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 1.34 2006/02/17 19:18:08 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 *, ufs_daddr_t, int, ufs_lbn_t, 171 long *); 172 static void deallocate_dependencies(struct buf *, struct inodedep *); 173 static void free_allocdirect(struct allocdirectlst *, 174 struct allocdirect *, int); 175 static int check_inode_unwritten(struct inodedep *); 176 static int free_inodedep(struct inodedep *); 177 static void handle_workitem_freeblocks(struct freeblks *); 178 static void merge_inode_lists(struct inodedep *); 179 static void setup_allocindir_phase2(struct buf *, struct inode *, 180 struct allocindir *); 181 static struct allocindir *newallocindir(struct inode *, int, ufs_daddr_t, 182 ufs_daddr_t); 183 static void handle_workitem_freefrag(struct freefrag *); 184 static struct freefrag *newfreefrag(struct inode *, ufs_daddr_t, long); 185 static void allocdirect_merge(struct allocdirectlst *, 186 struct allocdirect *, struct allocdirect *); 187 static struct bmsafemap *bmsafemap_lookup(struct buf *); 188 static int newblk_lookup(struct fs *, ufs_daddr_t, int, 189 struct newblk **); 190 static int inodedep_lookup(struct fs *, ino_t, int, struct inodedep **); 191 static int pagedep_lookup(struct inode *, ufs_lbn_t, int, 192 struct pagedep **); 193 static void pause_timer(void *); 194 static int request_cleanup(int, int); 195 static int process_worklist_item(struct mount *, int); 196 static void add_to_worklist(struct worklist *); 197 198 /* 199 * Exported softdep operations. 200 */ 201 static void softdep_disk_io_initiation(struct buf *); 202 static void softdep_disk_write_complete(struct buf *); 203 static void softdep_deallocate_dependencies(struct buf *); 204 static int softdep_fsync(struct vnode *); 205 static int softdep_process_worklist(struct mount *); 206 static void softdep_move_dependencies(struct buf *, struct buf *); 207 static int softdep_count_dependencies(struct buf *bp, int); 208 209 static struct bio_ops softdep_bioops = { 210 softdep_disk_io_initiation, /* io_start */ 211 softdep_disk_write_complete, /* io_complete */ 212 softdep_deallocate_dependencies, /* io_deallocate */ 213 softdep_fsync, /* io_fsync */ 214 softdep_process_worklist, /* io_sync */ 215 softdep_move_dependencies, /* io_movedeps */ 216 softdep_count_dependencies, /* io_countdeps */ 217 }; 218 219 /* 220 * Locking primitives. 221 * 222 * For a uniprocessor, all we need to do is protect against disk 223 * interrupts. For a multiprocessor, this lock would have to be 224 * a mutex. A single mutex is used throughout this file, though 225 * finer grain locking could be used if contention warranted it. 226 * 227 * For a multiprocessor, the sleep call would accept a lock and 228 * release it after the sleep processing was complete. In a uniprocessor 229 * implementation there is no such interlock, so we simple mark 230 * the places where it needs to be done with the `interlocked' form 231 * of the lock calls. Since the uniprocessor sleep already interlocks 232 * the spl, there is nothing that really needs to be done. 233 */ 234 #ifndef /* NOT */ DEBUG 235 static struct lockit { 236 } lk = { 0 }; 237 #define ACQUIRE_LOCK(lk) crit_enter_id("softupdates"); 238 #define FREE_LOCK(lk) crit_exit_id("softupdates"); 239 240 #else /* DEBUG */ 241 #define NOHOLDER ((struct thread *)-1) 242 #define SPECIAL_FLAG ((struct thread *)-2) 243 static struct lockit { 244 int lkt_spl; 245 struct thread *lkt_held; 246 } lk = { 0, NOHOLDER }; 247 static int lockcnt; 248 249 static void acquire_lock(struct lockit *); 250 static void free_lock(struct lockit *); 251 void softdep_panic(char *); 252 253 #define ACQUIRE_LOCK(lk) acquire_lock(lk) 254 #define FREE_LOCK(lk) free_lock(lk) 255 256 static void 257 acquire_lock(lk) 258 struct lockit *lk; 259 { 260 thread_t holder; 261 262 if (lk->lkt_held != NOHOLDER) { 263 holder = lk->lkt_held; 264 FREE_LOCK(lk); 265 if (holder == curthread) 266 panic("softdep_lock: locking against myself"); 267 else 268 panic("softdep_lock: lock held by %p", holder); 269 } 270 crit_enter_id("softupdates"); 271 lk->lkt_held = curthread; 272 lockcnt++; 273 } 274 275 static void 276 free_lock(lk) 277 struct lockit *lk; 278 { 279 280 if (lk->lkt_held == NOHOLDER) 281 panic("softdep_unlock: lock not held"); 282 lk->lkt_held = NOHOLDER; 283 crit_exit_id("softupdates"); 284 } 285 286 /* 287 * Function to release soft updates lock and panic. 288 */ 289 void 290 softdep_panic(msg) 291 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(lk, op, ident, flags, wmesg, timo) 316 struct lockit *lk; 317 int op; 318 void *ident; 319 int flags; 320 const char *wmesg; 321 int timo; 322 { 323 thread_t holder; 324 int s, retval; 325 326 s = lk->lkt_spl; 327 # ifdef DEBUG 328 if (lk->lkt_held == NOHOLDER) 329 panic("interlocked_sleep: lock not held"); 330 lk->lkt_held = NOHOLDER; 331 # endif /* DEBUG */ 332 switch (op) { 333 case SLEEP: 334 retval = tsleep(ident, flags, wmesg, timo); 335 break; 336 case LOCKBUF: 337 retval = BUF_LOCK((struct buf *)ident, flags); 338 break; 339 default: 340 panic("interlocked_sleep: unknown operation"); 341 } 342 # ifdef DEBUG 343 if (lk->lkt_held != NOHOLDER) { 344 holder = lk->lkt_held; 345 FREE_LOCK(lk); 346 if (holder == curthread) 347 panic("interlocked_sleep: locking against self"); 348 else 349 panic("interlocked_sleep: lock held by %p", holder); 350 } 351 lk->lkt_held = curthread; 352 lockcnt++; 353 # endif /* DEBUG */ 354 lk->lkt_spl = s; 355 return (retval); 356 } 357 358 /* 359 * Place holder for real semaphores. 360 */ 361 struct sema { 362 int value; 363 thread_t holder; 364 char *name; 365 int prio; 366 int timo; 367 }; 368 static void sema_init(struct sema *, char *, int, int); 369 static int sema_get(struct sema *, struct lockit *); 370 static void sema_release(struct sema *); 371 372 static void 373 sema_init(semap, name, prio, timo) 374 struct sema *semap; 375 char *name; 376 int prio, timo; 377 { 378 379 semap->holder = NOHOLDER; 380 semap->value = 0; 381 semap->name = name; 382 semap->prio = prio; 383 semap->timo = timo; 384 } 385 386 static int 387 sema_get(semap, interlock) 388 struct sema *semap; 389 struct lockit *interlock; 390 { 391 392 if (semap->value++ > 0) { 393 if (interlock != NULL) { 394 interlocked_sleep(interlock, SLEEP, (caddr_t)semap, 395 semap->prio, semap->name, semap->timo); 396 FREE_LOCK(interlock); 397 } else { 398 tsleep((caddr_t)semap, semap->prio, semap->name, 399 semap->timo); 400 } 401 return (0); 402 } 403 semap->holder = curthread; 404 if (interlock != NULL) 405 FREE_LOCK(interlock); 406 return (1); 407 } 408 409 static void 410 sema_release(semap) 411 struct sema *semap; 412 { 413 414 if (semap->value <= 0 || semap->holder != curthread) { 415 if (lk.lkt_held != NOHOLDER) 416 FREE_LOCK(&lk); 417 panic("sema_release: not held"); 418 } 419 if (--semap->value > 0) { 420 semap->value = 0; 421 wakeup(semap); 422 } 423 semap->holder = NOHOLDER; 424 } 425 426 /* 427 * Worklist queue management. 428 * These routines require that the lock be held. 429 */ 430 #ifndef /* NOT */ DEBUG 431 #define WORKLIST_INSERT(head, item) do { \ 432 (item)->wk_state |= ONWORKLIST; \ 433 LIST_INSERT_HEAD(head, item, wk_list); \ 434 } while (0) 435 #define WORKLIST_REMOVE(item) do { \ 436 (item)->wk_state &= ~ONWORKLIST; \ 437 LIST_REMOVE(item, wk_list); \ 438 } while (0) 439 #define WORKITEM_FREE(item, type) FREE(item, DtoM(type)) 440 441 #else /* DEBUG */ 442 static void worklist_insert(struct workhead *, struct worklist *); 443 static void worklist_remove(struct worklist *); 444 static void workitem_free(struct worklist *, int); 445 446 #define WORKLIST_INSERT(head, item) worklist_insert(head, item) 447 #define WORKLIST_REMOVE(item) worklist_remove(item) 448 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type) 449 450 static void 451 worklist_insert(head, item) 452 struct workhead *head; 453 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(item) 468 struct worklist *item; 469 { 470 471 if (lk.lkt_held == NOHOLDER) 472 panic("worklist_remove: lock not held"); 473 if ((item->wk_state & ONWORKLIST) == 0) { 474 FREE_LOCK(&lk); 475 panic("worklist_remove: not on list"); 476 } 477 item->wk_state &= ~ONWORKLIST; 478 LIST_REMOVE(item, wk_list); 479 } 480 481 static void 482 workitem_free(item, type) 483 struct worklist *item; 484 int type; 485 { 486 487 if (item->wk_state & ONWORKLIST) { 488 if (lk.lkt_held != NOHOLDER) 489 FREE_LOCK(&lk); 490 panic("workitem_free: still on list"); 491 } 492 if (item->wk_type != type) { 493 if (lk.lkt_held != NOHOLDER) 494 FREE_LOCK(&lk); 495 panic("workitem_free: type mismatch"); 496 } 497 FREE(item, DtoM(type)); 498 } 499 #endif /* DEBUG */ 500 501 /* 502 * Workitem queue management 503 */ 504 static struct workhead softdep_workitem_pending; 505 static int num_on_worklist; /* number of worklist items to be processed */ 506 static int softdep_worklist_busy; /* 1 => trying to do unmount */ 507 static int softdep_worklist_req; /* serialized waiters */ 508 static int max_softdeps; /* maximum number of structs before slowdown */ 509 static int tickdelay = 2; /* number of ticks to pause during slowdown */ 510 static int *stat_countp; /* statistic to count in proc_waiting timeout */ 511 static int proc_waiting; /* tracks whether we have a timeout posted */ 512 static struct callout handle; /* handle on posted proc_waiting timeout */ 513 static struct thread *filesys_syncer; /* proc of filesystem syncer process */ 514 static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 515 #define FLUSH_INODES 1 516 static int req_clear_remove; /* syncer process flush some freeblks */ 517 #define FLUSH_REMOVE 2 518 /* 519 * runtime statistics 520 */ 521 static int stat_worklist_push; /* number of worklist cleanups */ 522 static int stat_blk_limit_push; /* number of times block limit neared */ 523 static int stat_ino_limit_push; /* number of times inode limit neared */ 524 static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 525 static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 526 static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 527 static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 528 static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 529 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 530 static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 531 #ifdef DEBUG 532 #include <vm/vm.h> 533 #include <sys/sysctl.h> 534 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, ""); 535 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, ""); 536 SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,""); 537 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,""); 538 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,""); 539 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, ""); 540 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, ""); 541 SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, ""); 542 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, ""); 543 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, ""); 544 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, ""); 545 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, ""); 546 #endif /* DEBUG */ 547 548 /* 549 * Add an item to the end of the work queue. 550 * This routine requires that the lock be held. 551 * This is the only routine that adds items to the list. 552 * The following routine is the only one that removes items 553 * and does so in order from first to last. 554 */ 555 static void 556 add_to_worklist(wk) 557 struct worklist *wk; 558 { 559 static struct worklist *worklist_tail; 560 561 if (wk->wk_state & ONWORKLIST) { 562 if (lk.lkt_held != NOHOLDER) 563 FREE_LOCK(&lk); 564 panic("add_to_worklist: already on list"); 565 } 566 wk->wk_state |= ONWORKLIST; 567 if (LIST_FIRST(&softdep_workitem_pending) == NULL) 568 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list); 569 else 570 LIST_INSERT_AFTER(worklist_tail, wk, wk_list); 571 worklist_tail = wk; 572 num_on_worklist += 1; 573 } 574 575 /* 576 * Process that runs once per second to handle items in the background queue. 577 * 578 * Note that we ensure that everything is done in the order in which they 579 * appear in the queue. The code below depends on this property to ensure 580 * that blocks of a file are freed before the inode itself is freed. This 581 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 582 * until all the old ones have been purged from the dependency lists. 583 */ 584 static int 585 softdep_process_worklist(matchmnt) 586 struct mount *matchmnt; 587 { 588 thread_t td = curthread; 589 int matchcnt, loopcount; 590 long starttime; 591 592 /* 593 * Record the process identifier of our caller so that we can give 594 * this process preferential treatment in request_cleanup below. 595 */ 596 filesys_syncer = td; 597 matchcnt = 0; 598 599 /* 600 * There is no danger of having multiple processes run this 601 * code, but we have to single-thread it when softdep_flushfiles() 602 * is in operation to get an accurate count of the number of items 603 * related to its mount point that are in the list. 604 */ 605 if (matchmnt == NULL) { 606 if (softdep_worklist_busy < 0) 607 return(-1); 608 softdep_worklist_busy += 1; 609 } 610 611 /* 612 * If requested, try removing inode or removal dependencies. 613 */ 614 if (req_clear_inodedeps) { 615 clear_inodedeps(td); 616 req_clear_inodedeps -= 1; 617 wakeup_one(&proc_waiting); 618 } 619 if (req_clear_remove) { 620 clear_remove(td); 621 req_clear_remove -= 1; 622 wakeup_one(&proc_waiting); 623 } 624 loopcount = 1; 625 starttime = time_second; 626 while (num_on_worklist > 0) { 627 matchcnt += process_worklist_item(matchmnt, 0); 628 629 /* 630 * If a umount operation wants to run the worklist 631 * accurately, abort. 632 */ 633 if (softdep_worklist_req && matchmnt == NULL) { 634 matchcnt = -1; 635 break; 636 } 637 638 /* 639 * If requested, try removing inode or removal dependencies. 640 */ 641 if (req_clear_inodedeps) { 642 clear_inodedeps(td); 643 req_clear_inodedeps -= 1; 644 wakeup_one(&proc_waiting); 645 } 646 if (req_clear_remove) { 647 clear_remove(td); 648 req_clear_remove -= 1; 649 wakeup_one(&proc_waiting); 650 } 651 /* 652 * We do not generally want to stop for buffer space, but if 653 * we are really being a buffer hog, we will stop and wait. 654 */ 655 if (loopcount++ % 128 == 0) 656 bwillwrite(); 657 /* 658 * Never allow processing to run for more than one 659 * second. Otherwise the other syncer tasks may get 660 * excessively backlogged. 661 */ 662 if (starttime != time_second && matchmnt == NULL) { 663 matchcnt = -1; 664 break; 665 } 666 } 667 if (matchmnt == NULL) { 668 --softdep_worklist_busy; 669 if (softdep_worklist_req && softdep_worklist_busy == 0) 670 wakeup(&softdep_worklist_req); 671 } 672 return (matchcnt); 673 } 674 675 /* 676 * Process one item on the worklist. 677 */ 678 static int 679 process_worklist_item(matchmnt, flags) 680 struct mount *matchmnt; 681 int flags; 682 { 683 struct worklist *wk; 684 struct dirrem *dirrem; 685 struct fs *matchfs; 686 struct vnode *vp; 687 int matchcnt = 0; 688 689 matchfs = NULL; 690 if (matchmnt != NULL) 691 matchfs = VFSTOUFS(matchmnt)->um_fs; 692 ACQUIRE_LOCK(&lk); 693 /* 694 * Normally we just process each item on the worklist in order. 695 * However, if we are in a situation where we cannot lock any 696 * inodes, we have to skip over any dirrem requests whose 697 * vnodes are resident and locked. 698 */ 699 LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) { 700 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM) 701 break; 702 dirrem = WK_DIRREM(wk); 703 vp = ufs_ihashlookup(VFSTOUFS(dirrem->dm_mnt)->um_dev, 704 dirrem->dm_oldinum); 705 if (vp == NULL || !VOP_ISLOCKED(vp, curthread)) 706 break; 707 } 708 if (wk == 0) { 709 FREE_LOCK(&lk); 710 return (0); 711 } 712 WORKLIST_REMOVE(wk); 713 num_on_worklist -= 1; 714 FREE_LOCK(&lk); 715 switch (wk->wk_type) { 716 717 case D_DIRREM: 718 /* removal of a directory entry */ 719 if (WK_DIRREM(wk)->dm_mnt == matchmnt) 720 matchcnt += 1; 721 handle_workitem_remove(WK_DIRREM(wk)); 722 break; 723 724 case D_FREEBLKS: 725 /* releasing blocks and/or fragments from a file */ 726 if (WK_FREEBLKS(wk)->fb_fs == matchfs) 727 matchcnt += 1; 728 handle_workitem_freeblocks(WK_FREEBLKS(wk)); 729 break; 730 731 case D_FREEFRAG: 732 /* releasing a fragment when replaced as a file grows */ 733 if (WK_FREEFRAG(wk)->ff_fs == matchfs) 734 matchcnt += 1; 735 handle_workitem_freefrag(WK_FREEFRAG(wk)); 736 break; 737 738 case D_FREEFILE: 739 /* releasing an inode when its link count drops to 0 */ 740 if (WK_FREEFILE(wk)->fx_fs == matchfs) 741 matchcnt += 1; 742 handle_workitem_freefile(WK_FREEFILE(wk)); 743 break; 744 745 default: 746 panic("%s_process_worklist: Unknown type %s", 747 "softdep", TYPENAME(wk->wk_type)); 748 /* NOTREACHED */ 749 } 750 return (matchcnt); 751 } 752 753 /* 754 * Move dependencies from one buffer to another. 755 */ 756 static void 757 softdep_move_dependencies(oldbp, newbp) 758 struct buf *oldbp; 759 struct buf *newbp; 760 { 761 struct worklist *wk, *wktail; 762 763 if (LIST_FIRST(&newbp->b_dep) != NULL) 764 panic("softdep_move_dependencies: need merge code"); 765 wktail = 0; 766 ACQUIRE_LOCK(&lk); 767 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 768 LIST_REMOVE(wk, wk_list); 769 if (wktail == 0) 770 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 771 else 772 LIST_INSERT_AFTER(wktail, wk, wk_list); 773 wktail = wk; 774 } 775 FREE_LOCK(&lk); 776 } 777 778 /* 779 * Purge the work list of all items associated with a particular mount point. 780 */ 781 int 782 softdep_flushfiles(struct mount *oldmnt, int flags, struct thread *td) 783 { 784 struct vnode *devvp; 785 int error, loopcnt; 786 787 /* 788 * Await our turn to clear out the queue, then serialize access. 789 */ 790 while (softdep_worklist_busy != 0) { 791 softdep_worklist_req += 1; 792 tsleep(&softdep_worklist_req, 0, "softflush", 0); 793 softdep_worklist_req -= 1; 794 } 795 softdep_worklist_busy = -1; 796 797 if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0) { 798 softdep_worklist_busy = 0; 799 if (softdep_worklist_req) 800 wakeup(&softdep_worklist_req); 801 return (error); 802 } 803 /* 804 * Alternately flush the block device associated with the mount 805 * point and process any dependencies that the flushing 806 * creates. In theory, this loop can happen at most twice, 807 * but we give it a few extra just to be sure. 808 */ 809 devvp = VFSTOUFS(oldmnt)->um_devvp; 810 for (loopcnt = 10; loopcnt > 0; ) { 811 if (softdep_process_worklist(oldmnt) == 0) { 812 loopcnt--; 813 /* 814 * Do another flush in case any vnodes were brought in 815 * as part of the cleanup operations. 816 */ 817 if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0) 818 break; 819 /* 820 * If we still found nothing to do, we are really done. 821 */ 822 if (softdep_process_worklist(oldmnt) == 0) 823 break; 824 } 825 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, td); 826 error = VOP_FSYNC(devvp, MNT_WAIT, td); 827 VOP_UNLOCK(devvp, 0, td); 828 if (error) 829 break; 830 } 831 softdep_worklist_busy = 0; 832 if (softdep_worklist_req) 833 wakeup(&softdep_worklist_req); 834 835 /* 836 * If we are unmounting then it is an error to fail. If we 837 * are simply trying to downgrade to read-only, then filesystem 838 * activity can keep us busy forever, so we just fail with EBUSY. 839 */ 840 if (loopcnt == 0) { 841 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 842 panic("softdep_flushfiles: looping"); 843 error = EBUSY; 844 } 845 return (error); 846 } 847 848 /* 849 * Structure hashing. 850 * 851 * There are three types of structures that can be looked up: 852 * 1) pagedep structures identified by mount point, inode number, 853 * and logical block. 854 * 2) inodedep structures identified by mount point and inode number. 855 * 3) newblk structures identified by mount point and 856 * physical block number. 857 * 858 * The "pagedep" and "inodedep" dependency structures are hashed 859 * separately from the file blocks and inodes to which they correspond. 860 * This separation helps when the in-memory copy of an inode or 861 * file block must be replaced. It also obviates the need to access 862 * an inode or file page when simply updating (or de-allocating) 863 * dependency structures. Lookup of newblk structures is needed to 864 * find newly allocated blocks when trying to associate them with 865 * their allocdirect or allocindir structure. 866 * 867 * The lookup routines optionally create and hash a new instance when 868 * an existing entry is not found. 869 */ 870 #define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 871 #define NODELAY 0x0002 /* cannot do background work */ 872 873 /* 874 * Structures and routines associated with pagedep caching. 875 */ 876 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl; 877 u_long pagedep_hash; /* size of hash table - 1 */ 878 #define PAGEDEP_HASH(mp, inum, lbn) \ 879 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \ 880 pagedep_hash]) 881 static struct sema pagedep_in_progress; 882 883 /* 884 * Look up a pagedep. Return 1 if found, 0 if not found. 885 * If not found, allocate if DEPALLOC flag is passed. 886 * Found or allocated entry is returned in pagedeppp. 887 * This routine must be called with splbio interrupts blocked. 888 */ 889 static int 890 pagedep_lookup(ip, lbn, flags, pagedeppp) 891 struct inode *ip; 892 ufs_lbn_t lbn; 893 int flags; 894 struct pagedep **pagedeppp; 895 { 896 struct pagedep *pagedep; 897 struct pagedep_hashhead *pagedephd; 898 struct mount *mp; 899 int i; 900 901 #ifdef DEBUG 902 if (lk.lkt_held == NOHOLDER) 903 panic("pagedep_lookup: lock not held"); 904 #endif 905 mp = ITOV(ip)->v_mount; 906 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn); 907 top: 908 LIST_FOREACH(pagedep, pagedephd, pd_hash) 909 if (ip->i_number == pagedep->pd_ino && 910 lbn == pagedep->pd_lbn && 911 mp == pagedep->pd_mnt) 912 break; 913 if (pagedep) { 914 *pagedeppp = pagedep; 915 return (1); 916 } 917 if ((flags & DEPALLOC) == 0) { 918 *pagedeppp = NULL; 919 return (0); 920 } 921 if (sema_get(&pagedep_in_progress, &lk) == 0) { 922 ACQUIRE_LOCK(&lk); 923 goto top; 924 } 925 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP, 926 M_SOFTDEP_FLAGS); 927 bzero(pagedep, sizeof(struct pagedep)); 928 pagedep->pd_list.wk_type = D_PAGEDEP; 929 pagedep->pd_mnt = mp; 930 pagedep->pd_ino = ip->i_number; 931 pagedep->pd_lbn = lbn; 932 LIST_INIT(&pagedep->pd_dirremhd); 933 LIST_INIT(&pagedep->pd_pendinghd); 934 for (i = 0; i < DAHASHSZ; i++) 935 LIST_INIT(&pagedep->pd_diraddhd[i]); 936 ACQUIRE_LOCK(&lk); 937 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 938 sema_release(&pagedep_in_progress); 939 *pagedeppp = pagedep; 940 return (0); 941 } 942 943 /* 944 * Structures and routines associated with inodedep caching. 945 */ 946 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl; 947 static u_long inodedep_hash; /* size of hash table - 1 */ 948 static long num_inodedep; /* number of inodedep allocated */ 949 #define INODEDEP_HASH(fs, inum) \ 950 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash]) 951 static struct sema inodedep_in_progress; 952 953 /* 954 * Look up a inodedep. Return 1 if found, 0 if not found. 955 * If not found, allocate if DEPALLOC flag is passed. 956 * Found or allocated entry is returned in inodedeppp. 957 * This routine must be called with splbio interrupts blocked. 958 */ 959 static int 960 inodedep_lookup(fs, inum, flags, inodedeppp) 961 struct fs *fs; 962 ino_t inum; 963 int flags; 964 struct inodedep **inodedeppp; 965 { 966 struct inodedep *inodedep; 967 struct inodedep_hashhead *inodedephd; 968 int firsttry; 969 970 #ifdef DEBUG 971 if (lk.lkt_held == NOHOLDER) 972 panic("inodedep_lookup: lock not held"); 973 #endif 974 firsttry = 1; 975 inodedephd = INODEDEP_HASH(fs, inum); 976 top: 977 LIST_FOREACH(inodedep, inodedephd, id_hash) 978 if (inum == inodedep->id_ino && fs == inodedep->id_fs) 979 break; 980 if (inodedep) { 981 *inodedeppp = inodedep; 982 return (1); 983 } 984 if ((flags & DEPALLOC) == 0) { 985 *inodedeppp = NULL; 986 return (0); 987 } 988 /* 989 * If we are over our limit, try to improve the situation. 990 */ 991 if (num_inodedep > max_softdeps && firsttry && 992 speedup_syncer() == 0 && (flags & NODELAY) == 0 && 993 request_cleanup(FLUSH_INODES, 1)) { 994 firsttry = 0; 995 goto top; 996 } 997 if (sema_get(&inodedep_in_progress, &lk) == 0) { 998 ACQUIRE_LOCK(&lk); 999 goto top; 1000 } 1001 num_inodedep += 1; 1002 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep), 1003 M_INODEDEP, M_SOFTDEP_FLAGS); 1004 inodedep->id_list.wk_type = D_INODEDEP; 1005 inodedep->id_fs = fs; 1006 inodedep->id_ino = inum; 1007 inodedep->id_state = ALLCOMPLETE; 1008 inodedep->id_nlinkdelta = 0; 1009 inodedep->id_savedino = NULL; 1010 inodedep->id_savedsize = -1; 1011 inodedep->id_buf = NULL; 1012 LIST_INIT(&inodedep->id_pendinghd); 1013 LIST_INIT(&inodedep->id_inowait); 1014 LIST_INIT(&inodedep->id_bufwait); 1015 TAILQ_INIT(&inodedep->id_inoupdt); 1016 TAILQ_INIT(&inodedep->id_newinoupdt); 1017 ACQUIRE_LOCK(&lk); 1018 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 1019 sema_release(&inodedep_in_progress); 1020 *inodedeppp = inodedep; 1021 return (0); 1022 } 1023 1024 /* 1025 * Structures and routines associated with newblk caching. 1026 */ 1027 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl; 1028 u_long newblk_hash; /* size of hash table - 1 */ 1029 #define NEWBLK_HASH(fs, inum) \ 1030 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash]) 1031 static struct sema newblk_in_progress; 1032 1033 /* 1034 * Look up a newblk. Return 1 if found, 0 if not found. 1035 * If not found, allocate if DEPALLOC flag is passed. 1036 * Found or allocated entry is returned in newblkpp. 1037 */ 1038 static int 1039 newblk_lookup(fs, newblkno, flags, newblkpp) 1040 struct fs *fs; 1041 ufs_daddr_t newblkno; 1042 int flags; 1043 struct newblk **newblkpp; 1044 { 1045 struct newblk *newblk; 1046 struct newblk_hashhead *newblkhd; 1047 1048 newblkhd = NEWBLK_HASH(fs, newblkno); 1049 top: 1050 LIST_FOREACH(newblk, newblkhd, nb_hash) 1051 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs) 1052 break; 1053 if (newblk) { 1054 *newblkpp = newblk; 1055 return (1); 1056 } 1057 if ((flags & DEPALLOC) == 0) { 1058 *newblkpp = NULL; 1059 return (0); 1060 } 1061 if (sema_get(&newblk_in_progress, 0) == 0) 1062 goto top; 1063 MALLOC(newblk, struct newblk *, sizeof(struct newblk), 1064 M_NEWBLK, M_SOFTDEP_FLAGS); 1065 newblk->nb_state = 0; 1066 newblk->nb_fs = fs; 1067 newblk->nb_newblkno = newblkno; 1068 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 1069 sema_release(&newblk_in_progress); 1070 *newblkpp = newblk; 1071 return (0); 1072 } 1073 1074 /* 1075 * Executed during filesystem system initialization before 1076 * mounting any filesystems. 1077 */ 1078 void 1079 softdep_initialize() 1080 { 1081 callout_init(&handle); 1082 bioops = softdep_bioops; /* XXX hack */ 1083 1084 LIST_INIT(&mkdirlisthd); 1085 LIST_INIT(&softdep_workitem_pending); 1086 max_softdeps = min(desiredvnodes * 8, 1087 M_INODEDEP->ks_limit / (2 * sizeof(struct inodedep))); 1088 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 1089 &pagedep_hash); 1090 sema_init(&pagedep_in_progress, "pagedep", 0, 0); 1091 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash); 1092 sema_init(&inodedep_in_progress, "inodedep", 0, 0); 1093 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash); 1094 sema_init(&newblk_in_progress, "newblk", 0, 0); 1095 } 1096 1097 /* 1098 * Called at mount time to notify the dependency code that a 1099 * filesystem wishes to use it. 1100 */ 1101 int 1102 softdep_mount(devvp, mp, fs) 1103 struct vnode *devvp; 1104 struct mount *mp; 1105 struct fs *fs; 1106 { 1107 struct csum cstotal; 1108 struct cg *cgp; 1109 struct buf *bp; 1110 int error, cyl; 1111 1112 mp->mnt_flag &= ~MNT_ASYNC; 1113 mp->mnt_flag |= MNT_SOFTDEP; 1114 /* 1115 * When doing soft updates, the counters in the 1116 * superblock may have gotten out of sync, so we have 1117 * to scan the cylinder groups and recalculate them. 1118 */ 1119 if (fs->fs_clean != 0) 1120 return (0); 1121 bzero(&cstotal, sizeof cstotal); 1122 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 1123 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 1124 fs->fs_cgsize, &bp)) != 0) { 1125 brelse(bp); 1126 return (error); 1127 } 1128 cgp = (struct cg *)bp->b_data; 1129 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 1130 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 1131 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 1132 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 1133 fs->fs_cs(fs, cyl) = cgp->cg_cs; 1134 brelse(bp); 1135 } 1136 #ifdef DEBUG 1137 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 1138 printf("ffs_mountfs: superblock updated for soft updates\n"); 1139 #endif 1140 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 1141 return (0); 1142 } 1143 1144 /* 1145 * Protecting the freemaps (or bitmaps). 1146 * 1147 * To eliminate the need to execute fsck before mounting a filesystem 1148 * after a power failure, one must (conservatively) guarantee that the 1149 * on-disk copy of the bitmaps never indicate that a live inode or block is 1150 * free. So, when a block or inode is allocated, the bitmap should be 1151 * updated (on disk) before any new pointers. When a block or inode is 1152 * freed, the bitmap should not be updated until all pointers have been 1153 * reset. The latter dependency is handled by the delayed de-allocation 1154 * approach described below for block and inode de-allocation. The former 1155 * dependency is handled by calling the following procedure when a block or 1156 * inode is allocated. When an inode is allocated an "inodedep" is created 1157 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 1158 * Each "inodedep" is also inserted into the hash indexing structure so 1159 * that any additional link additions can be made dependent on the inode 1160 * allocation. 1161 * 1162 * The ufs filesystem maintains a number of free block counts (e.g., per 1163 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 1164 * in addition to the bitmaps. These counts are used to improve efficiency 1165 * during allocation and therefore must be consistent with the bitmaps. 1166 * There is no convenient way to guarantee post-crash consistency of these 1167 * counts with simple update ordering, for two main reasons: (1) The counts 1168 * and bitmaps for a single cylinder group block are not in the same disk 1169 * sector. If a disk write is interrupted (e.g., by power failure), one may 1170 * be written and the other not. (2) Some of the counts are located in the 1171 * superblock rather than the cylinder group block. So, we focus our soft 1172 * updates implementation on protecting the bitmaps. When mounting a 1173 * filesystem, we recompute the auxiliary counts from the bitmaps. 1174 */ 1175 1176 /* 1177 * Called just after updating the cylinder group block to allocate an inode. 1178 */ 1179 void 1180 softdep_setup_inomapdep(bp, ip, newinum) 1181 struct buf *bp; /* buffer for cylgroup block with inode map */ 1182 struct inode *ip; /* inode related to allocation */ 1183 ino_t newinum; /* new inode number being allocated */ 1184 { 1185 struct inodedep *inodedep; 1186 struct bmsafemap *bmsafemap; 1187 1188 /* 1189 * Create a dependency for the newly allocated inode. 1190 * Panic if it already exists as something is seriously wrong. 1191 * Otherwise add it to the dependency list for the buffer holding 1192 * the cylinder group map from which it was allocated. 1193 */ 1194 ACQUIRE_LOCK(&lk); 1195 if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) { 1196 FREE_LOCK(&lk); 1197 panic("softdep_setup_inomapdep: found inode"); 1198 } 1199 inodedep->id_buf = bp; 1200 inodedep->id_state &= ~DEPCOMPLETE; 1201 bmsafemap = bmsafemap_lookup(bp); 1202 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 1203 FREE_LOCK(&lk); 1204 } 1205 1206 /* 1207 * Called just after updating the cylinder group block to 1208 * allocate block or fragment. 1209 */ 1210 void 1211 softdep_setup_blkmapdep(bp, fs, newblkno) 1212 struct buf *bp; /* buffer for cylgroup block with block map */ 1213 struct fs *fs; /* filesystem doing allocation */ 1214 ufs_daddr_t newblkno; /* number of newly allocated block */ 1215 { 1216 struct newblk *newblk; 1217 struct bmsafemap *bmsafemap; 1218 1219 /* 1220 * Create a dependency for the newly allocated block. 1221 * Add it to the dependency list for the buffer holding 1222 * the cylinder group map from which it was allocated. 1223 */ 1224 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0) 1225 panic("softdep_setup_blkmapdep: found block"); 1226 ACQUIRE_LOCK(&lk); 1227 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp); 1228 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 1229 FREE_LOCK(&lk); 1230 } 1231 1232 /* 1233 * Find the bmsafemap associated with a cylinder group buffer. 1234 * If none exists, create one. The buffer must be locked when 1235 * this routine is called and this routine must be called with 1236 * splbio interrupts blocked. 1237 */ 1238 static struct bmsafemap * 1239 bmsafemap_lookup(bp) 1240 struct buf *bp; 1241 { 1242 struct bmsafemap *bmsafemap; 1243 struct worklist *wk; 1244 1245 #ifdef DEBUG 1246 if (lk.lkt_held == NOHOLDER) 1247 panic("bmsafemap_lookup: lock not held"); 1248 #endif 1249 LIST_FOREACH(wk, &bp->b_dep, wk_list) 1250 if (wk->wk_type == D_BMSAFEMAP) 1251 return (WK_BMSAFEMAP(wk)); 1252 FREE_LOCK(&lk); 1253 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap), 1254 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 1255 bmsafemap->sm_list.wk_type = D_BMSAFEMAP; 1256 bmsafemap->sm_list.wk_state = 0; 1257 bmsafemap->sm_buf = bp; 1258 LIST_INIT(&bmsafemap->sm_allocdirecthd); 1259 LIST_INIT(&bmsafemap->sm_allocindirhd); 1260 LIST_INIT(&bmsafemap->sm_inodedephd); 1261 LIST_INIT(&bmsafemap->sm_newblkhd); 1262 ACQUIRE_LOCK(&lk); 1263 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 1264 return (bmsafemap); 1265 } 1266 1267 /* 1268 * Direct block allocation dependencies. 1269 * 1270 * When a new block is allocated, the corresponding disk locations must be 1271 * initialized (with zeros or new data) before the on-disk inode points to 1272 * them. Also, the freemap from which the block was allocated must be 1273 * updated (on disk) before the inode's pointer. These two dependencies are 1274 * independent of each other and are needed for all file blocks and indirect 1275 * blocks that are pointed to directly by the inode. Just before the 1276 * "in-core" version of the inode is updated with a newly allocated block 1277 * number, a procedure (below) is called to setup allocation dependency 1278 * structures. These structures are removed when the corresponding 1279 * dependencies are satisfied or when the block allocation becomes obsolete 1280 * (i.e., the file is deleted, the block is de-allocated, or the block is a 1281 * fragment that gets upgraded). All of these cases are handled in 1282 * procedures described later. 1283 * 1284 * When a file extension causes a fragment to be upgraded, either to a larger 1285 * fragment or to a full block, the on-disk location may change (if the 1286 * previous fragment could not simply be extended). In this case, the old 1287 * fragment must be de-allocated, but not until after the inode's pointer has 1288 * been updated. In most cases, this is handled by later procedures, which 1289 * will construct a "freefrag" structure to be added to the workitem queue 1290 * when the inode update is complete (or obsolete). The main exception to 1291 * this is when an allocation occurs while a pending allocation dependency 1292 * (for the same block pointer) remains. This case is handled in the main 1293 * allocation dependency setup procedure by immediately freeing the 1294 * unreferenced fragments. 1295 */ 1296 void 1297 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 1298 struct inode *ip; /* inode to which block is being added */ 1299 ufs_lbn_t lbn; /* block pointer within inode */ 1300 ufs_daddr_t newblkno; /* disk block number being added */ 1301 ufs_daddr_t oldblkno; /* previous block number, 0 unless frag */ 1302 long newsize; /* size of new block */ 1303 long oldsize; /* size of new block */ 1304 struct buf *bp; /* bp for allocated block */ 1305 { 1306 struct allocdirect *adp, *oldadp; 1307 struct allocdirectlst *adphead; 1308 struct bmsafemap *bmsafemap; 1309 struct inodedep *inodedep; 1310 struct pagedep *pagedep; 1311 struct newblk *newblk; 1312 1313 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect), 1314 M_ALLOCDIRECT, M_SOFTDEP_FLAGS); 1315 bzero(adp, sizeof(struct allocdirect)); 1316 adp->ad_list.wk_type = D_ALLOCDIRECT; 1317 adp->ad_lbn = lbn; 1318 adp->ad_newblkno = newblkno; 1319 adp->ad_oldblkno = oldblkno; 1320 adp->ad_newsize = newsize; 1321 adp->ad_oldsize = oldsize; 1322 adp->ad_state = ATTACHED; 1323 if (newblkno == oldblkno) 1324 adp->ad_freefrag = NULL; 1325 else 1326 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1327 1328 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1329 panic("softdep_setup_allocdirect: lost block"); 1330 1331 ACQUIRE_LOCK(&lk); 1332 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep); 1333 adp->ad_inodedep = inodedep; 1334 1335 if (newblk->nb_state == DEPCOMPLETE) { 1336 adp->ad_state |= DEPCOMPLETE; 1337 adp->ad_buf = NULL; 1338 } else { 1339 bmsafemap = newblk->nb_bmsafemap; 1340 adp->ad_buf = bmsafemap->sm_buf; 1341 LIST_REMOVE(newblk, nb_deps); 1342 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1343 } 1344 LIST_REMOVE(newblk, nb_hash); 1345 FREE(newblk, M_NEWBLK); 1346 1347 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list); 1348 if (lbn >= NDADDR) { 1349 /* allocating an indirect block */ 1350 if (oldblkno != 0) { 1351 FREE_LOCK(&lk); 1352 panic("softdep_setup_allocdirect: non-zero indir"); 1353 } 1354 } else { 1355 /* 1356 * Allocating a direct block. 1357 * 1358 * If we are allocating a directory block, then we must 1359 * allocate an associated pagedep to track additions and 1360 * deletions. 1361 */ 1362 if ((ip->i_mode & IFMT) == IFDIR && 1363 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1364 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 1365 } 1366 /* 1367 * The list of allocdirects must be kept in sorted and ascending 1368 * order so that the rollback routines can quickly determine the 1369 * first uncommitted block (the size of the file stored on disk 1370 * ends at the end of the lowest committed fragment, or if there 1371 * are no fragments, at the end of the highest committed block). 1372 * Since files generally grow, the typical case is that the new 1373 * block is to be added at the end of the list. We speed this 1374 * special case by checking against the last allocdirect in the 1375 * list before laboriously traversing the list looking for the 1376 * insertion point. 1377 */ 1378 adphead = &inodedep->id_newinoupdt; 1379 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1380 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1381 /* insert at end of list */ 1382 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1383 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1384 allocdirect_merge(adphead, adp, oldadp); 1385 FREE_LOCK(&lk); 1386 return; 1387 } 1388 TAILQ_FOREACH(oldadp, adphead, ad_next) { 1389 if (oldadp->ad_lbn >= lbn) 1390 break; 1391 } 1392 if (oldadp == NULL) { 1393 FREE_LOCK(&lk); 1394 panic("softdep_setup_allocdirect: lost entry"); 1395 } 1396 /* insert in middle of list */ 1397 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1398 if (oldadp->ad_lbn == lbn) 1399 allocdirect_merge(adphead, adp, oldadp); 1400 FREE_LOCK(&lk); 1401 } 1402 1403 /* 1404 * Replace an old allocdirect dependency with a newer one. 1405 * This routine must be called with splbio interrupts blocked. 1406 */ 1407 static void 1408 allocdirect_merge(adphead, newadp, oldadp) 1409 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 1410 struct allocdirect *newadp; /* allocdirect being added */ 1411 struct allocdirect *oldadp; /* existing allocdirect being checked */ 1412 { 1413 struct freefrag *freefrag; 1414 1415 #ifdef DEBUG 1416 if (lk.lkt_held == NOHOLDER) 1417 panic("allocdirect_merge: lock not held"); 1418 #endif 1419 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 1420 newadp->ad_oldsize != oldadp->ad_newsize || 1421 newadp->ad_lbn >= NDADDR) { 1422 FREE_LOCK(&lk); 1423 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d", 1424 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn, 1425 NDADDR); 1426 } 1427 newadp->ad_oldblkno = oldadp->ad_oldblkno; 1428 newadp->ad_oldsize = oldadp->ad_oldsize; 1429 /* 1430 * If the old dependency had a fragment to free or had never 1431 * previously had a block allocated, then the new dependency 1432 * can immediately post its freefrag and adopt the old freefrag. 1433 * This action is done by swapping the freefrag dependencies. 1434 * The new dependency gains the old one's freefrag, and the 1435 * old one gets the new one and then immediately puts it on 1436 * the worklist when it is freed by free_allocdirect. It is 1437 * not possible to do this swap when the old dependency had a 1438 * non-zero size but no previous fragment to free. This condition 1439 * arises when the new block is an extension of the old block. 1440 * Here, the first part of the fragment allocated to the new 1441 * dependency is part of the block currently claimed on disk by 1442 * the old dependency, so cannot legitimately be freed until the 1443 * conditions for the new dependency are fulfilled. 1444 */ 1445 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 1446 freefrag = newadp->ad_freefrag; 1447 newadp->ad_freefrag = oldadp->ad_freefrag; 1448 oldadp->ad_freefrag = freefrag; 1449 } 1450 free_allocdirect(adphead, oldadp, 0); 1451 } 1452 1453 /* 1454 * Allocate a new freefrag structure if needed. 1455 */ 1456 static struct freefrag * 1457 newfreefrag(ip, blkno, size) 1458 struct inode *ip; 1459 ufs_daddr_t blkno; 1460 long size; 1461 { 1462 struct freefrag *freefrag; 1463 struct fs *fs; 1464 1465 if (blkno == 0) 1466 return (NULL); 1467 fs = ip->i_fs; 1468 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 1469 panic("newfreefrag: frag size"); 1470 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag), 1471 M_FREEFRAG, M_SOFTDEP_FLAGS); 1472 freefrag->ff_list.wk_type = D_FREEFRAG; 1473 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */ 1474 freefrag->ff_inum = ip->i_number; 1475 freefrag->ff_fs = fs; 1476 freefrag->ff_devvp = ip->i_devvp; 1477 freefrag->ff_blkno = blkno; 1478 freefrag->ff_fragsize = size; 1479 return (freefrag); 1480 } 1481 1482 /* 1483 * This workitem de-allocates fragments that were replaced during 1484 * file block allocation. 1485 */ 1486 static void 1487 handle_workitem_freefrag(freefrag) 1488 struct freefrag *freefrag; 1489 { 1490 struct inode tip; 1491 1492 tip.i_fs = freefrag->ff_fs; 1493 tip.i_devvp = freefrag->ff_devvp; 1494 tip.i_dev = freefrag->ff_devvp->v_rdev; 1495 tip.i_number = freefrag->ff_inum; 1496 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */ 1497 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize); 1498 FREE(freefrag, M_FREEFRAG); 1499 } 1500 1501 /* 1502 * Indirect block allocation dependencies. 1503 * 1504 * The same dependencies that exist for a direct block also exist when 1505 * a new block is allocated and pointed to by an entry in a block of 1506 * indirect pointers. The undo/redo states described above are also 1507 * used here. Because an indirect block contains many pointers that 1508 * may have dependencies, a second copy of the entire in-memory indirect 1509 * block is kept. The buffer cache copy is always completely up-to-date. 1510 * The second copy, which is used only as a source for disk writes, 1511 * contains only the safe pointers (i.e., those that have no remaining 1512 * update dependencies). The second copy is freed when all pointers 1513 * are safe. The cache is not allowed to replace indirect blocks with 1514 * pending update dependencies. If a buffer containing an indirect 1515 * block with dependencies is written, these routines will mark it 1516 * dirty again. It can only be successfully written once all the 1517 * dependencies are removed. The ffs_fsync routine in conjunction with 1518 * softdep_sync_metadata work together to get all the dependencies 1519 * removed so that a file can be successfully written to disk. Three 1520 * procedures are used when setting up indirect block pointer 1521 * dependencies. The division is necessary because of the organization 1522 * of the "balloc" routine and because of the distinction between file 1523 * pages and file metadata blocks. 1524 */ 1525 1526 /* 1527 * Allocate a new allocindir structure. 1528 */ 1529 static struct allocindir * 1530 newallocindir(ip, ptrno, newblkno, oldblkno) 1531 struct inode *ip; /* inode for file being extended */ 1532 int ptrno; /* offset of pointer in indirect block */ 1533 ufs_daddr_t newblkno; /* disk block number being added */ 1534 ufs_daddr_t oldblkno; /* previous block number, 0 if none */ 1535 { 1536 struct allocindir *aip; 1537 1538 MALLOC(aip, struct allocindir *, sizeof(struct allocindir), 1539 M_ALLOCINDIR, M_SOFTDEP_FLAGS); 1540 bzero(aip, sizeof(struct allocindir)); 1541 aip->ai_list.wk_type = D_ALLOCINDIR; 1542 aip->ai_state = ATTACHED; 1543 aip->ai_offset = ptrno; 1544 aip->ai_newblkno = newblkno; 1545 aip->ai_oldblkno = oldblkno; 1546 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize); 1547 return (aip); 1548 } 1549 1550 /* 1551 * Called just before setting an indirect block pointer 1552 * to a newly allocated file page. 1553 */ 1554 void 1555 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 1556 struct inode *ip; /* inode for file being extended */ 1557 ufs_lbn_t lbn; /* allocated block number within file */ 1558 struct buf *bp; /* buffer with indirect blk referencing page */ 1559 int ptrno; /* offset of pointer in indirect block */ 1560 ufs_daddr_t newblkno; /* disk block number being added */ 1561 ufs_daddr_t oldblkno; /* previous block number, 0 if none */ 1562 struct buf *nbp; /* buffer holding allocated page */ 1563 { 1564 struct allocindir *aip; 1565 struct pagedep *pagedep; 1566 1567 aip = newallocindir(ip, ptrno, newblkno, oldblkno); 1568 ACQUIRE_LOCK(&lk); 1569 /* 1570 * If we are allocating a directory page, then we must 1571 * allocate an associated pagedep to track additions and 1572 * deletions. 1573 */ 1574 if ((ip->i_mode & IFMT) == IFDIR && 1575 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1576 WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list); 1577 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1578 FREE_LOCK(&lk); 1579 setup_allocindir_phase2(bp, ip, aip); 1580 } 1581 1582 /* 1583 * Called just before setting an indirect block pointer to a 1584 * newly allocated indirect block. 1585 */ 1586 void 1587 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 1588 struct buf *nbp; /* newly allocated indirect block */ 1589 struct inode *ip; /* inode for file being extended */ 1590 struct buf *bp; /* indirect block referencing allocated block */ 1591 int ptrno; /* offset of pointer in indirect block */ 1592 ufs_daddr_t newblkno; /* disk block number being added */ 1593 { 1594 struct allocindir *aip; 1595 1596 aip = newallocindir(ip, ptrno, newblkno, 0); 1597 ACQUIRE_LOCK(&lk); 1598 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1599 FREE_LOCK(&lk); 1600 setup_allocindir_phase2(bp, ip, aip); 1601 } 1602 1603 /* 1604 * Called to finish the allocation of the "aip" allocated 1605 * by one of the two routines above. 1606 */ 1607 static void 1608 setup_allocindir_phase2(bp, ip, aip) 1609 struct buf *bp; /* in-memory copy of the indirect block */ 1610 struct inode *ip; /* inode for file being extended */ 1611 struct allocindir *aip; /* allocindir allocated by the above routines */ 1612 { 1613 struct worklist *wk; 1614 struct indirdep *indirdep, *newindirdep; 1615 struct bmsafemap *bmsafemap; 1616 struct allocindir *oldaip; 1617 struct freefrag *freefrag; 1618 struct newblk *newblk; 1619 1620 if (bp->b_lblkno >= 0) 1621 panic("setup_allocindir_phase2: not indir blk"); 1622 for (indirdep = NULL, newindirdep = NULL; ; ) { 1623 ACQUIRE_LOCK(&lk); 1624 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 1625 if (wk->wk_type != D_INDIRDEP) 1626 continue; 1627 indirdep = WK_INDIRDEP(wk); 1628 break; 1629 } 1630 if (indirdep == NULL && newindirdep) { 1631 indirdep = newindirdep; 1632 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 1633 newindirdep = NULL; 1634 } 1635 FREE_LOCK(&lk); 1636 if (indirdep) { 1637 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0, 1638 &newblk) == 0) 1639 panic("setup_allocindir: lost block"); 1640 ACQUIRE_LOCK(&lk); 1641 if (newblk->nb_state == DEPCOMPLETE) { 1642 aip->ai_state |= DEPCOMPLETE; 1643 aip->ai_buf = NULL; 1644 } else { 1645 bmsafemap = newblk->nb_bmsafemap; 1646 aip->ai_buf = bmsafemap->sm_buf; 1647 LIST_REMOVE(newblk, nb_deps); 1648 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd, 1649 aip, ai_deps); 1650 } 1651 LIST_REMOVE(newblk, nb_hash); 1652 FREE(newblk, M_NEWBLK); 1653 aip->ai_indirdep = indirdep; 1654 /* 1655 * Check to see if there is an existing dependency 1656 * for this block. If there is, merge the old 1657 * dependency into the new one. 1658 */ 1659 if (aip->ai_oldblkno == 0) 1660 oldaip = NULL; 1661 else 1662 1663 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) 1664 if (oldaip->ai_offset == aip->ai_offset) 1665 break; 1666 if (oldaip != NULL) { 1667 if (oldaip->ai_newblkno != aip->ai_oldblkno) { 1668 FREE_LOCK(&lk); 1669 panic("setup_allocindir_phase2: blkno"); 1670 } 1671 aip->ai_oldblkno = oldaip->ai_oldblkno; 1672 freefrag = oldaip->ai_freefrag; 1673 oldaip->ai_freefrag = aip->ai_freefrag; 1674 aip->ai_freefrag = freefrag; 1675 free_allocindir(oldaip, NULL); 1676 } 1677 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 1678 ((ufs_daddr_t *)indirdep->ir_savebp->b_data) 1679 [aip->ai_offset] = aip->ai_oldblkno; 1680 FREE_LOCK(&lk); 1681 } 1682 if (newindirdep) { 1683 /* 1684 * Avoid any possibility of data corruption by 1685 * ensuring that our old version is thrown away. 1686 */ 1687 newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 1688 brelse(newindirdep->ir_savebp); 1689 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP); 1690 } 1691 if (indirdep) 1692 break; 1693 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep), 1694 M_INDIRDEP, M_SOFTDEP_FLAGS); 1695 newindirdep->ir_list.wk_type = D_INDIRDEP; 1696 newindirdep->ir_state = ATTACHED; 1697 LIST_INIT(&newindirdep->ir_deplisthd); 1698 LIST_INIT(&newindirdep->ir_donehd); 1699 if (bp->b_bio2.bio_blkno == (daddr_t)-1) { 1700 VOP_BMAP(bp->b_vp, bp->b_bio1.bio_blkno, 1701 NULL, &bp->b_bio2.bio_blkno, 1702 NULL, NULL); 1703 } 1704 KKASSERT(bp->b_bio2.bio_blkno != (daddr_t)-1); 1705 newindirdep->ir_savebp = 1706 getblk(ip->i_devvp, bp->b_bio2.bio_blkno, bp->b_bcount, 1707 0, 0); 1708 BUF_KERNPROC(newindirdep->ir_savebp); 1709 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 1710 } 1711 } 1712 1713 /* 1714 * Block de-allocation dependencies. 1715 * 1716 * When blocks are de-allocated, the on-disk pointers must be nullified before 1717 * the blocks are made available for use by other files. (The true 1718 * requirement is that old pointers must be nullified before new on-disk 1719 * pointers are set. We chose this slightly more stringent requirement to 1720 * reduce complexity.) Our implementation handles this dependency by updating 1721 * the inode (or indirect block) appropriately but delaying the actual block 1722 * de-allocation (i.e., freemap and free space count manipulation) until 1723 * after the updated versions reach stable storage. After the disk is 1724 * updated, the blocks can be safely de-allocated whenever it is convenient. 1725 * This implementation handles only the common case of reducing a file's 1726 * length to zero. Other cases are handled by the conventional synchronous 1727 * write approach. 1728 * 1729 * The ffs implementation with which we worked double-checks 1730 * the state of the block pointers and file size as it reduces 1731 * a file's length. Some of this code is replicated here in our 1732 * soft updates implementation. The freeblks->fb_chkcnt field is 1733 * used to transfer a part of this information to the procedure 1734 * that eventually de-allocates the blocks. 1735 * 1736 * This routine should be called from the routine that shortens 1737 * a file's length, before the inode's size or block pointers 1738 * are modified. It will save the block pointer information for 1739 * later release and zero the inode so that the calling routine 1740 * can release it. 1741 */ 1742 struct softdep_setup_freeblocks_info { 1743 struct fs *fs; 1744 struct inode *ip; 1745 }; 1746 1747 static int softdep_setup_freeblocks_bp(struct buf *bp, void *data); 1748 1749 void 1750 softdep_setup_freeblocks(ip, length) 1751 struct inode *ip; /* The inode whose length is to be reduced */ 1752 off_t length; /* The new length for the file */ 1753 { 1754 struct softdep_setup_freeblocks_info info; 1755 struct freeblks *freeblks; 1756 struct inodedep *inodedep; 1757 struct allocdirect *adp; 1758 struct vnode *vp; 1759 struct buf *bp; 1760 struct fs *fs; 1761 int i, error, delay; 1762 int count; 1763 1764 fs = ip->i_fs; 1765 if (length != 0) 1766 panic("softde_setup_freeblocks: non-zero length"); 1767 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks), 1768 M_FREEBLKS, M_SOFTDEP_FLAGS); 1769 bzero(freeblks, sizeof(struct freeblks)); 1770 freeblks->fb_list.wk_type = D_FREEBLKS; 1771 freeblks->fb_state = ATTACHED; 1772 freeblks->fb_uid = ip->i_uid; 1773 freeblks->fb_previousinum = ip->i_number; 1774 freeblks->fb_devvp = ip->i_devvp; 1775 freeblks->fb_fs = fs; 1776 freeblks->fb_oldsize = ip->i_size; 1777 freeblks->fb_newsize = length; 1778 freeblks->fb_chkcnt = ip->i_blocks; 1779 for (i = 0; i < NDADDR; i++) { 1780 freeblks->fb_dblks[i] = ip->i_db[i]; 1781 ip->i_db[i] = 0; 1782 } 1783 for (i = 0; i < NIADDR; i++) { 1784 freeblks->fb_iblks[i] = ip->i_ib[i]; 1785 ip->i_ib[i] = 0; 1786 } 1787 ip->i_blocks = 0; 1788 ip->i_size = 0; 1789 /* 1790 * Push the zero'ed inode to to its disk buffer so that we are free 1791 * to delete its dependencies below. Once the dependencies are gone 1792 * the buffer can be safely released. 1793 */ 1794 if ((error = bread(ip->i_devvp, 1795 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 1796 (int)fs->fs_bsize, &bp)) != 0) 1797 softdep_error("softdep_setup_freeblocks", error); 1798 *((struct dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) = 1799 ip->i_din; 1800 /* 1801 * Find and eliminate any inode dependencies. 1802 */ 1803 ACQUIRE_LOCK(&lk); 1804 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep); 1805 if ((inodedep->id_state & IOSTARTED) != 0) { 1806 FREE_LOCK(&lk); 1807 panic("softdep_setup_freeblocks: inode busy"); 1808 } 1809 /* 1810 * Add the freeblks structure to the list of operations that 1811 * must await the zero'ed inode being written to disk. If we 1812 * still have a bitmap dependency (delay == 0), then the inode 1813 * has never been written to disk, so we can process the 1814 * freeblks below once we have deleted the dependencies. 1815 */ 1816 delay = (inodedep->id_state & DEPCOMPLETE); 1817 if (delay) 1818 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list); 1819 /* 1820 * Because the file length has been truncated to zero, any 1821 * pending block allocation dependency structures associated 1822 * with this inode are obsolete and can simply be de-allocated. 1823 * We must first merge the two dependency lists to get rid of 1824 * any duplicate freefrag structures, then purge the merged list. 1825 */ 1826 merge_inode_lists(inodedep); 1827 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 1828 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 1829 FREE_LOCK(&lk); 1830 bdwrite(bp); 1831 /* 1832 * We must wait for any I/O in progress to finish so that 1833 * all potential buffers on the dirty list will be visible. 1834 * Once they are all there, walk the list and get rid of 1835 * any dependencies. 1836 */ 1837 vp = ITOV(ip); 1838 ACQUIRE_LOCK(&lk); 1839 drain_output(vp, 1); 1840 1841 info.fs = fs; 1842 info.ip = ip; 1843 crit_enter(); 1844 do { 1845 count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 1846 softdep_setup_freeblocks_bp, &info); 1847 } while (count != 0); 1848 crit_exit(); 1849 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0) 1850 (void)free_inodedep(inodedep); 1851 1852 if (delay) { 1853 freeblks->fb_state |= DEPCOMPLETE; 1854 /* 1855 * If the inode with zeroed block pointers is now on disk 1856 * we can start freeing blocks. Add freeblks to the worklist 1857 * instead of calling handle_workitem_freeblocks directly as 1858 * it is more likely that additional IO is needed to complete 1859 * the request here than in the !delay case. 1860 */ 1861 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 1862 add_to_worklist(&freeblks->fb_list); 1863 } 1864 1865 FREE_LOCK(&lk); 1866 /* 1867 * If the inode has never been written to disk (delay == 0), 1868 * then we can process the freeblks now that we have deleted 1869 * the dependencies. 1870 */ 1871 if (!delay) 1872 handle_workitem_freeblocks(freeblks); 1873 } 1874 1875 static int 1876 softdep_setup_freeblocks_bp(struct buf *bp, void *data) 1877 { 1878 struct softdep_setup_freeblocks_info *info = data; 1879 struct inodedep *inodedep; 1880 1881 if (getdirtybuf(&bp, MNT_WAIT) == 0) { 1882 printf("softdep_setup_freeblocks_bp(1): caught bp %p going away\n", bp); 1883 return(-1); 1884 } 1885 if (bp->b_vp != ITOV(info->ip) || (bp->b_flags & B_DELWRI) == 0) { 1886 printf("softdep_setup_freeblocks_bp(2): caught bp %p going away\n", bp); 1887 BUF_UNLOCK(bp); 1888 return(-1); 1889 } 1890 (void) inodedep_lookup(info->fs, info->ip->i_number, 0, &inodedep); 1891 deallocate_dependencies(bp, inodedep); 1892 bp->b_flags |= B_INVAL | B_NOCACHE; 1893 FREE_LOCK(&lk); 1894 brelse(bp); 1895 ACQUIRE_LOCK(&lk); 1896 return(1); 1897 } 1898 1899 /* 1900 * Reclaim any dependency structures from a buffer that is about to 1901 * be reallocated to a new vnode. The buffer must be locked, thus, 1902 * no I/O completion operations can occur while we are manipulating 1903 * its associated dependencies. The mutex is held so that other I/O's 1904 * associated with related dependencies do not occur. 1905 */ 1906 static void 1907 deallocate_dependencies(bp, inodedep) 1908 struct buf *bp; 1909 struct inodedep *inodedep; 1910 { 1911 struct worklist *wk; 1912 struct indirdep *indirdep; 1913 struct allocindir *aip; 1914 struct pagedep *pagedep; 1915 struct dirrem *dirrem; 1916 struct diradd *dap; 1917 int i; 1918 1919 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 1920 switch (wk->wk_type) { 1921 1922 case D_INDIRDEP: 1923 indirdep = WK_INDIRDEP(wk); 1924 /* 1925 * None of the indirect pointers will ever be visible, 1926 * so they can simply be tossed. GOINGAWAY ensures 1927 * that allocated pointers will be saved in the buffer 1928 * cache until they are freed. Note that they will 1929 * only be able to be found by their physical address 1930 * since the inode mapping the logical address will 1931 * be gone. The save buffer used for the safe copy 1932 * was allocated in setup_allocindir_phase2 using 1933 * the physical address so it could be used for this 1934 * purpose. Hence we swap the safe copy with the real 1935 * copy, allowing the safe copy to be freed and holding 1936 * on to the real copy for later use in indir_trunc. 1937 * 1938 * NOTE: ir_savebp is relative to the block device 1939 * so b_bio1 contains the device block number. 1940 */ 1941 if (indirdep->ir_state & GOINGAWAY) { 1942 FREE_LOCK(&lk); 1943 panic("deallocate_dependencies: already gone"); 1944 } 1945 indirdep->ir_state |= GOINGAWAY; 1946 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 1947 free_allocindir(aip, inodedep); 1948 if (bp->b_bio1.bio_blkno >= 0 || 1949 bp->b_bio2.bio_blkno != indirdep->ir_savebp->b_bio1.bio_blkno) { 1950 FREE_LOCK(&lk); 1951 panic("deallocate_dependencies: not indir"); 1952 } 1953 bcopy(bp->b_data, indirdep->ir_savebp->b_data, 1954 bp->b_bcount); 1955 WORKLIST_REMOVE(wk); 1956 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk); 1957 continue; 1958 1959 case D_PAGEDEP: 1960 pagedep = WK_PAGEDEP(wk); 1961 /* 1962 * None of the directory additions will ever be 1963 * visible, so they can simply be tossed. 1964 */ 1965 for (i = 0; i < DAHASHSZ; i++) 1966 while ((dap = 1967 LIST_FIRST(&pagedep->pd_diraddhd[i]))) 1968 free_diradd(dap); 1969 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0) 1970 free_diradd(dap); 1971 /* 1972 * Copy any directory remove dependencies to the list 1973 * to be processed after the zero'ed inode is written. 1974 * If the inode has already been written, then they 1975 * can be dumped directly onto the work list. 1976 */ 1977 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 1978 LIST_REMOVE(dirrem, dm_next); 1979 dirrem->dm_dirinum = pagedep->pd_ino; 1980 if (inodedep == NULL || 1981 (inodedep->id_state & ALLCOMPLETE) == 1982 ALLCOMPLETE) 1983 add_to_worklist(&dirrem->dm_list); 1984 else 1985 WORKLIST_INSERT(&inodedep->id_bufwait, 1986 &dirrem->dm_list); 1987 } 1988 WORKLIST_REMOVE(&pagedep->pd_list); 1989 LIST_REMOVE(pagedep, pd_hash); 1990 WORKITEM_FREE(pagedep, D_PAGEDEP); 1991 continue; 1992 1993 case D_ALLOCINDIR: 1994 free_allocindir(WK_ALLOCINDIR(wk), inodedep); 1995 continue; 1996 1997 case D_ALLOCDIRECT: 1998 case D_INODEDEP: 1999 FREE_LOCK(&lk); 2000 panic("deallocate_dependencies: Unexpected type %s", 2001 TYPENAME(wk->wk_type)); 2002 /* NOTREACHED */ 2003 2004 default: 2005 FREE_LOCK(&lk); 2006 panic("deallocate_dependencies: Unknown type %s", 2007 TYPENAME(wk->wk_type)); 2008 /* NOTREACHED */ 2009 } 2010 } 2011 } 2012 2013 /* 2014 * Free an allocdirect. Generate a new freefrag work request if appropriate. 2015 * This routine must be called with splbio interrupts blocked. 2016 */ 2017 static void 2018 free_allocdirect(adphead, adp, delay) 2019 struct allocdirectlst *adphead; 2020 struct allocdirect *adp; 2021 int delay; 2022 { 2023 2024 #ifdef DEBUG 2025 if (lk.lkt_held == NOHOLDER) 2026 panic("free_allocdirect: lock not held"); 2027 #endif 2028 if ((adp->ad_state & DEPCOMPLETE) == 0) 2029 LIST_REMOVE(adp, ad_deps); 2030 TAILQ_REMOVE(adphead, adp, ad_next); 2031 if ((adp->ad_state & COMPLETE) == 0) 2032 WORKLIST_REMOVE(&adp->ad_list); 2033 if (adp->ad_freefrag != NULL) { 2034 if (delay) 2035 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 2036 &adp->ad_freefrag->ff_list); 2037 else 2038 add_to_worklist(&adp->ad_freefrag->ff_list); 2039 } 2040 WORKITEM_FREE(adp, D_ALLOCDIRECT); 2041 } 2042 2043 /* 2044 * Prepare an inode to be freed. The actual free operation is not 2045 * done until the zero'ed inode has been written to disk. 2046 */ 2047 void 2048 softdep_freefile(pvp, ino, mode) 2049 struct vnode *pvp; 2050 ino_t ino; 2051 int mode; 2052 { 2053 struct inode *ip = VTOI(pvp); 2054 struct inodedep *inodedep; 2055 struct freefile *freefile; 2056 2057 /* 2058 * This sets up the inode de-allocation dependency. 2059 */ 2060 MALLOC(freefile, struct freefile *, sizeof(struct freefile), 2061 M_FREEFILE, M_SOFTDEP_FLAGS); 2062 freefile->fx_list.wk_type = D_FREEFILE; 2063 freefile->fx_list.wk_state = 0; 2064 freefile->fx_mode = mode; 2065 freefile->fx_oldinum = ino; 2066 freefile->fx_devvp = ip->i_devvp; 2067 freefile->fx_fs = ip->i_fs; 2068 2069 /* 2070 * If the inodedep does not exist, then the zero'ed inode has 2071 * been written to disk. If the allocated inode has never been 2072 * written to disk, then the on-disk inode is zero'ed. In either 2073 * case we can free the file immediately. 2074 */ 2075 ACQUIRE_LOCK(&lk); 2076 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 || 2077 check_inode_unwritten(inodedep)) { 2078 FREE_LOCK(&lk); 2079 handle_workitem_freefile(freefile); 2080 return; 2081 } 2082 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 2083 FREE_LOCK(&lk); 2084 } 2085 2086 /* 2087 * Check to see if an inode has never been written to disk. If 2088 * so free the inodedep and return success, otherwise return failure. 2089 * This routine must be called with splbio interrupts blocked. 2090 * 2091 * If we still have a bitmap dependency, then the inode has never 2092 * been written to disk. Drop the dependency as it is no longer 2093 * necessary since the inode is being deallocated. We set the 2094 * ALLCOMPLETE flags since the bitmap now properly shows that the 2095 * inode is not allocated. Even if the inode is actively being 2096 * written, it has been rolled back to its zero'ed state, so we 2097 * are ensured that a zero inode is what is on the disk. For short 2098 * lived files, this change will usually result in removing all the 2099 * dependencies from the inode so that it can be freed immediately. 2100 */ 2101 static int 2102 check_inode_unwritten(inodedep) 2103 struct inodedep *inodedep; 2104 { 2105 2106 if ((inodedep->id_state & DEPCOMPLETE) != 0 || 2107 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2108 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2109 LIST_FIRST(&inodedep->id_inowait) != NULL || 2110 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2111 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2112 inodedep->id_nlinkdelta != 0) 2113 return (0); 2114 2115 /* 2116 * Another process might be in initiate_write_inodeblock 2117 * trying to allocate memory without holding "Softdep Lock". 2118 */ 2119 if ((inodedep->id_state & IOSTARTED) != 0 && 2120 inodedep->id_savedino == NULL) 2121 return(0); 2122 2123 inodedep->id_state |= ALLCOMPLETE; 2124 LIST_REMOVE(inodedep, id_deps); 2125 inodedep->id_buf = NULL; 2126 if (inodedep->id_state & ONWORKLIST) 2127 WORKLIST_REMOVE(&inodedep->id_list); 2128 if (inodedep->id_savedino != NULL) { 2129 FREE(inodedep->id_savedino, M_INODEDEP); 2130 inodedep->id_savedino = NULL; 2131 } 2132 if (free_inodedep(inodedep) == 0) { 2133 FREE_LOCK(&lk); 2134 panic("check_inode_unwritten: busy inode"); 2135 } 2136 return (1); 2137 } 2138 2139 /* 2140 * Try to free an inodedep structure. Return 1 if it could be freed. 2141 */ 2142 static int 2143 free_inodedep(inodedep) 2144 struct inodedep *inodedep; 2145 { 2146 2147 if ((inodedep->id_state & ONWORKLIST) != 0 || 2148 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 2149 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2150 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2151 LIST_FIRST(&inodedep->id_inowait) != NULL || 2152 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2153 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2154 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL) 2155 return (0); 2156 LIST_REMOVE(inodedep, id_hash); 2157 WORKITEM_FREE(inodedep, D_INODEDEP); 2158 num_inodedep -= 1; 2159 return (1); 2160 } 2161 2162 /* 2163 * This workitem routine performs the block de-allocation. 2164 * The workitem is added to the pending list after the updated 2165 * inode block has been written to disk. As mentioned above, 2166 * checks regarding the number of blocks de-allocated (compared 2167 * to the number of blocks allocated for the file) are also 2168 * performed in this function. 2169 */ 2170 static void 2171 handle_workitem_freeblocks(freeblks) 2172 struct freeblks *freeblks; 2173 { 2174 struct inode tip; 2175 ufs_daddr_t bn; 2176 struct fs *fs; 2177 int i, level, bsize; 2178 long nblocks, blocksreleased = 0; 2179 int error, allerror = 0; 2180 ufs_lbn_t baselbns[NIADDR], tmpval; 2181 2182 tip.i_number = freeblks->fb_previousinum; 2183 tip.i_devvp = freeblks->fb_devvp; 2184 tip.i_dev = freeblks->fb_devvp->v_rdev; 2185 tip.i_fs = freeblks->fb_fs; 2186 tip.i_size = freeblks->fb_oldsize; 2187 tip.i_uid = freeblks->fb_uid; 2188 fs = freeblks->fb_fs; 2189 tmpval = 1; 2190 baselbns[0] = NDADDR; 2191 for (i = 1; i < NIADDR; i++) { 2192 tmpval *= NINDIR(fs); 2193 baselbns[i] = baselbns[i - 1] + tmpval; 2194 } 2195 nblocks = btodb(fs->fs_bsize); 2196 blocksreleased = 0; 2197 /* 2198 * Indirect blocks first. 2199 */ 2200 for (level = (NIADDR - 1); level >= 0; level--) { 2201 if ((bn = freeblks->fb_iblks[level]) == 0) 2202 continue; 2203 if ((error = indir_trunc(&tip, fsbtodb(fs, bn), level, 2204 baselbns[level], &blocksreleased)) == 0) 2205 allerror = error; 2206 ffs_blkfree(&tip, bn, fs->fs_bsize); 2207 blocksreleased += nblocks; 2208 } 2209 /* 2210 * All direct blocks or frags. 2211 */ 2212 for (i = (NDADDR - 1); i >= 0; i--) { 2213 if ((bn = freeblks->fb_dblks[i]) == 0) 2214 continue; 2215 bsize = blksize(fs, &tip, i); 2216 ffs_blkfree(&tip, bn, bsize); 2217 blocksreleased += btodb(bsize); 2218 } 2219 2220 #ifdef DIAGNOSTIC 2221 if (freeblks->fb_chkcnt != blocksreleased) 2222 printf("handle_workitem_freeblocks: block count\n"); 2223 if (allerror) 2224 softdep_error("handle_workitem_freeblks", allerror); 2225 #endif /* DIAGNOSTIC */ 2226 WORKITEM_FREE(freeblks, D_FREEBLKS); 2227 } 2228 2229 /* 2230 * Release blocks associated with the inode ip and stored in the indirect 2231 * block dbn. If level is greater than SINGLE, the block is an indirect block 2232 * and recursive calls to indirtrunc must be used to cleanse other indirect 2233 * blocks. 2234 */ 2235 static int 2236 indir_trunc(ip, dbn, level, lbn, countp) 2237 struct inode *ip; 2238 ufs_daddr_t dbn; 2239 int level; 2240 ufs_lbn_t lbn; 2241 long *countp; 2242 { 2243 struct buf *bp; 2244 ufs_daddr_t *bap; 2245 ufs_daddr_t nb; 2246 struct fs *fs; 2247 struct worklist *wk; 2248 struct indirdep *indirdep; 2249 int i, lbnadd, nblocks; 2250 int error, allerror = 0; 2251 2252 fs = ip->i_fs; 2253 lbnadd = 1; 2254 for (i = level; i > 0; i--) 2255 lbnadd *= NINDIR(fs); 2256 /* 2257 * Get buffer of block pointers to be freed. This routine is not 2258 * called until the zero'ed inode has been written, so it is safe 2259 * to free blocks as they are encountered. Because the inode has 2260 * been zero'ed, calls to bmap on these blocks will fail. So, we 2261 * have to use the on-disk address and the block device for the 2262 * filesystem to look them up. If the file was deleted before its 2263 * indirect blocks were all written to disk, the routine that set 2264 * us up (deallocate_dependencies) will have arranged to leave 2265 * a complete copy of the indirect block in memory for our use. 2266 * Otherwise we have to read the blocks in from the disk. 2267 */ 2268 ACQUIRE_LOCK(&lk); 2269 if ((bp = incore(ip->i_devvp, dbn)) != NULL && 2270 (wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2271 if (wk->wk_type != D_INDIRDEP || 2272 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp || 2273 (indirdep->ir_state & GOINGAWAY) == 0) { 2274 FREE_LOCK(&lk); 2275 panic("indir_trunc: lost indirdep"); 2276 } 2277 WORKLIST_REMOVE(wk); 2278 WORKITEM_FREE(indirdep, D_INDIRDEP); 2279 if (LIST_FIRST(&bp->b_dep) != NULL) { 2280 FREE_LOCK(&lk); 2281 panic("indir_trunc: dangling dep"); 2282 } 2283 FREE_LOCK(&lk); 2284 } else { 2285 FREE_LOCK(&lk); 2286 error = bread(ip->i_devvp, dbn, (int)fs->fs_bsize, &bp); 2287 if (error) 2288 return (error); 2289 } 2290 /* 2291 * Recursively free indirect blocks. 2292 */ 2293 bap = (ufs_daddr_t *)bp->b_data; 2294 nblocks = btodb(fs->fs_bsize); 2295 for (i = NINDIR(fs) - 1; i >= 0; i--) { 2296 if ((nb = bap[i]) == 0) 2297 continue; 2298 if (level != 0) { 2299 if ((error = indir_trunc(ip, fsbtodb(fs, nb), 2300 level - 1, lbn + (i * lbnadd), countp)) != 0) 2301 allerror = error; 2302 } 2303 ffs_blkfree(ip, nb, fs->fs_bsize); 2304 *countp += nblocks; 2305 } 2306 bp->b_flags |= B_INVAL | B_NOCACHE; 2307 brelse(bp); 2308 return (allerror); 2309 } 2310 2311 /* 2312 * Free an allocindir. 2313 * This routine must be called with splbio interrupts blocked. 2314 */ 2315 static void 2316 free_allocindir(aip, inodedep) 2317 struct allocindir *aip; 2318 struct inodedep *inodedep; 2319 { 2320 struct freefrag *freefrag; 2321 2322 #ifdef DEBUG 2323 if (lk.lkt_held == NOHOLDER) 2324 panic("free_allocindir: lock not held"); 2325 #endif 2326 if ((aip->ai_state & DEPCOMPLETE) == 0) 2327 LIST_REMOVE(aip, ai_deps); 2328 if (aip->ai_state & ONWORKLIST) 2329 WORKLIST_REMOVE(&aip->ai_list); 2330 LIST_REMOVE(aip, ai_next); 2331 if ((freefrag = aip->ai_freefrag) != NULL) { 2332 if (inodedep == NULL) 2333 add_to_worklist(&freefrag->ff_list); 2334 else 2335 WORKLIST_INSERT(&inodedep->id_bufwait, 2336 &freefrag->ff_list); 2337 } 2338 WORKITEM_FREE(aip, D_ALLOCINDIR); 2339 } 2340 2341 /* 2342 * Directory entry addition dependencies. 2343 * 2344 * When adding a new directory entry, the inode (with its incremented link 2345 * count) must be written to disk before the directory entry's pointer to it. 2346 * Also, if the inode is newly allocated, the corresponding freemap must be 2347 * updated (on disk) before the directory entry's pointer. These requirements 2348 * are met via undo/redo on the directory entry's pointer, which consists 2349 * simply of the inode number. 2350 * 2351 * As directory entries are added and deleted, the free space within a 2352 * directory block can become fragmented. The ufs filesystem will compact 2353 * a fragmented directory block to make space for a new entry. When this 2354 * occurs, the offsets of previously added entries change. Any "diradd" 2355 * dependency structures corresponding to these entries must be updated with 2356 * the new offsets. 2357 */ 2358 2359 /* 2360 * This routine is called after the in-memory inode's link 2361 * count has been incremented, but before the directory entry's 2362 * pointer to the inode has been set. 2363 */ 2364 void 2365 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp) 2366 struct buf *bp; /* buffer containing directory block */ 2367 struct inode *dp; /* inode for directory */ 2368 off_t diroffset; /* offset of new entry in directory */ 2369 ino_t newinum; /* inode referenced by new directory entry */ 2370 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 2371 { 2372 int offset; /* offset of new entry within directory block */ 2373 ufs_lbn_t lbn; /* block in directory containing new entry */ 2374 struct fs *fs; 2375 struct diradd *dap; 2376 struct pagedep *pagedep; 2377 struct inodedep *inodedep; 2378 struct mkdir *mkdir1, *mkdir2; 2379 2380 /* 2381 * Whiteouts have no dependencies. 2382 */ 2383 if (newinum == WINO) { 2384 if (newdirbp != NULL) 2385 bdwrite(newdirbp); 2386 return; 2387 } 2388 2389 fs = dp->i_fs; 2390 lbn = lblkno(fs, diroffset); 2391 offset = blkoff(fs, diroffset); 2392 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, 2393 M_SOFTDEP_FLAGS); 2394 bzero(dap, sizeof(struct diradd)); 2395 dap->da_list.wk_type = D_DIRADD; 2396 dap->da_offset = offset; 2397 dap->da_newinum = newinum; 2398 dap->da_state = ATTACHED; 2399 if (newdirbp == NULL) { 2400 dap->da_state |= DEPCOMPLETE; 2401 ACQUIRE_LOCK(&lk); 2402 } else { 2403 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 2404 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2405 M_SOFTDEP_FLAGS); 2406 mkdir1->md_list.wk_type = D_MKDIR; 2407 mkdir1->md_state = MKDIR_BODY; 2408 mkdir1->md_diradd = dap; 2409 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2410 M_SOFTDEP_FLAGS); 2411 mkdir2->md_list.wk_type = D_MKDIR; 2412 mkdir2->md_state = MKDIR_PARENT; 2413 mkdir2->md_diradd = dap; 2414 /* 2415 * Dependency on "." and ".." being written to disk. 2416 */ 2417 mkdir1->md_buf = newdirbp; 2418 ACQUIRE_LOCK(&lk); 2419 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 2420 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list); 2421 FREE_LOCK(&lk); 2422 bdwrite(newdirbp); 2423 /* 2424 * Dependency on link count increase for parent directory 2425 */ 2426 ACQUIRE_LOCK(&lk); 2427 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0 2428 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2429 dap->da_state &= ~MKDIR_PARENT; 2430 WORKITEM_FREE(mkdir2, D_MKDIR); 2431 } else { 2432 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 2433 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 2434 } 2435 } 2436 /* 2437 * Link into parent directory pagedep to await its being written. 2438 */ 2439 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2440 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2441 dap->da_pagedep = pagedep; 2442 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 2443 da_pdlist); 2444 /* 2445 * Link into its inodedep. Put it on the id_bufwait list if the inode 2446 * is not yet written. If it is written, do the post-inode write 2447 * processing to put it on the id_pendinghd list. 2448 */ 2449 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep); 2450 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 2451 diradd_inode_written(dap, inodedep); 2452 else 2453 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2454 FREE_LOCK(&lk); 2455 } 2456 2457 /* 2458 * This procedure is called to change the offset of a directory 2459 * entry when compacting a directory block which must be owned 2460 * exclusively by the caller. Note that the actual entry movement 2461 * must be done in this procedure to ensure that no I/O completions 2462 * occur while the move is in progress. 2463 */ 2464 void 2465 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize) 2466 struct inode *dp; /* inode for directory */ 2467 caddr_t base; /* address of dp->i_offset */ 2468 caddr_t oldloc; /* address of old directory location */ 2469 caddr_t newloc; /* address of new directory location */ 2470 int entrysize; /* size of directory entry */ 2471 { 2472 int offset, oldoffset, newoffset; 2473 struct pagedep *pagedep; 2474 struct diradd *dap; 2475 ufs_lbn_t lbn; 2476 2477 ACQUIRE_LOCK(&lk); 2478 lbn = lblkno(dp->i_fs, dp->i_offset); 2479 offset = blkoff(dp->i_fs, dp->i_offset); 2480 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) 2481 goto done; 2482 oldoffset = offset + (oldloc - base); 2483 newoffset = offset + (newloc - base); 2484 2485 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) { 2486 if (dap->da_offset != oldoffset) 2487 continue; 2488 dap->da_offset = newoffset; 2489 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) 2490 break; 2491 LIST_REMOVE(dap, da_pdlist); 2492 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], 2493 dap, da_pdlist); 2494 break; 2495 } 2496 if (dap == NULL) { 2497 2498 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) { 2499 if (dap->da_offset == oldoffset) { 2500 dap->da_offset = newoffset; 2501 break; 2502 } 2503 } 2504 } 2505 done: 2506 bcopy(oldloc, newloc, entrysize); 2507 FREE_LOCK(&lk); 2508 } 2509 2510 /* 2511 * Free a diradd dependency structure. This routine must be called 2512 * with splbio interrupts blocked. 2513 */ 2514 static void 2515 free_diradd(dap) 2516 struct diradd *dap; 2517 { 2518 struct dirrem *dirrem; 2519 struct pagedep *pagedep; 2520 struct inodedep *inodedep; 2521 struct mkdir *mkdir, *nextmd; 2522 2523 #ifdef DEBUG 2524 if (lk.lkt_held == NOHOLDER) 2525 panic("free_diradd: lock not held"); 2526 #endif 2527 WORKLIST_REMOVE(&dap->da_list); 2528 LIST_REMOVE(dap, da_pdlist); 2529 if ((dap->da_state & DIRCHG) == 0) { 2530 pagedep = dap->da_pagedep; 2531 } else { 2532 dirrem = dap->da_previous; 2533 pagedep = dirrem->dm_pagedep; 2534 dirrem->dm_dirinum = pagedep->pd_ino; 2535 add_to_worklist(&dirrem->dm_list); 2536 } 2537 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum, 2538 0, &inodedep) != 0) 2539 (void) free_inodedep(inodedep); 2540 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2541 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 2542 nextmd = LIST_NEXT(mkdir, md_mkdirs); 2543 if (mkdir->md_diradd != dap) 2544 continue; 2545 dap->da_state &= ~mkdir->md_state; 2546 WORKLIST_REMOVE(&mkdir->md_list); 2547 LIST_REMOVE(mkdir, md_mkdirs); 2548 WORKITEM_FREE(mkdir, D_MKDIR); 2549 } 2550 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2551 FREE_LOCK(&lk); 2552 panic("free_diradd: unfound ref"); 2553 } 2554 } 2555 WORKITEM_FREE(dap, D_DIRADD); 2556 } 2557 2558 /* 2559 * Directory entry removal dependencies. 2560 * 2561 * When removing a directory entry, the entry's inode pointer must be 2562 * zero'ed on disk before the corresponding inode's link count is decremented 2563 * (possibly freeing the inode for re-use). This dependency is handled by 2564 * updating the directory entry but delaying the inode count reduction until 2565 * after the directory block has been written to disk. After this point, the 2566 * inode count can be decremented whenever it is convenient. 2567 */ 2568 2569 /* 2570 * This routine should be called immediately after removing 2571 * a directory entry. The inode's link count should not be 2572 * decremented by the calling procedure -- the soft updates 2573 * code will do this task when it is safe. 2574 */ 2575 void 2576 softdep_setup_remove(bp, dp, ip, isrmdir) 2577 struct buf *bp; /* buffer containing directory block */ 2578 struct inode *dp; /* inode for the directory being modified */ 2579 struct inode *ip; /* inode for directory entry being removed */ 2580 int isrmdir; /* indicates if doing RMDIR */ 2581 { 2582 struct dirrem *dirrem, *prevdirrem; 2583 2584 /* 2585 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 2586 */ 2587 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2588 2589 /* 2590 * If the COMPLETE flag is clear, then there were no active 2591 * entries and we want to roll back to a zeroed entry until 2592 * the new inode is committed to disk. If the COMPLETE flag is 2593 * set then we have deleted an entry that never made it to 2594 * disk. If the entry we deleted resulted from a name change, 2595 * then the old name still resides on disk. We cannot delete 2596 * its inode (returned to us in prevdirrem) until the zeroed 2597 * directory entry gets to disk. The new inode has never been 2598 * referenced on the disk, so can be deleted immediately. 2599 */ 2600 if ((dirrem->dm_state & COMPLETE) == 0) { 2601 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 2602 dm_next); 2603 FREE_LOCK(&lk); 2604 } else { 2605 if (prevdirrem != NULL) 2606 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 2607 prevdirrem, dm_next); 2608 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 2609 FREE_LOCK(&lk); 2610 handle_workitem_remove(dirrem); 2611 } 2612 } 2613 2614 /* 2615 * Allocate a new dirrem if appropriate and return it along with 2616 * its associated pagedep. Called without a lock, returns with lock. 2617 */ 2618 static long num_dirrem; /* number of dirrem allocated */ 2619 static struct dirrem * 2620 newdirrem(bp, dp, ip, isrmdir, prevdirremp) 2621 struct buf *bp; /* buffer containing directory block */ 2622 struct inode *dp; /* inode for the directory being modified */ 2623 struct inode *ip; /* inode for directory entry being removed */ 2624 int isrmdir; /* indicates if doing RMDIR */ 2625 struct dirrem **prevdirremp; /* previously referenced inode, if any */ 2626 { 2627 int offset; 2628 ufs_lbn_t lbn; 2629 struct diradd *dap; 2630 struct dirrem *dirrem; 2631 struct pagedep *pagedep; 2632 2633 /* 2634 * Whiteouts have no deletion dependencies. 2635 */ 2636 if (ip == NULL) 2637 panic("newdirrem: whiteout"); 2638 /* 2639 * If we are over our limit, try to improve the situation. 2640 * Limiting the number of dirrem structures will also limit 2641 * the number of freefile and freeblks structures. 2642 */ 2643 if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0) 2644 (void) request_cleanup(FLUSH_REMOVE, 0); 2645 num_dirrem += 1; 2646 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem), 2647 M_DIRREM, M_SOFTDEP_FLAGS); 2648 bzero(dirrem, sizeof(struct dirrem)); 2649 dirrem->dm_list.wk_type = D_DIRREM; 2650 dirrem->dm_state = isrmdir ? RMDIR : 0; 2651 dirrem->dm_mnt = ITOV(ip)->v_mount; 2652 dirrem->dm_oldinum = ip->i_number; 2653 *prevdirremp = NULL; 2654 2655 ACQUIRE_LOCK(&lk); 2656 lbn = lblkno(dp->i_fs, dp->i_offset); 2657 offset = blkoff(dp->i_fs, dp->i_offset); 2658 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2659 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2660 dirrem->dm_pagedep = pagedep; 2661 /* 2662 * Check for a diradd dependency for the same directory entry. 2663 * If present, then both dependencies become obsolete and can 2664 * be de-allocated. Check for an entry on both the pd_dirraddhd 2665 * list and the pd_pendinghd list. 2666 */ 2667 2668 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 2669 if (dap->da_offset == offset) 2670 break; 2671 if (dap == NULL) { 2672 2673 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 2674 if (dap->da_offset == offset) 2675 break; 2676 if (dap == NULL) 2677 return (dirrem); 2678 } 2679 /* 2680 * Must be ATTACHED at this point. 2681 */ 2682 if ((dap->da_state & ATTACHED) == 0) { 2683 FREE_LOCK(&lk); 2684 panic("newdirrem: not ATTACHED"); 2685 } 2686 if (dap->da_newinum != ip->i_number) { 2687 FREE_LOCK(&lk); 2688 panic("newdirrem: inum %"PRId64" should be %"PRId64, 2689 ip->i_number, dap->da_newinum); 2690 } 2691 /* 2692 * If we are deleting a changed name that never made it to disk, 2693 * then return the dirrem describing the previous inode (which 2694 * represents the inode currently referenced from this entry on disk). 2695 */ 2696 if ((dap->da_state & DIRCHG) != 0) { 2697 *prevdirremp = dap->da_previous; 2698 dap->da_state &= ~DIRCHG; 2699 dap->da_pagedep = pagedep; 2700 } 2701 /* 2702 * We are deleting an entry that never made it to disk. 2703 * Mark it COMPLETE so we can delete its inode immediately. 2704 */ 2705 dirrem->dm_state |= COMPLETE; 2706 free_diradd(dap); 2707 return (dirrem); 2708 } 2709 2710 /* 2711 * Directory entry change dependencies. 2712 * 2713 * Changing an existing directory entry requires that an add operation 2714 * be completed first followed by a deletion. The semantics for the addition 2715 * are identical to the description of adding a new entry above except 2716 * that the rollback is to the old inode number rather than zero. Once 2717 * the addition dependency is completed, the removal is done as described 2718 * in the removal routine above. 2719 */ 2720 2721 /* 2722 * This routine should be called immediately after changing 2723 * a directory entry. The inode's link count should not be 2724 * decremented by the calling procedure -- the soft updates 2725 * code will perform this task when it is safe. 2726 */ 2727 void 2728 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 2729 struct buf *bp; /* buffer containing directory block */ 2730 struct inode *dp; /* inode for the directory being modified */ 2731 struct inode *ip; /* inode for directory entry being removed */ 2732 ino_t newinum; /* new inode number for changed entry */ 2733 int isrmdir; /* indicates if doing RMDIR */ 2734 { 2735 int offset; 2736 struct diradd *dap = NULL; 2737 struct dirrem *dirrem, *prevdirrem; 2738 struct pagedep *pagedep; 2739 struct inodedep *inodedep; 2740 2741 offset = blkoff(dp->i_fs, dp->i_offset); 2742 2743 /* 2744 * Whiteouts do not need diradd dependencies. 2745 */ 2746 if (newinum != WINO) { 2747 MALLOC(dap, struct diradd *, sizeof(struct diradd), 2748 M_DIRADD, M_SOFTDEP_FLAGS); 2749 bzero(dap, sizeof(struct diradd)); 2750 dap->da_list.wk_type = D_DIRADD; 2751 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 2752 dap->da_offset = offset; 2753 dap->da_newinum = newinum; 2754 } 2755 2756 /* 2757 * Allocate a new dirrem and ACQUIRE_LOCK. 2758 */ 2759 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2760 pagedep = dirrem->dm_pagedep; 2761 /* 2762 * The possible values for isrmdir: 2763 * 0 - non-directory file rename 2764 * 1 - directory rename within same directory 2765 * inum - directory rename to new directory of given inode number 2766 * When renaming to a new directory, we are both deleting and 2767 * creating a new directory entry, so the link count on the new 2768 * directory should not change. Thus we do not need the followup 2769 * dirrem which is usually done in handle_workitem_remove. We set 2770 * the DIRCHG flag to tell handle_workitem_remove to skip the 2771 * followup dirrem. 2772 */ 2773 if (isrmdir > 1) 2774 dirrem->dm_state |= DIRCHG; 2775 2776 /* 2777 * Whiteouts have no additional dependencies, 2778 * so just put the dirrem on the correct list. 2779 */ 2780 if (newinum == WINO) { 2781 if ((dirrem->dm_state & COMPLETE) == 0) { 2782 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 2783 dm_next); 2784 } else { 2785 dirrem->dm_dirinum = pagedep->pd_ino; 2786 add_to_worklist(&dirrem->dm_list); 2787 } 2788 FREE_LOCK(&lk); 2789 return; 2790 } 2791 2792 /* 2793 * If the COMPLETE flag is clear, then there were no active 2794 * entries and we want to roll back to the previous inode until 2795 * the new inode is committed to disk. If the COMPLETE flag is 2796 * set, then we have deleted an entry that never made it to disk. 2797 * If the entry we deleted resulted from a name change, then the old 2798 * inode reference still resides on disk. Any rollback that we do 2799 * needs to be to that old inode (returned to us in prevdirrem). If 2800 * the entry we deleted resulted from a create, then there is 2801 * no entry on the disk, so we want to roll back to zero rather 2802 * than the uncommitted inode. In either of the COMPLETE cases we 2803 * want to immediately free the unwritten and unreferenced inode. 2804 */ 2805 if ((dirrem->dm_state & COMPLETE) == 0) { 2806 dap->da_previous = dirrem; 2807 } else { 2808 if (prevdirrem != NULL) { 2809 dap->da_previous = prevdirrem; 2810 } else { 2811 dap->da_state &= ~DIRCHG; 2812 dap->da_pagedep = pagedep; 2813 } 2814 dirrem->dm_dirinum = pagedep->pd_ino; 2815 add_to_worklist(&dirrem->dm_list); 2816 } 2817 /* 2818 * Link into its inodedep. Put it on the id_bufwait list if the inode 2819 * is not yet written. If it is written, do the post-inode write 2820 * processing to put it on the id_pendinghd list. 2821 */ 2822 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 || 2823 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2824 dap->da_state |= COMPLETE; 2825 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 2826 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 2827 } else { 2828 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 2829 dap, da_pdlist); 2830 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2831 } 2832 FREE_LOCK(&lk); 2833 } 2834 2835 /* 2836 * Called whenever the link count on an inode is changed. 2837 * It creates an inode dependency so that the new reference(s) 2838 * to the inode cannot be committed to disk until the updated 2839 * inode has been written. 2840 */ 2841 void 2842 softdep_change_linkcnt(ip) 2843 struct inode *ip; /* the inode with the increased link count */ 2844 { 2845 struct inodedep *inodedep; 2846 2847 ACQUIRE_LOCK(&lk); 2848 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 2849 if (ip->i_nlink < ip->i_effnlink) { 2850 FREE_LOCK(&lk); 2851 panic("softdep_change_linkcnt: bad delta"); 2852 } 2853 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2854 FREE_LOCK(&lk); 2855 } 2856 2857 /* 2858 * This workitem decrements the inode's link count. 2859 * If the link count reaches zero, the file is removed. 2860 */ 2861 static void 2862 handle_workitem_remove(dirrem) 2863 struct dirrem *dirrem; 2864 { 2865 struct thread *td = curthread; /* XXX */ 2866 struct inodedep *inodedep; 2867 struct vnode *vp; 2868 struct inode *ip; 2869 ino_t oldinum; 2870 int error; 2871 2872 if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) { 2873 softdep_error("handle_workitem_remove: vget", error); 2874 return; 2875 } 2876 ip = VTOI(vp); 2877 ACQUIRE_LOCK(&lk); 2878 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){ 2879 FREE_LOCK(&lk); 2880 panic("handle_workitem_remove: lost inodedep"); 2881 } 2882 /* 2883 * Normal file deletion. 2884 */ 2885 if ((dirrem->dm_state & RMDIR) == 0) { 2886 ip->i_nlink--; 2887 ip->i_flag |= IN_CHANGE; 2888 if (ip->i_nlink < ip->i_effnlink) { 2889 FREE_LOCK(&lk); 2890 panic("handle_workitem_remove: bad file delta"); 2891 } 2892 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2893 FREE_LOCK(&lk); 2894 vput(vp); 2895 num_dirrem -= 1; 2896 WORKITEM_FREE(dirrem, D_DIRREM); 2897 return; 2898 } 2899 /* 2900 * Directory deletion. Decrement reference count for both the 2901 * just deleted parent directory entry and the reference for ".". 2902 * Next truncate the directory to length zero. When the 2903 * truncation completes, arrange to have the reference count on 2904 * the parent decremented to account for the loss of "..". 2905 */ 2906 ip->i_nlink -= 2; 2907 ip->i_flag |= IN_CHANGE; 2908 if (ip->i_nlink < ip->i_effnlink) { 2909 FREE_LOCK(&lk); 2910 panic("handle_workitem_remove: bad dir delta"); 2911 } 2912 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2913 FREE_LOCK(&lk); 2914 if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, proc0.p_ucred, td)) != 0) 2915 softdep_error("handle_workitem_remove: truncate", error); 2916 /* 2917 * Rename a directory to a new parent. Since, we are both deleting 2918 * and creating a new directory entry, the link count on the new 2919 * directory should not change. Thus we skip the followup dirrem. 2920 */ 2921 if (dirrem->dm_state & DIRCHG) { 2922 vput(vp); 2923 num_dirrem -= 1; 2924 WORKITEM_FREE(dirrem, D_DIRREM); 2925 return; 2926 } 2927 /* 2928 * If the inodedep does not exist, then the zero'ed inode has 2929 * been written to disk. If the allocated inode has never been 2930 * written to disk, then the on-disk inode is zero'ed. In either 2931 * case we can remove the file immediately. 2932 */ 2933 ACQUIRE_LOCK(&lk); 2934 dirrem->dm_state = 0; 2935 oldinum = dirrem->dm_oldinum; 2936 dirrem->dm_oldinum = dirrem->dm_dirinum; 2937 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 || 2938 check_inode_unwritten(inodedep)) { 2939 FREE_LOCK(&lk); 2940 vput(vp); 2941 handle_workitem_remove(dirrem); 2942 return; 2943 } 2944 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 2945 FREE_LOCK(&lk); 2946 ip->i_flag |= IN_CHANGE; 2947 ffs_update(vp, 0); 2948 vput(vp); 2949 } 2950 2951 /* 2952 * Inode de-allocation dependencies. 2953 * 2954 * When an inode's link count is reduced to zero, it can be de-allocated. We 2955 * found it convenient to postpone de-allocation until after the inode is 2956 * written to disk with its new link count (zero). At this point, all of the 2957 * on-disk inode's block pointers are nullified and, with careful dependency 2958 * list ordering, all dependencies related to the inode will be satisfied and 2959 * the corresponding dependency structures de-allocated. So, if/when the 2960 * inode is reused, there will be no mixing of old dependencies with new 2961 * ones. This artificial dependency is set up by the block de-allocation 2962 * procedure above (softdep_setup_freeblocks) and completed by the 2963 * following procedure. 2964 */ 2965 static void 2966 handle_workitem_freefile(freefile) 2967 struct freefile *freefile; 2968 { 2969 struct vnode vp; 2970 struct inode tip; 2971 struct inodedep *idp; 2972 int error; 2973 2974 #ifdef DEBUG 2975 ACQUIRE_LOCK(&lk); 2976 error = inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp); 2977 FREE_LOCK(&lk); 2978 if (error) 2979 panic("handle_workitem_freefile: inodedep survived"); 2980 #endif 2981 tip.i_devvp = freefile->fx_devvp; 2982 tip.i_dev = freefile->fx_devvp->v_rdev; 2983 tip.i_fs = freefile->fx_fs; 2984 vp.v_data = &tip; 2985 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0) 2986 softdep_error("handle_workitem_freefile", error); 2987 WORKITEM_FREE(freefile, D_FREEFILE); 2988 } 2989 2990 /* 2991 * Helper function which unlinks marker element from work list and returns 2992 * the next element on the list. 2993 */ 2994 static __inline struct worklist * 2995 markernext(struct worklist *marker) 2996 { 2997 struct worklist *next; 2998 2999 next = LIST_NEXT(marker, wk_list); 3000 LIST_REMOVE(marker, wk_list); 3001 return next; 3002 } 3003 3004 /* 3005 * Disk writes. 3006 * 3007 * The dependency structures constructed above are most actively used when file 3008 * system blocks are written to disk. No constraints are placed on when a 3009 * block can be written, but unsatisfied update dependencies are made safe by 3010 * modifying (or replacing) the source memory for the duration of the disk 3011 * write. When the disk write completes, the memory block is again brought 3012 * up-to-date. 3013 * 3014 * In-core inode structure reclamation. 3015 * 3016 * Because there are a finite number of "in-core" inode structures, they are 3017 * reused regularly. By transferring all inode-related dependencies to the 3018 * in-memory inode block and indexing them separately (via "inodedep"s), we 3019 * can allow "in-core" inode structures to be reused at any time and avoid 3020 * any increase in contention. 3021 * 3022 * Called just before entering the device driver to initiate a new disk I/O. 3023 * The buffer must be locked, thus, no I/O completion operations can occur 3024 * while we are manipulating its associated dependencies. 3025 */ 3026 static void 3027 softdep_disk_io_initiation(bp) 3028 struct buf *bp; /* structure describing disk write to occur */ 3029 { 3030 struct worklist *wk; 3031 struct worklist marker; 3032 struct indirdep *indirdep; 3033 3034 /* 3035 * We only care about write operations. There should never 3036 * be dependencies for reads. 3037 */ 3038 if (bp->b_flags & B_READ) 3039 panic("softdep_disk_io_initiation: read"); 3040 3041 marker.wk_type = D_LAST + 1; /* Not a normal workitem */ 3042 3043 /* 3044 * Do any necessary pre-I/O processing. 3045 */ 3046 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = markernext(&marker)) { 3047 LIST_INSERT_AFTER(wk, &marker, wk_list); 3048 3049 switch (wk->wk_type) { 3050 3051 case D_PAGEDEP: 3052 initiate_write_filepage(WK_PAGEDEP(wk), bp); 3053 continue; 3054 3055 case D_INODEDEP: 3056 initiate_write_inodeblock(WK_INODEDEP(wk), bp); 3057 continue; 3058 3059 case D_INDIRDEP: 3060 indirdep = WK_INDIRDEP(wk); 3061 if (indirdep->ir_state & GOINGAWAY) 3062 panic("disk_io_initiation: indirdep gone"); 3063 /* 3064 * If there are no remaining dependencies, this 3065 * will be writing the real pointers, so the 3066 * dependency can be freed. 3067 */ 3068 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) { 3069 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 3070 brelse(indirdep->ir_savebp); 3071 /* inline expand WORKLIST_REMOVE(wk); */ 3072 wk->wk_state &= ~ONWORKLIST; 3073 LIST_REMOVE(wk, wk_list); 3074 WORKITEM_FREE(indirdep, D_INDIRDEP); 3075 continue; 3076 } 3077 /* 3078 * Replace up-to-date version with safe version. 3079 */ 3080 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount, 3081 M_INDIRDEP, M_SOFTDEP_FLAGS); 3082 ACQUIRE_LOCK(&lk); 3083 indirdep->ir_state &= ~ATTACHED; 3084 indirdep->ir_state |= UNDONE; 3085 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 3086 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 3087 bp->b_bcount); 3088 FREE_LOCK(&lk); 3089 continue; 3090 3091 case D_MKDIR: 3092 case D_BMSAFEMAP: 3093 case D_ALLOCDIRECT: 3094 case D_ALLOCINDIR: 3095 continue; 3096 3097 default: 3098 panic("handle_disk_io_initiation: Unexpected type %s", 3099 TYPENAME(wk->wk_type)); 3100 /* NOTREACHED */ 3101 } 3102 } 3103 } 3104 3105 /* 3106 * Called from within the procedure above to deal with unsatisfied 3107 * allocation dependencies in a directory. The buffer must be locked, 3108 * thus, no I/O completion operations can occur while we are 3109 * manipulating its associated dependencies. 3110 */ 3111 static void 3112 initiate_write_filepage(pagedep, bp) 3113 struct pagedep *pagedep; 3114 struct buf *bp; 3115 { 3116 struct diradd *dap; 3117 struct direct *ep; 3118 int i; 3119 3120 if (pagedep->pd_state & IOSTARTED) { 3121 /* 3122 * This can only happen if there is a driver that does not 3123 * understand chaining. Here biodone will reissue the call 3124 * to strategy for the incomplete buffers. 3125 */ 3126 printf("initiate_write_filepage: already started\n"); 3127 return; 3128 } 3129 pagedep->pd_state |= IOSTARTED; 3130 ACQUIRE_LOCK(&lk); 3131 for (i = 0; i < DAHASHSZ; i++) { 3132 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 3133 ep = (struct direct *) 3134 ((char *)bp->b_data + dap->da_offset); 3135 if (ep->d_ino != dap->da_newinum) { 3136 FREE_LOCK(&lk); 3137 panic("%s: dir inum %d != new %"PRId64, 3138 "initiate_write_filepage", 3139 ep->d_ino, dap->da_newinum); 3140 } 3141 if (dap->da_state & DIRCHG) 3142 ep->d_ino = dap->da_previous->dm_oldinum; 3143 else 3144 ep->d_ino = 0; 3145 dap->da_state &= ~ATTACHED; 3146 dap->da_state |= UNDONE; 3147 } 3148 } 3149 FREE_LOCK(&lk); 3150 } 3151 3152 /* 3153 * Called from within the procedure above to deal with unsatisfied 3154 * allocation dependencies in an inodeblock. The buffer must be 3155 * locked, thus, no I/O completion operations can occur while we 3156 * are manipulating its associated dependencies. 3157 */ 3158 static void 3159 initiate_write_inodeblock(inodedep, bp) 3160 struct inodedep *inodedep; 3161 struct buf *bp; /* The inode block */ 3162 { 3163 struct allocdirect *adp, *lastadp; 3164 struct dinode *dp; 3165 struct dinode *sip; 3166 struct fs *fs; 3167 ufs_lbn_t prevlbn = 0; 3168 int i, deplist; 3169 3170 if (inodedep->id_state & IOSTARTED) 3171 panic("initiate_write_inodeblock: already started"); 3172 inodedep->id_state |= IOSTARTED; 3173 fs = inodedep->id_fs; 3174 dp = (struct dinode *)bp->b_data + 3175 ino_to_fsbo(fs, inodedep->id_ino); 3176 /* 3177 * If the bitmap is not yet written, then the allocated 3178 * inode cannot be written to disk. 3179 */ 3180 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 3181 if (inodedep->id_savedino != NULL) 3182 panic("initiate_write_inodeblock: already doing I/O"); 3183 MALLOC(sip, struct dinode *, 3184 sizeof(struct dinode), M_INODEDEP, M_SOFTDEP_FLAGS); 3185 inodedep->id_savedino = sip; 3186 *inodedep->id_savedino = *dp; 3187 bzero((caddr_t)dp, sizeof(struct dinode)); 3188 dp->di_gen = inodedep->id_savedino->di_gen; 3189 return; 3190 } 3191 /* 3192 * If no dependencies, then there is nothing to roll back. 3193 */ 3194 inodedep->id_savedsize = dp->di_size; 3195 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 3196 return; 3197 /* 3198 * Set the dependencies to busy. 3199 */ 3200 ACQUIRE_LOCK(&lk); 3201 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3202 adp = TAILQ_NEXT(adp, ad_next)) { 3203 #ifdef DIAGNOSTIC 3204 if (deplist != 0 && prevlbn >= adp->ad_lbn) { 3205 FREE_LOCK(&lk); 3206 panic("softdep_write_inodeblock: lbn order"); 3207 } 3208 prevlbn = adp->ad_lbn; 3209 if (adp->ad_lbn < NDADDR && 3210 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) { 3211 FREE_LOCK(&lk); 3212 panic("%s: direct pointer #%ld mismatch %d != %d", 3213 "softdep_write_inodeblock", adp->ad_lbn, 3214 dp->di_db[adp->ad_lbn], adp->ad_newblkno); 3215 } 3216 if (adp->ad_lbn >= NDADDR && 3217 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) { 3218 FREE_LOCK(&lk); 3219 panic("%s: indirect pointer #%ld mismatch %d != %d", 3220 "softdep_write_inodeblock", adp->ad_lbn - NDADDR, 3221 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno); 3222 } 3223 deplist |= 1 << adp->ad_lbn; 3224 if ((adp->ad_state & ATTACHED) == 0) { 3225 FREE_LOCK(&lk); 3226 panic("softdep_write_inodeblock: Unknown state 0x%x", 3227 adp->ad_state); 3228 } 3229 #endif /* DIAGNOSTIC */ 3230 adp->ad_state &= ~ATTACHED; 3231 adp->ad_state |= UNDONE; 3232 } 3233 /* 3234 * The on-disk inode cannot claim to be any larger than the last 3235 * fragment that has been written. Otherwise, the on-disk inode 3236 * might have fragments that were not the last block in the file 3237 * which would corrupt the filesystem. 3238 */ 3239 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3240 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 3241 if (adp->ad_lbn >= NDADDR) 3242 break; 3243 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 3244 /* keep going until hitting a rollback to a frag */ 3245 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 3246 continue; 3247 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 3248 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 3249 #ifdef DIAGNOSTIC 3250 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) { 3251 FREE_LOCK(&lk); 3252 panic("softdep_write_inodeblock: lost dep1"); 3253 } 3254 #endif /* DIAGNOSTIC */ 3255 dp->di_db[i] = 0; 3256 } 3257 for (i = 0; i < NIADDR; i++) { 3258 #ifdef DIAGNOSTIC 3259 if (dp->di_ib[i] != 0 && 3260 (deplist & ((1 << NDADDR) << i)) == 0) { 3261 FREE_LOCK(&lk); 3262 panic("softdep_write_inodeblock: lost dep2"); 3263 } 3264 #endif /* DIAGNOSTIC */ 3265 dp->di_ib[i] = 0; 3266 } 3267 FREE_LOCK(&lk); 3268 return; 3269 } 3270 /* 3271 * If we have zero'ed out the last allocated block of the file, 3272 * roll back the size to the last currently allocated block. 3273 * We know that this last allocated block is a full-sized as 3274 * we already checked for fragments in the loop above. 3275 */ 3276 if (lastadp != NULL && 3277 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 3278 for (i = lastadp->ad_lbn; i >= 0; i--) 3279 if (dp->di_db[i] != 0) 3280 break; 3281 dp->di_size = (i + 1) * fs->fs_bsize; 3282 } 3283 /* 3284 * The only dependencies are for indirect blocks. 3285 * 3286 * The file size for indirect block additions is not guaranteed. 3287 * Such a guarantee would be non-trivial to achieve. The conventional 3288 * synchronous write implementation also does not make this guarantee. 3289 * Fsck should catch and fix discrepancies. Arguably, the file size 3290 * can be over-estimated without destroying integrity when the file 3291 * moves into the indirect blocks (i.e., is large). If we want to 3292 * postpone fsck, we are stuck with this argument. 3293 */ 3294 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 3295 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 3296 FREE_LOCK(&lk); 3297 } 3298 3299 /* 3300 * This routine is called during the completion interrupt 3301 * service routine for a disk write (from the procedure called 3302 * by the device driver to inform the filesystem caches of 3303 * a request completion). It should be called early in this 3304 * procedure, before the block is made available to other 3305 * processes or other routines are called. 3306 */ 3307 static void 3308 softdep_disk_write_complete(bp) 3309 struct buf *bp; /* describes the completed disk write */ 3310 { 3311 struct worklist *wk; 3312 struct workhead reattach; 3313 struct newblk *newblk; 3314 struct allocindir *aip; 3315 struct allocdirect *adp; 3316 struct indirdep *indirdep; 3317 struct inodedep *inodedep; 3318 struct bmsafemap *bmsafemap; 3319 3320 #ifdef DEBUG 3321 if (lk.lkt_held != NOHOLDER) 3322 panic("softdep_disk_write_complete: lock is held"); 3323 lk.lkt_held = SPECIAL_FLAG; 3324 #endif 3325 LIST_INIT(&reattach); 3326 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 3327 WORKLIST_REMOVE(wk); 3328 switch (wk->wk_type) { 3329 3330 case D_PAGEDEP: 3331 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 3332 WORKLIST_INSERT(&reattach, wk); 3333 continue; 3334 3335 case D_INODEDEP: 3336 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 3337 WORKLIST_INSERT(&reattach, wk); 3338 continue; 3339 3340 case D_BMSAFEMAP: 3341 bmsafemap = WK_BMSAFEMAP(wk); 3342 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) { 3343 newblk->nb_state |= DEPCOMPLETE; 3344 newblk->nb_bmsafemap = NULL; 3345 LIST_REMOVE(newblk, nb_deps); 3346 } 3347 while ((adp = 3348 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) { 3349 adp->ad_state |= DEPCOMPLETE; 3350 adp->ad_buf = NULL; 3351 LIST_REMOVE(adp, ad_deps); 3352 handle_allocdirect_partdone(adp); 3353 } 3354 while ((aip = 3355 LIST_FIRST(&bmsafemap->sm_allocindirhd))) { 3356 aip->ai_state |= DEPCOMPLETE; 3357 aip->ai_buf = NULL; 3358 LIST_REMOVE(aip, ai_deps); 3359 handle_allocindir_partdone(aip); 3360 } 3361 while ((inodedep = 3362 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { 3363 inodedep->id_state |= DEPCOMPLETE; 3364 LIST_REMOVE(inodedep, id_deps); 3365 inodedep->id_buf = NULL; 3366 } 3367 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 3368 continue; 3369 3370 case D_MKDIR: 3371 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 3372 continue; 3373 3374 case D_ALLOCDIRECT: 3375 adp = WK_ALLOCDIRECT(wk); 3376 adp->ad_state |= COMPLETE; 3377 handle_allocdirect_partdone(adp); 3378 continue; 3379 3380 case D_ALLOCINDIR: 3381 aip = WK_ALLOCINDIR(wk); 3382 aip->ai_state |= COMPLETE; 3383 handle_allocindir_partdone(aip); 3384 continue; 3385 3386 case D_INDIRDEP: 3387 indirdep = WK_INDIRDEP(wk); 3388 if (indirdep->ir_state & GOINGAWAY) { 3389 lk.lkt_held = NOHOLDER; 3390 panic("disk_write_complete: indirdep gone"); 3391 } 3392 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 3393 FREE(indirdep->ir_saveddata, M_INDIRDEP); 3394 indirdep->ir_saveddata = 0; 3395 indirdep->ir_state &= ~UNDONE; 3396 indirdep->ir_state |= ATTACHED; 3397 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 3398 handle_allocindir_partdone(aip); 3399 if (aip == LIST_FIRST(&indirdep->ir_donehd)) { 3400 lk.lkt_held = NOHOLDER; 3401 panic("disk_write_complete: not gone"); 3402 } 3403 } 3404 WORKLIST_INSERT(&reattach, wk); 3405 if ((bp->b_flags & B_DELWRI) == 0) 3406 stat_indir_blk_ptrs++; 3407 bdirty(bp); 3408 continue; 3409 3410 default: 3411 lk.lkt_held = NOHOLDER; 3412 panic("handle_disk_write_complete: Unknown type %s", 3413 TYPENAME(wk->wk_type)); 3414 /* NOTREACHED */ 3415 } 3416 } 3417 /* 3418 * Reattach any requests that must be redone. 3419 */ 3420 while ((wk = LIST_FIRST(&reattach)) != NULL) { 3421 WORKLIST_REMOVE(wk); 3422 WORKLIST_INSERT(&bp->b_dep, wk); 3423 } 3424 #ifdef DEBUG 3425 if (lk.lkt_held != SPECIAL_FLAG) 3426 panic("softdep_disk_write_complete: lock lost"); 3427 lk.lkt_held = NOHOLDER; 3428 #endif 3429 } 3430 3431 /* 3432 * Called from within softdep_disk_write_complete above. Note that 3433 * this routine is always called from interrupt level with further 3434 * splbio interrupts blocked. 3435 */ 3436 static void 3437 handle_allocdirect_partdone(adp) 3438 struct allocdirect *adp; /* the completed allocdirect */ 3439 { 3440 struct allocdirect *listadp; 3441 struct inodedep *inodedep; 3442 long bsize; 3443 3444 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3445 return; 3446 if (adp->ad_buf != NULL) { 3447 lk.lkt_held = NOHOLDER; 3448 panic("handle_allocdirect_partdone: dangling dep"); 3449 } 3450 /* 3451 * The on-disk inode cannot claim to be any larger than the last 3452 * fragment that has been written. Otherwise, the on-disk inode 3453 * might have fragments that were not the last block in the file 3454 * which would corrupt the filesystem. Thus, we cannot free any 3455 * allocdirects after one whose ad_oldblkno claims a fragment as 3456 * these blocks must be rolled back to zero before writing the inode. 3457 * We check the currently active set of allocdirects in id_inoupdt. 3458 */ 3459 inodedep = adp->ad_inodedep; 3460 bsize = inodedep->id_fs->fs_bsize; 3461 TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) { 3462 /* found our block */ 3463 if (listadp == adp) 3464 break; 3465 /* continue if ad_oldlbn is not a fragment */ 3466 if (listadp->ad_oldsize == 0 || 3467 listadp->ad_oldsize == bsize) 3468 continue; 3469 /* hit a fragment */ 3470 return; 3471 } 3472 /* 3473 * If we have reached the end of the current list without 3474 * finding the just finished dependency, then it must be 3475 * on the future dependency list. Future dependencies cannot 3476 * be freed until they are moved to the current list. 3477 */ 3478 if (listadp == NULL) { 3479 #ifdef DEBUG 3480 TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next) 3481 /* found our block */ 3482 if (listadp == adp) 3483 break; 3484 if (listadp == NULL) { 3485 lk.lkt_held = NOHOLDER; 3486 panic("handle_allocdirect_partdone: lost dep"); 3487 } 3488 #endif /* DEBUG */ 3489 return; 3490 } 3491 /* 3492 * If we have found the just finished dependency, then free 3493 * it along with anything that follows it that is complete. 3494 */ 3495 for (; adp; adp = listadp) { 3496 listadp = TAILQ_NEXT(adp, ad_next); 3497 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3498 return; 3499 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 3500 } 3501 } 3502 3503 /* 3504 * Called from within softdep_disk_write_complete above. Note that 3505 * this routine is always called from interrupt level with further 3506 * splbio interrupts blocked. 3507 */ 3508 static void 3509 handle_allocindir_partdone(aip) 3510 struct allocindir *aip; /* the completed allocindir */ 3511 { 3512 struct indirdep *indirdep; 3513 3514 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 3515 return; 3516 if (aip->ai_buf != NULL) { 3517 lk.lkt_held = NOHOLDER; 3518 panic("handle_allocindir_partdone: dangling dependency"); 3519 } 3520 indirdep = aip->ai_indirdep; 3521 if (indirdep->ir_state & UNDONE) { 3522 LIST_REMOVE(aip, ai_next); 3523 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 3524 return; 3525 } 3526 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 3527 aip->ai_newblkno; 3528 LIST_REMOVE(aip, ai_next); 3529 if (aip->ai_freefrag != NULL) 3530 add_to_worklist(&aip->ai_freefrag->ff_list); 3531 WORKITEM_FREE(aip, D_ALLOCINDIR); 3532 } 3533 3534 /* 3535 * Called from within softdep_disk_write_complete above to restore 3536 * in-memory inode block contents to their most up-to-date state. Note 3537 * that this routine is always called from interrupt level with further 3538 * splbio interrupts blocked. 3539 */ 3540 static int 3541 handle_written_inodeblock(inodedep, bp) 3542 struct inodedep *inodedep; 3543 struct buf *bp; /* buffer containing the inode block */ 3544 { 3545 struct worklist *wk, *filefree; 3546 struct allocdirect *adp, *nextadp; 3547 struct dinode *dp; 3548 int hadchanges; 3549 3550 if ((inodedep->id_state & IOSTARTED) == 0) { 3551 lk.lkt_held = NOHOLDER; 3552 panic("handle_written_inodeblock: not started"); 3553 } 3554 inodedep->id_state &= ~IOSTARTED; 3555 dp = (struct dinode *)bp->b_data + 3556 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 3557 /* 3558 * If we had to rollback the inode allocation because of 3559 * bitmaps being incomplete, then simply restore it. 3560 * Keep the block dirty so that it will not be reclaimed until 3561 * all associated dependencies have been cleared and the 3562 * corresponding updates written to disk. 3563 */ 3564 if (inodedep->id_savedino != NULL) { 3565 *dp = *inodedep->id_savedino; 3566 FREE(inodedep->id_savedino, M_INODEDEP); 3567 inodedep->id_savedino = NULL; 3568 if ((bp->b_flags & B_DELWRI) == 0) 3569 stat_inode_bitmap++; 3570 bdirty(bp); 3571 return (1); 3572 } 3573 inodedep->id_state |= COMPLETE; 3574 /* 3575 * Roll forward anything that had to be rolled back before 3576 * the inode could be updated. 3577 */ 3578 hadchanges = 0; 3579 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 3580 nextadp = TAILQ_NEXT(adp, ad_next); 3581 if (adp->ad_state & ATTACHED) { 3582 lk.lkt_held = NOHOLDER; 3583 panic("handle_written_inodeblock: new entry"); 3584 } 3585 if (adp->ad_lbn < NDADDR) { 3586 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) { 3587 lk.lkt_held = NOHOLDER; 3588 panic("%s: %s #%ld mismatch %d != %d", 3589 "handle_written_inodeblock", 3590 "direct pointer", adp->ad_lbn, 3591 dp->di_db[adp->ad_lbn], adp->ad_oldblkno); 3592 } 3593 dp->di_db[adp->ad_lbn] = adp->ad_newblkno; 3594 } else { 3595 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) { 3596 lk.lkt_held = NOHOLDER; 3597 panic("%s: %s #%ld allocated as %d", 3598 "handle_written_inodeblock", 3599 "indirect pointer", adp->ad_lbn - NDADDR, 3600 dp->di_ib[adp->ad_lbn - NDADDR]); 3601 } 3602 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno; 3603 } 3604 adp->ad_state &= ~UNDONE; 3605 adp->ad_state |= ATTACHED; 3606 hadchanges = 1; 3607 } 3608 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 3609 stat_direct_blk_ptrs++; 3610 /* 3611 * Reset the file size to its most up-to-date value. 3612 */ 3613 if (inodedep->id_savedsize == -1) { 3614 lk.lkt_held = NOHOLDER; 3615 panic("handle_written_inodeblock: bad size"); 3616 } 3617 if (dp->di_size != inodedep->id_savedsize) { 3618 dp->di_size = inodedep->id_savedsize; 3619 hadchanges = 1; 3620 } 3621 inodedep->id_savedsize = -1; 3622 /* 3623 * If there were any rollbacks in the inode block, then it must be 3624 * marked dirty so that its will eventually get written back in 3625 * its correct form. 3626 */ 3627 if (hadchanges) 3628 bdirty(bp); 3629 /* 3630 * Process any allocdirects that completed during the update. 3631 */ 3632 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 3633 handle_allocdirect_partdone(adp); 3634 /* 3635 * Process deallocations that were held pending until the 3636 * inode had been written to disk. Freeing of the inode 3637 * is delayed until after all blocks have been freed to 3638 * avoid creation of new <vfsid, inum, lbn> triples 3639 * before the old ones have been deleted. 3640 */ 3641 filefree = NULL; 3642 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 3643 WORKLIST_REMOVE(wk); 3644 switch (wk->wk_type) { 3645 3646 case D_FREEFILE: 3647 /* 3648 * We defer adding filefree to the worklist until 3649 * all other additions have been made to ensure 3650 * that it will be done after all the old blocks 3651 * have been freed. 3652 */ 3653 if (filefree != NULL) { 3654 lk.lkt_held = NOHOLDER; 3655 panic("handle_written_inodeblock: filefree"); 3656 } 3657 filefree = wk; 3658 continue; 3659 3660 case D_MKDIR: 3661 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 3662 continue; 3663 3664 case D_DIRADD: 3665 diradd_inode_written(WK_DIRADD(wk), inodedep); 3666 continue; 3667 3668 case D_FREEBLKS: 3669 wk->wk_state |= COMPLETE; 3670 if ((wk->wk_state & ALLCOMPLETE) != ALLCOMPLETE) 3671 continue; 3672 /* -- fall through -- */ 3673 case D_FREEFRAG: 3674 case D_DIRREM: 3675 add_to_worklist(wk); 3676 continue; 3677 3678 default: 3679 lk.lkt_held = NOHOLDER; 3680 panic("handle_written_inodeblock: Unknown type %s", 3681 TYPENAME(wk->wk_type)); 3682 /* NOTREACHED */ 3683 } 3684 } 3685 if (filefree != NULL) { 3686 if (free_inodedep(inodedep) == 0) { 3687 lk.lkt_held = NOHOLDER; 3688 panic("handle_written_inodeblock: live inodedep"); 3689 } 3690 add_to_worklist(filefree); 3691 return (0); 3692 } 3693 3694 /* 3695 * If no outstanding dependencies, free it. 3696 */ 3697 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0) 3698 return (0); 3699 return (hadchanges); 3700 } 3701 3702 /* 3703 * Process a diradd entry after its dependent inode has been written. 3704 * This routine must be called with splbio interrupts blocked. 3705 */ 3706 static void 3707 diradd_inode_written(dap, inodedep) 3708 struct diradd *dap; 3709 struct inodedep *inodedep; 3710 { 3711 struct pagedep *pagedep; 3712 3713 dap->da_state |= COMPLETE; 3714 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3715 if (dap->da_state & DIRCHG) 3716 pagedep = dap->da_previous->dm_pagedep; 3717 else 3718 pagedep = dap->da_pagedep; 3719 LIST_REMOVE(dap, da_pdlist); 3720 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3721 } 3722 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 3723 } 3724 3725 /* 3726 * Handle the completion of a mkdir dependency. 3727 */ 3728 static void 3729 handle_written_mkdir(mkdir, type) 3730 struct mkdir *mkdir; 3731 int type; 3732 { 3733 struct diradd *dap; 3734 struct pagedep *pagedep; 3735 3736 if (mkdir->md_state != type) { 3737 lk.lkt_held = NOHOLDER; 3738 panic("handle_written_mkdir: bad type"); 3739 } 3740 dap = mkdir->md_diradd; 3741 dap->da_state &= ~type; 3742 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 3743 dap->da_state |= DEPCOMPLETE; 3744 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3745 if (dap->da_state & DIRCHG) 3746 pagedep = dap->da_previous->dm_pagedep; 3747 else 3748 pagedep = dap->da_pagedep; 3749 LIST_REMOVE(dap, da_pdlist); 3750 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3751 } 3752 LIST_REMOVE(mkdir, md_mkdirs); 3753 WORKITEM_FREE(mkdir, D_MKDIR); 3754 } 3755 3756 /* 3757 * Called from within softdep_disk_write_complete above. 3758 * A write operation was just completed. Removed inodes can 3759 * now be freed and associated block pointers may be committed. 3760 * Note that this routine is always called from interrupt level 3761 * with further splbio interrupts blocked. 3762 */ 3763 static int 3764 handle_written_filepage(pagedep, bp) 3765 struct pagedep *pagedep; 3766 struct buf *bp; /* buffer containing the written page */ 3767 { 3768 struct dirrem *dirrem; 3769 struct diradd *dap, *nextdap; 3770 struct direct *ep; 3771 int i, chgs; 3772 3773 if ((pagedep->pd_state & IOSTARTED) == 0) { 3774 lk.lkt_held = NOHOLDER; 3775 panic("handle_written_filepage: not started"); 3776 } 3777 pagedep->pd_state &= ~IOSTARTED; 3778 /* 3779 * Process any directory removals that have been committed. 3780 */ 3781 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 3782 LIST_REMOVE(dirrem, dm_next); 3783 dirrem->dm_dirinum = pagedep->pd_ino; 3784 add_to_worklist(&dirrem->dm_list); 3785 } 3786 /* 3787 * Free any directory additions that have been committed. 3788 */ 3789 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 3790 free_diradd(dap); 3791 /* 3792 * Uncommitted directory entries must be restored. 3793 */ 3794 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 3795 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 3796 dap = nextdap) { 3797 nextdap = LIST_NEXT(dap, da_pdlist); 3798 if (dap->da_state & ATTACHED) { 3799 lk.lkt_held = NOHOLDER; 3800 panic("handle_written_filepage: attached"); 3801 } 3802 ep = (struct direct *) 3803 ((char *)bp->b_data + dap->da_offset); 3804 ep->d_ino = dap->da_newinum; 3805 dap->da_state &= ~UNDONE; 3806 dap->da_state |= ATTACHED; 3807 chgs = 1; 3808 /* 3809 * If the inode referenced by the directory has 3810 * been written out, then the dependency can be 3811 * moved to the pending list. 3812 */ 3813 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3814 LIST_REMOVE(dap, da_pdlist); 3815 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 3816 da_pdlist); 3817 } 3818 } 3819 } 3820 /* 3821 * If there were any rollbacks in the directory, then it must be 3822 * marked dirty so that its will eventually get written back in 3823 * its correct form. 3824 */ 3825 if (chgs) { 3826 if ((bp->b_flags & B_DELWRI) == 0) 3827 stat_dir_entry++; 3828 bdirty(bp); 3829 } 3830 /* 3831 * If no dependencies remain, the pagedep will be freed. 3832 * Otherwise it will remain to update the page before it 3833 * is written back to disk. 3834 */ 3835 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) { 3836 for (i = 0; i < DAHASHSZ; i++) 3837 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL) 3838 break; 3839 if (i == DAHASHSZ) { 3840 LIST_REMOVE(pagedep, pd_hash); 3841 WORKITEM_FREE(pagedep, D_PAGEDEP); 3842 return (0); 3843 } 3844 } 3845 return (1); 3846 } 3847 3848 /* 3849 * Writing back in-core inode structures. 3850 * 3851 * The filesystem only accesses an inode's contents when it occupies an 3852 * "in-core" inode structure. These "in-core" structures are separate from 3853 * the page frames used to cache inode blocks. Only the latter are 3854 * transferred to/from the disk. So, when the updated contents of the 3855 * "in-core" inode structure are copied to the corresponding in-memory inode 3856 * block, the dependencies are also transferred. The following procedure is 3857 * called when copying a dirty "in-core" inode to a cached inode block. 3858 */ 3859 3860 /* 3861 * Called when an inode is loaded from disk. If the effective link count 3862 * differed from the actual link count when it was last flushed, then we 3863 * need to ensure that the correct effective link count is put back. 3864 */ 3865 void 3866 softdep_load_inodeblock(ip) 3867 struct inode *ip; /* the "in_core" copy of the inode */ 3868 { 3869 struct inodedep *inodedep; 3870 3871 /* 3872 * Check for alternate nlink count. 3873 */ 3874 ip->i_effnlink = ip->i_nlink; 3875 ACQUIRE_LOCK(&lk); 3876 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3877 FREE_LOCK(&lk); 3878 return; 3879 } 3880 ip->i_effnlink -= inodedep->id_nlinkdelta; 3881 FREE_LOCK(&lk); 3882 } 3883 3884 /* 3885 * This routine is called just before the "in-core" inode 3886 * information is to be copied to the in-memory inode block. 3887 * Recall that an inode block contains several inodes. If 3888 * the force flag is set, then the dependencies will be 3889 * cleared so that the update can always be made. Note that 3890 * the buffer is locked when this routine is called, so we 3891 * will never be in the middle of writing the inode block 3892 * to disk. 3893 */ 3894 void 3895 softdep_update_inodeblock(ip, bp, waitfor) 3896 struct inode *ip; /* the "in_core" copy of the inode */ 3897 struct buf *bp; /* the buffer containing the inode block */ 3898 int waitfor; /* nonzero => update must be allowed */ 3899 { 3900 struct inodedep *inodedep; 3901 struct worklist *wk; 3902 int error, gotit; 3903 3904 /* 3905 * If the effective link count is not equal to the actual link 3906 * count, then we must track the difference in an inodedep while 3907 * the inode is (potentially) tossed out of the cache. Otherwise, 3908 * if there is no existing inodedep, then there are no dependencies 3909 * to track. 3910 */ 3911 ACQUIRE_LOCK(&lk); 3912 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3913 FREE_LOCK(&lk); 3914 if (ip->i_effnlink != ip->i_nlink) 3915 panic("softdep_update_inodeblock: bad link count"); 3916 return; 3917 } 3918 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) { 3919 FREE_LOCK(&lk); 3920 panic("softdep_update_inodeblock: bad delta"); 3921 } 3922 /* 3923 * Changes have been initiated. Anything depending on these 3924 * changes cannot occur until this inode has been written. 3925 */ 3926 inodedep->id_state &= ~COMPLETE; 3927 if ((inodedep->id_state & ONWORKLIST) == 0) 3928 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 3929 /* 3930 * Any new dependencies associated with the incore inode must 3931 * now be moved to the list associated with the buffer holding 3932 * the in-memory copy of the inode. Once merged process any 3933 * allocdirects that are completed by the merger. 3934 */ 3935 merge_inode_lists(inodedep); 3936 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL) 3937 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); 3938 /* 3939 * Now that the inode has been pushed into the buffer, the 3940 * operations dependent on the inode being written to disk 3941 * can be moved to the id_bufwait so that they will be 3942 * processed when the buffer I/O completes. 3943 */ 3944 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 3945 WORKLIST_REMOVE(wk); 3946 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 3947 } 3948 /* 3949 * Newly allocated inodes cannot be written until the bitmap 3950 * that allocates them have been written (indicated by 3951 * DEPCOMPLETE being set in id_state). If we are doing a 3952 * forced sync (e.g., an fsync on a file), we force the bitmap 3953 * to be written so that the update can be done. 3954 */ 3955 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) { 3956 FREE_LOCK(&lk); 3957 return; 3958 } 3959 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 3960 FREE_LOCK(&lk); 3961 if (gotit && 3962 (error = VOP_BWRITE(inodedep->id_buf->b_vp, inodedep->id_buf)) != 0) 3963 softdep_error("softdep_update_inodeblock: bwrite", error); 3964 if ((inodedep->id_state & DEPCOMPLETE) == 0) 3965 panic("softdep_update_inodeblock: update failed"); 3966 } 3967 3968 /* 3969 * Merge the new inode dependency list (id_newinoupdt) into the old 3970 * inode dependency list (id_inoupdt). This routine must be called 3971 * with splbio interrupts blocked. 3972 */ 3973 static void 3974 merge_inode_lists(inodedep) 3975 struct inodedep *inodedep; 3976 { 3977 struct allocdirect *listadp, *newadp; 3978 3979 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 3980 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) { 3981 if (listadp->ad_lbn < newadp->ad_lbn) { 3982 listadp = TAILQ_NEXT(listadp, ad_next); 3983 continue; 3984 } 3985 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 3986 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 3987 if (listadp->ad_lbn == newadp->ad_lbn) { 3988 allocdirect_merge(&inodedep->id_inoupdt, newadp, 3989 listadp); 3990 listadp = newadp; 3991 } 3992 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 3993 } 3994 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) { 3995 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 3996 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next); 3997 } 3998 } 3999 4000 /* 4001 * If we are doing an fsync, then we must ensure that any directory 4002 * entries for the inode have been written after the inode gets to disk. 4003 */ 4004 static int 4005 softdep_fsync(vp) 4006 struct vnode *vp; /* the "in_core" copy of the inode */ 4007 { 4008 struct inodedep *inodedep; 4009 struct pagedep *pagedep; 4010 struct worklist *wk; 4011 struct diradd *dap; 4012 struct mount *mnt; 4013 struct vnode *pvp; 4014 struct inode *ip; 4015 struct buf *bp; 4016 struct fs *fs; 4017 struct thread *td = curthread; /* XXX */ 4018 int error, flushparent; 4019 ino_t parentino; 4020 ufs_lbn_t lbn; 4021 4022 ip = VTOI(vp); 4023 fs = ip->i_fs; 4024 ACQUIRE_LOCK(&lk); 4025 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) { 4026 FREE_LOCK(&lk); 4027 return (0); 4028 } 4029 if (LIST_FIRST(&inodedep->id_inowait) != NULL || 4030 LIST_FIRST(&inodedep->id_bufwait) != NULL || 4031 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 4032 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) { 4033 FREE_LOCK(&lk); 4034 panic("softdep_fsync: pending ops"); 4035 } 4036 for (error = 0, flushparent = 0; ; ) { 4037 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 4038 break; 4039 if (wk->wk_type != D_DIRADD) { 4040 FREE_LOCK(&lk); 4041 panic("softdep_fsync: Unexpected type %s", 4042 TYPENAME(wk->wk_type)); 4043 } 4044 dap = WK_DIRADD(wk); 4045 /* 4046 * Flush our parent if this directory entry 4047 * has a MKDIR_PARENT dependency. 4048 */ 4049 if (dap->da_state & DIRCHG) 4050 pagedep = dap->da_previous->dm_pagedep; 4051 else 4052 pagedep = dap->da_pagedep; 4053 mnt = pagedep->pd_mnt; 4054 parentino = pagedep->pd_ino; 4055 lbn = pagedep->pd_lbn; 4056 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) { 4057 FREE_LOCK(&lk); 4058 panic("softdep_fsync: dirty"); 4059 } 4060 flushparent = dap->da_state & MKDIR_PARENT; 4061 /* 4062 * If we are being fsync'ed as part of vgone'ing this vnode, 4063 * then we will not be able to release and recover the 4064 * vnode below, so we just have to give up on writing its 4065 * directory entry out. It will eventually be written, just 4066 * not now, but then the user was not asking to have it 4067 * written, so we are not breaking any promises. 4068 */ 4069 if (vp->v_flag & VRECLAIMED) 4070 break; 4071 /* 4072 * We prevent deadlock by always fetching inodes from the 4073 * root, moving down the directory tree. Thus, when fetching 4074 * our parent directory, we must unlock ourselves before 4075 * requesting the lock on our parent. See the comment in 4076 * ufs_lookup for details on possible races. 4077 */ 4078 FREE_LOCK(&lk); 4079 VOP_UNLOCK(vp, 0, td); 4080 error = VFS_VGET(mnt, parentino, &pvp); 4081 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 4082 if (error != 0) 4083 return (error); 4084 if (flushparent) { 4085 if ((error = UFS_UPDATE(pvp, 1)) != 0) { 4086 vput(pvp); 4087 return (error); 4088 } 4089 } 4090 /* 4091 * Flush directory page containing the inode's name. 4092 */ 4093 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), &bp); 4094 if (error == 0) 4095 error = VOP_BWRITE(bp->b_vp, bp); 4096 vput(pvp); 4097 if (error != 0) 4098 return (error); 4099 ACQUIRE_LOCK(&lk); 4100 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) 4101 break; 4102 } 4103 FREE_LOCK(&lk); 4104 return (0); 4105 } 4106 4107 /* 4108 * Flush all the dirty bitmaps associated with the block device 4109 * before flushing the rest of the dirty blocks so as to reduce 4110 * the number of dependencies that will have to be rolled back. 4111 */ 4112 static int softdep_fsync_mountdev_bp(struct buf *bp, void *data); 4113 4114 void 4115 softdep_fsync_mountdev(vp) 4116 struct vnode *vp; 4117 { 4118 if (!vn_isdisk(vp, NULL)) 4119 panic("softdep_fsync_mountdev: vnode not a disk"); 4120 ACQUIRE_LOCK(&lk); 4121 crit_enter(); 4122 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 4123 softdep_fsync_mountdev_bp, vp); 4124 crit_exit(); 4125 drain_output(vp, 1); 4126 FREE_LOCK(&lk); 4127 } 4128 4129 static int 4130 softdep_fsync_mountdev_bp(struct buf *bp, void *data) 4131 { 4132 struct worklist *wk; 4133 struct vnode *vp = data; 4134 4135 /* 4136 * If it is already scheduled, skip to the next buffer. 4137 */ 4138 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) 4139 return(0); 4140 if (bp->b_vp != vp || (bp->b_flags & B_DELWRI) == 0) { 4141 BUF_UNLOCK(bp); 4142 printf("softdep_fsync_mountdev_bp: warning, buffer %p ripped out from under vnode %p\n", bp, vp); 4143 return(0); 4144 } 4145 /* 4146 * We are only interested in bitmaps with outstanding 4147 * dependencies. 4148 */ 4149 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 4150 wk->wk_type != D_BMSAFEMAP || 4151 (bp->b_xflags & BX_BKGRDINPROG)) { 4152 BUF_UNLOCK(bp); 4153 return(0); 4154 } 4155 bremfree(bp); 4156 FREE_LOCK(&lk); 4157 (void) bawrite(bp); 4158 ACQUIRE_LOCK(&lk); 4159 return(0); 4160 } 4161 4162 /* 4163 * This routine is called when we are trying to synchronously flush a 4164 * file. This routine must eliminate any filesystem metadata dependencies 4165 * so that the syncing routine can succeed by pushing the dirty blocks 4166 * associated with the file. If any I/O errors occur, they are returned. 4167 */ 4168 struct softdep_sync_metadata_info { 4169 struct vnode *vp; 4170 int waitfor; 4171 }; 4172 4173 static int softdep_sync_metadata_bp(struct buf *bp, void *data); 4174 4175 int 4176 softdep_sync_metadata(struct vnode *vp, struct thread *td) 4177 { 4178 struct softdep_sync_metadata_info info; 4179 int error, waitfor; 4180 4181 /* 4182 * Check whether this vnode is involved in a filesystem 4183 * that is doing soft dependency processing. 4184 */ 4185 if (!vn_isdisk(vp, NULL)) { 4186 if (!DOINGSOFTDEP(vp)) 4187 return (0); 4188 } else 4189 if (vp->v_rdev->si_mountpoint == NULL || 4190 (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0) 4191 return (0); 4192 /* 4193 * Ensure that any direct block dependencies have been cleared. 4194 */ 4195 ACQUIRE_LOCK(&lk); 4196 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) { 4197 FREE_LOCK(&lk); 4198 return (error); 4199 } 4200 /* 4201 * For most files, the only metadata dependencies are the 4202 * cylinder group maps that allocate their inode or blocks. 4203 * The block allocation dependencies can be found by traversing 4204 * the dependency lists for any buffers that remain on their 4205 * dirty buffer list. The inode allocation dependency will 4206 * be resolved when the inode is updated with MNT_WAIT. 4207 * This work is done in two passes. The first pass grabs most 4208 * of the buffers and begins asynchronously writing them. The 4209 * only way to wait for these asynchronous writes is to sleep 4210 * on the filesystem vnode which may stay busy for a long time 4211 * if the filesystem is active. So, instead, we make a second 4212 * pass over the dependencies blocking on each write. In the 4213 * usual case we will be blocking against a write that we 4214 * initiated, so when it is done the dependency will have been 4215 * resolved. Thus the second pass is expected to end quickly. 4216 */ 4217 waitfor = MNT_NOWAIT; 4218 top: 4219 /* 4220 * We must wait for any I/O in progress to finish so that 4221 * all potential buffers on the dirty list will be visible. 4222 */ 4223 drain_output(vp, 1); 4224 info.vp = vp; 4225 info.waitfor = waitfor; 4226 crit_enter(); 4227 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 4228 softdep_sync_metadata_bp, &info); 4229 crit_exit(); 4230 if (error < 0) { 4231 FREE_LOCK(&lk); 4232 return(-error); /* error code */ 4233 } 4234 4235 /* 4236 * The brief unlock is to allow any pent up dependency 4237 * processing to be done. Then proceed with the second pass. 4238 */ 4239 if (waitfor == MNT_NOWAIT) { 4240 waitfor = MNT_WAIT; 4241 FREE_LOCK(&lk); 4242 ACQUIRE_LOCK(&lk); 4243 goto top; 4244 } 4245 4246 /* 4247 * If we have managed to get rid of all the dirty buffers, 4248 * then we are done. For certain directories and block 4249 * devices, we may need to do further work. 4250 * 4251 * We must wait for any I/O in progress to finish so that 4252 * all potential buffers on the dirty list will be visible. 4253 */ 4254 drain_output(vp, 1); 4255 if (RB_EMPTY(&vp->v_rbdirty_tree)) { 4256 FREE_LOCK(&lk); 4257 return (0); 4258 } 4259 4260 FREE_LOCK(&lk); 4261 /* 4262 * If we are trying to sync a block device, some of its buffers may 4263 * contain metadata that cannot be written until the contents of some 4264 * partially written files have been written to disk. The only easy 4265 * way to accomplish this is to sync the entire filesystem (luckily 4266 * this happens rarely). 4267 */ 4268 if (vn_isdisk(vp, NULL) && 4269 vp->v_rdev && 4270 vp->v_rdev->si_mountpoint && !VOP_ISLOCKED(vp, NULL) && 4271 (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT, td)) != 0) 4272 return (error); 4273 return (0); 4274 } 4275 4276 static int 4277 softdep_sync_metadata_bp(struct buf *bp, void *data) 4278 { 4279 struct softdep_sync_metadata_info *info = data; 4280 struct pagedep *pagedep; 4281 struct allocdirect *adp; 4282 struct allocindir *aip; 4283 struct worklist *wk; 4284 struct buf *nbp; 4285 int error; 4286 int i; 4287 4288 if (getdirtybuf(&bp, MNT_WAIT) == 0) { 4289 printf("softdep_sync_metadata_bp(1): caught buf %p going away\n", bp); 4290 return (1); 4291 } 4292 if (bp->b_vp != info->vp || (bp->b_flags & B_DELWRI) == 0) { 4293 printf("softdep_sync_metadata_bp(2): caught buf %p going away vp %p\n", bp, info->vp); 4294 BUF_UNLOCK(bp); 4295 return(1); 4296 } 4297 4298 /* 4299 * As we hold the buffer locked, none of its dependencies 4300 * will disappear. 4301 */ 4302 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 4303 switch (wk->wk_type) { 4304 4305 case D_ALLOCDIRECT: 4306 adp = WK_ALLOCDIRECT(wk); 4307 if (adp->ad_state & DEPCOMPLETE) 4308 break; 4309 nbp = adp->ad_buf; 4310 if (getdirtybuf(&nbp, info->waitfor) == 0) 4311 break; 4312 FREE_LOCK(&lk); 4313 if (info->waitfor == MNT_NOWAIT) { 4314 bawrite(nbp); 4315 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 4316 bawrite(bp); 4317 ACQUIRE_LOCK(&lk); 4318 return (-error); 4319 } 4320 ACQUIRE_LOCK(&lk); 4321 break; 4322 4323 case D_ALLOCINDIR: 4324 aip = WK_ALLOCINDIR(wk); 4325 if (aip->ai_state & DEPCOMPLETE) 4326 break; 4327 nbp = aip->ai_buf; 4328 if (getdirtybuf(&nbp, info->waitfor) == 0) 4329 break; 4330 FREE_LOCK(&lk); 4331 if (info->waitfor == MNT_NOWAIT) { 4332 bawrite(nbp); 4333 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 4334 bawrite(bp); 4335 ACQUIRE_LOCK(&lk); 4336 return (-error); 4337 } 4338 ACQUIRE_LOCK(&lk); 4339 break; 4340 4341 case D_INDIRDEP: 4342 restart: 4343 4344 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) { 4345 if (aip->ai_state & DEPCOMPLETE) 4346 continue; 4347 nbp = aip->ai_buf; 4348 if (getdirtybuf(&nbp, MNT_WAIT) == 0) 4349 goto restart; 4350 FREE_LOCK(&lk); 4351 if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 4352 bawrite(bp); 4353 ACQUIRE_LOCK(&lk); 4354 return (-error); 4355 } 4356 ACQUIRE_LOCK(&lk); 4357 goto restart; 4358 } 4359 break; 4360 4361 case D_INODEDEP: 4362 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs, 4363 WK_INODEDEP(wk)->id_ino)) != 0) { 4364 FREE_LOCK(&lk); 4365 bawrite(bp); 4366 ACQUIRE_LOCK(&lk); 4367 return (-error); 4368 } 4369 break; 4370 4371 case D_PAGEDEP: 4372 /* 4373 * We are trying to sync a directory that may 4374 * have dependencies on both its own metadata 4375 * and/or dependencies on the inodes of any 4376 * recently allocated files. We walk its diradd 4377 * lists pushing out the associated inode. 4378 */ 4379 pagedep = WK_PAGEDEP(wk); 4380 for (i = 0; i < DAHASHSZ; i++) { 4381 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 4382 continue; 4383 if ((error = 4384 flush_pagedep_deps(info->vp, 4385 pagedep->pd_mnt, 4386 &pagedep->pd_diraddhd[i]))) { 4387 FREE_LOCK(&lk); 4388 bawrite(bp); 4389 ACQUIRE_LOCK(&lk); 4390 return (-error); 4391 } 4392 } 4393 break; 4394 4395 case D_MKDIR: 4396 /* 4397 * This case should never happen if the vnode has 4398 * been properly sync'ed. However, if this function 4399 * is used at a place where the vnode has not yet 4400 * been sync'ed, this dependency can show up. So, 4401 * rather than panic, just flush it. 4402 */ 4403 nbp = WK_MKDIR(wk)->md_buf; 4404 if (getdirtybuf(&nbp, info->waitfor) == 0) 4405 break; 4406 FREE_LOCK(&lk); 4407 if (info->waitfor == MNT_NOWAIT) { 4408 bawrite(nbp); 4409 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 4410 bawrite(bp); 4411 ACQUIRE_LOCK(&lk); 4412 return (-error); 4413 } 4414 ACQUIRE_LOCK(&lk); 4415 break; 4416 4417 case D_BMSAFEMAP: 4418 /* 4419 * This case should never happen if the vnode has 4420 * been properly sync'ed. However, if this function 4421 * is used at a place where the vnode has not yet 4422 * been sync'ed, this dependency can show up. So, 4423 * rather than panic, just flush it. 4424 * 4425 * nbp can wind up == bp if a device node for the 4426 * same filesystem is being fsynced at the same time, 4427 * leading to a panic if we don't catch the case. 4428 */ 4429 nbp = WK_BMSAFEMAP(wk)->sm_buf; 4430 if (nbp == bp) 4431 break; 4432 if (getdirtybuf(&nbp, info->waitfor) == 0) 4433 break; 4434 FREE_LOCK(&lk); 4435 if (info->waitfor == MNT_NOWAIT) { 4436 bawrite(nbp); 4437 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 4438 bawrite(bp); 4439 ACQUIRE_LOCK(&lk); 4440 return (-error); 4441 } 4442 ACQUIRE_LOCK(&lk); 4443 break; 4444 4445 default: 4446 FREE_LOCK(&lk); 4447 panic("softdep_sync_metadata: Unknown type %s", 4448 TYPENAME(wk->wk_type)); 4449 /* NOTREACHED */ 4450 } 4451 } 4452 FREE_LOCK(&lk); 4453 bawrite(bp); 4454 ACQUIRE_LOCK(&lk); 4455 return(0); 4456 } 4457 4458 /* 4459 * Flush the dependencies associated with an inodedep. 4460 * Called with splbio blocked. 4461 */ 4462 static int 4463 flush_inodedep_deps(fs, ino) 4464 struct fs *fs; 4465 ino_t ino; 4466 { 4467 struct inodedep *inodedep; 4468 struct allocdirect *adp; 4469 int error, waitfor; 4470 struct buf *bp; 4471 4472 /* 4473 * This work is done in two passes. The first pass grabs most 4474 * of the buffers and begins asynchronously writing them. The 4475 * only way to wait for these asynchronous writes is to sleep 4476 * on the filesystem vnode which may stay busy for a long time 4477 * if the filesystem is active. So, instead, we make a second 4478 * pass over the dependencies blocking on each write. In the 4479 * usual case we will be blocking against a write that we 4480 * initiated, so when it is done the dependency will have been 4481 * resolved. Thus the second pass is expected to end quickly. 4482 * We give a brief window at the top of the loop to allow 4483 * any pending I/O to complete. 4484 */ 4485 for (waitfor = MNT_NOWAIT; ; ) { 4486 FREE_LOCK(&lk); 4487 ACQUIRE_LOCK(&lk); 4488 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4489 return (0); 4490 TAILQ_FOREACH(adp, &inodedep->id_inoupdt, ad_next) { 4491 if (adp->ad_state & DEPCOMPLETE) 4492 continue; 4493 bp = adp->ad_buf; 4494 if (getdirtybuf(&bp, waitfor) == 0) { 4495 if (waitfor == MNT_NOWAIT) 4496 continue; 4497 break; 4498 } 4499 FREE_LOCK(&lk); 4500 if (waitfor == MNT_NOWAIT) { 4501 bawrite(bp); 4502 } else if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) { 4503 ACQUIRE_LOCK(&lk); 4504 return (error); 4505 } 4506 ACQUIRE_LOCK(&lk); 4507 break; 4508 } 4509 if (adp != NULL) 4510 continue; 4511 TAILQ_FOREACH(adp, &inodedep->id_newinoupdt, ad_next) { 4512 if (adp->ad_state & DEPCOMPLETE) 4513 continue; 4514 bp = adp->ad_buf; 4515 if (getdirtybuf(&bp, waitfor) == 0) { 4516 if (waitfor == MNT_NOWAIT) 4517 continue; 4518 break; 4519 } 4520 FREE_LOCK(&lk); 4521 if (waitfor == MNT_NOWAIT) { 4522 bawrite(bp); 4523 } else if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) { 4524 ACQUIRE_LOCK(&lk); 4525 return (error); 4526 } 4527 ACQUIRE_LOCK(&lk); 4528 break; 4529 } 4530 if (adp != NULL) 4531 continue; 4532 /* 4533 * If pass2, we are done, otherwise do pass 2. 4534 */ 4535 if (waitfor == MNT_WAIT) 4536 break; 4537 waitfor = MNT_WAIT; 4538 } 4539 /* 4540 * Try freeing inodedep in case all dependencies have been removed. 4541 */ 4542 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0) 4543 (void) free_inodedep(inodedep); 4544 return (0); 4545 } 4546 4547 /* 4548 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 4549 * Called with splbio blocked. 4550 */ 4551 static int 4552 flush_pagedep_deps(pvp, mp, diraddhdp) 4553 struct vnode *pvp; 4554 struct mount *mp; 4555 struct diraddhd *diraddhdp; 4556 { 4557 struct thread *td = curthread; /* XXX */ 4558 struct inodedep *inodedep; 4559 struct ufsmount *ump; 4560 struct diradd *dap; 4561 struct vnode *vp; 4562 int gotit, error = 0; 4563 struct buf *bp; 4564 ino_t inum; 4565 4566 ump = VFSTOUFS(mp); 4567 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 4568 /* 4569 * Flush ourselves if this directory entry 4570 * has a MKDIR_PARENT dependency. 4571 */ 4572 if (dap->da_state & MKDIR_PARENT) { 4573 FREE_LOCK(&lk); 4574 if ((error = UFS_UPDATE(pvp, 1)) != 0) 4575 break; 4576 ACQUIRE_LOCK(&lk); 4577 /* 4578 * If that cleared dependencies, go on to next. 4579 */ 4580 if (dap != LIST_FIRST(diraddhdp)) 4581 continue; 4582 if (dap->da_state & MKDIR_PARENT) { 4583 FREE_LOCK(&lk); 4584 panic("flush_pagedep_deps: MKDIR_PARENT"); 4585 } 4586 } 4587 /* 4588 * A newly allocated directory must have its "." and 4589 * ".." entries written out before its name can be 4590 * committed in its parent. We do not want or need 4591 * the full semantics of a synchronous VOP_FSYNC as 4592 * that may end up here again, once for each directory 4593 * level in the filesystem. Instead, we push the blocks 4594 * and wait for them to clear. We have to fsync twice 4595 * because the first call may choose to defer blocks 4596 * that still have dependencies, but deferral will 4597 * happen at most once. 4598 */ 4599 inum = dap->da_newinum; 4600 if (dap->da_state & MKDIR_BODY) { 4601 FREE_LOCK(&lk); 4602 if ((error = VFS_VGET(mp, inum, &vp)) != 0) 4603 break; 4604 if ((error=VOP_FSYNC(vp, MNT_NOWAIT, td)) || 4605 (error=VOP_FSYNC(vp, MNT_NOWAIT, td))) { 4606 vput(vp); 4607 break; 4608 } 4609 drain_output(vp, 0); 4610 vput(vp); 4611 ACQUIRE_LOCK(&lk); 4612 /* 4613 * If that cleared dependencies, go on to next. 4614 */ 4615 if (dap != LIST_FIRST(diraddhdp)) 4616 continue; 4617 if (dap->da_state & MKDIR_BODY) { 4618 FREE_LOCK(&lk); 4619 panic("flush_pagedep_deps: MKDIR_BODY"); 4620 } 4621 } 4622 /* 4623 * Flush the inode on which the directory entry depends. 4624 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 4625 * the only remaining dependency is that the updated inode 4626 * count must get pushed to disk. The inode has already 4627 * been pushed into its inode buffer (via VOP_UPDATE) at 4628 * the time of the reference count change. So we need only 4629 * locate that buffer, ensure that there will be no rollback 4630 * caused by a bitmap dependency, then write the inode buffer. 4631 */ 4632 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) { 4633 FREE_LOCK(&lk); 4634 panic("flush_pagedep_deps: lost inode"); 4635 } 4636 /* 4637 * If the inode still has bitmap dependencies, 4638 * push them to disk. 4639 */ 4640 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4641 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 4642 FREE_LOCK(&lk); 4643 if (gotit && 4644 (error = VOP_BWRITE(inodedep->id_buf->b_vp, 4645 inodedep->id_buf)) != 0) 4646 break; 4647 ACQUIRE_LOCK(&lk); 4648 if (dap != LIST_FIRST(diraddhdp)) 4649 continue; 4650 } 4651 /* 4652 * If the inode is still sitting in a buffer waiting 4653 * to be written, push it to disk. 4654 */ 4655 FREE_LOCK(&lk); 4656 if ((error = bread(ump->um_devvp, 4657 fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), 4658 (int)ump->um_fs->fs_bsize, &bp)) != 0) 4659 break; 4660 if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) 4661 break; 4662 ACQUIRE_LOCK(&lk); 4663 /* 4664 * If we have failed to get rid of all the dependencies 4665 * then something is seriously wrong. 4666 */ 4667 if (dap == LIST_FIRST(diraddhdp)) { 4668 FREE_LOCK(&lk); 4669 panic("flush_pagedep_deps: flush failed"); 4670 } 4671 } 4672 if (error) 4673 ACQUIRE_LOCK(&lk); 4674 return (error); 4675 } 4676 4677 /* 4678 * A large burst of file addition or deletion activity can drive the 4679 * memory load excessively high. First attempt to slow things down 4680 * using the techniques below. If that fails, this routine requests 4681 * the offending operations to fall back to running synchronously 4682 * until the memory load returns to a reasonable level. 4683 */ 4684 int 4685 softdep_slowdown(vp) 4686 struct vnode *vp; 4687 { 4688 int max_softdeps_hard; 4689 4690 max_softdeps_hard = max_softdeps * 11 / 10; 4691 if (num_dirrem < max_softdeps_hard / 2 && 4692 num_inodedep < max_softdeps_hard) 4693 return (0); 4694 stat_sync_limit_hit += 1; 4695 return (1); 4696 } 4697 4698 /* 4699 * If memory utilization has gotten too high, deliberately slow things 4700 * down and speed up the I/O processing. 4701 */ 4702 static int 4703 request_cleanup(resource, islocked) 4704 int resource; 4705 int islocked; 4706 { 4707 struct thread *td = curthread; /* XXX */ 4708 4709 /* 4710 * We never hold up the filesystem syncer process. 4711 */ 4712 if (td == filesys_syncer) 4713 return (0); 4714 /* 4715 * First check to see if the work list has gotten backlogged. 4716 * If it has, co-opt this process to help clean up two entries. 4717 * Because this process may hold inodes locked, we cannot 4718 * handle any remove requests that might block on a locked 4719 * inode as that could lead to deadlock. 4720 */ 4721 if (num_on_worklist > max_softdeps / 10) { 4722 if (islocked) 4723 FREE_LOCK(&lk); 4724 process_worklist_item(NULL, LK_NOWAIT); 4725 process_worklist_item(NULL, LK_NOWAIT); 4726 stat_worklist_push += 2; 4727 if (islocked) 4728 ACQUIRE_LOCK(&lk); 4729 return(1); 4730 } 4731 4732 /* 4733 * If we are resource constrained on inode dependencies, try 4734 * flushing some dirty inodes. Otherwise, we are constrained 4735 * by file deletions, so try accelerating flushes of directories 4736 * with removal dependencies. We would like to do the cleanup 4737 * here, but we probably hold an inode locked at this point and 4738 * that might deadlock against one that we try to clean. So, 4739 * the best that we can do is request the syncer daemon to do 4740 * the cleanup for us. 4741 */ 4742 switch (resource) { 4743 4744 case FLUSH_INODES: 4745 stat_ino_limit_push += 1; 4746 req_clear_inodedeps += 1; 4747 stat_countp = &stat_ino_limit_hit; 4748 break; 4749 4750 case FLUSH_REMOVE: 4751 stat_blk_limit_push += 1; 4752 req_clear_remove += 1; 4753 stat_countp = &stat_blk_limit_hit; 4754 break; 4755 4756 default: 4757 if (islocked) 4758 FREE_LOCK(&lk); 4759 panic("request_cleanup: unknown type"); 4760 } 4761 /* 4762 * Hopefully the syncer daemon will catch up and awaken us. 4763 * We wait at most tickdelay before proceeding in any case. 4764 */ 4765 if (islocked == 0) 4766 ACQUIRE_LOCK(&lk); 4767 crit_enter(); 4768 proc_waiting += 1; 4769 if (!callout_active(&handle)) 4770 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2, 4771 pause_timer, NULL); 4772 interlocked_sleep(&lk, SLEEP, (caddr_t)&proc_waiting, 0, 4773 "softupdate", 0); 4774 proc_waiting -= 1; 4775 crit_exit(); 4776 if (islocked == 0) 4777 FREE_LOCK(&lk); 4778 return (1); 4779 } 4780 4781 /* 4782 * Awaken processes pausing in request_cleanup and clear proc_waiting 4783 * to indicate that there is no longer a timer running. 4784 */ 4785 void 4786 pause_timer(arg) 4787 void *arg; 4788 { 4789 *stat_countp += 1; 4790 wakeup_one(&proc_waiting); 4791 if (proc_waiting > 0) 4792 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2, 4793 pause_timer, NULL); 4794 else 4795 callout_deactivate(&handle); 4796 } 4797 4798 /* 4799 * Flush out a directory with at least one removal dependency in an effort to 4800 * reduce the number of dirrem, freefile, and freeblks dependency structures. 4801 */ 4802 static void 4803 clear_remove(struct thread *td) 4804 { 4805 struct pagedep_hashhead *pagedephd; 4806 struct pagedep *pagedep; 4807 static int next = 0; 4808 struct mount *mp; 4809 struct vnode *vp; 4810 int error, cnt; 4811 ino_t ino; 4812 4813 ACQUIRE_LOCK(&lk); 4814 for (cnt = 0; cnt < pagedep_hash; cnt++) { 4815 pagedephd = &pagedep_hashtbl[next++]; 4816 if (next >= pagedep_hash) 4817 next = 0; 4818 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 4819 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL) 4820 continue; 4821 mp = pagedep->pd_mnt; 4822 ino = pagedep->pd_ino; 4823 FREE_LOCK(&lk); 4824 if ((error = VFS_VGET(mp, ino, &vp)) != 0) { 4825 softdep_error("clear_remove: vget", error); 4826 return; 4827 } 4828 if ((error = VOP_FSYNC(vp, MNT_NOWAIT, td))) 4829 softdep_error("clear_remove: fsync", error); 4830 drain_output(vp, 0); 4831 vput(vp); 4832 return; 4833 } 4834 } 4835 FREE_LOCK(&lk); 4836 } 4837 4838 /* 4839 * Clear out a block of dirty inodes in an effort to reduce 4840 * the number of inodedep dependency structures. 4841 */ 4842 struct clear_inodedeps_info { 4843 struct fs *fs; 4844 struct mount *mp; 4845 }; 4846 4847 static int 4848 clear_inodedeps_mountlist_callback(struct mount *mp, void *data) 4849 { 4850 struct clear_inodedeps_info *info = data; 4851 4852 if ((mp->mnt_flag & MNT_SOFTDEP) && info->fs == VFSTOUFS(mp)->um_fs) { 4853 info->mp = mp; 4854 return(-1); 4855 } 4856 return(0); 4857 } 4858 4859 static void 4860 clear_inodedeps(struct thread *td) 4861 { 4862 struct clear_inodedeps_info info; 4863 struct inodedep_hashhead *inodedephd; 4864 struct inodedep *inodedep; 4865 static int next = 0; 4866 struct vnode *vp; 4867 struct fs *fs; 4868 int error, cnt; 4869 ino_t firstino, lastino, ino; 4870 4871 ACQUIRE_LOCK(&lk); 4872 /* 4873 * Pick a random inode dependency to be cleared. 4874 * We will then gather up all the inodes in its block 4875 * that have dependencies and flush them out. 4876 */ 4877 for (cnt = 0; cnt < inodedep_hash; cnt++) { 4878 inodedephd = &inodedep_hashtbl[next++]; 4879 if (next >= inodedep_hash) 4880 next = 0; 4881 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 4882 break; 4883 } 4884 if (inodedep == NULL) { 4885 FREE_LOCK(&lk); 4886 return; 4887 } 4888 /* 4889 * Ugly code to find mount point given pointer to superblock. 4890 */ 4891 fs = inodedep->id_fs; 4892 info.mp = NULL; 4893 info.fs = fs; 4894 mountlist_scan(clear_inodedeps_mountlist_callback, 4895 &info, MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 4896 /* 4897 * Find the last inode in the block with dependencies. 4898 */ 4899 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 4900 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 4901 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0) 4902 break; 4903 /* 4904 * Asynchronously push all but the last inode with dependencies. 4905 * Synchronously push the last inode with dependencies to ensure 4906 * that the inode block gets written to free up the inodedeps. 4907 */ 4908 for (ino = firstino; ino <= lastino; ino++) { 4909 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4910 continue; 4911 FREE_LOCK(&lk); 4912 if ((error = VFS_VGET(info.mp, ino, &vp)) != 0) { 4913 softdep_error("clear_inodedeps: vget", error); 4914 return; 4915 } 4916 if (ino == lastino) { 4917 if ((error = VOP_FSYNC(vp, MNT_WAIT, td))) 4918 softdep_error("clear_inodedeps: fsync1", error); 4919 } else { 4920 if ((error = VOP_FSYNC(vp, MNT_NOWAIT, td))) 4921 softdep_error("clear_inodedeps: fsync2", error); 4922 drain_output(vp, 0); 4923 } 4924 vput(vp); 4925 ACQUIRE_LOCK(&lk); 4926 } 4927 FREE_LOCK(&lk); 4928 } 4929 4930 /* 4931 * Function to determine if the buffer has outstanding dependencies 4932 * that will cause a roll-back if the buffer is written. If wantcount 4933 * is set, return number of dependencies, otherwise just yes or no. 4934 */ 4935 static int 4936 softdep_count_dependencies(bp, wantcount) 4937 struct buf *bp; 4938 int wantcount; 4939 { 4940 struct worklist *wk; 4941 struct inodedep *inodedep; 4942 struct indirdep *indirdep; 4943 struct allocindir *aip; 4944 struct pagedep *pagedep; 4945 struct diradd *dap; 4946 int i, retval; 4947 4948 retval = 0; 4949 ACQUIRE_LOCK(&lk); 4950 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 4951 switch (wk->wk_type) { 4952 4953 case D_INODEDEP: 4954 inodedep = WK_INODEDEP(wk); 4955 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4956 /* bitmap allocation dependency */ 4957 retval += 1; 4958 if (!wantcount) 4959 goto out; 4960 } 4961 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 4962 /* direct block pointer dependency */ 4963 retval += 1; 4964 if (!wantcount) 4965 goto out; 4966 } 4967 continue; 4968 4969 case D_INDIRDEP: 4970 indirdep = WK_INDIRDEP(wk); 4971 4972 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 4973 /* indirect block pointer dependency */ 4974 retval += 1; 4975 if (!wantcount) 4976 goto out; 4977 } 4978 continue; 4979 4980 case D_PAGEDEP: 4981 pagedep = WK_PAGEDEP(wk); 4982 for (i = 0; i < DAHASHSZ; i++) { 4983 4984 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 4985 /* directory entry dependency */ 4986 retval += 1; 4987 if (!wantcount) 4988 goto out; 4989 } 4990 } 4991 continue; 4992 4993 case D_BMSAFEMAP: 4994 case D_ALLOCDIRECT: 4995 case D_ALLOCINDIR: 4996 case D_MKDIR: 4997 /* never a dependency on these blocks */ 4998 continue; 4999 5000 default: 5001 FREE_LOCK(&lk); 5002 panic("softdep_check_for_rollback: Unexpected type %s", 5003 TYPENAME(wk->wk_type)); 5004 /* NOTREACHED */ 5005 } 5006 } 5007 out: 5008 FREE_LOCK(&lk); 5009 return retval; 5010 } 5011 5012 /* 5013 * Acquire exclusive access to a buffer. 5014 * Must be called with splbio blocked. 5015 * Return 1 if buffer was acquired. 5016 */ 5017 static int 5018 getdirtybuf(bpp, waitfor) 5019 struct buf **bpp; 5020 int waitfor; 5021 { 5022 struct buf *bp; 5023 int error; 5024 5025 for (;;) { 5026 if ((bp = *bpp) == NULL) 5027 return (0); 5028 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 5029 if ((bp->b_xflags & BX_BKGRDINPROG) == 0) 5030 break; 5031 BUF_UNLOCK(bp); 5032 if (waitfor != MNT_WAIT) 5033 return (0); 5034 bp->b_xflags |= BX_BKGRDWAIT; 5035 interlocked_sleep(&lk, SLEEP, &bp->b_xflags, 0, 5036 "getbuf", 0); 5037 continue; 5038 } 5039 if (waitfor != MNT_WAIT) 5040 return (0); 5041 error = interlocked_sleep(&lk, LOCKBUF, bp, 5042 LK_EXCLUSIVE | LK_SLEEPFAIL, 0, 0); 5043 if (error != ENOLCK) { 5044 FREE_LOCK(&lk); 5045 panic("getdirtybuf: inconsistent lock"); 5046 } 5047 } 5048 if ((bp->b_flags & B_DELWRI) == 0) { 5049 BUF_UNLOCK(bp); 5050 return (0); 5051 } 5052 bremfree(bp); 5053 return (1); 5054 } 5055 5056 /* 5057 * Wait for pending output on a vnode to complete. 5058 * Must be called with vnode locked. 5059 */ 5060 static void 5061 drain_output(vp, islocked) 5062 struct vnode *vp; 5063 int islocked; 5064 { 5065 5066 if (!islocked) 5067 ACQUIRE_LOCK(&lk); 5068 while (vp->v_track_write.bk_active) { 5069 vp->v_track_write.bk_waitflag = 1; 5070 interlocked_sleep(&lk, SLEEP, &vp->v_track_write, 5071 0, "drainvp", 0); 5072 } 5073 if (!islocked) 5074 FREE_LOCK(&lk); 5075 } 5076 5077 /* 5078 * Called whenever a buffer that is being invalidated or reallocated 5079 * contains dependencies. This should only happen if an I/O error has 5080 * occurred. The routine is called with the buffer locked. 5081 */ 5082 static void 5083 softdep_deallocate_dependencies(bp) 5084 struct buf *bp; 5085 { 5086 5087 if ((bp->b_flags & B_ERROR) == 0) 5088 panic("softdep_deallocate_dependencies: dangling deps"); 5089 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntfromname, bp->b_error); 5090 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 5091 } 5092 5093 /* 5094 * Function to handle asynchronous write errors in the filesystem. 5095 */ 5096 void 5097 softdep_error(func, error) 5098 char *func; 5099 int error; 5100 { 5101 5102 /* XXX should do something better! */ 5103 printf("%s: got error %d while accessing filesystem\n", func, error); 5104 } 5105