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 1.54 2007/11/07 00:46:39 dillon Exp $ 41 */ 42 43 /* 44 * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide. 45 */ 46 #ifndef DIAGNOSTIC 47 #define DIAGNOSTIC 48 #endif 49 #ifndef DEBUG 50 #define DEBUG 51 #endif 52 53 #include <sys/param.h> 54 #include <sys/kernel.h> 55 #include <sys/systm.h> 56 #include <sys/buf.h> 57 #include <sys/malloc.h> 58 #include <sys/mount.h> 59 #include <sys/proc.h> 60 #include <sys/syslog.h> 61 #include <sys/vnode.h> 62 #include <sys/conf.h> 63 #include <sys/buf2.h> 64 #include <machine/inttypes.h> 65 #include "dir.h" 66 #include "quota.h" 67 #include "inode.h" 68 #include "ufsmount.h" 69 #include "fs.h" 70 #include "softdep.h" 71 #include "ffs_extern.h" 72 #include "ufs_extern.h" 73 74 #include <sys/thread2.h> 75 76 /* 77 * These definitions need to be adapted to the system to which 78 * this file is being ported. 79 */ 80 /* 81 * malloc types defined for the softdep system. 82 */ 83 MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies"); 84 MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies"); 85 MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation"); 86 MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map"); 87 MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode"); 88 MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies"); 89 MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block"); 90 MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode"); 91 MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode"); 92 MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated"); 93 MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry"); 94 MALLOC_DEFINE(M_MKDIR, "mkdir","New directory"); 95 MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted"); 96 97 #define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE) 98 99 #define D_PAGEDEP 0 100 #define D_INODEDEP 1 101 #define D_NEWBLK 2 102 #define D_BMSAFEMAP 3 103 #define D_ALLOCDIRECT 4 104 #define D_INDIRDEP 5 105 #define D_ALLOCINDIR 6 106 #define D_FREEFRAG 7 107 #define D_FREEBLKS 8 108 #define D_FREEFILE 9 109 #define D_DIRADD 10 110 #define D_MKDIR 11 111 #define D_DIRREM 12 112 #define D_LAST D_DIRREM 113 114 /* 115 * translate from workitem type to memory type 116 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 117 */ 118 static struct malloc_type *memtype[] = { 119 M_PAGEDEP, 120 M_INODEDEP, 121 M_NEWBLK, 122 M_BMSAFEMAP, 123 M_ALLOCDIRECT, 124 M_INDIRDEP, 125 M_ALLOCINDIR, 126 M_FREEFRAG, 127 M_FREEBLKS, 128 M_FREEFILE, 129 M_DIRADD, 130 M_MKDIR, 131 M_DIRREM 132 }; 133 134 #define DtoM(type) (memtype[type]) 135 136 /* 137 * Names of malloc types. 138 */ 139 #define TYPENAME(type) \ 140 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???") 141 /* 142 * End system adaptaion definitions. 143 */ 144 145 /* 146 * Internal function prototypes. 147 */ 148 static void softdep_error(char *, int); 149 static void drain_output(struct vnode *, int); 150 static int getdirtybuf(struct buf **, int); 151 static void clear_remove(struct thread *); 152 static void clear_inodedeps(struct thread *); 153 static int flush_pagedep_deps(struct vnode *, struct mount *, 154 struct diraddhd *); 155 static int flush_inodedep_deps(struct fs *, ino_t); 156 static int handle_written_filepage(struct pagedep *, struct buf *); 157 static void diradd_inode_written(struct diradd *, struct inodedep *); 158 static int handle_written_inodeblock(struct inodedep *, struct buf *); 159 static void handle_allocdirect_partdone(struct allocdirect *); 160 static void handle_allocindir_partdone(struct allocindir *); 161 static void initiate_write_filepage(struct pagedep *, struct buf *); 162 static void handle_written_mkdir(struct mkdir *, int); 163 static void initiate_write_inodeblock(struct inodedep *, struct buf *); 164 static void handle_workitem_freefile(struct freefile *); 165 static void handle_workitem_remove(struct dirrem *); 166 static struct dirrem *newdirrem(struct buf *, struct inode *, 167 struct inode *, int, struct dirrem **); 168 static void free_diradd(struct diradd *); 169 static void free_allocindir(struct allocindir *, struct inodedep *); 170 static int indir_trunc (struct inode *, off_t, int, ufs_lbn_t, long *); 171 static void deallocate_dependencies(struct buf *, struct inodedep *); 172 static void free_allocdirect(struct allocdirectlst *, 173 struct allocdirect *, int); 174 static int check_inode_unwritten(struct inodedep *); 175 static int free_inodedep(struct inodedep *); 176 static void handle_workitem_freeblocks(struct freeblks *); 177 static void merge_inode_lists(struct inodedep *); 178 static void setup_allocindir_phase2(struct buf *, struct inode *, 179 struct allocindir *); 180 static struct allocindir *newallocindir(struct inode *, int, ufs_daddr_t, 181 ufs_daddr_t); 182 static void handle_workitem_freefrag(struct freefrag *); 183 static struct freefrag *newfreefrag(struct inode *, ufs_daddr_t, long); 184 static void allocdirect_merge(struct allocdirectlst *, 185 struct allocdirect *, struct allocdirect *); 186 static struct bmsafemap *bmsafemap_lookup(struct buf *); 187 static int newblk_lookup(struct fs *, ufs_daddr_t, int, 188 struct newblk **); 189 static int inodedep_lookup(struct fs *, ino_t, int, struct inodedep **); 190 static int pagedep_lookup(struct inode *, ufs_lbn_t, int, 191 struct pagedep **); 192 static void pause_timer(void *); 193 static int request_cleanup(int, int); 194 static int process_worklist_item(struct mount *, int); 195 static void add_to_worklist(struct worklist *); 196 197 /* 198 * Exported softdep operations. 199 */ 200 static void softdep_disk_io_initiation(struct buf *); 201 static void softdep_disk_write_complete(struct buf *); 202 static void softdep_deallocate_dependencies(struct buf *); 203 static int softdep_fsync(struct vnode *); 204 static int softdep_process_worklist(struct mount *); 205 static void softdep_move_dependencies(struct buf *, struct buf *); 206 static int softdep_count_dependencies(struct buf *bp, int); 207 static int softdep_checkread(struct buf *bp); 208 static int softdep_checkwrite(struct buf *bp); 209 210 static struct bio_ops softdep_bioops = { 211 .io_start = softdep_disk_io_initiation, 212 .io_complete = softdep_disk_write_complete, 213 .io_deallocate = softdep_deallocate_dependencies, 214 .io_fsync = softdep_fsync, 215 .io_sync = softdep_process_worklist, 216 .io_movedeps = softdep_move_dependencies, 217 .io_countdeps = softdep_count_dependencies, 218 .io_checkread = softdep_checkread, 219 .io_checkwrite = softdep_checkwrite 220 }; 221 222 /* 223 * Locking primitives. 224 * 225 * For a uniprocessor, all we need to do is protect against disk 226 * interrupts. For a multiprocessor, this lock would have to be 227 * a mutex. A single mutex is used throughout this file, though 228 * finer grain locking could be used if contention warranted it. 229 * 230 * For a multiprocessor, the sleep call would accept a lock and 231 * release it after the sleep processing was complete. In a uniprocessor 232 * implementation there is no such interlock, so we simple mark 233 * the places where it needs to be done with the `interlocked' form 234 * of the lock calls. Since the uniprocessor sleep already interlocks 235 * the spl, there is nothing that really needs to be done. 236 */ 237 #ifndef /* NOT */ DEBUG 238 static struct lockit { 239 } lk = { 0 }; 240 #define ACQUIRE_LOCK(lk) crit_enter_id("softupdates"); 241 #define FREE_LOCK(lk) crit_exit_id("softupdates"); 242 243 #else /* DEBUG */ 244 #define NOHOLDER ((struct thread *)-1) 245 #define SPECIAL_FLAG ((struct thread *)-2) 246 static struct lockit { 247 int lkt_spl; 248 struct thread *lkt_held; 249 } lk = { 0, NOHOLDER }; 250 static int lockcnt; 251 252 static void acquire_lock(struct lockit *); 253 static void free_lock(struct lockit *); 254 void softdep_panic(char *); 255 256 #define ACQUIRE_LOCK(lk) acquire_lock(lk) 257 #define FREE_LOCK(lk) free_lock(lk) 258 259 static void 260 acquire_lock(struct lockit *lk) 261 { 262 thread_t holder; 263 264 if (lk->lkt_held != NOHOLDER) { 265 holder = lk->lkt_held; 266 FREE_LOCK(lk); 267 if (holder == curthread) 268 panic("softdep_lock: locking against myself"); 269 else 270 panic("softdep_lock: lock held by %p", holder); 271 } 272 crit_enter_id("softupdates"); 273 lk->lkt_held = curthread; 274 lockcnt++; 275 } 276 277 static void 278 free_lock(struct lockit *lk) 279 { 280 281 if (lk->lkt_held == NOHOLDER) 282 panic("softdep_unlock: lock not held"); 283 lk->lkt_held = NOHOLDER; 284 crit_exit_id("softupdates"); 285 } 286 287 /* 288 * Function to release soft updates lock and panic. 289 */ 290 void 291 softdep_panic(char *msg) 292 { 293 294 if (lk.lkt_held != NOHOLDER) 295 FREE_LOCK(&lk); 296 panic(msg); 297 } 298 #endif /* DEBUG */ 299 300 static int interlocked_sleep(struct lockit *, int, void *, int, 301 const char *, int); 302 303 /* 304 * When going to sleep, we must save our SPL so that it does 305 * not get lost if some other process uses the lock while we 306 * are sleeping. We restore it after we have slept. This routine 307 * wraps the interlocking with functions that sleep. The list 308 * below enumerates the available set of operations. 309 */ 310 #define UNKNOWN 0 311 #define SLEEP 1 312 #define LOCKBUF 2 313 314 static int 315 interlocked_sleep(struct lockit *lk, int op, void *ident, int flags, 316 const char *wmesg, int timo) 317 { 318 thread_t holder; 319 int s, retval; 320 321 s = lk->lkt_spl; 322 # ifdef DEBUG 323 if (lk->lkt_held == NOHOLDER) 324 panic("interlocked_sleep: lock not held"); 325 lk->lkt_held = NOHOLDER; 326 # endif /* DEBUG */ 327 switch (op) { 328 case SLEEP: 329 retval = tsleep(ident, flags, wmesg, timo); 330 break; 331 case LOCKBUF: 332 retval = BUF_LOCK((struct buf *)ident, flags); 333 break; 334 default: 335 panic("interlocked_sleep: unknown operation"); 336 } 337 # ifdef DEBUG 338 if (lk->lkt_held != NOHOLDER) { 339 holder = lk->lkt_held; 340 FREE_LOCK(lk); 341 if (holder == curthread) 342 panic("interlocked_sleep: locking against self"); 343 else 344 panic("interlocked_sleep: lock held by %p", holder); 345 } 346 lk->lkt_held = curthread; 347 lockcnt++; 348 # endif /* DEBUG */ 349 lk->lkt_spl = s; 350 return (retval); 351 } 352 353 /* 354 * Place holder for real semaphores. 355 */ 356 struct sema { 357 int value; 358 thread_t holder; 359 char *name; 360 int prio; 361 int timo; 362 }; 363 static void sema_init(struct sema *, char *, int, int); 364 static int sema_get(struct sema *, struct lockit *); 365 static void sema_release(struct sema *); 366 367 static void 368 sema_init(struct sema *semap, char *name, int prio, int timo) 369 { 370 371 semap->holder = NOHOLDER; 372 semap->value = 0; 373 semap->name = name; 374 semap->prio = prio; 375 semap->timo = timo; 376 } 377 378 static int 379 sema_get(struct sema *semap, struct lockit *interlock) 380 { 381 382 if (semap->value++ > 0) { 383 if (interlock != NULL) { 384 interlocked_sleep(interlock, SLEEP, (caddr_t)semap, 385 semap->prio, semap->name, semap->timo); 386 FREE_LOCK(interlock); 387 } else { 388 tsleep((caddr_t)semap, semap->prio, semap->name, 389 semap->timo); 390 } 391 return (0); 392 } 393 semap->holder = curthread; 394 if (interlock != NULL) 395 FREE_LOCK(interlock); 396 return (1); 397 } 398 399 static void 400 sema_release(struct sema *semap) 401 { 402 403 if (semap->value <= 0 || semap->holder != curthread) { 404 if (lk.lkt_held != NOHOLDER) 405 FREE_LOCK(&lk); 406 panic("sema_release: not held"); 407 } 408 if (--semap->value > 0) { 409 semap->value = 0; 410 wakeup(semap); 411 } 412 semap->holder = NOHOLDER; 413 } 414 415 /* 416 * Worklist queue management. 417 * These routines require that the lock be held. 418 */ 419 #ifndef /* NOT */ DEBUG 420 #define WORKLIST_INSERT(head, item) do { \ 421 (item)->wk_state |= ONWORKLIST; \ 422 LIST_INSERT_HEAD(head, item, wk_list); \ 423 } while (0) 424 425 #define WORKLIST_INSERT_BP(bp, item) do { \ 426 (item)->wk_state |= ONWORKLIST; \ 427 (bp)->b_ops = &softdep_bioops; \ 428 LIST_INSERT_HEAD(&(bp)->b_dep, item, wk_list); \ 429 } while (0) 430 431 #define WORKLIST_REMOVE(item) do { \ 432 (item)->wk_state &= ~ONWORKLIST; \ 433 LIST_REMOVE(item, wk_list); \ 434 } while (0) 435 436 #define WORKITEM_FREE(item, type) FREE(item, DtoM(type)) 437 438 #else /* DEBUG */ 439 static void worklist_insert(struct workhead *, struct worklist *); 440 static void worklist_remove(struct worklist *); 441 static void workitem_free(struct worklist *, int); 442 443 #define WORKLIST_INSERT_BP(bp, item) do { \ 444 (bp)->b_ops = &softdep_bioops; \ 445 worklist_insert(&(bp)->b_dep, item); \ 446 } while (0) 447 448 #define WORKLIST_INSERT(head, item) worklist_insert(head, item) 449 #define WORKLIST_REMOVE(item) worklist_remove(item) 450 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type) 451 452 static void 453 worklist_insert(struct workhead *head, struct worklist *item) 454 { 455 456 if (lk.lkt_held == NOHOLDER) 457 panic("worklist_insert: lock not held"); 458 if (item->wk_state & ONWORKLIST) { 459 FREE_LOCK(&lk); 460 panic("worklist_insert: already on list"); 461 } 462 item->wk_state |= ONWORKLIST; 463 LIST_INSERT_HEAD(head, item, wk_list); 464 } 465 466 static void 467 worklist_remove(struct worklist *item) 468 { 469 470 if (lk.lkt_held == NOHOLDER) 471 panic("worklist_remove: lock not held"); 472 if ((item->wk_state & ONWORKLIST) == 0) { 473 FREE_LOCK(&lk); 474 panic("worklist_remove: not on list"); 475 } 476 item->wk_state &= ~ONWORKLIST; 477 LIST_REMOVE(item, wk_list); 478 } 479 480 static void 481 workitem_free(struct worklist *item, int type) 482 { 483 484 if (item->wk_state & ONWORKLIST) { 485 if (lk.lkt_held != NOHOLDER) 486 FREE_LOCK(&lk); 487 panic("workitem_free: still on list"); 488 } 489 if (item->wk_type != type) { 490 if (lk.lkt_held != NOHOLDER) 491 FREE_LOCK(&lk); 492 panic("workitem_free: type mismatch"); 493 } 494 FREE(item, DtoM(type)); 495 } 496 #endif /* DEBUG */ 497 498 /* 499 * Workitem queue management 500 */ 501 static struct workhead softdep_workitem_pending; 502 static int num_on_worklist; /* number of worklist items to be processed */ 503 static int softdep_worklist_busy; /* 1 => trying to do unmount */ 504 static int softdep_worklist_req; /* serialized waiters */ 505 static int max_softdeps; /* maximum number of structs before slowdown */ 506 static int tickdelay = 2; /* number of ticks to pause during slowdown */ 507 static int *stat_countp; /* statistic to count in proc_waiting timeout */ 508 static int proc_waiting; /* tracks whether we have a timeout posted */ 509 static struct callout handle; /* handle on posted proc_waiting timeout */ 510 static struct thread *filesys_syncer; /* proc of filesystem syncer process */ 511 static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 512 #define FLUSH_INODES 1 513 static int req_clear_remove; /* syncer process flush some freeblks */ 514 #define FLUSH_REMOVE 2 515 /* 516 * runtime statistics 517 */ 518 static int stat_worklist_push; /* number of worklist cleanups */ 519 static int stat_blk_limit_push; /* number of times block limit neared */ 520 static int stat_ino_limit_push; /* number of times inode limit neared */ 521 static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 522 static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 523 static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 524 static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 525 static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 526 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 527 static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 528 #ifdef DEBUG 529 #include <vm/vm.h> 530 #include <sys/sysctl.h> 531 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, ""); 532 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, ""); 533 SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,""); 534 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,""); 535 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,""); 536 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, ""); 537 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, ""); 538 SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, ""); 539 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, ""); 540 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, ""); 541 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, ""); 542 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, ""); 543 #endif /* DEBUG */ 544 545 /* 546 * Add an item to the end of the work queue. 547 * This routine requires that the lock be held. 548 * This is the only routine that adds items to the list. 549 * The following routine is the only one that removes items 550 * and does so in order from first to last. 551 */ 552 static void 553 add_to_worklist(struct worklist *wk) 554 { 555 static struct worklist *worklist_tail; 556 557 if (wk->wk_state & ONWORKLIST) { 558 if (lk.lkt_held != NOHOLDER) 559 FREE_LOCK(&lk); 560 panic("add_to_worklist: already on list"); 561 } 562 wk->wk_state |= ONWORKLIST; 563 if (LIST_FIRST(&softdep_workitem_pending) == NULL) 564 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list); 565 else 566 LIST_INSERT_AFTER(worklist_tail, wk, wk_list); 567 worklist_tail = wk; 568 num_on_worklist += 1; 569 } 570 571 /* 572 * Process that runs once per second to handle items in the background queue. 573 * 574 * Note that we ensure that everything is done in the order in which they 575 * appear in the queue. The code below depends on this property to ensure 576 * that blocks of a file are freed before the inode itself is freed. This 577 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 578 * until all the old ones have been purged from the dependency lists. 579 */ 580 static int 581 softdep_process_worklist(struct mount *matchmnt) 582 { 583 thread_t td = curthread; 584 int matchcnt, loopcount; 585 long starttime; 586 587 /* 588 * Record the process identifier of our caller so that we can give 589 * this process preferential treatment in request_cleanup below. 590 */ 591 filesys_syncer = td; 592 matchcnt = 0; 593 594 /* 595 * There is no danger of having multiple processes run this 596 * code, but we have to single-thread it when softdep_flushfiles() 597 * is in operation to get an accurate count of the number of items 598 * related to its mount point that are in the list. 599 */ 600 if (matchmnt == NULL) { 601 if (softdep_worklist_busy < 0) 602 return(-1); 603 softdep_worklist_busy += 1; 604 } 605 606 /* 607 * If requested, try removing inode or removal dependencies. 608 */ 609 if (req_clear_inodedeps) { 610 clear_inodedeps(td); 611 req_clear_inodedeps -= 1; 612 wakeup_one(&proc_waiting); 613 } 614 if (req_clear_remove) { 615 clear_remove(td); 616 req_clear_remove -= 1; 617 wakeup_one(&proc_waiting); 618 } 619 loopcount = 1; 620 starttime = time_second; 621 while (num_on_worklist > 0) { 622 matchcnt += process_worklist_item(matchmnt, 0); 623 624 /* 625 * If a umount operation wants to run the worklist 626 * accurately, abort. 627 */ 628 if (softdep_worklist_req && matchmnt == NULL) { 629 matchcnt = -1; 630 break; 631 } 632 633 /* 634 * If requested, try removing inode or removal dependencies. 635 */ 636 if (req_clear_inodedeps) { 637 clear_inodedeps(td); 638 req_clear_inodedeps -= 1; 639 wakeup_one(&proc_waiting); 640 } 641 if (req_clear_remove) { 642 clear_remove(td); 643 req_clear_remove -= 1; 644 wakeup_one(&proc_waiting); 645 } 646 /* 647 * We do not generally want to stop for buffer space, but if 648 * we are really being a buffer hog, we will stop and wait. 649 */ 650 if (loopcount++ % 128 == 0) 651 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); 1360 bzero(adp, sizeof(struct allocdirect)); 1361 adp->ad_list.wk_type = D_ALLOCDIRECT; 1362 adp->ad_lbn = lbn; 1363 adp->ad_newblkno = newblkno; 1364 adp->ad_oldblkno = oldblkno; 1365 adp->ad_newsize = newsize; 1366 adp->ad_oldsize = oldsize; 1367 adp->ad_state = ATTACHED; 1368 if (newblkno == oldblkno) 1369 adp->ad_freefrag = NULL; 1370 else 1371 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1372 1373 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1374 panic("softdep_setup_allocdirect: lost block"); 1375 1376 ACQUIRE_LOCK(&lk); 1377 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep); 1378 adp->ad_inodedep = inodedep; 1379 1380 if (newblk->nb_state == DEPCOMPLETE) { 1381 adp->ad_state |= DEPCOMPLETE; 1382 adp->ad_buf = NULL; 1383 } else { 1384 bmsafemap = newblk->nb_bmsafemap; 1385 adp->ad_buf = bmsafemap->sm_buf; 1386 LIST_REMOVE(newblk, nb_deps); 1387 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1388 } 1389 LIST_REMOVE(newblk, nb_hash); 1390 FREE(newblk, M_NEWBLK); 1391 1392 WORKLIST_INSERT_BP(bp, &adp->ad_list); 1393 if (lbn >= NDADDR) { 1394 /* allocating an indirect block */ 1395 if (oldblkno != 0) { 1396 FREE_LOCK(&lk); 1397 panic("softdep_setup_allocdirect: non-zero indir"); 1398 } 1399 } else { 1400 /* 1401 * Allocating a direct block. 1402 * 1403 * If we are allocating a directory block, then we must 1404 * allocate an associated pagedep to track additions and 1405 * deletions. 1406 */ 1407 if ((ip->i_mode & IFMT) == IFDIR && 1408 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) { 1409 WORKLIST_INSERT_BP(bp, &pagedep->pd_list); 1410 } 1411 } 1412 /* 1413 * The list of allocdirects must be kept in sorted and ascending 1414 * order so that the rollback routines can quickly determine the 1415 * first uncommitted block (the size of the file stored on disk 1416 * ends at the end of the lowest committed fragment, or if there 1417 * are no fragments, at the end of the highest committed block). 1418 * Since files generally grow, the typical case is that the new 1419 * block is to be added at the end of the list. We speed this 1420 * special case by checking against the last allocdirect in the 1421 * list before laboriously traversing the list looking for the 1422 * insertion point. 1423 */ 1424 adphead = &inodedep->id_newinoupdt; 1425 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1426 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1427 /* insert at end of list */ 1428 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1429 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1430 allocdirect_merge(adphead, adp, oldadp); 1431 FREE_LOCK(&lk); 1432 return; 1433 } 1434 TAILQ_FOREACH(oldadp, adphead, ad_next) { 1435 if (oldadp->ad_lbn >= lbn) 1436 break; 1437 } 1438 if (oldadp == NULL) { 1439 FREE_LOCK(&lk); 1440 panic("softdep_setup_allocdirect: lost entry"); 1441 } 1442 /* insert in middle of list */ 1443 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1444 if (oldadp->ad_lbn == lbn) 1445 allocdirect_merge(adphead, adp, oldadp); 1446 FREE_LOCK(&lk); 1447 } 1448 1449 /* 1450 * Replace an old allocdirect dependency with a newer one. 1451 * This routine must be called with splbio interrupts blocked. 1452 * 1453 * Parameters: 1454 * adphead: head of list holding allocdirects 1455 * newadp: allocdirect being added 1456 * oldadp: existing allocdirect being checked 1457 */ 1458 static void 1459 allocdirect_merge(struct allocdirectlst *adphead, 1460 struct allocdirect *newadp, 1461 struct allocdirect *oldadp) 1462 { 1463 struct freefrag *freefrag; 1464 1465 #ifdef DEBUG 1466 if (lk.lkt_held == NOHOLDER) 1467 panic("allocdirect_merge: lock not held"); 1468 #endif 1469 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 1470 newadp->ad_oldsize != oldadp->ad_newsize || 1471 newadp->ad_lbn >= NDADDR) { 1472 FREE_LOCK(&lk); 1473 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d", 1474 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn, 1475 NDADDR); 1476 } 1477 newadp->ad_oldblkno = oldadp->ad_oldblkno; 1478 newadp->ad_oldsize = oldadp->ad_oldsize; 1479 /* 1480 * If the old dependency had a fragment to free or had never 1481 * previously had a block allocated, then the new dependency 1482 * can immediately post its freefrag and adopt the old freefrag. 1483 * This action is done by swapping the freefrag dependencies. 1484 * The new dependency gains the old one's freefrag, and the 1485 * old one gets the new one and then immediately puts it on 1486 * the worklist when it is freed by free_allocdirect. It is 1487 * not possible to do this swap when the old dependency had a 1488 * non-zero size but no previous fragment to free. This condition 1489 * arises when the new block is an extension of the old block. 1490 * Here, the first part of the fragment allocated to the new 1491 * dependency is part of the block currently claimed on disk by 1492 * the old dependency, so cannot legitimately be freed until the 1493 * conditions for the new dependency are fulfilled. 1494 */ 1495 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 1496 freefrag = newadp->ad_freefrag; 1497 newadp->ad_freefrag = oldadp->ad_freefrag; 1498 oldadp->ad_freefrag = freefrag; 1499 } 1500 free_allocdirect(adphead, oldadp, 0); 1501 } 1502 1503 /* 1504 * Allocate a new freefrag structure if needed. 1505 */ 1506 static struct freefrag * 1507 newfreefrag(struct inode *ip, ufs_daddr_t blkno, long size) 1508 { 1509 struct freefrag *freefrag; 1510 struct fs *fs; 1511 1512 if (blkno == 0) 1513 return (NULL); 1514 fs = ip->i_fs; 1515 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 1516 panic("newfreefrag: frag size"); 1517 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag), 1518 M_FREEFRAG, M_SOFTDEP_FLAGS); 1519 freefrag->ff_list.wk_type = D_FREEFRAG; 1520 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */ 1521 freefrag->ff_inum = ip->i_number; 1522 freefrag->ff_fs = fs; 1523 freefrag->ff_devvp = ip->i_devvp; 1524 freefrag->ff_blkno = blkno; 1525 freefrag->ff_fragsize = size; 1526 return (freefrag); 1527 } 1528 1529 /* 1530 * This workitem de-allocates fragments that were replaced during 1531 * file block allocation. 1532 */ 1533 static void 1534 handle_workitem_freefrag(struct freefrag *freefrag) 1535 { 1536 struct inode tip; 1537 1538 tip.i_fs = freefrag->ff_fs; 1539 tip.i_devvp = freefrag->ff_devvp; 1540 tip.i_dev = freefrag->ff_devvp->v_rdev; 1541 tip.i_number = freefrag->ff_inum; 1542 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */ 1543 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize); 1544 FREE(freefrag, M_FREEFRAG); 1545 } 1546 1547 /* 1548 * Indirect block allocation dependencies. 1549 * 1550 * The same dependencies that exist for a direct block also exist when 1551 * a new block is allocated and pointed to by an entry in a block of 1552 * indirect pointers. The undo/redo states described above are also 1553 * used here. Because an indirect block contains many pointers that 1554 * may have dependencies, a second copy of the entire in-memory indirect 1555 * block is kept. The buffer cache copy is always completely up-to-date. 1556 * The second copy, which is used only as a source for disk writes, 1557 * contains only the safe pointers (i.e., those that have no remaining 1558 * update dependencies). The second copy is freed when all pointers 1559 * are safe. The cache is not allowed to replace indirect blocks with 1560 * pending update dependencies. If a buffer containing an indirect 1561 * block with dependencies is written, these routines will mark it 1562 * dirty again. It can only be successfully written once all the 1563 * dependencies are removed. The ffs_fsync routine in conjunction with 1564 * softdep_sync_metadata work together to get all the dependencies 1565 * removed so that a file can be successfully written to disk. Three 1566 * procedures are used when setting up indirect block pointer 1567 * dependencies. The division is necessary because of the organization 1568 * of the "balloc" routine and because of the distinction between file 1569 * pages and file metadata blocks. 1570 */ 1571 1572 /* 1573 * Allocate a new allocindir structure. 1574 * 1575 * Parameters: 1576 * ip: inode for file being extended 1577 * ptrno: offset of pointer in indirect block 1578 * newblkno: disk block number being added 1579 * oldblkno: previous block number, 0 if none 1580 */ 1581 static struct allocindir * 1582 newallocindir(struct inode *ip, int ptrno, ufs_daddr_t newblkno, 1583 ufs_daddr_t oldblkno) 1584 { 1585 struct allocindir *aip; 1586 1587 MALLOC(aip, struct allocindir *, sizeof(struct allocindir), 1588 M_ALLOCINDIR, M_SOFTDEP_FLAGS); 1589 bzero(aip, sizeof(struct allocindir)); 1590 aip->ai_list.wk_type = D_ALLOCINDIR; 1591 aip->ai_state = ATTACHED; 1592 aip->ai_offset = ptrno; 1593 aip->ai_newblkno = newblkno; 1594 aip->ai_oldblkno = oldblkno; 1595 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize); 1596 return (aip); 1597 } 1598 1599 /* 1600 * Called just before setting an indirect block pointer 1601 * to a newly allocated file page. 1602 * 1603 * Parameters: 1604 * ip: inode for file being extended 1605 * lbn: allocated block number within file 1606 * bp: buffer with indirect blk referencing page 1607 * ptrno: offset of pointer in indirect block 1608 * newblkno: disk block number being added 1609 * oldblkno: previous block number, 0 if none 1610 * nbp: buffer holding allocated page 1611 */ 1612 void 1613 softdep_setup_allocindir_page(struct inode *ip, ufs_lbn_t lbn, 1614 struct buf *bp, int ptrno, 1615 ufs_daddr_t newblkno, ufs_daddr_t oldblkno, 1616 struct buf *nbp) 1617 { 1618 struct allocindir *aip; 1619 struct pagedep *pagedep; 1620 1621 aip = newallocindir(ip, ptrno, newblkno, oldblkno); 1622 ACQUIRE_LOCK(&lk); 1623 /* 1624 * If we are allocating a directory page, then we must 1625 * allocate an associated pagedep to track additions and 1626 * deletions. 1627 */ 1628 if ((ip->i_mode & IFMT) == IFDIR && 1629 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1630 WORKLIST_INSERT_BP(nbp, &pagedep->pd_list); 1631 WORKLIST_INSERT_BP(nbp, &aip->ai_list); 1632 FREE_LOCK(&lk); 1633 setup_allocindir_phase2(bp, ip, aip); 1634 } 1635 1636 /* 1637 * Called just before setting an indirect block pointer to a 1638 * newly allocated indirect block. 1639 * Parameters: 1640 * nbp: newly allocated indirect block 1641 * ip: inode for file being extended 1642 * bp: indirect block referencing allocated block 1643 * ptrno: offset of pointer in indirect block 1644 * newblkno: disk block number being added 1645 */ 1646 void 1647 softdep_setup_allocindir_meta(struct buf *nbp, struct inode *ip, 1648 struct buf *bp, int ptrno, 1649 ufs_daddr_t newblkno) 1650 { 1651 struct allocindir *aip; 1652 1653 aip = newallocindir(ip, ptrno, newblkno, 0); 1654 ACQUIRE_LOCK(&lk); 1655 WORKLIST_INSERT_BP(nbp, &aip->ai_list); 1656 FREE_LOCK(&lk); 1657 setup_allocindir_phase2(bp, ip, aip); 1658 } 1659 1660 /* 1661 * Called to finish the allocation of the "aip" allocated 1662 * by one of the two routines above. 1663 * 1664 * Parameters: 1665 * bp: in-memory copy of the indirect block 1666 * ip: inode for file being extended 1667 * aip: allocindir allocated by the above routines 1668 */ 1669 static void 1670 setup_allocindir_phase2(struct buf *bp, struct inode *ip, 1671 struct allocindir *aip) 1672 { 1673 struct worklist *wk; 1674 struct indirdep *indirdep, *newindirdep; 1675 struct bmsafemap *bmsafemap; 1676 struct allocindir *oldaip; 1677 struct freefrag *freefrag; 1678 struct newblk *newblk; 1679 1680 if (bp->b_loffset >= 0) 1681 panic("setup_allocindir_phase2: not indir blk"); 1682 for (indirdep = NULL, newindirdep = NULL; ; ) { 1683 ACQUIRE_LOCK(&lk); 1684 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 1685 if (wk->wk_type != D_INDIRDEP) 1686 continue; 1687 indirdep = WK_INDIRDEP(wk); 1688 break; 1689 } 1690 if (indirdep == NULL && newindirdep) { 1691 indirdep = newindirdep; 1692 WORKLIST_INSERT_BP(bp, &indirdep->ir_list); 1693 newindirdep = NULL; 1694 } 1695 FREE_LOCK(&lk); 1696 if (indirdep) { 1697 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0, 1698 &newblk) == 0) 1699 panic("setup_allocindir: lost block"); 1700 ACQUIRE_LOCK(&lk); 1701 if (newblk->nb_state == DEPCOMPLETE) { 1702 aip->ai_state |= DEPCOMPLETE; 1703 aip->ai_buf = NULL; 1704 } else { 1705 bmsafemap = newblk->nb_bmsafemap; 1706 aip->ai_buf = bmsafemap->sm_buf; 1707 LIST_REMOVE(newblk, nb_deps); 1708 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd, 1709 aip, ai_deps); 1710 } 1711 LIST_REMOVE(newblk, nb_hash); 1712 FREE(newblk, M_NEWBLK); 1713 aip->ai_indirdep = indirdep; 1714 /* 1715 * Check to see if there is an existing dependency 1716 * for this block. If there is, merge the old 1717 * dependency into the new one. 1718 */ 1719 if (aip->ai_oldblkno == 0) 1720 oldaip = NULL; 1721 else 1722 1723 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) 1724 if (oldaip->ai_offset == aip->ai_offset) 1725 break; 1726 if (oldaip != NULL) { 1727 if (oldaip->ai_newblkno != aip->ai_oldblkno) { 1728 FREE_LOCK(&lk); 1729 panic("setup_allocindir_phase2: blkno"); 1730 } 1731 aip->ai_oldblkno = oldaip->ai_oldblkno; 1732 freefrag = oldaip->ai_freefrag; 1733 oldaip->ai_freefrag = aip->ai_freefrag; 1734 aip->ai_freefrag = freefrag; 1735 free_allocindir(oldaip, NULL); 1736 } 1737 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 1738 ((ufs_daddr_t *)indirdep->ir_savebp->b_data) 1739 [aip->ai_offset] = aip->ai_oldblkno; 1740 FREE_LOCK(&lk); 1741 } 1742 if (newindirdep) { 1743 /* 1744 * Avoid any possibility of data corruption by 1745 * ensuring that our old version is thrown away. 1746 */ 1747 newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 1748 brelse(newindirdep->ir_savebp); 1749 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP); 1750 } 1751 if (indirdep) 1752 break; 1753 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep), 1754 M_INDIRDEP, M_SOFTDEP_FLAGS); 1755 newindirdep->ir_list.wk_type = D_INDIRDEP; 1756 newindirdep->ir_state = ATTACHED; 1757 LIST_INIT(&newindirdep->ir_deplisthd); 1758 LIST_INIT(&newindirdep->ir_donehd); 1759 if (bp->b_bio2.bio_offset == NOOFFSET) { 1760 VOP_BMAP(bp->b_vp, bp->b_bio1.bio_offset, 1761 &bp->b_bio2.bio_offset, NULL, NULL); 1762 } 1763 KKASSERT(bp->b_bio2.bio_offset != NOOFFSET); 1764 newindirdep->ir_savebp = getblk(ip->i_devvp, 1765 bp->b_bio2.bio_offset, 1766 bp->b_bcount, 0, 0); 1767 BUF_KERNPROC(newindirdep->ir_savebp); 1768 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 1769 } 1770 } 1771 1772 /* 1773 * Block de-allocation dependencies. 1774 * 1775 * When blocks are de-allocated, the on-disk pointers must be nullified before 1776 * the blocks are made available for use by other files. (The true 1777 * requirement is that old pointers must be nullified before new on-disk 1778 * pointers are set. We chose this slightly more stringent requirement to 1779 * reduce complexity.) Our implementation handles this dependency by updating 1780 * the inode (or indirect block) appropriately but delaying the actual block 1781 * de-allocation (i.e., freemap and free space count manipulation) until 1782 * after the updated versions reach stable storage. After the disk is 1783 * updated, the blocks can be safely de-allocated whenever it is convenient. 1784 * This implementation handles only the common case of reducing a file's 1785 * length to zero. Other cases are handled by the conventional synchronous 1786 * write approach. 1787 * 1788 * The ffs implementation with which we worked double-checks 1789 * the state of the block pointers and file size as it reduces 1790 * a file's length. Some of this code is replicated here in our 1791 * soft updates implementation. The freeblks->fb_chkcnt field is 1792 * used to transfer a part of this information to the procedure 1793 * that eventually de-allocates the blocks. 1794 * 1795 * This routine should be called from the routine that shortens 1796 * a file's length, before the inode's size or block pointers 1797 * are modified. It will save the block pointer information for 1798 * later release and zero the inode so that the calling routine 1799 * can release it. 1800 */ 1801 struct softdep_setup_freeblocks_info { 1802 struct fs *fs; 1803 struct inode *ip; 1804 }; 1805 1806 static int softdep_setup_freeblocks_bp(struct buf *bp, void *data); 1807 1808 /* 1809 * Parameters: 1810 * ip: The inode whose length is to be reduced 1811 * length: The new length for the file 1812 */ 1813 void 1814 softdep_setup_freeblocks(struct inode *ip, off_t length) 1815 { 1816 struct softdep_setup_freeblocks_info info; 1817 struct freeblks *freeblks; 1818 struct inodedep *inodedep; 1819 struct allocdirect *adp; 1820 struct vnode *vp; 1821 struct buf *bp; 1822 struct fs *fs; 1823 int i, error, delay; 1824 int count; 1825 1826 fs = ip->i_fs; 1827 if (length != 0) 1828 panic("softde_setup_freeblocks: non-zero length"); 1829 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks), 1830 M_FREEBLKS, M_SOFTDEP_FLAGS); 1831 bzero(freeblks, sizeof(struct freeblks)); 1832 freeblks->fb_list.wk_type = D_FREEBLKS; 1833 freeblks->fb_state = ATTACHED; 1834 freeblks->fb_uid = ip->i_uid; 1835 freeblks->fb_previousinum = ip->i_number; 1836 freeblks->fb_devvp = ip->i_devvp; 1837 freeblks->fb_fs = fs; 1838 freeblks->fb_oldsize = ip->i_size; 1839 freeblks->fb_newsize = length; 1840 freeblks->fb_chkcnt = ip->i_blocks; 1841 for (i = 0; i < NDADDR; i++) { 1842 freeblks->fb_dblks[i] = ip->i_db[i]; 1843 ip->i_db[i] = 0; 1844 } 1845 for (i = 0; i < NIADDR; i++) { 1846 freeblks->fb_iblks[i] = ip->i_ib[i]; 1847 ip->i_ib[i] = 0; 1848 } 1849 ip->i_blocks = 0; 1850 ip->i_size = 0; 1851 /* 1852 * Push the zero'ed inode to to its disk buffer so that we are free 1853 * to delete its dependencies below. Once the dependencies are gone 1854 * the buffer can be safely released. 1855 */ 1856 if ((error = bread(ip->i_devvp, 1857 fsbtodoff(fs, ino_to_fsba(fs, ip->i_number)), 1858 (int)fs->fs_bsize, &bp)) != 0) 1859 softdep_error("softdep_setup_freeblocks", error); 1860 *((struct ufs1_dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) = 1861 ip->i_din; 1862 /* 1863 * Find and eliminate any inode dependencies. 1864 */ 1865 ACQUIRE_LOCK(&lk); 1866 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep); 1867 if ((inodedep->id_state & IOSTARTED) != 0) { 1868 FREE_LOCK(&lk); 1869 panic("softdep_setup_freeblocks: inode busy"); 1870 } 1871 /* 1872 * Add the freeblks structure to the list of operations that 1873 * must await the zero'ed inode being written to disk. If we 1874 * still have a bitmap dependency (delay == 0), then the inode 1875 * has never been written to disk, so we can process the 1876 * freeblks below once we have deleted the dependencies. 1877 */ 1878 delay = (inodedep->id_state & DEPCOMPLETE); 1879 if (delay) 1880 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list); 1881 /* 1882 * Because the file length has been truncated to zero, any 1883 * pending block allocation dependency structures associated 1884 * with this inode are obsolete and can simply be de-allocated. 1885 * We must first merge the two dependency lists to get rid of 1886 * any duplicate freefrag structures, then purge the merged list. 1887 */ 1888 merge_inode_lists(inodedep); 1889 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 1890 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 1891 FREE_LOCK(&lk); 1892 bdwrite(bp); 1893 /* 1894 * We must wait for any I/O in progress to finish so that 1895 * all potential buffers on the dirty list will be visible. 1896 * Once they are all there, walk the list and get rid of 1897 * any dependencies. 1898 */ 1899 vp = ITOV(ip); 1900 ACQUIRE_LOCK(&lk); 1901 drain_output(vp, 1); 1902 1903 info.fs = fs; 1904 info.ip = ip; 1905 do { 1906 count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 1907 softdep_setup_freeblocks_bp, &info); 1908 } while (count != 0); 1909 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0) 1910 (void)free_inodedep(inodedep); 1911 1912 if (delay) { 1913 freeblks->fb_state |= DEPCOMPLETE; 1914 /* 1915 * If the inode with zeroed block pointers is now on disk 1916 * we can start freeing blocks. Add freeblks to the worklist 1917 * instead of calling handle_workitem_freeblocks directly as 1918 * it is more likely that additional IO is needed to complete 1919 * the request here than in the !delay case. 1920 */ 1921 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 1922 add_to_worklist(&freeblks->fb_list); 1923 } 1924 1925 FREE_LOCK(&lk); 1926 /* 1927 * If the inode has never been written to disk (delay == 0), 1928 * then we can process the freeblks now that we have deleted 1929 * the dependencies. 1930 */ 1931 if (!delay) 1932 handle_workitem_freeblocks(freeblks); 1933 } 1934 1935 static int 1936 softdep_setup_freeblocks_bp(struct buf *bp, void *data) 1937 { 1938 struct softdep_setup_freeblocks_info *info = data; 1939 struct inodedep *inodedep; 1940 1941 if (getdirtybuf(&bp, MNT_WAIT) == 0) { 1942 kprintf("softdep_setup_freeblocks_bp(1): caught bp %p going away\n", bp); 1943 return(-1); 1944 } 1945 if (bp->b_vp != ITOV(info->ip) || (bp->b_flags & B_DELWRI) == 0) { 1946 kprintf("softdep_setup_freeblocks_bp(2): caught bp %p going away\n", bp); 1947 BUF_UNLOCK(bp); 1948 return(-1); 1949 } 1950 (void) inodedep_lookup(info->fs, info->ip->i_number, 0, &inodedep); 1951 deallocate_dependencies(bp, inodedep); 1952 bp->b_flags |= B_INVAL | B_NOCACHE; 1953 FREE_LOCK(&lk); 1954 brelse(bp); 1955 ACQUIRE_LOCK(&lk); 1956 return(1); 1957 } 1958 1959 /* 1960 * Reclaim any dependency structures from a buffer that is about to 1961 * be reallocated to a new vnode. The buffer must be locked, thus, 1962 * no I/O completion operations can occur while we are manipulating 1963 * its associated dependencies. The mutex is held so that other I/O's 1964 * associated with related dependencies do not occur. 1965 */ 1966 static void 1967 deallocate_dependencies(struct buf *bp, struct inodedep *inodedep) 1968 { 1969 struct worklist *wk; 1970 struct indirdep *indirdep; 1971 struct allocindir *aip; 1972 struct pagedep *pagedep; 1973 struct dirrem *dirrem; 1974 struct diradd *dap; 1975 int i; 1976 1977 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 1978 switch (wk->wk_type) { 1979 1980 case D_INDIRDEP: 1981 indirdep = WK_INDIRDEP(wk); 1982 /* 1983 * None of the indirect pointers will ever be visible, 1984 * so they can simply be tossed. GOINGAWAY ensures 1985 * that allocated pointers will be saved in the buffer 1986 * cache until they are freed. Note that they will 1987 * only be able to be found by their physical address 1988 * since the inode mapping the logical address will 1989 * be gone. The save buffer used for the safe copy 1990 * was allocated in setup_allocindir_phase2 using 1991 * the physical address so it could be used for this 1992 * purpose. Hence we swap the safe copy with the real 1993 * copy, allowing the safe copy to be freed and holding 1994 * on to the real copy for later use in indir_trunc. 1995 * 1996 * NOTE: ir_savebp is relative to the block device 1997 * so b_bio1 contains the device block number. 1998 */ 1999 if (indirdep->ir_state & GOINGAWAY) { 2000 FREE_LOCK(&lk); 2001 panic("deallocate_dependencies: already gone"); 2002 } 2003 indirdep->ir_state |= GOINGAWAY; 2004 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 2005 free_allocindir(aip, inodedep); 2006 if (bp->b_bio1.bio_offset >= 0 || 2007 bp->b_bio2.bio_offset != indirdep->ir_savebp->b_bio1.bio_offset) { 2008 FREE_LOCK(&lk); 2009 panic("deallocate_dependencies: not indir"); 2010 } 2011 bcopy(bp->b_data, indirdep->ir_savebp->b_data, 2012 bp->b_bcount); 2013 WORKLIST_REMOVE(wk); 2014 WORKLIST_INSERT_BP(indirdep->ir_savebp, wk); 2015 continue; 2016 2017 case D_PAGEDEP: 2018 pagedep = WK_PAGEDEP(wk); 2019 /* 2020 * None of the directory additions will ever be 2021 * visible, so they can simply be tossed. 2022 */ 2023 for (i = 0; i < DAHASHSZ; i++) 2024 while ((dap = 2025 LIST_FIRST(&pagedep->pd_diraddhd[i]))) 2026 free_diradd(dap); 2027 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0) 2028 free_diradd(dap); 2029 /* 2030 * Copy any directory remove dependencies to the list 2031 * to be processed after the zero'ed inode is written. 2032 * If the inode has already been written, then they 2033 * can be dumped directly onto the work list. 2034 */ 2035 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 2036 LIST_REMOVE(dirrem, dm_next); 2037 dirrem->dm_dirinum = pagedep->pd_ino; 2038 if (inodedep == NULL || 2039 (inodedep->id_state & ALLCOMPLETE) == 2040 ALLCOMPLETE) 2041 add_to_worklist(&dirrem->dm_list); 2042 else 2043 WORKLIST_INSERT(&inodedep->id_bufwait, 2044 &dirrem->dm_list); 2045 } 2046 WORKLIST_REMOVE(&pagedep->pd_list); 2047 LIST_REMOVE(pagedep, pd_hash); 2048 WORKITEM_FREE(pagedep, D_PAGEDEP); 2049 continue; 2050 2051 case D_ALLOCINDIR: 2052 free_allocindir(WK_ALLOCINDIR(wk), inodedep); 2053 continue; 2054 2055 case D_ALLOCDIRECT: 2056 case D_INODEDEP: 2057 FREE_LOCK(&lk); 2058 panic("deallocate_dependencies: Unexpected type %s", 2059 TYPENAME(wk->wk_type)); 2060 /* NOTREACHED */ 2061 2062 default: 2063 FREE_LOCK(&lk); 2064 panic("deallocate_dependencies: Unknown type %s", 2065 TYPENAME(wk->wk_type)); 2066 /* NOTREACHED */ 2067 } 2068 } 2069 } 2070 2071 /* 2072 * Free an allocdirect. Generate a new freefrag work request if appropriate. 2073 * This routine must be called with splbio interrupts blocked. 2074 */ 2075 static void 2076 free_allocdirect(struct allocdirectlst *adphead, 2077 struct allocdirect *adp, int delay) 2078 { 2079 2080 #ifdef DEBUG 2081 if (lk.lkt_held == NOHOLDER) 2082 panic("free_allocdirect: lock not held"); 2083 #endif 2084 if ((adp->ad_state & DEPCOMPLETE) == 0) 2085 LIST_REMOVE(adp, ad_deps); 2086 TAILQ_REMOVE(adphead, adp, ad_next); 2087 if ((adp->ad_state & COMPLETE) == 0) 2088 WORKLIST_REMOVE(&adp->ad_list); 2089 if (adp->ad_freefrag != NULL) { 2090 if (delay) 2091 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 2092 &adp->ad_freefrag->ff_list); 2093 else 2094 add_to_worklist(&adp->ad_freefrag->ff_list); 2095 } 2096 WORKITEM_FREE(adp, D_ALLOCDIRECT); 2097 } 2098 2099 /* 2100 * Prepare an inode to be freed. The actual free operation is not 2101 * done until the zero'ed inode has been written to disk. 2102 */ 2103 void 2104 softdep_freefile(struct vnode *pvp, ino_t ino, int mode) 2105 { 2106 struct inode *ip = VTOI(pvp); 2107 struct inodedep *inodedep; 2108 struct freefile *freefile; 2109 2110 /* 2111 * This sets up the inode de-allocation dependency. 2112 */ 2113 MALLOC(freefile, struct freefile *, sizeof(struct freefile), 2114 M_FREEFILE, M_SOFTDEP_FLAGS); 2115 freefile->fx_list.wk_type = D_FREEFILE; 2116 freefile->fx_list.wk_state = 0; 2117 freefile->fx_mode = mode; 2118 freefile->fx_oldinum = ino; 2119 freefile->fx_devvp = ip->i_devvp; 2120 freefile->fx_fs = ip->i_fs; 2121 2122 /* 2123 * If the inodedep does not exist, then the zero'ed inode has 2124 * been written to disk. If the allocated inode has never been 2125 * written to disk, then the on-disk inode is zero'ed. In either 2126 * case we can free the file immediately. 2127 */ 2128 ACQUIRE_LOCK(&lk); 2129 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 || 2130 check_inode_unwritten(inodedep)) { 2131 FREE_LOCK(&lk); 2132 handle_workitem_freefile(freefile); 2133 return; 2134 } 2135 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 2136 FREE_LOCK(&lk); 2137 } 2138 2139 /* 2140 * Check to see if an inode has never been written to disk. If 2141 * so free the inodedep and return success, otherwise return failure. 2142 * This routine must be called with splbio interrupts blocked. 2143 * 2144 * If we still have a bitmap dependency, then the inode has never 2145 * been written to disk. Drop the dependency as it is no longer 2146 * necessary since the inode is being deallocated. We set the 2147 * ALLCOMPLETE flags since the bitmap now properly shows that the 2148 * inode is not allocated. Even if the inode is actively being 2149 * written, it has been rolled back to its zero'ed state, so we 2150 * are ensured that a zero inode is what is on the disk. For short 2151 * lived files, this change will usually result in removing all the 2152 * dependencies from the inode so that it can be freed immediately. 2153 */ 2154 static int 2155 check_inode_unwritten(struct inodedep *inodedep) 2156 { 2157 2158 if ((inodedep->id_state & DEPCOMPLETE) != 0 || 2159 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2160 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2161 LIST_FIRST(&inodedep->id_inowait) != NULL || 2162 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2163 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2164 inodedep->id_nlinkdelta != 0) 2165 return (0); 2166 2167 /* 2168 * Another process might be in initiate_write_inodeblock 2169 * trying to allocate memory without holding "Softdep Lock". 2170 */ 2171 if ((inodedep->id_state & IOSTARTED) != 0 && 2172 inodedep->id_savedino == NULL) 2173 return(0); 2174 2175 inodedep->id_state |= ALLCOMPLETE; 2176 LIST_REMOVE(inodedep, id_deps); 2177 inodedep->id_buf = NULL; 2178 if (inodedep->id_state & ONWORKLIST) 2179 WORKLIST_REMOVE(&inodedep->id_list); 2180 if (inodedep->id_savedino != NULL) { 2181 FREE(inodedep->id_savedino, M_INODEDEP); 2182 inodedep->id_savedino = NULL; 2183 } 2184 if (free_inodedep(inodedep) == 0) { 2185 FREE_LOCK(&lk); 2186 panic("check_inode_unwritten: busy inode"); 2187 } 2188 return (1); 2189 } 2190 2191 /* 2192 * Try to free an inodedep structure. Return 1 if it could be freed. 2193 */ 2194 static int 2195 free_inodedep(struct inodedep *inodedep) 2196 { 2197 2198 if ((inodedep->id_state & ONWORKLIST) != 0 || 2199 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 2200 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2201 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2202 LIST_FIRST(&inodedep->id_inowait) != NULL || 2203 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2204 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2205 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL) 2206 return (0); 2207 LIST_REMOVE(inodedep, id_hash); 2208 WORKITEM_FREE(inodedep, D_INODEDEP); 2209 num_inodedep -= 1; 2210 return (1); 2211 } 2212 2213 /* 2214 * This workitem routine performs the block de-allocation. 2215 * The workitem is added to the pending list after the updated 2216 * inode block has been written to disk. As mentioned above, 2217 * checks regarding the number of blocks de-allocated (compared 2218 * to the number of blocks allocated for the file) are also 2219 * performed in this function. 2220 */ 2221 static void 2222 handle_workitem_freeblocks(struct freeblks *freeblks) 2223 { 2224 struct inode tip; 2225 ufs_daddr_t bn; 2226 struct fs *fs; 2227 int i, level, bsize; 2228 long nblocks, blocksreleased = 0; 2229 int error, allerror = 0; 2230 ufs_lbn_t baselbns[NIADDR], tmpval; 2231 2232 tip.i_number = freeblks->fb_previousinum; 2233 tip.i_devvp = freeblks->fb_devvp; 2234 tip.i_dev = freeblks->fb_devvp->v_rdev; 2235 tip.i_fs = freeblks->fb_fs; 2236 tip.i_size = freeblks->fb_oldsize; 2237 tip.i_uid = freeblks->fb_uid; 2238 fs = freeblks->fb_fs; 2239 tmpval = 1; 2240 baselbns[0] = NDADDR; 2241 for (i = 1; i < NIADDR; i++) { 2242 tmpval *= NINDIR(fs); 2243 baselbns[i] = baselbns[i - 1] + tmpval; 2244 } 2245 nblocks = btodb(fs->fs_bsize); 2246 blocksreleased = 0; 2247 /* 2248 * Indirect blocks first. 2249 */ 2250 for (level = (NIADDR - 1); level >= 0; level--) { 2251 if ((bn = freeblks->fb_iblks[level]) == 0) 2252 continue; 2253 if ((error = indir_trunc(&tip, fsbtodoff(fs, bn), level, 2254 baselbns[level], &blocksreleased)) == 0) 2255 allerror = error; 2256 ffs_blkfree(&tip, bn, fs->fs_bsize); 2257 blocksreleased += nblocks; 2258 } 2259 /* 2260 * All direct blocks or frags. 2261 */ 2262 for (i = (NDADDR - 1); i >= 0; i--) { 2263 if ((bn = freeblks->fb_dblks[i]) == 0) 2264 continue; 2265 bsize = blksize(fs, &tip, i); 2266 ffs_blkfree(&tip, bn, bsize); 2267 blocksreleased += btodb(bsize); 2268 } 2269 2270 #ifdef DIAGNOSTIC 2271 if (freeblks->fb_chkcnt != blocksreleased) 2272 kprintf("handle_workitem_freeblocks: block count\n"); 2273 if (allerror) 2274 softdep_error("handle_workitem_freeblks", allerror); 2275 #endif /* DIAGNOSTIC */ 2276 WORKITEM_FREE(freeblks, D_FREEBLKS); 2277 } 2278 2279 /* 2280 * Release blocks associated with the inode ip and stored in the indirect 2281 * block at doffset. If level is greater than SINGLE, the block is an 2282 * indirect block and recursive calls to indirtrunc must be used to 2283 * cleanse other indirect blocks. 2284 */ 2285 static int 2286 indir_trunc(struct inode *ip, off_t doffset, int level, ufs_lbn_t lbn, 2287 long *countp) 2288 { 2289 struct buf *bp; 2290 ufs_daddr_t *bap; 2291 ufs_daddr_t nb; 2292 struct fs *fs; 2293 struct worklist *wk; 2294 struct indirdep *indirdep; 2295 int i, lbnadd, nblocks; 2296 int error, allerror = 0; 2297 2298 fs = ip->i_fs; 2299 lbnadd = 1; 2300 for (i = level; i > 0; i--) 2301 lbnadd *= NINDIR(fs); 2302 /* 2303 * Get buffer of block pointers to be freed. This routine is not 2304 * called until the zero'ed inode has been written, so it is safe 2305 * to free blocks as they are encountered. Because the inode has 2306 * been zero'ed, calls to bmap on these blocks will fail. So, we 2307 * have to use the on-disk address and the block device for the 2308 * filesystem to look them up. If the file was deleted before its 2309 * indirect blocks were all written to disk, the routine that set 2310 * us up (deallocate_dependencies) will have arranged to leave 2311 * a complete copy of the indirect block in memory for our use. 2312 * Otherwise we have to read the blocks in from the disk. 2313 */ 2314 ACQUIRE_LOCK(&lk); 2315 if ((bp = findblk(ip->i_devvp, doffset)) != NULL && 2316 (wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2317 /* 2318 * bp must be ir_savebp, which is held locked for our use. 2319 */ 2320 if (wk->wk_type != D_INDIRDEP || 2321 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp || 2322 (indirdep->ir_state & GOINGAWAY) == 0) { 2323 FREE_LOCK(&lk); 2324 panic("indir_trunc: lost indirdep"); 2325 } 2326 WORKLIST_REMOVE(wk); 2327 WORKITEM_FREE(indirdep, D_INDIRDEP); 2328 if (LIST_FIRST(&bp->b_dep) != NULL) { 2329 FREE_LOCK(&lk); 2330 panic("indir_trunc: dangling dep"); 2331 } 2332 FREE_LOCK(&lk); 2333 } else { 2334 FREE_LOCK(&lk); 2335 error = bread(ip->i_devvp, doffset, (int)fs->fs_bsize, &bp); 2336 if (error) 2337 return (error); 2338 } 2339 /* 2340 * Recursively free indirect blocks. 2341 */ 2342 bap = (ufs_daddr_t *)bp->b_data; 2343 nblocks = btodb(fs->fs_bsize); 2344 for (i = NINDIR(fs) - 1; i >= 0; i--) { 2345 if ((nb = bap[i]) == 0) 2346 continue; 2347 if (level != 0) { 2348 if ((error = indir_trunc(ip, fsbtodoff(fs, nb), 2349 level - 1, lbn + (i * lbnadd), countp)) != 0) 2350 allerror = error; 2351 } 2352 ffs_blkfree(ip, nb, fs->fs_bsize); 2353 *countp += nblocks; 2354 } 2355 bp->b_flags |= B_INVAL | B_NOCACHE; 2356 brelse(bp); 2357 return (allerror); 2358 } 2359 2360 /* 2361 * Free an allocindir. 2362 * This routine must be called with splbio interrupts blocked. 2363 */ 2364 static void 2365 free_allocindir(struct allocindir *aip, struct inodedep *inodedep) 2366 { 2367 struct freefrag *freefrag; 2368 2369 #ifdef DEBUG 2370 if (lk.lkt_held == NOHOLDER) 2371 panic("free_allocindir: lock not held"); 2372 #endif 2373 if ((aip->ai_state & DEPCOMPLETE) == 0) 2374 LIST_REMOVE(aip, ai_deps); 2375 if (aip->ai_state & ONWORKLIST) 2376 WORKLIST_REMOVE(&aip->ai_list); 2377 LIST_REMOVE(aip, ai_next); 2378 if ((freefrag = aip->ai_freefrag) != NULL) { 2379 if (inodedep == NULL) 2380 add_to_worklist(&freefrag->ff_list); 2381 else 2382 WORKLIST_INSERT(&inodedep->id_bufwait, 2383 &freefrag->ff_list); 2384 } 2385 WORKITEM_FREE(aip, D_ALLOCINDIR); 2386 } 2387 2388 /* 2389 * Directory entry addition dependencies. 2390 * 2391 * When adding a new directory entry, the inode (with its incremented link 2392 * count) must be written to disk before the directory entry's pointer to it. 2393 * Also, if the inode is newly allocated, the corresponding freemap must be 2394 * updated (on disk) before the directory entry's pointer. These requirements 2395 * are met via undo/redo on the directory entry's pointer, which consists 2396 * simply of the inode number. 2397 * 2398 * As directory entries are added and deleted, the free space within a 2399 * directory block can become fragmented. The ufs filesystem will compact 2400 * a fragmented directory block to make space for a new entry. When this 2401 * occurs, the offsets of previously added entries change. Any "diradd" 2402 * dependency structures corresponding to these entries must be updated with 2403 * the new offsets. 2404 */ 2405 2406 /* 2407 * This routine is called after the in-memory inode's link 2408 * count has been incremented, but before the directory entry's 2409 * pointer to the inode has been set. 2410 * 2411 * Parameters: 2412 * bp: buffer containing directory block 2413 * dp: inode for directory 2414 * diroffset: offset of new entry in directory 2415 * newinum: inode referenced by new directory entry 2416 * newdirbp: non-NULL => contents of new mkdir 2417 */ 2418 void 2419 softdep_setup_directory_add(struct buf *bp, struct inode *dp, off_t diroffset, 2420 ino_t newinum, struct buf *newdirbp) 2421 { 2422 int offset; /* offset of new entry within directory block */ 2423 ufs_lbn_t lbn; /* block in directory containing new entry */ 2424 struct fs *fs; 2425 struct diradd *dap; 2426 struct pagedep *pagedep; 2427 struct inodedep *inodedep; 2428 struct mkdir *mkdir1, *mkdir2; 2429 2430 /* 2431 * Whiteouts have no dependencies. 2432 */ 2433 if (newinum == WINO) { 2434 if (newdirbp != NULL) 2435 bdwrite(newdirbp); 2436 return; 2437 } 2438 2439 fs = dp->i_fs; 2440 lbn = lblkno(fs, diroffset); 2441 offset = blkoff(fs, diroffset); 2442 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, 2443 M_SOFTDEP_FLAGS); 2444 bzero(dap, sizeof(struct diradd)); 2445 dap->da_list.wk_type = D_DIRADD; 2446 dap->da_offset = offset; 2447 dap->da_newinum = newinum; 2448 dap->da_state = ATTACHED; 2449 if (newdirbp == NULL) { 2450 dap->da_state |= DEPCOMPLETE; 2451 ACQUIRE_LOCK(&lk); 2452 } else { 2453 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 2454 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2455 M_SOFTDEP_FLAGS); 2456 mkdir1->md_list.wk_type = D_MKDIR; 2457 mkdir1->md_state = MKDIR_BODY; 2458 mkdir1->md_diradd = dap; 2459 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2460 M_SOFTDEP_FLAGS); 2461 mkdir2->md_list.wk_type = D_MKDIR; 2462 mkdir2->md_state = MKDIR_PARENT; 2463 mkdir2->md_diradd = dap; 2464 /* 2465 * Dependency on "." and ".." being written to disk. 2466 */ 2467 mkdir1->md_buf = newdirbp; 2468 ACQUIRE_LOCK(&lk); 2469 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 2470 WORKLIST_INSERT_BP(newdirbp, &mkdir1->md_list); 2471 FREE_LOCK(&lk); 2472 bdwrite(newdirbp); 2473 /* 2474 * Dependency on link count increase for parent directory 2475 */ 2476 ACQUIRE_LOCK(&lk); 2477 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0 2478 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2479 dap->da_state &= ~MKDIR_PARENT; 2480 WORKITEM_FREE(mkdir2, D_MKDIR); 2481 } else { 2482 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 2483 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 2484 } 2485 } 2486 /* 2487 * Link into parent directory pagedep to await its being written. 2488 */ 2489 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2490 WORKLIST_INSERT_BP(bp, &pagedep->pd_list); 2491 dap->da_pagedep = pagedep; 2492 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 2493 da_pdlist); 2494 /* 2495 * Link into its inodedep. Put it on the id_bufwait list if the inode 2496 * is not yet written. If it is written, do the post-inode write 2497 * processing to put it on the id_pendinghd list. 2498 */ 2499 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep); 2500 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 2501 diradd_inode_written(dap, inodedep); 2502 else 2503 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2504 FREE_LOCK(&lk); 2505 } 2506 2507 /* 2508 * This procedure is called to change the offset of a directory 2509 * entry when compacting a directory block which must be owned 2510 * exclusively by the caller. Note that the actual entry movement 2511 * must be done in this procedure to ensure that no I/O completions 2512 * occur while the move is in progress. 2513 * 2514 * Parameters: 2515 * dp: inode for directory 2516 * base: address of dp->i_offset 2517 * oldloc: address of old directory location 2518 * newloc: address of new directory location 2519 * entrysize: size of directory entry 2520 */ 2521 void 2522 softdep_change_directoryentry_offset(struct inode *dp, caddr_t base, 2523 caddr_t oldloc, caddr_t newloc, 2524 int entrysize) 2525 { 2526 int offset, oldoffset, newoffset; 2527 struct pagedep *pagedep; 2528 struct diradd *dap; 2529 ufs_lbn_t lbn; 2530 2531 ACQUIRE_LOCK(&lk); 2532 lbn = lblkno(dp->i_fs, dp->i_offset); 2533 offset = blkoff(dp->i_fs, dp->i_offset); 2534 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) 2535 goto done; 2536 oldoffset = offset + (oldloc - base); 2537 newoffset = offset + (newloc - base); 2538 2539 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) { 2540 if (dap->da_offset != oldoffset) 2541 continue; 2542 dap->da_offset = newoffset; 2543 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) 2544 break; 2545 LIST_REMOVE(dap, da_pdlist); 2546 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], 2547 dap, da_pdlist); 2548 break; 2549 } 2550 if (dap == NULL) { 2551 2552 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) { 2553 if (dap->da_offset == oldoffset) { 2554 dap->da_offset = newoffset; 2555 break; 2556 } 2557 } 2558 } 2559 done: 2560 bcopy(oldloc, newloc, entrysize); 2561 FREE_LOCK(&lk); 2562 } 2563 2564 /* 2565 * Free a diradd dependency structure. This routine must be called 2566 * with splbio interrupts blocked. 2567 */ 2568 static void 2569 free_diradd(struct diradd *dap) 2570 { 2571 struct dirrem *dirrem; 2572 struct pagedep *pagedep; 2573 struct inodedep *inodedep; 2574 struct mkdir *mkdir, *nextmd; 2575 2576 #ifdef DEBUG 2577 if (lk.lkt_held == NOHOLDER) 2578 panic("free_diradd: lock not held"); 2579 #endif 2580 WORKLIST_REMOVE(&dap->da_list); 2581 LIST_REMOVE(dap, da_pdlist); 2582 if ((dap->da_state & DIRCHG) == 0) { 2583 pagedep = dap->da_pagedep; 2584 } else { 2585 dirrem = dap->da_previous; 2586 pagedep = dirrem->dm_pagedep; 2587 dirrem->dm_dirinum = pagedep->pd_ino; 2588 add_to_worklist(&dirrem->dm_list); 2589 } 2590 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum, 2591 0, &inodedep) != 0) 2592 (void) free_inodedep(inodedep); 2593 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2594 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 2595 nextmd = LIST_NEXT(mkdir, md_mkdirs); 2596 if (mkdir->md_diradd != dap) 2597 continue; 2598 dap->da_state &= ~mkdir->md_state; 2599 WORKLIST_REMOVE(&mkdir->md_list); 2600 LIST_REMOVE(mkdir, md_mkdirs); 2601 WORKITEM_FREE(mkdir, D_MKDIR); 2602 } 2603 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2604 FREE_LOCK(&lk); 2605 panic("free_diradd: unfound ref"); 2606 } 2607 } 2608 WORKITEM_FREE(dap, D_DIRADD); 2609 } 2610 2611 /* 2612 * Directory entry removal dependencies. 2613 * 2614 * When removing a directory entry, the entry's inode pointer must be 2615 * zero'ed on disk before the corresponding inode's link count is decremented 2616 * (possibly freeing the inode for re-use). This dependency is handled by 2617 * updating the directory entry but delaying the inode count reduction until 2618 * after the directory block has been written to disk. After this point, the 2619 * inode count can be decremented whenever it is convenient. 2620 */ 2621 2622 /* 2623 * This routine should be called immediately after removing 2624 * a directory entry. The inode's link count should not be 2625 * decremented by the calling procedure -- the soft updates 2626 * code will do this task when it is safe. 2627 * 2628 * Parameters: 2629 * bp: buffer containing directory block 2630 * dp: inode for the directory being modified 2631 * ip: inode for directory entry being removed 2632 * isrmdir: indicates if doing RMDIR 2633 */ 2634 void 2635 softdep_setup_remove(struct buf *bp, struct inode *dp, struct inode *ip, 2636 int isrmdir) 2637 { 2638 struct dirrem *dirrem, *prevdirrem; 2639 2640 /* 2641 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 2642 */ 2643 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2644 2645 /* 2646 * If the COMPLETE flag is clear, then there were no active 2647 * entries and we want to roll back to a zeroed entry until 2648 * the new inode is committed to disk. If the COMPLETE flag is 2649 * set then we have deleted an entry that never made it to 2650 * disk. If the entry we deleted resulted from a name change, 2651 * then the old name still resides on disk. We cannot delete 2652 * its inode (returned to us in prevdirrem) until the zeroed 2653 * directory entry gets to disk. The new inode has never been 2654 * referenced on the disk, so can be deleted immediately. 2655 */ 2656 if ((dirrem->dm_state & COMPLETE) == 0) { 2657 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 2658 dm_next); 2659 FREE_LOCK(&lk); 2660 } else { 2661 if (prevdirrem != NULL) 2662 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 2663 prevdirrem, dm_next); 2664 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 2665 FREE_LOCK(&lk); 2666 handle_workitem_remove(dirrem); 2667 } 2668 } 2669 2670 /* 2671 * Allocate a new dirrem if appropriate and return it along with 2672 * its associated pagedep. Called without a lock, returns with lock. 2673 */ 2674 static long num_dirrem; /* number of dirrem allocated */ 2675 2676 /* 2677 * Parameters: 2678 * bp: buffer containing directory block 2679 * dp: inode for the directory being modified 2680 * ip: inode for directory entry being removed 2681 * isrmdir: indicates if doing RMDIR 2682 * prevdirremp: previously referenced inode, if any 2683 */ 2684 static struct dirrem * 2685 newdirrem(struct buf *bp, struct inode *dp, struct inode *ip, 2686 int isrmdir, struct dirrem **prevdirremp) 2687 { 2688 int offset; 2689 ufs_lbn_t lbn; 2690 struct diradd *dap; 2691 struct dirrem *dirrem; 2692 struct pagedep *pagedep; 2693 2694 /* 2695 * Whiteouts have no deletion dependencies. 2696 */ 2697 if (ip == NULL) 2698 panic("newdirrem: whiteout"); 2699 /* 2700 * If we are over our limit, try to improve the situation. 2701 * Limiting the number of dirrem structures will also limit 2702 * the number of freefile and freeblks structures. 2703 */ 2704 if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0) 2705 (void) request_cleanup(FLUSH_REMOVE, 0); 2706 num_dirrem += 1; 2707 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem), 2708 M_DIRREM, M_SOFTDEP_FLAGS); 2709 bzero(dirrem, sizeof(struct dirrem)); 2710 dirrem->dm_list.wk_type = D_DIRREM; 2711 dirrem->dm_state = isrmdir ? RMDIR : 0; 2712 dirrem->dm_mnt = ITOV(ip)->v_mount; 2713 dirrem->dm_oldinum = ip->i_number; 2714 *prevdirremp = NULL; 2715 2716 ACQUIRE_LOCK(&lk); 2717 lbn = lblkno(dp->i_fs, dp->i_offset); 2718 offset = blkoff(dp->i_fs, dp->i_offset); 2719 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2720 WORKLIST_INSERT_BP(bp, &pagedep->pd_list); 2721 dirrem->dm_pagedep = pagedep; 2722 /* 2723 * Check for a diradd dependency for the same directory entry. 2724 * If present, then both dependencies become obsolete and can 2725 * be de-allocated. Check for an entry on both the pd_dirraddhd 2726 * list and the pd_pendinghd list. 2727 */ 2728 2729 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 2730 if (dap->da_offset == offset) 2731 break; 2732 if (dap == NULL) { 2733 2734 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 2735 if (dap->da_offset == offset) 2736 break; 2737 if (dap == NULL) 2738 return (dirrem); 2739 } 2740 /* 2741 * Must be ATTACHED at this point. 2742 */ 2743 if ((dap->da_state & ATTACHED) == 0) { 2744 FREE_LOCK(&lk); 2745 panic("newdirrem: not ATTACHED"); 2746 } 2747 if (dap->da_newinum != ip->i_number) { 2748 FREE_LOCK(&lk); 2749 panic("newdirrem: inum %"PRId64" should be %"PRId64, 2750 ip->i_number, dap->da_newinum); 2751 } 2752 /* 2753 * If we are deleting a changed name that never made it to disk, 2754 * then return the dirrem describing the previous inode (which 2755 * represents the inode currently referenced from this entry on disk). 2756 */ 2757 if ((dap->da_state & DIRCHG) != 0) { 2758 *prevdirremp = dap->da_previous; 2759 dap->da_state &= ~DIRCHG; 2760 dap->da_pagedep = pagedep; 2761 } 2762 /* 2763 * We are deleting an entry that never made it to disk. 2764 * Mark it COMPLETE so we can delete its inode immediately. 2765 */ 2766 dirrem->dm_state |= COMPLETE; 2767 free_diradd(dap); 2768 return (dirrem); 2769 } 2770 2771 /* 2772 * Directory entry change dependencies. 2773 * 2774 * Changing an existing directory entry requires that an add operation 2775 * be completed first followed by a deletion. The semantics for the addition 2776 * are identical to the description of adding a new entry above except 2777 * that the rollback is to the old inode number rather than zero. Once 2778 * the addition dependency is completed, the removal is done as described 2779 * in the removal routine above. 2780 */ 2781 2782 /* 2783 * This routine should be called immediately after changing 2784 * a directory entry. The inode's link count should not be 2785 * decremented by the calling procedure -- the soft updates 2786 * code will perform this task when it is safe. 2787 * 2788 * Parameters: 2789 * bp: buffer containing directory block 2790 * dp: inode for the directory being modified 2791 * ip: inode for directory entry being removed 2792 * newinum: new inode number for changed entry 2793 * isrmdir: indicates if doing RMDIR 2794 */ 2795 void 2796 softdep_setup_directory_change(struct buf *bp, struct inode *dp, 2797 struct inode *ip, ino_t newinum, 2798 int isrmdir) 2799 { 2800 int offset; 2801 struct diradd *dap = NULL; 2802 struct dirrem *dirrem, *prevdirrem; 2803 struct pagedep *pagedep; 2804 struct inodedep *inodedep; 2805 2806 offset = blkoff(dp->i_fs, dp->i_offset); 2807 2808 /* 2809 * Whiteouts do not need diradd dependencies. 2810 */ 2811 if (newinum != WINO) { 2812 MALLOC(dap, struct diradd *, sizeof(struct diradd), 2813 M_DIRADD, M_SOFTDEP_FLAGS); 2814 bzero(dap, sizeof(struct diradd)); 2815 dap->da_list.wk_type = D_DIRADD; 2816 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 2817 dap->da_offset = offset; 2818 dap->da_newinum = newinum; 2819 } 2820 2821 /* 2822 * Allocate a new dirrem and ACQUIRE_LOCK. 2823 */ 2824 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2825 pagedep = dirrem->dm_pagedep; 2826 /* 2827 * The possible values for isrmdir: 2828 * 0 - non-directory file rename 2829 * 1 - directory rename within same directory 2830 * inum - directory rename to new directory of given inode number 2831 * When renaming to a new directory, we are both deleting and 2832 * creating a new directory entry, so the link count on the new 2833 * directory should not change. Thus we do not need the followup 2834 * dirrem which is usually done in handle_workitem_remove. We set 2835 * the DIRCHG flag to tell handle_workitem_remove to skip the 2836 * followup dirrem. 2837 */ 2838 if (isrmdir > 1) 2839 dirrem->dm_state |= DIRCHG; 2840 2841 /* 2842 * Whiteouts have no additional dependencies, 2843 * so just put the dirrem on the correct list. 2844 */ 2845 if (newinum == WINO) { 2846 if ((dirrem->dm_state & COMPLETE) == 0) { 2847 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 2848 dm_next); 2849 } else { 2850 dirrem->dm_dirinum = pagedep->pd_ino; 2851 add_to_worklist(&dirrem->dm_list); 2852 } 2853 FREE_LOCK(&lk); 2854 return; 2855 } 2856 2857 /* 2858 * If the COMPLETE flag is clear, then there were no active 2859 * entries and we want to roll back to the previous inode until 2860 * the new inode is committed to disk. If the COMPLETE flag is 2861 * set, then we have deleted an entry that never made it to disk. 2862 * If the entry we deleted resulted from a name change, then the old 2863 * inode reference still resides on disk. Any rollback that we do 2864 * needs to be to that old inode (returned to us in prevdirrem). If 2865 * the entry we deleted resulted from a create, then there is 2866 * no entry on the disk, so we want to roll back to zero rather 2867 * than the uncommitted inode. In either of the COMPLETE cases we 2868 * want to immediately free the unwritten and unreferenced inode. 2869 */ 2870 if ((dirrem->dm_state & COMPLETE) == 0) { 2871 dap->da_previous = dirrem; 2872 } else { 2873 if (prevdirrem != NULL) { 2874 dap->da_previous = prevdirrem; 2875 } else { 2876 dap->da_state &= ~DIRCHG; 2877 dap->da_pagedep = pagedep; 2878 } 2879 dirrem->dm_dirinum = pagedep->pd_ino; 2880 add_to_worklist(&dirrem->dm_list); 2881 } 2882 /* 2883 * Link into its inodedep. Put it on the id_bufwait list if the inode 2884 * is not yet written. If it is written, do the post-inode write 2885 * processing to put it on the id_pendinghd list. 2886 */ 2887 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 || 2888 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2889 dap->da_state |= COMPLETE; 2890 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 2891 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 2892 } else { 2893 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 2894 dap, da_pdlist); 2895 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2896 } 2897 FREE_LOCK(&lk); 2898 } 2899 2900 /* 2901 * Called whenever the link count on an inode is changed. 2902 * It creates an inode dependency so that the new reference(s) 2903 * to the inode cannot be committed to disk until the updated 2904 * inode has been written. 2905 * 2906 * Parameters: 2907 * ip: the inode with the increased link count 2908 */ 2909 void 2910 softdep_change_linkcnt(struct inode *ip) 2911 { 2912 struct inodedep *inodedep; 2913 2914 ACQUIRE_LOCK(&lk); 2915 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 2916 if (ip->i_nlink < ip->i_effnlink) { 2917 FREE_LOCK(&lk); 2918 panic("softdep_change_linkcnt: bad delta"); 2919 } 2920 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2921 FREE_LOCK(&lk); 2922 } 2923 2924 /* 2925 * This workitem decrements the inode's link count. 2926 * If the link count reaches zero, the file is removed. 2927 */ 2928 static void 2929 handle_workitem_remove(struct dirrem *dirrem) 2930 { 2931 struct inodedep *inodedep; 2932 struct vnode *vp; 2933 struct inode *ip; 2934 ino_t oldinum; 2935 int error; 2936 2937 if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) { 2938 softdep_error("handle_workitem_remove: vget", error); 2939 return; 2940 } 2941 ip = VTOI(vp); 2942 ACQUIRE_LOCK(&lk); 2943 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){ 2944 FREE_LOCK(&lk); 2945 panic("handle_workitem_remove: lost inodedep"); 2946 } 2947 /* 2948 * Normal file deletion. 2949 */ 2950 if ((dirrem->dm_state & RMDIR) == 0) { 2951 ip->i_nlink--; 2952 ip->i_flag |= IN_CHANGE; 2953 if (ip->i_nlink < ip->i_effnlink) { 2954 FREE_LOCK(&lk); 2955 panic("handle_workitem_remove: bad file delta"); 2956 } 2957 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2958 FREE_LOCK(&lk); 2959 vput(vp); 2960 num_dirrem -= 1; 2961 WORKITEM_FREE(dirrem, D_DIRREM); 2962 return; 2963 } 2964 /* 2965 * Directory deletion. Decrement reference count for both the 2966 * just deleted parent directory entry and the reference for ".". 2967 * Next truncate the directory to length zero. When the 2968 * truncation completes, arrange to have the reference count on 2969 * the parent decremented to account for the loss of "..". 2970 */ 2971 ip->i_nlink -= 2; 2972 ip->i_flag |= IN_CHANGE; 2973 if (ip->i_nlink < ip->i_effnlink) { 2974 FREE_LOCK(&lk); 2975 panic("handle_workitem_remove: bad dir delta"); 2976 } 2977 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2978 FREE_LOCK(&lk); 2979 if ((error = ffs_truncate(vp, (off_t)0, 0, proc0.p_ucred)) != 0) 2980 softdep_error("handle_workitem_remove: truncate", error); 2981 /* 2982 * Rename a directory to a new parent. Since, we are both deleting 2983 * and creating a new directory entry, the link count on the new 2984 * directory should not change. Thus we skip the followup dirrem. 2985 */ 2986 if (dirrem->dm_state & DIRCHG) { 2987 vput(vp); 2988 num_dirrem -= 1; 2989 WORKITEM_FREE(dirrem, D_DIRREM); 2990 return; 2991 } 2992 /* 2993 * If the inodedep does not exist, then the zero'ed inode has 2994 * been written to disk. If the allocated inode has never been 2995 * written to disk, then the on-disk inode is zero'ed. In either 2996 * case we can remove the file immediately. 2997 */ 2998 ACQUIRE_LOCK(&lk); 2999 dirrem->dm_state = 0; 3000 oldinum = dirrem->dm_oldinum; 3001 dirrem->dm_oldinum = dirrem->dm_dirinum; 3002 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 || 3003 check_inode_unwritten(inodedep)) { 3004 FREE_LOCK(&lk); 3005 vput(vp); 3006 handle_workitem_remove(dirrem); 3007 return; 3008 } 3009 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 3010 FREE_LOCK(&lk); 3011 ip->i_flag |= IN_CHANGE; 3012 ffs_update(vp, 0); 3013 vput(vp); 3014 } 3015 3016 /* 3017 * Inode de-allocation dependencies. 3018 * 3019 * When an inode's link count is reduced to zero, it can be de-allocated. We 3020 * found it convenient to postpone de-allocation until after the inode is 3021 * written to disk with its new link count (zero). At this point, all of the 3022 * on-disk inode's block pointers are nullified and, with careful dependency 3023 * list ordering, all dependencies related to the inode will be satisfied and 3024 * the corresponding dependency structures de-allocated. So, if/when the 3025 * inode is reused, there will be no mixing of old dependencies with new 3026 * ones. This artificial dependency is set up by the block de-allocation 3027 * procedure above (softdep_setup_freeblocks) and completed by the 3028 * following procedure. 3029 */ 3030 static void 3031 handle_workitem_freefile(struct freefile *freefile) 3032 { 3033 struct vnode vp; 3034 struct inode tip; 3035 struct inodedep *idp; 3036 int error; 3037 3038 #ifdef DEBUG 3039 ACQUIRE_LOCK(&lk); 3040 error = inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp); 3041 FREE_LOCK(&lk); 3042 if (error) 3043 panic("handle_workitem_freefile: inodedep survived"); 3044 #endif 3045 tip.i_devvp = freefile->fx_devvp; 3046 tip.i_dev = freefile->fx_devvp->v_rdev; 3047 tip.i_fs = freefile->fx_fs; 3048 vp.v_data = &tip; 3049 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0) 3050 softdep_error("handle_workitem_freefile", error); 3051 WORKITEM_FREE(freefile, D_FREEFILE); 3052 } 3053 3054 /* 3055 * Helper function which unlinks marker element from work list and returns 3056 * the next element on the list. 3057 */ 3058 static __inline struct worklist * 3059 markernext(struct worklist *marker) 3060 { 3061 struct worklist *next; 3062 3063 next = LIST_NEXT(marker, wk_list); 3064 LIST_REMOVE(marker, wk_list); 3065 return next; 3066 } 3067 3068 /* 3069 * checkread, checkwrite 3070 * 3071 */ 3072 static int 3073 softdep_checkread(struct buf *bp) 3074 { 3075 return(0); 3076 } 3077 3078 static int 3079 softdep_checkwrite(struct buf *bp) 3080 { 3081 return(0); 3082 } 3083 3084 /* 3085 * Disk writes. 3086 * 3087 * The dependency structures constructed above are most actively used when file 3088 * system blocks are written to disk. No constraints are placed on when a 3089 * block can be written, but unsatisfied update dependencies are made safe by 3090 * modifying (or replacing) the source memory for the duration of the disk 3091 * write. When the disk write completes, the memory block is again brought 3092 * up-to-date. 3093 * 3094 * In-core inode structure reclamation. 3095 * 3096 * Because there are a finite number of "in-core" inode structures, they are 3097 * reused regularly. By transferring all inode-related dependencies to the 3098 * in-memory inode block and indexing them separately (via "inodedep"s), we 3099 * can allow "in-core" inode structures to be reused at any time and avoid 3100 * any increase in contention. 3101 * 3102 * Called just before entering the device driver to initiate a new disk I/O. 3103 * The buffer must be locked, thus, no I/O completion operations can occur 3104 * while we are manipulating its associated dependencies. 3105 * 3106 * Parameters: 3107 * bp: structure describing disk write to occur 3108 */ 3109 static void 3110 softdep_disk_io_initiation(struct buf *bp) 3111 { 3112 struct worklist *wk; 3113 struct worklist marker; 3114 struct indirdep *indirdep; 3115 3116 /* 3117 * We only care about write operations. There should never 3118 * be dependencies for reads. 3119 */ 3120 if (bp->b_cmd == BUF_CMD_READ) 3121 panic("softdep_disk_io_initiation: read"); 3122 3123 marker.wk_type = D_LAST + 1; /* Not a normal workitem */ 3124 3125 /* 3126 * Do any necessary pre-I/O processing. 3127 */ 3128 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = markernext(&marker)) { 3129 LIST_INSERT_AFTER(wk, &marker, wk_list); 3130 3131 switch (wk->wk_type) { 3132 3133 case D_PAGEDEP: 3134 initiate_write_filepage(WK_PAGEDEP(wk), bp); 3135 continue; 3136 3137 case D_INODEDEP: 3138 initiate_write_inodeblock(WK_INODEDEP(wk), bp); 3139 continue; 3140 3141 case D_INDIRDEP: 3142 indirdep = WK_INDIRDEP(wk); 3143 if (indirdep->ir_state & GOINGAWAY) 3144 panic("disk_io_initiation: indirdep gone"); 3145 /* 3146 * If there are no remaining dependencies, this 3147 * will be writing the real pointers, so the 3148 * dependency can be freed. 3149 */ 3150 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) { 3151 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 3152 brelse(indirdep->ir_savebp); 3153 /* inline expand WORKLIST_REMOVE(wk); */ 3154 wk->wk_state &= ~ONWORKLIST; 3155 LIST_REMOVE(wk, wk_list); 3156 WORKITEM_FREE(indirdep, D_INDIRDEP); 3157 continue; 3158 } 3159 /* 3160 * Replace up-to-date version with safe version. 3161 */ 3162 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount, 3163 M_INDIRDEP, M_SOFTDEP_FLAGS); 3164 ACQUIRE_LOCK(&lk); 3165 indirdep->ir_state &= ~ATTACHED; 3166 indirdep->ir_state |= UNDONE; 3167 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 3168 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 3169 bp->b_bcount); 3170 FREE_LOCK(&lk); 3171 continue; 3172 3173 case D_MKDIR: 3174 case D_BMSAFEMAP: 3175 case D_ALLOCDIRECT: 3176 case D_ALLOCINDIR: 3177 continue; 3178 3179 default: 3180 panic("handle_disk_io_initiation: Unexpected type %s", 3181 TYPENAME(wk->wk_type)); 3182 /* NOTREACHED */ 3183 } 3184 } 3185 } 3186 3187 /* 3188 * Called from within the procedure above to deal with unsatisfied 3189 * allocation dependencies in a directory. The buffer must be locked, 3190 * thus, no I/O completion operations can occur while we are 3191 * manipulating its associated dependencies. 3192 */ 3193 static void 3194 initiate_write_filepage(struct pagedep *pagedep, struct buf *bp) 3195 { 3196 struct diradd *dap; 3197 struct direct *ep; 3198 int i; 3199 3200 if (pagedep->pd_state & IOSTARTED) { 3201 /* 3202 * This can only happen if there is a driver that does not 3203 * understand chaining. Here biodone will reissue the call 3204 * to strategy for the incomplete buffers. 3205 */ 3206 kprintf("initiate_write_filepage: already started\n"); 3207 return; 3208 } 3209 pagedep->pd_state |= IOSTARTED; 3210 ACQUIRE_LOCK(&lk); 3211 for (i = 0; i < DAHASHSZ; i++) { 3212 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 3213 ep = (struct direct *) 3214 ((char *)bp->b_data + dap->da_offset); 3215 if (ep->d_ino != dap->da_newinum) { 3216 FREE_LOCK(&lk); 3217 panic("%s: dir inum %d != new %"PRId64, 3218 "initiate_write_filepage", 3219 ep->d_ino, dap->da_newinum); 3220 } 3221 if (dap->da_state & DIRCHG) 3222 ep->d_ino = dap->da_previous->dm_oldinum; 3223 else 3224 ep->d_ino = 0; 3225 dap->da_state &= ~ATTACHED; 3226 dap->da_state |= UNDONE; 3227 } 3228 } 3229 FREE_LOCK(&lk); 3230 } 3231 3232 /* 3233 * Called from within the procedure above to deal with unsatisfied 3234 * allocation dependencies in an inodeblock. The buffer must be 3235 * locked, thus, no I/O completion operations can occur while we 3236 * are manipulating its associated dependencies. 3237 * 3238 * Parameters: 3239 * bp: The inode block 3240 */ 3241 static void 3242 initiate_write_inodeblock(struct inodedep *inodedep, struct buf *bp) 3243 { 3244 struct allocdirect *adp, *lastadp; 3245 struct ufs1_dinode *dp; 3246 struct ufs1_dinode *sip; 3247 struct fs *fs; 3248 ufs_lbn_t prevlbn = 0; 3249 int i, deplist; 3250 3251 if (inodedep->id_state & IOSTARTED) 3252 panic("initiate_write_inodeblock: already started"); 3253 inodedep->id_state |= IOSTARTED; 3254 fs = inodedep->id_fs; 3255 dp = (struct ufs1_dinode *)bp->b_data + 3256 ino_to_fsbo(fs, inodedep->id_ino); 3257 /* 3258 * If the bitmap is not yet written, then the allocated 3259 * inode cannot be written to disk. 3260 */ 3261 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 3262 if (inodedep->id_savedino != NULL) 3263 panic("initiate_write_inodeblock: already doing I/O"); 3264 MALLOC(sip, struct ufs1_dinode *, 3265 sizeof(struct ufs1_dinode), M_INODEDEP, M_SOFTDEP_FLAGS); 3266 inodedep->id_savedino = sip; 3267 *inodedep->id_savedino = *dp; 3268 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 3269 dp->di_gen = inodedep->id_savedino->di_gen; 3270 return; 3271 } 3272 /* 3273 * If no dependencies, then there is nothing to roll back. 3274 */ 3275 inodedep->id_savedsize = dp->di_size; 3276 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 3277 return; 3278 /* 3279 * Set the dependencies to busy. 3280 */ 3281 ACQUIRE_LOCK(&lk); 3282 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3283 adp = TAILQ_NEXT(adp, ad_next)) { 3284 #ifdef DIAGNOSTIC 3285 if (deplist != 0 && prevlbn >= adp->ad_lbn) { 3286 FREE_LOCK(&lk); 3287 panic("softdep_write_inodeblock: lbn order"); 3288 } 3289 prevlbn = adp->ad_lbn; 3290 if (adp->ad_lbn < NDADDR && 3291 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) { 3292 FREE_LOCK(&lk); 3293 panic("%s: direct pointer #%ld mismatch %d != %d", 3294 "softdep_write_inodeblock", adp->ad_lbn, 3295 dp->di_db[adp->ad_lbn], adp->ad_newblkno); 3296 } 3297 if (adp->ad_lbn >= NDADDR && 3298 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) { 3299 FREE_LOCK(&lk); 3300 panic("%s: indirect pointer #%ld mismatch %d != %d", 3301 "softdep_write_inodeblock", adp->ad_lbn - NDADDR, 3302 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno); 3303 } 3304 deplist |= 1 << adp->ad_lbn; 3305 if ((adp->ad_state & ATTACHED) == 0) { 3306 FREE_LOCK(&lk); 3307 panic("softdep_write_inodeblock: Unknown state 0x%x", 3308 adp->ad_state); 3309 } 3310 #endif /* DIAGNOSTIC */ 3311 adp->ad_state &= ~ATTACHED; 3312 adp->ad_state |= UNDONE; 3313 } 3314 /* 3315 * The on-disk inode cannot claim to be any larger than the last 3316 * fragment that has been written. Otherwise, the on-disk inode 3317 * might have fragments that were not the last block in the file 3318 * which would corrupt the filesystem. 3319 */ 3320 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3321 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 3322 if (adp->ad_lbn >= NDADDR) 3323 break; 3324 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 3325 /* keep going until hitting a rollback to a frag */ 3326 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 3327 continue; 3328 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 3329 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 3330 #ifdef DIAGNOSTIC 3331 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) { 3332 FREE_LOCK(&lk); 3333 panic("softdep_write_inodeblock: lost dep1"); 3334 } 3335 #endif /* DIAGNOSTIC */ 3336 dp->di_db[i] = 0; 3337 } 3338 for (i = 0; i < NIADDR; i++) { 3339 #ifdef DIAGNOSTIC 3340 if (dp->di_ib[i] != 0 && 3341 (deplist & ((1 << NDADDR) << i)) == 0) { 3342 FREE_LOCK(&lk); 3343 panic("softdep_write_inodeblock: lost dep2"); 3344 } 3345 #endif /* DIAGNOSTIC */ 3346 dp->di_ib[i] = 0; 3347 } 3348 FREE_LOCK(&lk); 3349 return; 3350 } 3351 /* 3352 * If we have zero'ed out the last allocated block of the file, 3353 * roll back the size to the last currently allocated block. 3354 * We know that this last allocated block is a full-sized as 3355 * we already checked for fragments in the loop above. 3356 */ 3357 if (lastadp != NULL && 3358 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 3359 for (i = lastadp->ad_lbn; i >= 0; i--) 3360 if (dp->di_db[i] != 0) 3361 break; 3362 dp->di_size = (i + 1) * fs->fs_bsize; 3363 } 3364 /* 3365 * The only dependencies are for indirect blocks. 3366 * 3367 * The file size for indirect block additions is not guaranteed. 3368 * Such a guarantee would be non-trivial to achieve. The conventional 3369 * synchronous write implementation also does not make this guarantee. 3370 * Fsck should catch and fix discrepancies. Arguably, the file size 3371 * can be over-estimated without destroying integrity when the file 3372 * moves into the indirect blocks (i.e., is large). If we want to 3373 * postpone fsck, we are stuck with this argument. 3374 */ 3375 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 3376 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 3377 FREE_LOCK(&lk); 3378 } 3379 3380 /* 3381 * This routine is called during the completion interrupt 3382 * service routine for a disk write (from the procedure called 3383 * by the device driver to inform the filesystem caches of 3384 * a request completion). It should be called early in this 3385 * procedure, before the block is made available to other 3386 * processes or other routines are called. 3387 * 3388 * Parameters: 3389 * bp: describes the completed disk write 3390 */ 3391 static void 3392 softdep_disk_write_complete(struct buf *bp) 3393 { 3394 struct worklist *wk; 3395 struct workhead reattach; 3396 struct newblk *newblk; 3397 struct allocindir *aip; 3398 struct allocdirect *adp; 3399 struct indirdep *indirdep; 3400 struct inodedep *inodedep; 3401 struct bmsafemap *bmsafemap; 3402 3403 #ifdef DEBUG 3404 if (lk.lkt_held != NOHOLDER) 3405 panic("softdep_disk_write_complete: lock is held"); 3406 lk.lkt_held = SPECIAL_FLAG; 3407 #endif 3408 LIST_INIT(&reattach); 3409 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 3410 WORKLIST_REMOVE(wk); 3411 switch (wk->wk_type) { 3412 3413 case D_PAGEDEP: 3414 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 3415 WORKLIST_INSERT(&reattach, wk); 3416 continue; 3417 3418 case D_INODEDEP: 3419 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 3420 WORKLIST_INSERT(&reattach, wk); 3421 continue; 3422 3423 case D_BMSAFEMAP: 3424 bmsafemap = WK_BMSAFEMAP(wk); 3425 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) { 3426 newblk->nb_state |= DEPCOMPLETE; 3427 newblk->nb_bmsafemap = NULL; 3428 LIST_REMOVE(newblk, nb_deps); 3429 } 3430 while ((adp = 3431 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) { 3432 adp->ad_state |= DEPCOMPLETE; 3433 adp->ad_buf = NULL; 3434 LIST_REMOVE(adp, ad_deps); 3435 handle_allocdirect_partdone(adp); 3436 } 3437 while ((aip = 3438 LIST_FIRST(&bmsafemap->sm_allocindirhd))) { 3439 aip->ai_state |= DEPCOMPLETE; 3440 aip->ai_buf = NULL; 3441 LIST_REMOVE(aip, ai_deps); 3442 handle_allocindir_partdone(aip); 3443 } 3444 while ((inodedep = 3445 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { 3446 inodedep->id_state |= DEPCOMPLETE; 3447 LIST_REMOVE(inodedep, id_deps); 3448 inodedep->id_buf = NULL; 3449 } 3450 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 3451 continue; 3452 3453 case D_MKDIR: 3454 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 3455 continue; 3456 3457 case D_ALLOCDIRECT: 3458 adp = WK_ALLOCDIRECT(wk); 3459 adp->ad_state |= COMPLETE; 3460 handle_allocdirect_partdone(adp); 3461 continue; 3462 3463 case D_ALLOCINDIR: 3464 aip = WK_ALLOCINDIR(wk); 3465 aip->ai_state |= COMPLETE; 3466 handle_allocindir_partdone(aip); 3467 continue; 3468 3469 case D_INDIRDEP: 3470 indirdep = WK_INDIRDEP(wk); 3471 if (indirdep->ir_state & GOINGAWAY) { 3472 lk.lkt_held = NOHOLDER; 3473 panic("disk_write_complete: indirdep gone"); 3474 } 3475 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 3476 FREE(indirdep->ir_saveddata, M_INDIRDEP); 3477 indirdep->ir_saveddata = 0; 3478 indirdep->ir_state &= ~UNDONE; 3479 indirdep->ir_state |= ATTACHED; 3480 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 3481 handle_allocindir_partdone(aip); 3482 if (aip == LIST_FIRST(&indirdep->ir_donehd)) { 3483 lk.lkt_held = NOHOLDER; 3484 panic("disk_write_complete: not gone"); 3485 } 3486 } 3487 WORKLIST_INSERT(&reattach, wk); 3488 if ((bp->b_flags & B_DELWRI) == 0) 3489 stat_indir_blk_ptrs++; 3490 bdirty(bp); 3491 continue; 3492 3493 default: 3494 lk.lkt_held = NOHOLDER; 3495 panic("handle_disk_write_complete: Unknown type %s", 3496 TYPENAME(wk->wk_type)); 3497 /* NOTREACHED */ 3498 } 3499 } 3500 /* 3501 * Reattach any requests that must be redone. 3502 */ 3503 while ((wk = LIST_FIRST(&reattach)) != NULL) { 3504 WORKLIST_REMOVE(wk); 3505 WORKLIST_INSERT_BP(bp, wk); 3506 } 3507 #ifdef DEBUG 3508 if (lk.lkt_held != SPECIAL_FLAG) 3509 panic("softdep_disk_write_complete: lock lost"); 3510 lk.lkt_held = NOHOLDER; 3511 #endif 3512 } 3513 3514 /* 3515 * Called from within softdep_disk_write_complete above. Note that 3516 * this routine is always called from interrupt level with further 3517 * splbio interrupts blocked. 3518 * 3519 * Parameters: 3520 * adp: the completed allocdirect 3521 */ 3522 static void 3523 handle_allocdirect_partdone(struct allocdirect *adp) 3524 { 3525 struct allocdirect *listadp; 3526 struct inodedep *inodedep; 3527 long bsize; 3528 3529 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3530 return; 3531 if (adp->ad_buf != NULL) { 3532 lk.lkt_held = NOHOLDER; 3533 panic("handle_allocdirect_partdone: dangling dep"); 3534 } 3535 /* 3536 * The on-disk inode cannot claim to be any larger than the last 3537 * fragment that has been written. Otherwise, the on-disk inode 3538 * might have fragments that were not the last block in the file 3539 * which would corrupt the filesystem. Thus, we cannot free any 3540 * allocdirects after one whose ad_oldblkno claims a fragment as 3541 * these blocks must be rolled back to zero before writing the inode. 3542 * We check the currently active set of allocdirects in id_inoupdt. 3543 */ 3544 inodedep = adp->ad_inodedep; 3545 bsize = inodedep->id_fs->fs_bsize; 3546 TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) { 3547 /* found our block */ 3548 if (listadp == adp) 3549 break; 3550 /* continue if ad_oldlbn is not a fragment */ 3551 if (listadp->ad_oldsize == 0 || 3552 listadp->ad_oldsize == bsize) 3553 continue; 3554 /* hit a fragment */ 3555 return; 3556 } 3557 /* 3558 * If we have reached the end of the current list without 3559 * finding the just finished dependency, then it must be 3560 * on the future dependency list. Future dependencies cannot 3561 * be freed until they are moved to the current list. 3562 */ 3563 if (listadp == NULL) { 3564 #ifdef DEBUG 3565 TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next) 3566 /* found our block */ 3567 if (listadp == adp) 3568 break; 3569 if (listadp == NULL) { 3570 lk.lkt_held = NOHOLDER; 3571 panic("handle_allocdirect_partdone: lost dep"); 3572 } 3573 #endif /* DEBUG */ 3574 return; 3575 } 3576 /* 3577 * If we have found the just finished dependency, then free 3578 * it along with anything that follows it that is complete. 3579 */ 3580 for (; adp; adp = listadp) { 3581 listadp = TAILQ_NEXT(adp, ad_next); 3582 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3583 return; 3584 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 3585 } 3586 } 3587 3588 /* 3589 * Called from within softdep_disk_write_complete above. Note that 3590 * this routine is always called from interrupt level with further 3591 * splbio interrupts blocked. 3592 * 3593 * Parameters: 3594 * aip: the completed allocindir 3595 */ 3596 static void 3597 handle_allocindir_partdone(struct allocindir *aip) 3598 { 3599 struct indirdep *indirdep; 3600 3601 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 3602 return; 3603 if (aip->ai_buf != NULL) { 3604 lk.lkt_held = NOHOLDER; 3605 panic("handle_allocindir_partdone: dangling dependency"); 3606 } 3607 indirdep = aip->ai_indirdep; 3608 if (indirdep->ir_state & UNDONE) { 3609 LIST_REMOVE(aip, ai_next); 3610 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 3611 return; 3612 } 3613 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 3614 aip->ai_newblkno; 3615 LIST_REMOVE(aip, ai_next); 3616 if (aip->ai_freefrag != NULL) 3617 add_to_worklist(&aip->ai_freefrag->ff_list); 3618 WORKITEM_FREE(aip, D_ALLOCINDIR); 3619 } 3620 3621 /* 3622 * Called from within softdep_disk_write_complete above to restore 3623 * in-memory inode block contents to their most up-to-date state. Note 3624 * that this routine is always called from interrupt level with further 3625 * splbio interrupts blocked. 3626 * 3627 * Parameters: 3628 * bp: buffer containing the inode block 3629 */ 3630 static int 3631 handle_written_inodeblock(struct inodedep *inodedep, struct buf *bp) 3632 { 3633 struct worklist *wk, *filefree; 3634 struct allocdirect *adp, *nextadp; 3635 struct ufs1_dinode *dp; 3636 int hadchanges; 3637 3638 if ((inodedep->id_state & IOSTARTED) == 0) { 3639 lk.lkt_held = NOHOLDER; 3640 panic("handle_written_inodeblock: not started"); 3641 } 3642 inodedep->id_state &= ~IOSTARTED; 3643 dp = (struct ufs1_dinode *)bp->b_data + 3644 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 3645 /* 3646 * If we had to rollback the inode allocation because of 3647 * bitmaps being incomplete, then simply restore it. 3648 * Keep the block dirty so that it will not be reclaimed until 3649 * all associated dependencies have been cleared and the 3650 * corresponding updates written to disk. 3651 */ 3652 if (inodedep->id_savedino != NULL) { 3653 *dp = *inodedep->id_savedino; 3654 FREE(inodedep->id_savedino, M_INODEDEP); 3655 inodedep->id_savedino = NULL; 3656 if ((bp->b_flags & B_DELWRI) == 0) 3657 stat_inode_bitmap++; 3658 bdirty(bp); 3659 return (1); 3660 } 3661 inodedep->id_state |= COMPLETE; 3662 /* 3663 * Roll forward anything that had to be rolled back before 3664 * the inode could be updated. 3665 */ 3666 hadchanges = 0; 3667 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 3668 nextadp = TAILQ_NEXT(adp, ad_next); 3669 if (adp->ad_state & ATTACHED) { 3670 lk.lkt_held = NOHOLDER; 3671 panic("handle_written_inodeblock: new entry"); 3672 } 3673 if (adp->ad_lbn < NDADDR) { 3674 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) { 3675 lk.lkt_held = NOHOLDER; 3676 panic("%s: %s #%ld mismatch %d != %d", 3677 "handle_written_inodeblock", 3678 "direct pointer", adp->ad_lbn, 3679 dp->di_db[adp->ad_lbn], adp->ad_oldblkno); 3680 } 3681 dp->di_db[adp->ad_lbn] = adp->ad_newblkno; 3682 } else { 3683 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) { 3684 lk.lkt_held = NOHOLDER; 3685 panic("%s: %s #%ld allocated as %d", 3686 "handle_written_inodeblock", 3687 "indirect pointer", adp->ad_lbn - NDADDR, 3688 dp->di_ib[adp->ad_lbn - NDADDR]); 3689 } 3690 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno; 3691 } 3692 adp->ad_state &= ~UNDONE; 3693 adp->ad_state |= ATTACHED; 3694 hadchanges = 1; 3695 } 3696 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 3697 stat_direct_blk_ptrs++; 3698 /* 3699 * Reset the file size to its most up-to-date value. 3700 */ 3701 if (inodedep->id_savedsize == -1) { 3702 lk.lkt_held = NOHOLDER; 3703 panic("handle_written_inodeblock: bad size"); 3704 } 3705 if (dp->di_size != inodedep->id_savedsize) { 3706 dp->di_size = inodedep->id_savedsize; 3707 hadchanges = 1; 3708 } 3709 inodedep->id_savedsize = -1; 3710 /* 3711 * If there were any rollbacks in the inode block, then it must be 3712 * marked dirty so that its will eventually get written back in 3713 * its correct form. 3714 */ 3715 if (hadchanges) 3716 bdirty(bp); 3717 /* 3718 * Process any allocdirects that completed during the update. 3719 */ 3720 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 3721 handle_allocdirect_partdone(adp); 3722 /* 3723 * Process deallocations that were held pending until the 3724 * inode had been written to disk. Freeing of the inode 3725 * is delayed until after all blocks have been freed to 3726 * avoid creation of new <vfsid, inum, lbn> triples 3727 * before the old ones have been deleted. 3728 */ 3729 filefree = NULL; 3730 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 3731 WORKLIST_REMOVE(wk); 3732 switch (wk->wk_type) { 3733 3734 case D_FREEFILE: 3735 /* 3736 * We defer adding filefree to the worklist until 3737 * all other additions have been made to ensure 3738 * that it will be done after all the old blocks 3739 * have been freed. 3740 */ 3741 if (filefree != NULL) { 3742 lk.lkt_held = NOHOLDER; 3743 panic("handle_written_inodeblock: filefree"); 3744 } 3745 filefree = wk; 3746 continue; 3747 3748 case D_MKDIR: 3749 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 3750 continue; 3751 3752 case D_DIRADD: 3753 diradd_inode_written(WK_DIRADD(wk), inodedep); 3754 continue; 3755 3756 case D_FREEBLKS: 3757 wk->wk_state |= COMPLETE; 3758 if ((wk->wk_state & ALLCOMPLETE) != ALLCOMPLETE) 3759 continue; 3760 /* -- fall through -- */ 3761 case D_FREEFRAG: 3762 case D_DIRREM: 3763 add_to_worklist(wk); 3764 continue; 3765 3766 default: 3767 lk.lkt_held = NOHOLDER; 3768 panic("handle_written_inodeblock: Unknown type %s", 3769 TYPENAME(wk->wk_type)); 3770 /* NOTREACHED */ 3771 } 3772 } 3773 if (filefree != NULL) { 3774 if (free_inodedep(inodedep) == 0) { 3775 lk.lkt_held = NOHOLDER; 3776 panic("handle_written_inodeblock: live inodedep"); 3777 } 3778 add_to_worklist(filefree); 3779 return (0); 3780 } 3781 3782 /* 3783 * If no outstanding dependencies, free it. 3784 */ 3785 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0) 3786 return (0); 3787 return (hadchanges); 3788 } 3789 3790 /* 3791 * Process a diradd entry after its dependent inode has been written. 3792 * This routine must be called with splbio interrupts blocked. 3793 */ 3794 static void 3795 diradd_inode_written(struct diradd *dap, struct inodedep *inodedep) 3796 { 3797 struct pagedep *pagedep; 3798 3799 dap->da_state |= COMPLETE; 3800 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3801 if (dap->da_state & DIRCHG) 3802 pagedep = dap->da_previous->dm_pagedep; 3803 else 3804 pagedep = dap->da_pagedep; 3805 LIST_REMOVE(dap, da_pdlist); 3806 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3807 } 3808 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 3809 } 3810 3811 /* 3812 * Handle the completion of a mkdir dependency. 3813 */ 3814 static void 3815 handle_written_mkdir(struct mkdir *mkdir, int type) 3816 { 3817 struct diradd *dap; 3818 struct pagedep *pagedep; 3819 3820 if (mkdir->md_state != type) { 3821 lk.lkt_held = NOHOLDER; 3822 panic("handle_written_mkdir: bad type"); 3823 } 3824 dap = mkdir->md_diradd; 3825 dap->da_state &= ~type; 3826 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 3827 dap->da_state |= DEPCOMPLETE; 3828 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3829 if (dap->da_state & DIRCHG) 3830 pagedep = dap->da_previous->dm_pagedep; 3831 else 3832 pagedep = dap->da_pagedep; 3833 LIST_REMOVE(dap, da_pdlist); 3834 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3835 } 3836 LIST_REMOVE(mkdir, md_mkdirs); 3837 WORKITEM_FREE(mkdir, D_MKDIR); 3838 } 3839 3840 /* 3841 * Called from within softdep_disk_write_complete above. 3842 * A write operation was just completed. Removed inodes can 3843 * now be freed and associated block pointers may be committed. 3844 * Note that this routine is always called from interrupt level 3845 * with further splbio interrupts blocked. 3846 * 3847 * Parameters: 3848 * bp: buffer containing the written page 3849 */ 3850 static int 3851 handle_written_filepage(struct pagedep *pagedep, struct buf *bp) 3852 { 3853 struct dirrem *dirrem; 3854 struct diradd *dap, *nextdap; 3855 struct direct *ep; 3856 int i, chgs; 3857 3858 if ((pagedep->pd_state & IOSTARTED) == 0) { 3859 lk.lkt_held = NOHOLDER; 3860 panic("handle_written_filepage: not started"); 3861 } 3862 pagedep->pd_state &= ~IOSTARTED; 3863 /* 3864 * Process any directory removals that have been committed. 3865 */ 3866 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 3867 LIST_REMOVE(dirrem, dm_next); 3868 dirrem->dm_dirinum = pagedep->pd_ino; 3869 add_to_worklist(&dirrem->dm_list); 3870 } 3871 /* 3872 * Free any directory additions that have been committed. 3873 */ 3874 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 3875 free_diradd(dap); 3876 /* 3877 * Uncommitted directory entries must be restored. 3878 */ 3879 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 3880 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 3881 dap = nextdap) { 3882 nextdap = LIST_NEXT(dap, da_pdlist); 3883 if (dap->da_state & ATTACHED) { 3884 lk.lkt_held = NOHOLDER; 3885 panic("handle_written_filepage: attached"); 3886 } 3887 ep = (struct direct *) 3888 ((char *)bp->b_data + dap->da_offset); 3889 ep->d_ino = dap->da_newinum; 3890 dap->da_state &= ~UNDONE; 3891 dap->da_state |= ATTACHED; 3892 chgs = 1; 3893 /* 3894 * If the inode referenced by the directory has 3895 * been written out, then the dependency can be 3896 * moved to the pending list. 3897 */ 3898 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3899 LIST_REMOVE(dap, da_pdlist); 3900 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 3901 da_pdlist); 3902 } 3903 } 3904 } 3905 /* 3906 * If there were any rollbacks in the directory, then it must be 3907 * marked dirty so that its will eventually get written back in 3908 * its correct form. 3909 */ 3910 if (chgs) { 3911 if ((bp->b_flags & B_DELWRI) == 0) 3912 stat_dir_entry++; 3913 bdirty(bp); 3914 } 3915 /* 3916 * If no dependencies remain, the pagedep will be freed. 3917 * Otherwise it will remain to update the page before it 3918 * is written back to disk. 3919 */ 3920 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) { 3921 for (i = 0; i < DAHASHSZ; i++) 3922 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL) 3923 break; 3924 if (i == DAHASHSZ) { 3925 LIST_REMOVE(pagedep, pd_hash); 3926 WORKITEM_FREE(pagedep, D_PAGEDEP); 3927 return (0); 3928 } 3929 } 3930 return (1); 3931 } 3932 3933 /* 3934 * Writing back in-core inode structures. 3935 * 3936 * The filesystem only accesses an inode's contents when it occupies an 3937 * "in-core" inode structure. These "in-core" structures are separate from 3938 * the page frames used to cache inode blocks. Only the latter are 3939 * transferred to/from the disk. So, when the updated contents of the 3940 * "in-core" inode structure are copied to the corresponding in-memory inode 3941 * block, the dependencies are also transferred. The following procedure is 3942 * called when copying a dirty "in-core" inode to a cached inode block. 3943 */ 3944 3945 /* 3946 * Called when an inode is loaded from disk. If the effective link count 3947 * differed from the actual link count when it was last flushed, then we 3948 * need to ensure that the correct effective link count is put back. 3949 * 3950 * Parameters: 3951 * ip: the "in_core" copy of the inode 3952 */ 3953 void 3954 softdep_load_inodeblock(struct inode *ip) 3955 { 3956 struct inodedep *inodedep; 3957 3958 /* 3959 * Check for alternate nlink count. 3960 */ 3961 ip->i_effnlink = ip->i_nlink; 3962 ACQUIRE_LOCK(&lk); 3963 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3964 FREE_LOCK(&lk); 3965 return; 3966 } 3967 ip->i_effnlink -= inodedep->id_nlinkdelta; 3968 FREE_LOCK(&lk); 3969 } 3970 3971 /* 3972 * This routine is called just before the "in-core" inode 3973 * information is to be copied to the in-memory inode block. 3974 * Recall that an inode block contains several inodes. If 3975 * the force flag is set, then the dependencies will be 3976 * cleared so that the update can always be made. Note that 3977 * the buffer is locked when this routine is called, so we 3978 * will never be in the middle of writing the inode block 3979 * to disk. 3980 * 3981 * Parameters: 3982 * ip: the "in_core" copy of the inode 3983 * bp: the buffer containing the inode block 3984 * waitfor: nonzero => update must be allowed 3985 */ 3986 void 3987 softdep_update_inodeblock(struct inode *ip, struct buf *bp, 3988 int waitfor) 3989 { 3990 struct inodedep *inodedep; 3991 struct worklist *wk; 3992 int error, gotit; 3993 3994 /* 3995 * If the effective link count is not equal to the actual link 3996 * count, then we must track the difference in an inodedep while 3997 * the inode is (potentially) tossed out of the cache. Otherwise, 3998 * if there is no existing inodedep, then there are no dependencies 3999 * to track. 4000 */ 4001 ACQUIRE_LOCK(&lk); 4002 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 4003 FREE_LOCK(&lk); 4004 if (ip->i_effnlink != ip->i_nlink) 4005 panic("softdep_update_inodeblock: bad link count"); 4006 return; 4007 } 4008 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) { 4009 FREE_LOCK(&lk); 4010 panic("softdep_update_inodeblock: bad delta"); 4011 } 4012 /* 4013 * Changes have been initiated. Anything depending on these 4014 * changes cannot occur until this inode has been written. 4015 */ 4016 inodedep->id_state &= ~COMPLETE; 4017 if ((inodedep->id_state & ONWORKLIST) == 0) 4018 WORKLIST_INSERT_BP(bp, &inodedep->id_list); 4019 /* 4020 * Any new dependencies associated with the incore inode must 4021 * now be moved to the list associated with the buffer holding 4022 * the in-memory copy of the inode. Once merged process any 4023 * allocdirects that are completed by the merger. 4024 */ 4025 merge_inode_lists(inodedep); 4026 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL) 4027 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); 4028 /* 4029 * Now that the inode has been pushed into the buffer, the 4030 * operations dependent on the inode being written to disk 4031 * can be moved to the id_bufwait so that they will be 4032 * processed when the buffer I/O completes. 4033 */ 4034 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 4035 WORKLIST_REMOVE(wk); 4036 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 4037 } 4038 /* 4039 * Newly allocated inodes cannot be written until the bitmap 4040 * that allocates them have been written (indicated by 4041 * DEPCOMPLETE being set in id_state). If we are doing a 4042 * forced sync (e.g., an fsync on a file), we force the bitmap 4043 * to be written so that the update can be done. 4044 */ 4045 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) { 4046 FREE_LOCK(&lk); 4047 return; 4048 } 4049 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 4050 FREE_LOCK(&lk); 4051 if (gotit && 4052 (error = bwrite(inodedep->id_buf)) != 0) 4053 softdep_error("softdep_update_inodeblock: bwrite", error); 4054 } 4055 4056 /* 4057 * Merge the new inode dependency list (id_newinoupdt) into the old 4058 * inode dependency list (id_inoupdt). This routine must be called 4059 * with splbio interrupts blocked. 4060 */ 4061 static void 4062 merge_inode_lists(struct inodedep *inodedep) 4063 { 4064 struct allocdirect *listadp, *newadp; 4065 4066 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 4067 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) { 4068 if (listadp->ad_lbn < newadp->ad_lbn) { 4069 listadp = TAILQ_NEXT(listadp, ad_next); 4070 continue; 4071 } 4072 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 4073 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 4074 if (listadp->ad_lbn == newadp->ad_lbn) { 4075 allocdirect_merge(&inodedep->id_inoupdt, newadp, 4076 listadp); 4077 listadp = newadp; 4078 } 4079 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 4080 } 4081 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) { 4082 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 4083 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next); 4084 } 4085 } 4086 4087 /* 4088 * If we are doing an fsync, then we must ensure that any directory 4089 * entries for the inode have been written after the inode gets to disk. 4090 * 4091 * Parameters: 4092 * vp: the "in_core" copy of the inode 4093 */ 4094 static int 4095 softdep_fsync(struct vnode *vp) 4096 { 4097 struct inodedep *inodedep; 4098 struct pagedep *pagedep; 4099 struct worklist *wk; 4100 struct diradd *dap; 4101 struct mount *mnt; 4102 struct vnode *pvp; 4103 struct inode *ip; 4104 struct buf *bp; 4105 struct fs *fs; 4106 int error, flushparent; 4107 ino_t parentino; 4108 ufs_lbn_t lbn; 4109 4110 /* 4111 * Move check from original kernel code, possibly not needed any 4112 * more with the per-mount bioops. 4113 */ 4114 if ((vp->v_mount->mnt_flag & MNT_SOFTDEP) == 0) 4115 return (0); 4116 4117 ip = VTOI(vp); 4118 fs = ip->i_fs; 4119 ACQUIRE_LOCK(&lk); 4120 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) { 4121 FREE_LOCK(&lk); 4122 return (0); 4123 } 4124 if (LIST_FIRST(&inodedep->id_inowait) != NULL || 4125 LIST_FIRST(&inodedep->id_bufwait) != NULL || 4126 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 4127 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) { 4128 FREE_LOCK(&lk); 4129 panic("softdep_fsync: pending ops"); 4130 } 4131 for (error = 0, flushparent = 0; ; ) { 4132 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 4133 break; 4134 if (wk->wk_type != D_DIRADD) { 4135 FREE_LOCK(&lk); 4136 panic("softdep_fsync: Unexpected type %s", 4137 TYPENAME(wk->wk_type)); 4138 } 4139 dap = WK_DIRADD(wk); 4140 /* 4141 * Flush our parent if this directory entry 4142 * has a MKDIR_PARENT dependency. 4143 */ 4144 if (dap->da_state & DIRCHG) 4145 pagedep = dap->da_previous->dm_pagedep; 4146 else 4147 pagedep = dap->da_pagedep; 4148 mnt = pagedep->pd_mnt; 4149 parentino = pagedep->pd_ino; 4150 lbn = pagedep->pd_lbn; 4151 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) { 4152 FREE_LOCK(&lk); 4153 panic("softdep_fsync: dirty"); 4154 } 4155 flushparent = dap->da_state & MKDIR_PARENT; 4156 /* 4157 * If we are being fsync'ed as part of vgone'ing this vnode, 4158 * then we will not be able to release and recover the 4159 * vnode below, so we just have to give up on writing its 4160 * directory entry out. It will eventually be written, just 4161 * not now, but then the user was not asking to have it 4162 * written, so we are not breaking any promises. 4163 */ 4164 if (vp->v_flag & VRECLAIMED) 4165 break; 4166 /* 4167 * We prevent deadlock by always fetching inodes from the 4168 * root, moving down the directory tree. Thus, when fetching 4169 * our parent directory, we must unlock ourselves before 4170 * requesting the lock on our parent. See the comment in 4171 * ufs_lookup for details on possible races. 4172 */ 4173 FREE_LOCK(&lk); 4174 vn_unlock(vp); 4175 error = VFS_VGET(mnt, parentino, &pvp); 4176 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 4177 if (error != 0) 4178 return (error); 4179 if (flushparent) { 4180 if ((error = ffs_update(pvp, 1)) != 0) { 4181 vput(pvp); 4182 return (error); 4183 } 4184 } 4185 /* 4186 * Flush directory page containing the inode's name. 4187 */ 4188 error = bread(pvp, lblktodoff(fs, lbn), blksize(fs, VTOI(pvp), lbn), &bp); 4189 if (error == 0) 4190 error = bwrite(bp); 4191 vput(pvp); 4192 if (error != 0) 4193 return (error); 4194 ACQUIRE_LOCK(&lk); 4195 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) 4196 break; 4197 } 4198 FREE_LOCK(&lk); 4199 return (0); 4200 } 4201 4202 /* 4203 * Flush all the dirty bitmaps associated with the block device 4204 * before flushing the rest of the dirty blocks so as to reduce 4205 * the number of dependencies that will have to be rolled back. 4206 */ 4207 static int softdep_fsync_mountdev_bp(struct buf *bp, void *data); 4208 4209 void 4210 softdep_fsync_mountdev(struct vnode *vp) 4211 { 4212 if (!vn_isdisk(vp, NULL)) 4213 panic("softdep_fsync_mountdev: vnode not a disk"); 4214 ACQUIRE_LOCK(&lk); 4215 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 4216 softdep_fsync_mountdev_bp, vp); 4217 drain_output(vp, 1); 4218 FREE_LOCK(&lk); 4219 } 4220 4221 static int 4222 softdep_fsync_mountdev_bp(struct buf *bp, void *data) 4223 { 4224 struct worklist *wk; 4225 struct vnode *vp = data; 4226 4227 /* 4228 * If it is already scheduled, skip to the next buffer. 4229 */ 4230 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) 4231 return(0); 4232 if (bp->b_vp != vp || (bp->b_flags & B_DELWRI) == 0) { 4233 BUF_UNLOCK(bp); 4234 kprintf("softdep_fsync_mountdev_bp: warning, buffer %p ripped out from under vnode %p\n", bp, vp); 4235 return(0); 4236 } 4237 /* 4238 * We are only interested in bitmaps with outstanding 4239 * dependencies. 4240 */ 4241 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 4242 wk->wk_type != D_BMSAFEMAP) { 4243 BUF_UNLOCK(bp); 4244 return(0); 4245 } 4246 bremfree(bp); 4247 FREE_LOCK(&lk); 4248 (void) bawrite(bp); 4249 ACQUIRE_LOCK(&lk); 4250 return(0); 4251 } 4252 4253 /* 4254 * This routine is called when we are trying to synchronously flush a 4255 * file. This routine must eliminate any filesystem metadata dependencies 4256 * so that the syncing routine can succeed by pushing the dirty blocks 4257 * associated with the file. If any I/O errors occur, they are returned. 4258 */ 4259 struct softdep_sync_metadata_info { 4260 struct vnode *vp; 4261 int waitfor; 4262 }; 4263 4264 static int softdep_sync_metadata_bp(struct buf *bp, void *data); 4265 4266 int 4267 softdep_sync_metadata(struct vnode *vp, struct thread *td) 4268 { 4269 struct softdep_sync_metadata_info info; 4270 int error, waitfor; 4271 4272 /* 4273 * Check whether this vnode is involved in a filesystem 4274 * that is doing soft dependency processing. 4275 */ 4276 if (!vn_isdisk(vp, NULL)) { 4277 if (!DOINGSOFTDEP(vp)) 4278 return (0); 4279 } else 4280 if (vp->v_rdev->si_mountpoint == NULL || 4281 (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0) 4282 return (0); 4283 /* 4284 * Ensure that any direct block dependencies have been cleared. 4285 */ 4286 ACQUIRE_LOCK(&lk); 4287 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) { 4288 FREE_LOCK(&lk); 4289 return (error); 4290 } 4291 /* 4292 * For most files, the only metadata dependencies are the 4293 * cylinder group maps that allocate their inode or blocks. 4294 * The block allocation dependencies can be found by traversing 4295 * the dependency lists for any buffers that remain on their 4296 * dirty buffer list. The inode allocation dependency will 4297 * be resolved when the inode is updated with MNT_WAIT. 4298 * This work is done in two passes. The first pass grabs most 4299 * of the buffers and begins asynchronously writing them. The 4300 * only way to wait for these asynchronous writes is to sleep 4301 * on the filesystem vnode which may stay busy for a long time 4302 * if the filesystem is active. So, instead, we make a second 4303 * pass over the dependencies blocking on each write. In the 4304 * usual case we will be blocking against a write that we 4305 * initiated, so when it is done the dependency will have been 4306 * resolved. Thus the second pass is expected to end quickly. 4307 */ 4308 waitfor = MNT_NOWAIT; 4309 top: 4310 /* 4311 * We must wait for any I/O in progress to finish so that 4312 * all potential buffers on the dirty list will be visible. 4313 */ 4314 drain_output(vp, 1); 4315 info.vp = vp; 4316 info.waitfor = waitfor; 4317 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 4318 softdep_sync_metadata_bp, &info); 4319 if (error < 0) { 4320 FREE_LOCK(&lk); 4321 return(-error); /* error code */ 4322 } 4323 4324 /* 4325 * The brief unlock is to allow any pent up dependency 4326 * processing to be done. Then proceed with the second pass. 4327 */ 4328 if (waitfor == MNT_NOWAIT) { 4329 waitfor = MNT_WAIT; 4330 FREE_LOCK(&lk); 4331 ACQUIRE_LOCK(&lk); 4332 goto top; 4333 } 4334 4335 /* 4336 * If we have managed to get rid of all the dirty buffers, 4337 * then we are done. For certain directories and block 4338 * devices, we may need to do further work. 4339 * 4340 * We must wait for any I/O in progress to finish so that 4341 * all potential buffers on the dirty list will be visible. 4342 */ 4343 drain_output(vp, 1); 4344 if (RB_EMPTY(&vp->v_rbdirty_tree)) { 4345 FREE_LOCK(&lk); 4346 return (0); 4347 } 4348 4349 FREE_LOCK(&lk); 4350 /* 4351 * If we are trying to sync a block device, some of its buffers may 4352 * contain metadata that cannot be written until the contents of some 4353 * partially written files have been written to disk. The only easy 4354 * way to accomplish this is to sync the entire filesystem (luckily 4355 * this happens rarely). 4356 */ 4357 if (vn_isdisk(vp, NULL) && 4358 vp->v_rdev && 4359 vp->v_rdev->si_mountpoint && !vn_islocked(vp) && 4360 (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT)) != 0) 4361 return (error); 4362 return (0); 4363 } 4364 4365 static int 4366 softdep_sync_metadata_bp(struct buf *bp, void *data) 4367 { 4368 struct softdep_sync_metadata_info *info = data; 4369 struct pagedep *pagedep; 4370 struct allocdirect *adp; 4371 struct allocindir *aip; 4372 struct worklist *wk; 4373 struct buf *nbp; 4374 int error; 4375 int i; 4376 4377 if (getdirtybuf(&bp, MNT_WAIT) == 0) { 4378 kprintf("softdep_sync_metadata_bp(1): caught buf %p going away\n", bp); 4379 return (1); 4380 } 4381 if (bp->b_vp != info->vp || (bp->b_flags & B_DELWRI) == 0) { 4382 kprintf("softdep_sync_metadata_bp(2): caught buf %p going away vp %p\n", bp, info->vp); 4383 BUF_UNLOCK(bp); 4384 return(1); 4385 } 4386 4387 /* 4388 * As we hold the buffer locked, none of its dependencies 4389 * will disappear. 4390 */ 4391 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 4392 switch (wk->wk_type) { 4393 4394 case D_ALLOCDIRECT: 4395 adp = WK_ALLOCDIRECT(wk); 4396 if (adp->ad_state & DEPCOMPLETE) 4397 break; 4398 nbp = adp->ad_buf; 4399 if (getdirtybuf(&nbp, info->waitfor) == 0) 4400 break; 4401 FREE_LOCK(&lk); 4402 if (info->waitfor == MNT_NOWAIT) { 4403 bawrite(nbp); 4404 } else if ((error = bwrite(nbp)) != 0) { 4405 bawrite(bp); 4406 ACQUIRE_LOCK(&lk); 4407 return (-error); 4408 } 4409 ACQUIRE_LOCK(&lk); 4410 break; 4411 4412 case D_ALLOCINDIR: 4413 aip = WK_ALLOCINDIR(wk); 4414 if (aip->ai_state & DEPCOMPLETE) 4415 break; 4416 nbp = aip->ai_buf; 4417 if (getdirtybuf(&nbp, info->waitfor) == 0) 4418 break; 4419 FREE_LOCK(&lk); 4420 if (info->waitfor == MNT_NOWAIT) { 4421 bawrite(nbp); 4422 } else if ((error = bwrite(nbp)) != 0) { 4423 bawrite(bp); 4424 ACQUIRE_LOCK(&lk); 4425 return (-error); 4426 } 4427 ACQUIRE_LOCK(&lk); 4428 break; 4429 4430 case D_INDIRDEP: 4431 restart: 4432 4433 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) { 4434 if (aip->ai_state & DEPCOMPLETE) 4435 continue; 4436 nbp = aip->ai_buf; 4437 if (getdirtybuf(&nbp, MNT_WAIT) == 0) 4438 goto restart; 4439 FREE_LOCK(&lk); 4440 if ((error = bwrite(nbp)) != 0) { 4441 bawrite(bp); 4442 ACQUIRE_LOCK(&lk); 4443 return (-error); 4444 } 4445 ACQUIRE_LOCK(&lk); 4446 goto restart; 4447 } 4448 break; 4449 4450 case D_INODEDEP: 4451 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs, 4452 WK_INODEDEP(wk)->id_ino)) != 0) { 4453 FREE_LOCK(&lk); 4454 bawrite(bp); 4455 ACQUIRE_LOCK(&lk); 4456 return (-error); 4457 } 4458 break; 4459 4460 case D_PAGEDEP: 4461 /* 4462 * We are trying to sync a directory that may 4463 * have dependencies on both its own metadata 4464 * and/or dependencies on the inodes of any 4465 * recently allocated files. We walk its diradd 4466 * lists pushing out the associated inode. 4467 */ 4468 pagedep = WK_PAGEDEP(wk); 4469 for (i = 0; i < DAHASHSZ; i++) { 4470 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 4471 continue; 4472 if ((error = 4473 flush_pagedep_deps(info->vp, 4474 pagedep->pd_mnt, 4475 &pagedep->pd_diraddhd[i]))) { 4476 FREE_LOCK(&lk); 4477 bawrite(bp); 4478 ACQUIRE_LOCK(&lk); 4479 return (-error); 4480 } 4481 } 4482 break; 4483 4484 case D_MKDIR: 4485 /* 4486 * This case should never happen if the vnode has 4487 * been properly sync'ed. However, if this function 4488 * is used at a place where the vnode has not yet 4489 * been sync'ed, this dependency can show up. So, 4490 * rather than panic, just flush it. 4491 */ 4492 nbp = WK_MKDIR(wk)->md_buf; 4493 if (getdirtybuf(&nbp, info->waitfor) == 0) 4494 break; 4495 FREE_LOCK(&lk); 4496 if (info->waitfor == MNT_NOWAIT) { 4497 bawrite(nbp); 4498 } else if ((error = bwrite(nbp)) != 0) { 4499 bawrite(bp); 4500 ACQUIRE_LOCK(&lk); 4501 return (-error); 4502 } 4503 ACQUIRE_LOCK(&lk); 4504 break; 4505 4506 case D_BMSAFEMAP: 4507 /* 4508 * This case should never happen if the vnode has 4509 * been properly sync'ed. However, if this function 4510 * is used at a place where the vnode has not yet 4511 * been sync'ed, this dependency can show up. So, 4512 * rather than panic, just flush it. 4513 * 4514 * nbp can wind up == bp if a device node for the 4515 * same filesystem is being fsynced at the same time, 4516 * leading to a panic if we don't catch the case. 4517 */ 4518 nbp = WK_BMSAFEMAP(wk)->sm_buf; 4519 if (nbp == bp) 4520 break; 4521 if (getdirtybuf(&nbp, info->waitfor) == 0) 4522 break; 4523 FREE_LOCK(&lk); 4524 if (info->waitfor == MNT_NOWAIT) { 4525 bawrite(nbp); 4526 } else if ((error = bwrite(nbp)) != 0) { 4527 bawrite(bp); 4528 ACQUIRE_LOCK(&lk); 4529 return (-error); 4530 } 4531 ACQUIRE_LOCK(&lk); 4532 break; 4533 4534 default: 4535 FREE_LOCK(&lk); 4536 panic("softdep_sync_metadata: Unknown type %s", 4537 TYPENAME(wk->wk_type)); 4538 /* NOTREACHED */ 4539 } 4540 } 4541 FREE_LOCK(&lk); 4542 bawrite(bp); 4543 ACQUIRE_LOCK(&lk); 4544 return(0); 4545 } 4546 4547 /* 4548 * Flush the dependencies associated with an inodedep. 4549 * Called with splbio blocked. 4550 */ 4551 static int 4552 flush_inodedep_deps(struct fs *fs, ino_t ino) 4553 { 4554 struct inodedep *inodedep; 4555 struct allocdirect *adp; 4556 int error, waitfor; 4557 struct buf *bp; 4558 4559 /* 4560 * This work is done in two passes. The first pass grabs most 4561 * of the buffers and begins asynchronously writing them. The 4562 * only way to wait for these asynchronous writes is to sleep 4563 * on the filesystem vnode which may stay busy for a long time 4564 * if the filesystem is active. So, instead, we make a second 4565 * pass over the dependencies blocking on each write. In the 4566 * usual case we will be blocking against a write that we 4567 * initiated, so when it is done the dependency will have been 4568 * resolved. Thus the second pass is expected to end quickly. 4569 * We give a brief window at the top of the loop to allow 4570 * any pending I/O to complete. 4571 */ 4572 for (waitfor = MNT_NOWAIT; ; ) { 4573 FREE_LOCK(&lk); 4574 ACQUIRE_LOCK(&lk); 4575 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4576 return (0); 4577 TAILQ_FOREACH(adp, &inodedep->id_inoupdt, ad_next) { 4578 if (adp->ad_state & DEPCOMPLETE) 4579 continue; 4580 bp = adp->ad_buf; 4581 if (getdirtybuf(&bp, waitfor) == 0) { 4582 if (waitfor == MNT_NOWAIT) 4583 continue; 4584 break; 4585 } 4586 FREE_LOCK(&lk); 4587 if (waitfor == MNT_NOWAIT) { 4588 bawrite(bp); 4589 } else if ((error = bwrite(bp)) != 0) { 4590 ACQUIRE_LOCK(&lk); 4591 return (error); 4592 } 4593 ACQUIRE_LOCK(&lk); 4594 break; 4595 } 4596 if (adp != NULL) 4597 continue; 4598 TAILQ_FOREACH(adp, &inodedep->id_newinoupdt, ad_next) { 4599 if (adp->ad_state & DEPCOMPLETE) 4600 continue; 4601 bp = adp->ad_buf; 4602 if (getdirtybuf(&bp, waitfor) == 0) { 4603 if (waitfor == MNT_NOWAIT) 4604 continue; 4605 break; 4606 } 4607 FREE_LOCK(&lk); 4608 if (waitfor == MNT_NOWAIT) { 4609 bawrite(bp); 4610 } else if ((error = bwrite(bp)) != 0) { 4611 ACQUIRE_LOCK(&lk); 4612 return (error); 4613 } 4614 ACQUIRE_LOCK(&lk); 4615 break; 4616 } 4617 if (adp != NULL) 4618 continue; 4619 /* 4620 * If pass2, we are done, otherwise do pass 2. 4621 */ 4622 if (waitfor == MNT_WAIT) 4623 break; 4624 waitfor = MNT_WAIT; 4625 } 4626 /* 4627 * Try freeing inodedep in case all dependencies have been removed. 4628 */ 4629 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0) 4630 (void) free_inodedep(inodedep); 4631 return (0); 4632 } 4633 4634 /* 4635 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 4636 * Called with splbio blocked. 4637 */ 4638 static int 4639 flush_pagedep_deps(struct vnode *pvp, struct mount *mp, 4640 struct diraddhd *diraddhdp) 4641 { 4642 struct inodedep *inodedep; 4643 struct ufsmount *ump; 4644 struct diradd *dap; 4645 struct vnode *vp; 4646 int gotit, error = 0; 4647 struct buf *bp; 4648 ino_t inum; 4649 4650 ump = VFSTOUFS(mp); 4651 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 4652 /* 4653 * Flush ourselves if this directory entry 4654 * has a MKDIR_PARENT dependency. 4655 */ 4656 if (dap->da_state & MKDIR_PARENT) { 4657 FREE_LOCK(&lk); 4658 if ((error = ffs_update(pvp, 1)) != 0) 4659 break; 4660 ACQUIRE_LOCK(&lk); 4661 /* 4662 * If that cleared dependencies, go on to next. 4663 */ 4664 if (dap != LIST_FIRST(diraddhdp)) 4665 continue; 4666 if (dap->da_state & MKDIR_PARENT) { 4667 FREE_LOCK(&lk); 4668 panic("flush_pagedep_deps: MKDIR_PARENT"); 4669 } 4670 } 4671 /* 4672 * A newly allocated directory must have its "." and 4673 * ".." entries written out before its name can be 4674 * committed in its parent. We do not want or need 4675 * the full semantics of a synchronous VOP_FSYNC as 4676 * that may end up here again, once for each directory 4677 * level in the filesystem. Instead, we push the blocks 4678 * and wait for them to clear. We have to fsync twice 4679 * because the first call may choose to defer blocks 4680 * that still have dependencies, but deferral will 4681 * happen at most once. 4682 */ 4683 inum = dap->da_newinum; 4684 if (dap->da_state & MKDIR_BODY) { 4685 FREE_LOCK(&lk); 4686 if ((error = VFS_VGET(mp, inum, &vp)) != 0) 4687 break; 4688 if ((error=VOP_FSYNC(vp, MNT_NOWAIT)) || 4689 (error=VOP_FSYNC(vp, MNT_NOWAIT))) { 4690 vput(vp); 4691 break; 4692 } 4693 drain_output(vp, 0); 4694 vput(vp); 4695 ACQUIRE_LOCK(&lk); 4696 /* 4697 * If that cleared dependencies, go on to next. 4698 */ 4699 if (dap != LIST_FIRST(diraddhdp)) 4700 continue; 4701 if (dap->da_state & MKDIR_BODY) { 4702 FREE_LOCK(&lk); 4703 panic("flush_pagedep_deps: MKDIR_BODY"); 4704 } 4705 } 4706 /* 4707 * Flush the inode on which the directory entry depends. 4708 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 4709 * the only remaining dependency is that the updated inode 4710 * count must get pushed to disk. The inode has already 4711 * been pushed into its inode buffer (via VOP_UPDATE) at 4712 * the time of the reference count change. So we need only 4713 * locate that buffer, ensure that there will be no rollback 4714 * caused by a bitmap dependency, then write the inode buffer. 4715 */ 4716 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) { 4717 FREE_LOCK(&lk); 4718 panic("flush_pagedep_deps: lost inode"); 4719 } 4720 /* 4721 * If the inode still has bitmap dependencies, 4722 * push them to disk. 4723 */ 4724 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4725 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 4726 FREE_LOCK(&lk); 4727 if (gotit && (error = bwrite(inodedep->id_buf)) != 0) 4728 break; 4729 ACQUIRE_LOCK(&lk); 4730 if (dap != LIST_FIRST(diraddhdp)) 4731 continue; 4732 } 4733 /* 4734 * If the inode is still sitting in a buffer waiting 4735 * to be written, push it to disk. 4736 */ 4737 FREE_LOCK(&lk); 4738 if ((error = bread(ump->um_devvp, 4739 fsbtodoff(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), 4740 (int)ump->um_fs->fs_bsize, &bp)) != 0) 4741 break; 4742 if ((error = bwrite(bp)) != 0) 4743 break; 4744 ACQUIRE_LOCK(&lk); 4745 /* 4746 * If we have failed to get rid of all the dependencies 4747 * then something is seriously wrong. 4748 */ 4749 if (dap == LIST_FIRST(diraddhdp)) { 4750 FREE_LOCK(&lk); 4751 panic("flush_pagedep_deps: flush failed"); 4752 } 4753 } 4754 if (error) 4755 ACQUIRE_LOCK(&lk); 4756 return (error); 4757 } 4758 4759 /* 4760 * A large burst of file addition or deletion activity can drive the 4761 * memory load excessively high. First attempt to slow things down 4762 * using the techniques below. If that fails, this routine requests 4763 * the offending operations to fall back to running synchronously 4764 * until the memory load returns to a reasonable level. 4765 */ 4766 int 4767 softdep_slowdown(struct vnode *vp) 4768 { 4769 int max_softdeps_hard; 4770 4771 max_softdeps_hard = max_softdeps * 11 / 10; 4772 if (num_dirrem < max_softdeps_hard / 2 && 4773 num_inodedep < max_softdeps_hard) 4774 return (0); 4775 stat_sync_limit_hit += 1; 4776 return (1); 4777 } 4778 4779 /* 4780 * If memory utilization has gotten too high, deliberately slow things 4781 * down and speed up the I/O processing. 4782 */ 4783 static int 4784 request_cleanup(int resource, int islocked) 4785 { 4786 struct thread *td = curthread; /* XXX */ 4787 4788 /* 4789 * We never hold up the filesystem syncer process. 4790 */ 4791 if (td == filesys_syncer) 4792 return (0); 4793 /* 4794 * First check to see if the work list has gotten backlogged. 4795 * If it has, co-opt this process to help clean up two entries. 4796 * Because this process may hold inodes locked, we cannot 4797 * handle any remove requests that might block on a locked 4798 * inode as that could lead to deadlock. 4799 */ 4800 if (num_on_worklist > max_softdeps / 10) { 4801 if (islocked) 4802 FREE_LOCK(&lk); 4803 process_worklist_item(NULL, LK_NOWAIT); 4804 process_worklist_item(NULL, LK_NOWAIT); 4805 stat_worklist_push += 2; 4806 if (islocked) 4807 ACQUIRE_LOCK(&lk); 4808 return(1); 4809 } 4810 4811 /* 4812 * If we are resource constrained on inode dependencies, try 4813 * flushing some dirty inodes. Otherwise, we are constrained 4814 * by file deletions, so try accelerating flushes of directories 4815 * with removal dependencies. We would like to do the cleanup 4816 * here, but we probably hold an inode locked at this point and 4817 * that might deadlock against one that we try to clean. So, 4818 * the best that we can do is request the syncer daemon to do 4819 * the cleanup for us. 4820 */ 4821 switch (resource) { 4822 4823 case FLUSH_INODES: 4824 stat_ino_limit_push += 1; 4825 req_clear_inodedeps += 1; 4826 stat_countp = &stat_ino_limit_hit; 4827 break; 4828 4829 case FLUSH_REMOVE: 4830 stat_blk_limit_push += 1; 4831 req_clear_remove += 1; 4832 stat_countp = &stat_blk_limit_hit; 4833 break; 4834 4835 default: 4836 if (islocked) 4837 FREE_LOCK(&lk); 4838 panic("request_cleanup: unknown type"); 4839 } 4840 /* 4841 * Hopefully the syncer daemon will catch up and awaken us. 4842 * We wait at most tickdelay before proceeding in any case. 4843 */ 4844 if (islocked == 0) 4845 ACQUIRE_LOCK(&lk); 4846 proc_waiting += 1; 4847 if (!callout_active(&handle)) 4848 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2, 4849 pause_timer, NULL); 4850 interlocked_sleep(&lk, SLEEP, (caddr_t)&proc_waiting, 0, 4851 "softupdate", 0); 4852 proc_waiting -= 1; 4853 if (islocked == 0) 4854 FREE_LOCK(&lk); 4855 return (1); 4856 } 4857 4858 /* 4859 * Awaken processes pausing in request_cleanup and clear proc_waiting 4860 * to indicate that there is no longer a timer running. 4861 */ 4862 void 4863 pause_timer(void *arg) 4864 { 4865 *stat_countp += 1; 4866 wakeup_one(&proc_waiting); 4867 if (proc_waiting > 0) 4868 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2, 4869 pause_timer, NULL); 4870 else 4871 callout_deactivate(&handle); 4872 } 4873 4874 /* 4875 * Flush out a directory with at least one removal dependency in an effort to 4876 * reduce the number of dirrem, freefile, and freeblks dependency structures. 4877 */ 4878 static void 4879 clear_remove(struct thread *td) 4880 { 4881 struct pagedep_hashhead *pagedephd; 4882 struct pagedep *pagedep; 4883 static int next = 0; 4884 struct mount *mp; 4885 struct vnode *vp; 4886 int error, cnt; 4887 ino_t ino; 4888 4889 ACQUIRE_LOCK(&lk); 4890 for (cnt = 0; cnt < pagedep_hash; cnt++) { 4891 pagedephd = &pagedep_hashtbl[next++]; 4892 if (next >= pagedep_hash) 4893 next = 0; 4894 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 4895 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL) 4896 continue; 4897 mp = pagedep->pd_mnt; 4898 ino = pagedep->pd_ino; 4899 FREE_LOCK(&lk); 4900 if ((error = VFS_VGET(mp, ino, &vp)) != 0) { 4901 softdep_error("clear_remove: vget", error); 4902 return; 4903 } 4904 if ((error = VOP_FSYNC(vp, MNT_NOWAIT))) 4905 softdep_error("clear_remove: fsync", error); 4906 drain_output(vp, 0); 4907 vput(vp); 4908 return; 4909 } 4910 } 4911 FREE_LOCK(&lk); 4912 } 4913 4914 /* 4915 * Clear out a block of dirty inodes in an effort to reduce 4916 * the number of inodedep dependency structures. 4917 */ 4918 struct clear_inodedeps_info { 4919 struct fs *fs; 4920 struct mount *mp; 4921 }; 4922 4923 static int 4924 clear_inodedeps_mountlist_callback(struct mount *mp, void *data) 4925 { 4926 struct clear_inodedeps_info *info = data; 4927 4928 if ((mp->mnt_flag & MNT_SOFTDEP) && info->fs == VFSTOUFS(mp)->um_fs) { 4929 info->mp = mp; 4930 return(-1); 4931 } 4932 return(0); 4933 } 4934 4935 static void 4936 clear_inodedeps(struct thread *td) 4937 { 4938 struct clear_inodedeps_info info; 4939 struct inodedep_hashhead *inodedephd; 4940 struct inodedep *inodedep; 4941 static int next = 0; 4942 struct vnode *vp; 4943 struct fs *fs; 4944 int error, cnt; 4945 ino_t firstino, lastino, ino; 4946 4947 ACQUIRE_LOCK(&lk); 4948 /* 4949 * Pick a random inode dependency to be cleared. 4950 * We will then gather up all the inodes in its block 4951 * that have dependencies and flush them out. 4952 */ 4953 for (cnt = 0; cnt < inodedep_hash; cnt++) { 4954 inodedephd = &inodedep_hashtbl[next++]; 4955 if (next >= inodedep_hash) 4956 next = 0; 4957 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 4958 break; 4959 } 4960 if (inodedep == NULL) { 4961 FREE_LOCK(&lk); 4962 return; 4963 } 4964 /* 4965 * Ugly code to find mount point given pointer to superblock. 4966 */ 4967 fs = inodedep->id_fs; 4968 info.mp = NULL; 4969 info.fs = fs; 4970 mountlist_scan(clear_inodedeps_mountlist_callback, 4971 &info, MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 4972 /* 4973 * Find the last inode in the block with dependencies. 4974 */ 4975 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 4976 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 4977 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0) 4978 break; 4979 /* 4980 * Asynchronously push all but the last inode with dependencies. 4981 * Synchronously push the last inode with dependencies to ensure 4982 * that the inode block gets written to free up the inodedeps. 4983 */ 4984 for (ino = firstino; ino <= lastino; ino++) { 4985 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4986 continue; 4987 FREE_LOCK(&lk); 4988 if ((error = VFS_VGET(info.mp, ino, &vp)) != 0) { 4989 softdep_error("clear_inodedeps: vget", error); 4990 return; 4991 } 4992 if (ino == lastino) { 4993 if ((error = VOP_FSYNC(vp, MNT_WAIT))) 4994 softdep_error("clear_inodedeps: fsync1", error); 4995 } else { 4996 if ((error = VOP_FSYNC(vp, MNT_NOWAIT))) 4997 softdep_error("clear_inodedeps: fsync2", error); 4998 drain_output(vp, 0); 4999 } 5000 vput(vp); 5001 ACQUIRE_LOCK(&lk); 5002 } 5003 FREE_LOCK(&lk); 5004 } 5005 5006 /* 5007 * Function to determine if the buffer has outstanding dependencies 5008 * that will cause a roll-back if the buffer is written. If wantcount 5009 * is set, return number of dependencies, otherwise just yes or no. 5010 */ 5011 static int 5012 softdep_count_dependencies(struct buf *bp, int wantcount) 5013 { 5014 struct worklist *wk; 5015 struct inodedep *inodedep; 5016 struct indirdep *indirdep; 5017 struct allocindir *aip; 5018 struct pagedep *pagedep; 5019 struct diradd *dap; 5020 int i, retval; 5021 5022 retval = 0; 5023 ACQUIRE_LOCK(&lk); 5024 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5025 switch (wk->wk_type) { 5026 5027 case D_INODEDEP: 5028 inodedep = WK_INODEDEP(wk); 5029 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 5030 /* bitmap allocation dependency */ 5031 retval += 1; 5032 if (!wantcount) 5033 goto out; 5034 } 5035 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 5036 /* direct block pointer dependency */ 5037 retval += 1; 5038 if (!wantcount) 5039 goto out; 5040 } 5041 continue; 5042 5043 case D_INDIRDEP: 5044 indirdep = WK_INDIRDEP(wk); 5045 5046 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 5047 /* indirect block pointer dependency */ 5048 retval += 1; 5049 if (!wantcount) 5050 goto out; 5051 } 5052 continue; 5053 5054 case D_PAGEDEP: 5055 pagedep = WK_PAGEDEP(wk); 5056 for (i = 0; i < DAHASHSZ; i++) { 5057 5058 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 5059 /* directory entry dependency */ 5060 retval += 1; 5061 if (!wantcount) 5062 goto out; 5063 } 5064 } 5065 continue; 5066 5067 case D_BMSAFEMAP: 5068 case D_ALLOCDIRECT: 5069 case D_ALLOCINDIR: 5070 case D_MKDIR: 5071 /* never a dependency on these blocks */ 5072 continue; 5073 5074 default: 5075 FREE_LOCK(&lk); 5076 panic("softdep_check_for_rollback: Unexpected type %s", 5077 TYPENAME(wk->wk_type)); 5078 /* NOTREACHED */ 5079 } 5080 } 5081 out: 5082 FREE_LOCK(&lk); 5083 return retval; 5084 } 5085 5086 /* 5087 * Acquire exclusive access to a buffer. 5088 * Must be called with splbio blocked. 5089 * Return 1 if buffer was acquired. 5090 */ 5091 static int 5092 getdirtybuf(struct buf **bpp, int waitfor) 5093 { 5094 struct buf *bp; 5095 int error; 5096 5097 for (;;) { 5098 if ((bp = *bpp) == NULL) 5099 return (0); 5100 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) 5101 break; 5102 if (waitfor != MNT_WAIT) 5103 return (0); 5104 error = interlocked_sleep(&lk, LOCKBUF, bp, 5105 LK_EXCLUSIVE | LK_SLEEPFAIL, 0, 0); 5106 if (error != ENOLCK) { 5107 FREE_LOCK(&lk); 5108 panic("getdirtybuf: inconsistent lock"); 5109 } 5110 } 5111 if ((bp->b_flags & B_DELWRI) == 0) { 5112 BUF_UNLOCK(bp); 5113 return (0); 5114 } 5115 bremfree(bp); 5116 return (1); 5117 } 5118 5119 /* 5120 * Wait for pending output on a vnode to complete. 5121 * Must be called with vnode locked. 5122 */ 5123 static void 5124 drain_output(struct vnode *vp, int islocked) 5125 { 5126 5127 if (!islocked) 5128 ACQUIRE_LOCK(&lk); 5129 while (vp->v_track_write.bk_active) { 5130 vp->v_track_write.bk_waitflag = 1; 5131 interlocked_sleep(&lk, SLEEP, &vp->v_track_write, 5132 0, "drainvp", 0); 5133 } 5134 if (!islocked) 5135 FREE_LOCK(&lk); 5136 } 5137 5138 /* 5139 * Called whenever a buffer that is being invalidated or reallocated 5140 * contains dependencies. This should only happen if an I/O error has 5141 * occurred. The routine is called with the buffer locked. 5142 */ 5143 static void 5144 softdep_deallocate_dependencies(struct buf *bp) 5145 { 5146 if ((bp->b_flags & B_ERROR) == 0) 5147 panic("softdep_deallocate_dependencies: dangling deps"); 5148 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntfromname, bp->b_error); 5149 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 5150 } 5151 5152 /* 5153 * Function to handle asynchronous write errors in the filesystem. 5154 */ 5155 void 5156 softdep_error(char *func, int error) 5157 { 5158 5159 /* XXX should do something better! */ 5160 kprintf("%s: got error %d while accessing filesystem\n", func, error); 5161 } 5162