done after all the old blocks 3733 * have been freed. 3734 */ 3735 if (filefree != NULL) { 3736 lk.lkt_held = NOHOLDER; 3737 panic("handle_written_inodeblock: filefree"); 3738 } 3739 filefree = wk; 3740 continue; 3741 3742 case D_MKDIR: 3743 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 3744 continue; 3745 3746 case D_DIRADD: 3747 diradd_inode_written(WK_DIRADD(wk), inodedep); 3748 continue; 3749 3750 case D_FREEBLKS: 3751 wk->wk_state |= COMPLETE; 3752 if ((wk->wk_state & ALLCOMPLETE) != ALLCOMPLETE) 3753 continue; 3754 /* -- fall through -- */ 3755 case D_FREEFRAG: 3756 case D_DIRREM: 3757 add_to_worklist(wk); 3758 continue; 3759 3760 default: 3761 lk.lkt_held = NOHOLDER; 3762 panic("handle_written_inodeblock: Unknown type %s", 3763 TYPENAME(wk->wk_type)); 3764 /* NOTREACHED */ 3765 } 3766 } 3767 if (filefree != NULL) { 3768 if (free_inodedep(inodedep) == 0) { 3769 lk.lkt_held = NOHOLDER; 3770 panic("handle_written_inodeblock: live inodedep"); 3771 } 3772 add_to_worklist(filefree); 3773 return (0); 3774 } 3775 3776 /* 3777 * If no outstanding dependencies, free it. 3778 */ 3779 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0) 3780 return (0); 3781 return (hadchanges); 3782 } 3783 3784 /* 3785 * Process a diradd entry after its dependent inode has been written. 3786 * This routine must be called with splbio interrupts blocked. 3787 */ 3788 static void 3789 diradd_inode_written(struct diradd *dap, struct inodedep *inodedep) 3790 { 3791 struct pagedep *pagedep; 3792 3793 dap->da_state |= COMPLETE; 3794 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3795 if (dap->da_state & DIRCHG) 3796 pagedep = dap->da_previous->dm_pagedep; 3797 else 3798 pagedep = dap->da_pagedep; 3799 LIST_REMOVE(dap, da_pdlist); 3800 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3801 } 3802 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 3803 } 3804 3805 /* 3806 * Handle the completion of a mkdir dependency. 3807 */ 3808 static void 3809 handle_written_mkdir(struct mkdir *mkdir, int type) 3810 { 3811 struct diradd *dap; 3812 struct pagedep *pagedep; 3813 3814 if (mkdir->md_state != type) { 3815 lk.lkt_held = NOHOLDER; 3816 panic("handle_written_mkdir: bad type"); 3817 } 3818 dap = mkdir->md_diradd; 3819 dap->da_state &= ~type; 3820 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 3821 dap->da_state |= DEPCOMPLETE; 3822 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3823 if (dap->da_state & DIRCHG) 3824 pagedep = dap->da_previous->dm_pagedep; 3825 else 3826 pagedep = dap->da_pagedep; 3827 LIST_REMOVE(dap, da_pdlist); 3828 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3829 } 3830 LIST_REMOVE(mkdir, md_mkdirs); 3831 WORKITEM_FREE(mkdir, D_MKDIR); 3832 } 3833 3834 /* 3835 * Called from within softdep_disk_write_complete above. 3836 * A write operation was just completed. Removed inodes can 3837 * now be freed and associated block pointers may be committed. 3838 * Note that this routine is always called from interrupt level 3839 * with further splbio interrupts blocked. 3840 * 3841 * Parameters: 3842 * bp: buffer containing the written page 3843 */ 3844 static int 3845 handle_written_filepage(struct pagedep *pagedep, struct buf *bp) 3846 { 3847 struct dirrem *dirrem; 3848 struct diradd *dap, *nextdap; 3849 struct direct *ep; 3850 int i, chgs; 3851 3852 if ((pagedep->pd_state & IOSTARTED) == 0) { 3853 lk.lkt_held = NOHOLDER; 3854 panic("handle_written_filepage: not started"); 3855 } 3856 pagedep->pd_state &= ~IOSTARTED; 3857 /* 3858 * Process any directory removals that have been committed. 3859 */ 3860 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 3861 LIST_REMOVE(dirrem, dm_next); 3862 dirrem->dm_dirinum = pagedep->pd_ino; 3863 add_to_worklist(&dirrem->dm_list); 3864 } 3865 /* 3866 * Free any directory additions that have been committed. 3867 */ 3868 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 3869 free_diradd(dap); 3870 /* 3871 * Uncommitted directory entries must be restored. 3872 */ 3873 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 3874 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 3875 dap = nextdap) { 3876 nextdap = LIST_NEXT(dap, da_pdlist); 3877 if (dap->da_state & ATTACHED) { 3878 lk.lkt_held = NOHOLDER; 3879 panic("handle_written_filepage: attached"); 3880 } 3881 ep = (struct direct *) 3882 ((char *)bp->b_data + dap->da_offset); 3883 ep->d_ino = dap->da_newinum; 3884 dap->da_state &= ~UNDONE; 3885 dap->da_state |= ATTACHED; 3886 chgs = 1; 3887 /* 3888 * If the inode referenced by the directory has 3889 * been written out, then the dependency can be 3890 * moved to the pending list. 3891 */ 3892 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3893 LIST_REMOVE(dap, da_pdlist); 3894 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 3895 da_pdlist); 3896 } 3897 } 3898 } 3899 /* 3900 * If there were any rollbacks in the directory, then it must be 3901 * marked dirty so that its will eventually get written back in 3902 * its correct form. 3903 */ 3904 if (chgs) { 3905 if ((bp->b_flags & B_DELWRI) == 0) 3906 stat_dir_entry++; 3907 bdirty(bp); 3908 } 3909 /* 3910 * If no dependencies remain, the pagedep will be freed. 3911 * Otherwise it will remain to update the page before it 3912 * is written back to disk. 3913 */ 3914 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) { 3915 for (i = 0; i < DAHASHSZ; i++) 3916 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL) 3917 break; 3918 if (i == DAHASHSZ) { 3919 LIST_REMOVE(pagedep, pd_hash); 3920 WORKITEM_FREE(pagedep, D_PAGEDEP); 3921 return (0); 3922 } 3923 } 3924 return (1); 3925 } 3926 3927 /* 3928 * Writing back in-core inode structures. 3929 * 3930 * The filesystem only accesses an inode's contents when it occupies an 3931 * "in-core" inode structure. These "in-core" structures are separate from 3932 * the page frames used to cache inode blocks. Only the latter are 3933 * transferred to/from the disk. So, when the updated contents of the 3934 * "in-core" inode structure are copied to the corresponding in-memory inode 3935 * block, the dependencies are also transferred. The following procedure is 3936 * called when copying a dirty "in-core" inode to a cached inode block. 3937 */ 3938 3939 /* 3940 * Called when an inode is loaded from disk. If the effective link count 3941 * differed from the actual link count when it was last flushed, then we 3942 * need to ensure that the correct effective link count is put back. 3943 * 3944 * Parameters: 3945 * ip: the "in_core" copy of the inode 3946 */ 3947 void 3948 softdep_load_inodeblock(struct inode *ip) 3949 { 3950 struct inodedep *inodedep; 3951 3952 /* 3953 * Check for alternate nlink count. 3954 */ 3955 ip->i_effnlink = ip->i_nlink; 3956 ACQUIRE_LOCK(&lk); 3957 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3958 FREE_LOCK(&lk); 3959 return; 3960 } 3961 ip->i_effnlink -= inodedep->id_nlinkdelta; 3962 FREE_LOCK(&lk); 3963 } 3964 3965 /* 3966 * This routine is called just before the "in-core" inode 3967 * information is to be copied to the in-memory inode block. 3968 * Recall that an inode block contains several inodes. If 3969 * the force flag is set, then the dependencies will be 3970 * cleared so that the update can always be made. Note that 3971 * the buffer is locked when this routine is called, so we 3972 * will never be in the middle of writing the inode block 3973 * to disk. 3974 * 3975 * Parameters: 3976 * ip: the "in_core" copy of the inode 3977 * bp: the buffer containing the inode block 3978 * waitfor: nonzero => update must be allowed 3979 */ 3980 void 3981 softdep_update_inodeblock(struct inode *ip, struct buf *bp, 3982 int waitfor) 3983 { 3984 struct inodedep *inodedep; 3985 struct worklist *wk; 3986 int error, gotit; 3987 3988 /* 3989 * If the effective link count is not equal to the actual link 3990 * count, then we must track the difference in an inodedep while 3991 * the inode is (potentially) tossed out of the cache. Otherwise, 3992 * if there is no existing inodedep, then there are no dependencies 3993 * to track. 3994 */ 3995 ACQUIRE_LOCK(&lk); 3996 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3997 FREE_LOCK(&lk); 3998 if (ip->i_effnlink != ip->i_nlink) 3999 panic("softdep_update_inodeblock: bad link count"); 4000 return; 4001 } 4002 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) { 4003 FREE_LOCK(&lk); 4004 panic("softdep_update_inodeblock: bad delta"); 4005 } 4006 /* 4007 * Changes have been initiated. Anything depending on these 4008 * changes cannot occur until this inode has been written. 4009 */ 4010 inodedep->id_state &= ~COMPLETE; 4011 if ((inodedep->id_state & ONWORKLIST) == 0) 4012 WORKLIST_INSERT_BP(bp, &inodedep->id_list); 4013 /* 4014 * Any new dependencies associated with the incore inode must 4015 * now be moved to the list associated with the buffer holding 4016 * the in-memory copy of the inode. Once merged process any 4017 * allocdirects that are completed by the merger. 4018 */ 4019 merge_inode_lists(inodedep); 4020 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL) 4021 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); 4022 /* 4023 * Now that the inode has been pushed into the buffer, the 4024 * operations dependent on the inode being written to disk 4025 * can be moved to the id_bufwait so that they will be 4026 * processed when the buffer I/O completes. 4027 */ 4028 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 4029 WORKLIST_REMOVE(wk); 4030 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 4031 } 4032 /* 4033 * Newly allocated inodes cannot be written until the bitmap 4034 * that allocates them have been written (indicated by 4035 * DEPCOMPLETE being set in id_state). If we are doing a 4036 * forced sync (e.g., an fsync on a file), we force the bitmap 4037 * to be written so that the update can be done. 4038 */ 4039 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) { 4040 FREE_LOCK(&lk); 4041 return; 4042 } 4043 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 4044 FREE_LOCK(&lk); 4045 if (gotit && 4046 (error = bwrite(inodedep->id_buf)) != 0) 4047 softdep_error("softdep_update_inodeblock: bwrite", error); 4048 } 4049 4050 /* 4051 * Merge the new inode dependency list (id_newinoupdt) into the old 4052 * inode dependency list (id_inoupdt). This routine must be called 4053 * with splbio interrupts blocked. 4054 */ 4055 static void 4056 merge_inode_lists(struct inodedep *inodedep) 4057 { 4058 struct allocdirect *listadp, *newadp; 4059 4060 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 4061 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) { 4062 if (listadp->ad_lbn < newadp->ad_lbn) { 4063 listadp = TAILQ_NEXT(listadp, ad_next); 4064 continue; 4065 } 4066 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 4067 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 4068 if (listadp->ad_lbn == newadp->ad_lbn) { 4069 allocdirect_merge(&inodedep->id_inoupdt, newadp, 4070 listadp); 4071 listadp = newadp; 4072 } 4073 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 4074 } 4075 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) { 4076 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 4077 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next); 4078 } 4079 } 4080 4081 /* 4082 * If we are doing an fsync, then we must ensure that any directory 4083 * entries for the inode have been written after the inode gets to disk. 4084 * 4085 * Parameters: 4086 * vp: the "in_core" copy of the inode 4087 */ 4088 static int 4089 softdep_fsync(struct vnode *vp) 4090 { 4091 struct inodedep *inodedep; 4092 struct pagedep *pagedep; 4093 struct worklist *wk; 4094 struct diradd *dap; 4095 struct mount *mnt; 4096 struct vnode *pvp; 4097 struct inode *ip; 4098 struct buf *bp; 4099 struct fs *fs; 4100 int error, flushparent; 4101 ino_t parentino; 4102 ufs_lbn_t lbn; 4103 4104 /* 4105 * Move check from original kernel code, possibly not needed any 4106 * more with the per-mount bioops. 4107 */ 4108 if ((vp->v_mount->mnt_flag & MNT_SOFTDEP) == 0) 4109 return (0); 4110 4111 ip = VTOI(vp); 4112 fs = ip->i_fs; 4113 ACQUIRE_LOCK(&lk); 4114 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) { 4115 FREE_LOCK(&lk); 4116 return (0); 4117 } 4118 if (LIST_FIRST(&inodedep->id_inowait) != NULL || 4119 LIST_FIRST(&inodedep->id_bufwait) != NULL || 4120 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 4121 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) { 4122 FREE_LOCK(&lk); 4123 panic("softdep_fsync: pending ops"); 4124 } 4125 for (error = 0, flushparent = 0; ; ) { 4126 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 4127 break; 4128 if (wk->wk_type != D_DIRADD) { 4129 FREE_LOCK(&lk); 4130 panic("softdep_fsync: Unexpected type %s", 4131 TYPENAME(wk->wk_type)); 4132 } 4133 dap = WK_DIRADD(wk); 4134 /* 4135 * Flush our parent if this directory entry 4136 * has a MKDIR_PARENT dependency. 4137 */ 4138 if (dap->da_state & DIRCHG) 4139 pagedep = dap->da_previous->dm_pagedep; 4140 else 4141 pagedep = dap->da_pagedep; 4142 mnt = pagedep->pd_mnt; 4143 parentino = pagedep->pd_ino; 4144 lbn = pagedep->pd_lbn; 4145 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) { 4146 FREE_LOCK(&lk); 4147 panic("softdep_fsync: dirty"); 4148 } 4149 flushparent = dap->da_state & MKDIR_PARENT; 4150 /* 4151 * If we are being fsync'ed as part of vgone'ing this vnode, 4152 * then we will not be able to release and recover the 4153 * vnode below, so we just have to give up on writing its 4154 * directory entry out. It will eventually be written, just 4155 * not now, but then the user was not asking to have it 4156 * written, so we are not breaking any promises. 4157 */ 4158 if (vp->v_flag & VRECLAIMED) 4159 break; 4160 /* 4161 * We prevent deadlock by always fetching inodes from the 4162 * root, moving down the directory tree. Thus, when fetching 4163 * our parent directory, we must unlock ourselves before 4164 * requesting the lock on our parent. See the comment in 4165 * ufs_lookup for details on possible races. 4166 */ 4167 FREE_LOCK(&lk); 4168 vn_unlock(vp); 4169 error = VFS_VGET(mnt, parentino, &pvp); 4170 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4171 if (error != 0) 4172 return (error); 4173 if (flushparent) { 4174 if ((error = ffs_update(pvp, 1)) != 0) { 4175 vput(pvp); 4176 return (error); 4177 } 4178 } 4179 /* 4180 * Flush directory page containing the inode's name. 4181 */ 4182 error = bread(pvp, lblktodoff(fs, lbn), blksize(fs, VTOI(pvp), lbn), &bp); 4183 if (error == 0) 4184 error = bwrite(bp); 4185 vput(pvp); 4186 if (error != 0) 4187 return (error); 4188 ACQUIRE_LOCK(&lk); 4189 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) 4190 break; 4191 } 4192 FREE_LOCK(&lk); 4193 return (0); 4194 } 4195 4196 /* 4197 * Flush all the dirty bitmaps associated with the block device 4198 * before flushing the rest of the dirty blocks so as to reduce 4199 * the number of dependencies that will have to be rolled back. 4200 */ 4201 static int softdep_fsync_mountdev_bp(struct buf *bp, void *data); 4202 4203 void 4204 softdep_fsync_mountdev(struct vnode *vp) 4205 { 4206 if (!vn_isdisk(vp, NULL)) 4207 panic("softdep_fsync_mountdev: vnode not a disk"); 4208 ACQUIRE_LOCK(&lk); 4209 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 4210 softdep_fsync_mountdev_bp, vp); 4211 drain_output(vp, 1); 4212 FREE_LOCK(&lk); 4213 } 4214 4215 static int 4216 softdep_fsync_mountdev_bp(struct buf *bp, void *data) 4217 { 4218 struct worklist *wk; 4219 struct vnode *vp = data; 4220 4221 /* 4222 * If it is already scheduled, skip to the next buffer. 4223 */ 4224 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) 4225 return(0); 4226 if (bp->b_vp != vp || (bp->b_flags & B_DELWRI) == 0) { 4227 BUF_UNLOCK(bp); 4228 kprintf("softdep_fsync_mountdev_bp: warning, buffer %p ripped out from under vnode %p\n", bp, vp); 4229 return(0); 4230 } 4231 /* 4232 * We are only interested in bitmaps with outstanding 4233 * dependencies. 4234 */ 4235 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 4236 wk->wk_type != D_BMSAFEMAP) { 4237 BUF_UNLOCK(bp); 4238 return(0); 4239 } 4240 bremfree(bp); 4241 FREE_LOCK(&lk); 4242 (void) bawrite(bp); 4243 ACQUIRE_LOCK(&lk); 4244 return(0); 4245 } 4246 4247 /* 4248 * This routine is called when we are trying to synchronously flush a 4249 * file. This routine must eliminate any filesystem metadata dependencies 4250 * so that the syncing routine can succeed by pushing the dirty blocks 4251 * associated with the file. If any I/O errors occur, they are returned. 4252 */ 4253 struct softdep_sync_metadata_info { 4254 struct vnode *vp; 4255 int waitfor; 4256 }; 4257 4258 static int softdep_sync_metadata_bp(struct buf *bp, void *data); 4259 4260 int 4261 softdep_sync_metadata(struct vnode *vp, struct thread *td) 4262 { 4263 struct softdep_sync_metadata_info info; 4264 int error, waitfor; 4265 4266 /* 4267 * Check whether this vnode is involved in a filesystem 4268 * that is doing soft dependency processing. 4269 */ 4270 if (!vn_isdisk(vp, NULL)) { 4271 if (!DOINGSOFTDEP(vp)) 4272 return (0); 4273 } else 4274 if (vp->v_rdev->si_mountpoint == NULL || 4275 (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0) 4276 return (0); 4277 /* 4278 * Ensure that any direct block dependencies have been cleared. 4279 */ 4280 ACQUIRE_LOCK(&lk); 4281 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) { 4282 FREE_LOCK(&lk); 4283 return (error); 4284 } 4285 /* 4286 * For most files, the only metadata dependencies are the 4287 * cylinder group maps that allocate their inode or blocks. 4288 * The block allocation dependencies can be found by traversing 4289 * the dependency lists for any buffers that remain on their 4290 * dirty buffer list. The inode allocation dependency will 4291 * be resolved when the inode is updated with MNT_WAIT. 4292 * This work is done in two passes. The first pass grabs most 4293 * of the buffers and begins asynchronously writing them. The 4294 * only way to wait for these asynchronous writes is to sleep 4295 * on the filesystem vnode which may stay busy for a long time 4296 * if the filesystem is active. So, instead, we make a second 4297 * pass over the dependencies blocking on each write. In the 4298 * usual case we will be blocking against a write that we 4299 * initiated, so when it is done the dependency will have been 4300 * resolved. Thus the second pass is expected to end quickly. 4301 */ 4302 waitfor = MNT_NOWAIT; 4303 top: 4304 /* 4305 * We must wait for any I/O in progress to finish so that 4306 * all potential buffers on the dirty list will be visible. 4307 */ 4308 drain_output(vp, 1); 4309 info.vp = vp; 4310 info.waitfor = waitfor; 4311 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 4312 softdep_sync_metadata_bp, &info); 4313 if (error < 0) { 4314 FREE_LOCK(&lk); 4315 return(-error); /* error code */ 4316 } 4317 4318 /* 4319 * The brief unlock is to allow any pent up dependency 4320 * processing to be done. Then proceed with the second pass. 4321 */ 4322 if (waitfor == MNT_NOWAIT) { 4323 waitfor = MNT_WAIT; 4324 FREE_LOCK(&lk); 4325 ACQUIRE_LOCK(&lk); 4326 goto top; 4327 } 4328 4329 /* 4330 * If we have managed to get rid of all the dirty buffers, 4331 * then we are done. For certain directories and block 4332 * devices, we may need to do further work. 4333 * 4334 * We must wait for any I/O in progress to finish so that 4335 * all potential buffers on the dirty list will be visible. 4336 */ 4337 drain_output(vp, 1); 4338 if (RB_EMPTY(&vp->v_rbdirty_tree)) { 4339 FREE_LOCK(&lk); 4340 return (0); 4341 } 4342 4343 FREE_LOCK(&lk); 4344 /* 4345 * If we are trying to sync a block device, some of its buffers may 4346 * contain metadata that cannot be written until the contents of some 4347 * partially written files have been written to disk. The only easy 4348 * way to accomplish this is to sync the entire filesystem (luckily 4349 * this happens rarely). 4350 */ 4351 if (vn_isdisk(vp, NULL) && 4352 vp->v_rdev && 4353 vp->v_rdev->si_mountpoint && !vn_islocked(vp) && 4354 (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT)) != 0) 4355 return (error); 4356 return (0); 4357 } 4358 4359 static int 4360 softdep_sync_metadata_bp(struct buf *bp, void *data) 4361 { 4362 struct softdep_sync_metadata_info *info = data; 4363 struct pagedep *pagedep; 4364 struct allocdirect *adp; 4365 struct allocindir *aip; 4366 struct worklist *wk; 4367 struct buf *nbp; 4368 int error; 4369 int i; 4370 4371 if (getdirtybuf(&bp, MNT_WAIT) == 0) { 4372 kprintf("softdep_sync_metadata_bp(1): caught buf %p going away\n", bp); 4373 return (1); 4374 } 4375 if (bp->b_vp != info->vp || (bp->b_flags & B_DELWRI) == 0) { 4376 kprintf("softdep_sync_metadata_bp(2): caught buf %p going away vp %p\n", bp, info->vp); 4377 BUF_UNLOCK(bp); 4378 return(1); 4379 } 4380 4381 /* 4382 * As we hold the buffer locked, none of its dependencies 4383 * will disappear. 4384 */ 4385 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 4386 switch (wk->wk_type) { 4387 4388 case D_ALLOCDIRECT: 4389 adp = WK_ALLOCDIRECT(wk); 4390 if (adp->ad_state & DEPCOMPLETE) 4391 break; 4392 nbp = adp->ad_buf; 4393 if (getdirtybuf(&nbp, info->waitfor) == 0) 4394 break; 4395 FREE_LOCK(&lk); 4396 if (info->waitfor == MNT_NOWAIT) { 4397 bawrite(nbp); 4398 } else if ((error = bwrite(nbp)) != 0) { 4399 bawrite(bp); 4400 ACQUIRE_LOCK(&lk); 4401 return (-error); 4402 } 4403 ACQUIRE_LOCK(&lk); 4404 break; 4405 4406 case D_ALLOCINDIR: 4407 aip = WK_ALLOCINDIR(wk); 4408 if (aip->ai_state & DEPCOMPLETE) 4409 break; 4410 nbp = aip->ai_buf; 4411 if (getdirtybuf(&nbp, info->waitfor) == 0) 4412 break; 4413 FREE_LOCK(&lk); 4414 if (info->waitfor == MNT_NOWAIT) { 4415 bawrite(nbp); 4416 } else if ((error = bwrite(nbp)) != 0) { 4417 bawrite(bp); 4418 ACQUIRE_LOCK(&lk); 4419 return (-error); 4420 } 4421 ACQUIRE_LOCK(&lk); 4422 break; 4423 4424 case D_INDIRDEP: 4425 restart: 4426 4427 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) { 4428 if (aip->ai_state & DEPCOMPLETE) 4429 continue; 4430 nbp = aip->ai_buf; 4431 if (getdirtybuf(&nbp, MNT_WAIT) == 0) 4432 goto restart; 4433 FREE_LOCK(&lk); 4434 if ((error = bwrite(nbp)) != 0) { 4435 bawrite(bp); 4436 ACQUIRE_LOCK(&lk); 4437 return (-error); 4438 } 4439 ACQUIRE_LOCK(&lk); 4440 goto restart; 4441 } 4442 break; 4443 4444 case D_INODEDEP: 4445 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs, 4446 WK_INODEDEP(wk)->id_ino)) != 0) { 4447 FREE_LOCK(&lk); 4448 bawrite(bp); 4449 ACQUIRE_LOCK(&lk); 4450 return (-error); 4451 } 4452 break; 4453 4454 case D_PAGEDEP: 4455 /* 4456 * We are trying to sync a directory that may 4457 * have dependencies on both its own metadata 4458 * and/or dependencies on the inodes of any 4459 * recently allocated files. We walk its diradd 4460 * lists pushing out the associated inode. 4461 */ 4462 pagedep = WK_PAGEDEP(wk); 4463 for (i = 0; i < DAHASHSZ; i++) { 4464 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 4465 continue; 4466 if ((error = 4467 flush_pagedep_deps(info->vp, 4468 pagedep->pd_mnt, 4469 &pagedep->pd_diraddhd[i]))) { 4470 FREE_LOCK(&lk); 4471 bawrite(bp); 4472 ACQUIRE_LOCK(&lk); 4473 return (-error); 4474 } 4475 } 4476 break; 4477 4478 case D_MKDIR: 4479 /* 4480 * This case should never happen if the vnode has 4481 * been properly sync'ed. However, if this function 4482 * is used at a place where the vnode has not yet 4483 * been sync'ed, this dependency can show up. So, 4484 * rather than panic, just flush it. 4485 */ 4486 nbp = WK_MKDIR(wk)->md_buf; 4487 if (getdirtybuf(&nbp, info->waitfor) == 0) 4488 break; 4489 FREE_LOCK(&lk); 4490 if (info->waitfor == MNT_NOWAIT) { 4491 bawrite(nbp); 4492 } else if ((error = bwrite(nbp)) != 0) { 4493 bawrite(bp); 4494 ACQUIRE_LOCK(&lk); 4495 return (-error); 4496 } 4497 ACQUIRE_LOCK(&lk); 4498 break; 4499 4500 case D_BMSAFEMAP: 4501 /* 4502 * This case should never happen if the vnode has 4503 * been properly sync'ed. However, if this function 4504 * is used at a place where the vnode has not yet 4505 * been sync'ed, this dependency can show up. So, 4506 * rather than panic, just flush it. 4507 * 4508 * nbp can wind up == bp if a device node for the 4509 * same filesystem is being fsynced at the same time, 4510 * leading to a panic if we don't catch the case. 4511 */ 4512 nbp = WK_BMSAFEMAP(wk)->sm_buf; 4513 if (nbp == bp) 4514 break; 4515 if (getdirtybuf(&nbp, info->waitfor) == 0) 4516 break; 4517 FREE_LOCK(&lk); 4518 if (info->waitfor == MNT_NOWAIT) { 4519 bawrite(nbp); 4520 } else if ((error = bwrite(nbp)) != 0) { 4521 bawrite(bp); 4522 ACQUIRE_LOCK(&lk); 4523 return (-error); 4524 } 4525 ACQUIRE_LOCK(&lk); 4526 break; 4527 4528 default: 4529 FREE_LOCK(&lk); 4530 panic("softdep_sync_metadata: Unknown type %s", 4531 TYPENAME(wk->wk_type)); 4532 /* NOTREACHED */ 4533 } 4534 } 4535 FREE_LOCK(&lk); 4536 bawrite(bp); 4537 ACQUIRE_LOCK(&lk); 4538 return(0); 4539 } 4540 4541 /* 4542 * Flush the dependencies associated with an inodedep. 4543 * Called with splbio blocked. 4544 */ 4545 static int 4546 flush_inodedep_deps(struct fs *fs, ino_t ino) 4547 { 4548 struct inodedep *inodedep; 4549 struct allocdirect *adp; 4550 int error, waitfor; 4551 struct buf *bp; 4552 4553 /* 4554 * This work is done in two passes. The first pass grabs most 4555 * of the buffers and begins asynchronously writing them. The 4556 * only way to wait for these asynchronous writes is to sleep 4557 * on the filesystem vnode which may stay busy for a long time 4558 * if the filesystem is active. So, instead, we make a second 4559 * pass over the dependencies blocking on each write. In the 4560 * usual case we will be blocking against a write that we 4561 * initiated, so when it is done the dependency will have been 4562 * resolved. Thus the second pass is expected to end quickly. 4563 * We give a brief window at the top of the loop to allow 4564 * any pending I/O to complete. 4565 */ 4566 for (waitfor = MNT_NOWAIT; ; ) { 4567 FREE_LOCK(&lk); 4568 ACQUIRE_LOCK(&lk); 4569 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4570 return (0); 4571 TAILQ_FOREACH(adp, &inodedep->id_inoupdt, ad_next) { 4572 if (adp->ad_state & DEPCOMPLETE) 4573 continue; 4574 bp = adp->ad_buf; 4575 if (getdirtybuf(&bp, waitfor) == 0) { 4576 if (waitfor == MNT_NOWAIT) 4577 continue; 4578 break; 4579 } 4580 FREE_LOCK(&lk); 4581 if (waitfor == MNT_NOWAIT) { 4582 bawrite(bp); 4583 } else if ((error = bwrite(bp)) != 0) { 4584 ACQUIRE_LOCK(&lk); 4585 return (error); 4586 } 4587 ACQUIRE_LOCK(&lk); 4588 break; 4589 } 4590 if (adp != NULL) 4591 continue; 4592 TAILQ_FOREACH(adp, &inodedep->id_newinoupdt, ad_next) { 4593 if (adp->ad_state & DEPCOMPLETE) 4594 continue; 4595 bp = adp->ad_buf; 4596 if (getdirtybuf(&bp, waitfor) == 0) { 4597 if (waitfor == MNT_NOWAIT) 4598 continue; 4599 break; 4600 } 4601 FREE_LOCK(&lk); 4602 if (waitfor == MNT_NOWAIT) { 4603 bawrite(bp); 4604 } else if ((error = bwrite(bp)) != 0) { 4605 ACQUIRE_LOCK(&lk); 4606 return (error); 4607 } 4608 ACQUIRE_LOCK(&lk); 4609 break; 4610 } 4611 if (adp != NULL) 4612 continue; 4613 /* 4614 * If pass2, we are done, otherwise do pass 2. 4615 */ 4616 if (waitfor == MNT_WAIT) 4617 break; 4618 waitfor = MNT_WAIT; 4619 } 4620 /* 4621 * Try freeing inodedep in case all dependencies have been removed. 4622 */ 4623 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0) 4624 (void) free_inodedep(inodedep); 4625 return (0); 4626 } 4627 4628 /* 4629 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 4630 * Called with splbio blocked. 4631 */ 4632 static int 4633 flush_pagedep_deps(struct vnode *pvp, struct mount *mp, 4634 struct diraddhd *diraddhdp) 4635 { 4636 struct inodedep *inodedep; 4637 struct ufsmount *ump; 4638 struct diradd *dap; 4639 struct vnode *vp; 4640 int gotit, error = 0; 4641 struct buf *bp; 4642 ino_t inum; 4643 4644 ump = VFSTOUFS(mp); 4645 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 4646 /* 4647 * Flush ourselves if this directory entry 4648 * has a MKDIR_PARENT dependency. 4649 */ 4650 if (dap->da_state & MKDIR_PARENT) { 4651 FREE_LOCK(&lk); 4652 if ((error = ffs_update(pvp, 1)) != 0) 4653 break; 4654 ACQUIRE_LOCK(&lk); 4655 /* 4656 * If that cleared dependencies, go on to next. 4657 */ 4658 if (dap != LIST_FIRST(diraddhdp)) 4659 continue; 4660 if (dap->da_state & MKDIR_PARENT) { 4661 FREE_LOCK(&lk); 4662 panic("flush_pagedep_deps: MKDIR_PARENT"); 4663 } 4664 } 4665 /* 4666 * A newly allocated directory must have its "." and 4667 * ".." entries written out before its name can be 4668 * committed in its parent. We do not want or need 4669 * the full semantics of a synchronous VOP_FSYNC as 4670 * that may end up here again, once for each directory 4671 * level in the filesystem. Instead, we push the blocks 4672 * and wait for them to clear. We have to fsync twice 4673 * because the first call may choose to defer blocks 4674 * that still have dependencies, but deferral will 4675 * happen at most once. 4676 */ 4677 inum = dap->da_newinum; 4678 if (dap->da_state & MKDIR_BODY) { 4679 FREE_LOCK(&lk); 4680 if ((error = VFS_VGET(mp, inum, &vp)) != 0) 4681 break; 4682 if ((error=VOP_FSYNC(vp, MNT_NOWAIT)) || 4683 (error=VOP_FSYNC(vp, MNT_NOWAIT))) { 4684 vput(vp); 4685 break; 4686 } 4687 drain_output(vp, 0); 4688 vput(vp); 4689 ACQUIRE_LOCK(&lk); 4690 /* 4691 * If that cleared dependencies, go on to next. 4692 */ 4693 if (dap != LIST_FIRST(diraddhdp)) 4694 continue; 4695 if (dap->da_state & MKDIR_BODY) { 4696 FREE_LOCK(&lk); 4697 panic("flush_pagedep_deps: MKDIR_BODY"); 4698 } 4699 } 4700 /* 4701 * Flush the inode on which the directory entry depends. 4702 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 4703 * the only remaining dependency is that the updated inode 4704 * count must get pushed to disk. The inode has already 4705 * been pushed into its inode buffer (via VOP_UPDATE) at 4706 * the time of the reference count change. So we need only 4707 * locate that buffer, ensure that there will be no rollback 4708 * caused by a bitmap dependency, then write the inode buffer. 4709 */ 4710 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) { 4711 FREE_LOCK(&lk); 4712 panic("flush_pagedep_deps: lost inode"); 4713 } 4714 /* 4715 * If the inode still has bitmap dependencies, 4716 * push them to disk. 4717 */ 4718 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4719 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 4720 FREE_LOCK(&lk); 4721 if (gotit && (error = bwrite(inodedep->id_buf)) != 0) 4722 break; 4723 ACQUIRE_LOCK(&lk); 4724 if (dap != LIST_FIRST(diraddhdp)) 4725 continue; 4726 } 4727 /* 4728 * If the inode is still sitting in a buffer waiting 4729 * to be written, push it to disk. 4730 */ 4731 FREE_LOCK(&lk); 4732 if ((error = bread(ump->um_devvp, 4733 fsbtodoff(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), 4734 (int)ump->um_fs->fs_bsize, &bp)) != 0) 4735 break; 4736 if ((error = bwrite(bp)) != 0) 4737 break; 4738 ACQUIRE_LOCK(&lk); 4739 /* 4740 * If we have failed to get rid of all the dependencies 4741 * then something is seriously wrong. 4742 */ 4743 if (dap == LIST_FIRST(diraddhdp)) { 4744 FREE_LOCK(&lk); 4745 panic("flush_pagedep_deps: flush failed"); 4746 } 4747 } 4748 if (error) 4749 ACQUIRE_LOCK(&lk); 4750 return (error); 4751 } 4752 4753 /* 4754 * A large burst of file addition or deletion activity can drive the 4755 * memory load excessively high. First attempt to slow things down 4756 * using the techniques below. If that fails, this routine requests 4757 * the offending operations to fall back to running synchronously 4758 * until the memory load returns to a reasonable level. 4759 */ 4760 int 4761 softdep_slowdown(struct vnode *vp) 4762 { 4763 int max_softdeps_hard; 4764 4765 max_softdeps_hard = max_softdeps * 11 / 10; 4766 if (num_dirrem < max_softdeps_hard / 2 && 4767 num_inodedep < max_softdeps_hard) 4768 return (0); 4769 stat_sync_limit_hit += 1; 4770 return (1); 4771 } 4772 4773 /* 4774 * If memory utilization has gotten too high, deliberately slow things 4775 * down and speed up the I/O processing. 4776 */ 4777 static int 4778 request_cleanup(int resource, int islocked) 4779 { 4780 struct thread *td = curthread; /* XXX */ 4781 4782 /* 4783 * We never hold up the filesystem syncer process. 4784 */ 4785 if (td == filesys_syncer) 4786 return (0); 4787 /* 4788 * First check to see if the work list has gotten backlogged. 4789 * If it has, co-opt this process to help clean up two entries. 4790 * Because this process may hold inodes locked, we cannot 4791 * handle any remove requests that might block on a locked 4792 * inode as that could lead to deadlock. 4793 */ 4794 if (num_on_worklist > max_softdeps / 10) { 4795 if (islocked) 4796 FREE_LOCK(&lk); 4797 process_worklist_item(NULL, LK_NOWAIT); 4798 process_worklist_item(NULL, LK_NOWAIT); 4799 stat_worklist_push += 2; 4800 if (islocked) 4801 ACQUIRE_LOCK(&lk); 4802 return(1); 4803 } 4804 4805 /* 4806 * If we are resource constrained on inode dependencies, try 4807 * flushing some dirty inodes. Otherwise, we are constrained 4808 * by file deletions, so try accelerating flushes of directories 4809 * with removal dependencies. We would like to do the cleanup 4810 * here, but we probably hold an inode locked at this point and 4811 * that might deadlock against one that we try to clean. So, 4812 * the best that we can do is request the syncer daemon to do 4813 * the cleanup for us. 4814 */ 4815 switch (resource) { 4816 4817 case FLUSH_INODES: 4818 stat_ino_limit_push += 1; 4819 req_clear_inodedeps += 1; 4820 stat_countp = &stat_ino_limit_hit; 4821 break; 4822 4823 case FLUSH_REMOVE: 4824 stat_blk_limit_push += 1; 4825 req_clear_remove += 1; 4826 stat_countp = &stat_blk_limit_hit; 4827 break; 4828 4829 default: 4830 if (islocked) 4831 FREE_LOCK(&lk); 4832 panic("request_cleanup: unknown type"); 4833 } 4834 /* 4835 * Hopefully the syncer daemon will catch up and awaken us. 4836 * We wait at most tickdelay before proceeding in any case. 4837 */ 4838 if (islocked == 0) 4839 ACQUIRE_LOCK(&lk); 4840 proc_waiting += 1; 4841 if (!callout_active(&handle)) 4842 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2, 4843 pause_timer, NULL); 4844 interlocked_sleep(&lk, SLEEP, (caddr_t)&proc_waiting, 0, 4845 "softupdate", 0); 4846 proc_waiting -= 1; 4847 if (islocked == 0) 4848 FREE_LOCK(&lk); 4849 return (1); 4850 } 4851 4852 /* 4853 * Awaken processes pausing in request_cleanup and clear proc_waiting 4854 * to indicate that there is no longer a timer running. 4855 */ 4856 void 4857 pause_timer(void *arg) 4858 { 4859 *stat_countp += 1; 4860 wakeup_one(&proc_waiting); 4861 if (proc_waiting > 0) 4862 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2, 4863 pause_timer, NULL); 4864 else 4865 callout_deactivate(&handle); 4866 } 4867 4868 /* 4869 * Flush out a directory with at least one removal dependency in an effort to 4870 * reduce the number of dirrem, freefile, and freeblks dependency structures. 4871 */ 4872 static void 4873 clear_remove(struct thread *td) 4874 { 4875 struct pagedep_hashhead *pagedephd; 4876 struct pagedep *pagedep; 4877 static int next = 0; 4878 struct mount *mp; 4879 struct vnode *vp; 4880 int error, cnt; 4881 ino_t ino; 4882 4883 ACQUIRE_LOCK(&lk); 4884 for (cnt = 0; cnt < pagedep_hash; cnt++) { 4885 pagedephd = &pagedep_hashtbl[next++]; 4886 if (next >= pagedep_hash) 4887 next = 0; 4888 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 4889 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL) 4890 continue; 4891 mp = pagedep->pd_mnt; 4892 ino = pagedep->pd_ino; 4893 FREE_LOCK(&lk); 4894 if ((error = VFS_VGET(mp, ino, &vp)) != 0) { 4895 softdep_error("clear_remove: vget", error); 4896 return; 4897 } 4898 if ((error = VOP_FSYNC(vp, MNT_NOWAIT))) 4899 softdep_error("clear_remove: fsync", error); 4900 drain_output(vp, 0); 4901 vput(vp); 4902 return; 4903 } 4904 } 4905 FREE_LOCK(&lk); 4906 } 4907 4908 /* 4909 * Clear out a block of dirty inodes in an effort to reduce 4910 * the number of inodedep dependency structures. 4911 */ 4912 struct clear_inodedeps_info { 4913 struct fs *fs; 4914 struct mount *mp; 4915 }; 4916 4917 static int 4918 clear_inodedeps_mountlist_callback(struct mount *mp, void *data) 4919 { 4920 struct clear_inodedeps_info *info = data; 4921 4922 if ((mp->mnt_flag & MNT_SOFTDEP) && info->fs == VFSTOUFS(mp)->um_fs) { 4923 info->mp = mp; 4924 return(-1); 4925 } 4926 return(0); 4927 } 4928 4929 static void 4930 clear_inodedeps(struct thread *td) 4931 { 4932 struct clear_inodedeps_info info; 4933 struct inodedep_hashhead *inodedephd; 4934 struct inodedep *inodedep; 4935 static int next = 0; 4936 struct vnode *vp; 4937 struct fs *fs; 4938 int error, cnt; 4939 ino_t firstino, lastino, ino; 4940 4941 ACQUIRE_LOCK(&lk); 4942 /* 4943 * Pick a random inode dependency to be cleared. 4944 * We will then gather up all the inodes in its block 4945 * that have dependencies and flush them out. 4946 */ 4947 for (cnt = 0; cnt < inodedep_hash; cnt++) { 4948 inodedephd = &inodedep_hashtbl[next++]; 4949 if (next >= inodedep_hash) 4950 next = 0; 4951 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 4952 break; 4953 } 4954 if (inodedep == NULL) { 4955 FREE_LOCK(&lk); 4956 return; 4957 } 4958 /* 4959 * Ugly code to find mount point given pointer to superblock. 4960 */ 4961 fs = inodedep->id_fs; 4962 info.mp = NULL; 4963 info.fs = fs; 4964 mountlist_scan(clear_inodedeps_mountlist_callback, 4965 &info, MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 4966 /* 4967 * Find the last inode in the block with dependencies. 4968 */ 4969 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 4970 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 4971 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0) 4972 break; 4973 /* 4974 * Asynchronously push all but the last inode with dependencies. 4975 * Synchronously push the last inode with dependencies to ensure 4976 * that the inode block gets written to free up the inodedeps. 4977 */ 4978 for (ino = firstino; ino <= lastino; ino++) { 4979 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4980 continue; 4981 FREE_LOCK(&lk); 4982 if ((error = VFS_VGET(info.mp, ino, &vp)) != 0) { 4983 softdep_error("clear_inodedeps: vget", error); 4984 return; 4985 } 4986 if (ino == lastino) { 4987 if ((error = VOP_FSYNC(vp, MNT_WAIT))) 4988 softdep_error("clear_inodedeps: fsync1", error); 4989 } else { 4990 if ((error = VOP_FSYNC(vp, MNT_NOWAIT))) 4991 softdep_error("clear_inodedeps: fsync2", error); 4992 drain_output(vp, 0); 4993 } 4994 vput(vp); 4995 ACQUIRE_LOCK(&lk); 4996 } 4997 FREE_LOCK(&lk); 4998 } 4999 5000 /* 5001 * Function to determine if the buffer has outstanding dependencies 5002 * that will cause a roll-back if the buffer is written. If wantcount 5003 * is set, return number of dependencies, otherwise just yes or no. 5004 */ 5005 static int 5006 softdep_count_dependencies(struct buf *bp, int wantcount) 5007 { 5008 struct worklist *wk; 5009 struct inodedep *inodedep; 5010 struct indirdep *indirdep; 5011 struct allocindir *aip; 5012 struct pagedep *pagedep; 5013 struct diradd *dap; 5014 int i, retval; 5015 5016 retval = 0; 5017 ACQUIRE_LOCK(&lk); 5018 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5019 switch (wk->wk_type) { 5020 5021 case D_INODEDEP: 5022 inodedep = WK_INODEDEP(wk); 5023 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 5024 /* bitmap allocation dependency */ 5025 retval += 1; 5026 if (!wantcount) 5027 goto out; 5028 } 5029 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 5030 /* direct block pointer dependency */ 5031 retval += 1; 5032 if (!wantcount) 5033 goto out; 5034 } 5035 continue; 5036 5037 case D_INDIRDEP: 5038 indirdep = WK_INDIRDEP(wk); 5039 5040 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 5041 /* indirect block pointer dependency */ 5042 retval += 1; 5043 if (!wantcount) 5044 goto out; 5045 } 5046 continue; 5047 5048 case D_PAGEDEP: 5049 pagedep = WK_PAGEDEP(wk); 5050 for (i = 0; i < DAHASHSZ; i++) { 5051 5052 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 5053 /* directory entry dependency */ 5054 retval += 1; 5055 if (!wantcount) 5056 goto out; 5057 } 5058 } 5059 continue; 5060 5061 case D_BMSAFEMAP: 5062 case D_ALLOCDIRECT: 5063 case D_ALLOCINDIR: 5064 case D_MKDIR: 5065 /* never a dependency on these blocks */ 5066 continue; 5067 5068 default: 5069 FREE_LOCK(&lk); 5070 panic("softdep_check_for_rollback: Unexpected type %s", 5071 TYPENAME(wk->wk_type)); 5072 /* NOTREACHED */ 5073 } 5074 } 5075 out: 5076 FREE_LOCK(&lk); 5077 return retval; 5078 } 5079 5080 /* 5081 * Acquire exclusive access to a buffer. 5082 * Must be called with splbio blocked. 5083 * Return 1 if buffer was acquired. 5084 */ 5085 static int 5086 getdirtybuf(struct buf **bpp, int waitfor) 5087 { 5088 struct buf *bp; 5089 int error; 5090 5091 for (;;) { 5092 if ((bp = *bpp) == NULL) 5093 return (0); 5094 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) 5095 break; 5096 if (waitfor != MNT_WAIT) 5097 return (0); 5098 error = interlocked_sleep(&lk, LOCKBUF, bp, 5099 LK_EXCLUSIVE | LK_SLEEPFAIL, 0, 0); 5100 if (error != ENOLCK) { 5101 FREE_LOCK(&lk); 5102 panic("getdirtybuf: inconsistent lock"); 5103 } 5104 } 5105 if ((bp->b_flags & B_DELWRI) == 0) { 5106 BUF_UNLOCK(bp); 5107 return (0); 5108 } 5109 bremfree(bp); 5110 return (1); 5111 } 5112 5113 /* 5114 * Wait for pending output on a vnode to complete. 5115 * Must be called with vnode locked. 5116 */ 5117 static void 5118 drain_output(struct vnode *vp, int islocked) 5119 { 5120 5121 if (!islocked) 5122 ACQUIRE_LOCK(&lk); 5123 while (vp->v_track_write.bk_active) { 5124 vp->v_track_write.bk_waitflag = 1; 5125 interlocked_sleep(&lk, SLEEP, &vp->v_track_write, 5126 0, "drainvp", 0); 5127 } 5128 if (!islocked) 5129 FREE_LOCK(&lk); 5130 } 5131 5132 /* 5133 * Called whenever a buffer that is being invalidated or reallocated 5134 * contains dependencies. This should only happen if an I/O error has 5135 * occurred. The routine is called with the buffer locked. 5136 */ 5137 static void 5138 softdep_deallocate_dependencies(struct buf *bp) 5139 { 5140 if ((bp->b_flags & B_ERROR) == 0) 5141 panic("softdep_deallocate_dependencies: dangling deps"); 5142 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntfromname, bp->b_error); 5143 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 5144 } 5145 5146 /* 5147 * Function to handle asynchronous write errors in the filesystem. 5148 */ 5149 void 5150 softdep_error(char *func, int error) 5151 { 5152 5153 /* XXX should do something better! */ 5154 kprintf("%s: got error %d while accessing filesystem\n", func, error); 5155 } 5156