1 /* 2 * Copyright 1998 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 * The following are the copyrights and redistribution conditions that 10 * apply to this copy of the soft update software. For a license 11 * to use, redistribute or sell the soft update software under 12 * conditions other than those described here, please contact the 13 * author at one of the following addresses: 14 * 15 * Marshall Kirk McKusick mckusick@mckusick.com 16 * 1614 Oxford Street +1-510-843-9542 17 * Berkeley, CA 94709-1608 18 * USA 19 * 20 * Redistribution and use in source and binary forms, with or without 21 * modification, are permitted provided that the following conditions 22 * are met: 23 * 24 * 1. Redistributions of source code must retain the above copyright 25 * notice, this list of conditions and the following disclaimer. 26 * 2. Redistributions in binary form must reproduce the above copyright 27 * notice, this list of conditions and the following disclaimer in the 28 * documentation and/or other materials provided with the distribution. 29 * 3. None of the names of McKusick, Ganger, Patt, or the University of 30 * Michigan may be used to endorse or promote products derived from 31 * this software without specific prior written permission. 32 * 4. Redistributions in any form must be accompanied by information on 33 * how to obtain complete source code for any accompanying software 34 * that uses this software. This source code must either be included 35 * in the distribution or be available for no more than the cost of 36 * distribution plus a nominal fee, and must be freely redistributable 37 * under reasonable conditions. For an executable file, complete 38 * source code means the source code for all modules it contains. 39 * It does not mean source code for modules or files that typically 40 * accompany the operating system on which the executable file runs, 41 * e.g., standard library modules or system header files. 42 * 43 * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY 44 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 45 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 46 * DISCLAIMED. IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR 47 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 48 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 49 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 50 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 51 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 52 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 53 * SUCH DAMAGE. 54 * 55 * from: @(#)ffs_softdep.c 9.40 (McKusick) 6/15/99 56 * $FreeBSD$ 57 */ 58 59 /* 60 * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide. 61 */ 62 #ifndef DIAGNOSTIC 63 #define DIAGNOSTIC 64 #endif 65 #ifndef DEBUG 66 #define DEBUG 67 #endif 68 69 #include <sys/param.h> 70 #include <sys/kernel.h> 71 #include <sys/systm.h> 72 #include <sys/buf.h> 73 #include <sys/malloc.h> 74 #include <sys/mount.h> 75 #include <sys/proc.h> 76 #include <sys/syslog.h> 77 #include <sys/vnode.h> 78 #include <sys/conf.h> 79 #include <ufs/ufs/dir.h> 80 #include <ufs/ufs/quota.h> 81 #include <ufs/ufs/inode.h> 82 #include <ufs/ufs/ufsmount.h> 83 #include <ufs/ffs/fs.h> 84 #include <ufs/ffs/softdep.h> 85 #include <ufs/ffs/ffs_extern.h> 86 #include <ufs/ufs/ufs_extern.h> 87 88 /* 89 * These definitions need to be adapted to the system to which 90 * this file is being ported. 91 */ 92 /* 93 * malloc types defined for the softdep system. 94 */ 95 MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies"); 96 MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies"); 97 MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation"); 98 MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map"); 99 MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode"); 100 MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies"); 101 MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block"); 102 MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode"); 103 MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode"); 104 MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated"); 105 MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry"); 106 MALLOC_DEFINE(M_MKDIR, "mkdir","New directory"); 107 MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted"); 108 109 #define D_PAGEDEP 0 110 #define D_INODEDEP 1 111 #define D_NEWBLK 2 112 #define D_BMSAFEMAP 3 113 #define D_ALLOCDIRECT 4 114 #define D_INDIRDEP 5 115 #define D_ALLOCINDIR 6 116 #define D_FREEFRAG 7 117 #define D_FREEBLKS 8 118 #define D_FREEFILE 9 119 #define D_DIRADD 10 120 #define D_MKDIR 11 121 #define D_DIRREM 12 122 #define D_LAST D_DIRREM 123 124 /* 125 * translate from workitem type to memory type 126 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 127 */ 128 static struct malloc_type *memtype[] = { 129 M_PAGEDEP, 130 M_INODEDEP, 131 M_NEWBLK, 132 M_BMSAFEMAP, 133 M_ALLOCDIRECT, 134 M_INDIRDEP, 135 M_ALLOCINDIR, 136 M_FREEFRAG, 137 M_FREEBLKS, 138 M_FREEFILE, 139 M_DIRADD, 140 M_MKDIR, 141 M_DIRREM 142 }; 143 144 #define DtoM(type) (memtype[type]) 145 146 /* 147 * Names of malloc types. 148 */ 149 #define TYPENAME(type) \ 150 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???") 151 #define CURPROC curproc 152 /* 153 * End system adaptaion definitions. 154 */ 155 156 /* 157 * Internal function prototypes. 158 */ 159 static void softdep_error __P((char *, int)); 160 static void drain_output __P((struct vnode *, int)); 161 static int getdirtybuf __P((struct buf **, int)); 162 static void clear_remove __P((struct proc *)); 163 static void clear_inodedeps __P((struct proc *)); 164 static int flush_pagedep_deps __P((struct vnode *, struct mount *, 165 struct diraddhd *)); 166 static int flush_inodedep_deps __P((struct fs *, ino_t)); 167 static int handle_written_filepage __P((struct pagedep *, struct buf *)); 168 static void diradd_inode_written __P((struct diradd *, struct inodedep *)); 169 static int handle_written_inodeblock __P((struct inodedep *, struct buf *)); 170 static void handle_allocdirect_partdone __P((struct allocdirect *)); 171 static void handle_allocindir_partdone __P((struct allocindir *)); 172 static void initiate_write_filepage __P((struct pagedep *, struct buf *)); 173 static void handle_written_mkdir __P((struct mkdir *, int)); 174 static void initiate_write_inodeblock __P((struct inodedep *, struct buf *)); 175 static void handle_workitem_freefile __P((struct freefile *)); 176 static void handle_workitem_remove __P((struct dirrem *)); 177 static struct dirrem *newdirrem __P((struct buf *, struct inode *, 178 struct inode *, int)); 179 static void free_diradd __P((struct diradd *)); 180 static void free_allocindir __P((struct allocindir *, struct inodedep *)); 181 static int indir_trunc __P((struct inode *, ufs_daddr_t, int, ufs_lbn_t, 182 long *)); 183 static void deallocate_dependencies __P((struct buf *, struct inodedep *)); 184 static void free_allocdirect __P((struct allocdirectlst *, 185 struct allocdirect *, int)); 186 static int free_inodedep __P((struct inodedep *)); 187 static void handle_workitem_freeblocks __P((struct freeblks *)); 188 static void merge_inode_lists __P((struct inodedep *)); 189 static void setup_allocindir_phase2 __P((struct buf *, struct inode *, 190 struct allocindir *)); 191 static struct allocindir *newallocindir __P((struct inode *, int, ufs_daddr_t, 192 ufs_daddr_t)); 193 static void handle_workitem_freefrag __P((struct freefrag *)); 194 static struct freefrag *newfreefrag __P((struct inode *, ufs_daddr_t, long)); 195 static void allocdirect_merge __P((struct allocdirectlst *, 196 struct allocdirect *, struct allocdirect *)); 197 static struct bmsafemap *bmsafemap_lookup __P((struct buf *)); 198 static int newblk_lookup __P((struct fs *, ufs_daddr_t, int, 199 struct newblk **)); 200 static int inodedep_lookup __P((struct fs *, ino_t, int, struct inodedep **)); 201 static int pagedep_lookup __P((struct inode *, ufs_lbn_t, int, 202 struct pagedep **)); 203 static void pause_timer __P((void *)); 204 static int request_cleanup __P((int, int)); 205 static void add_to_worklist __P((struct worklist *)); 206 207 /* 208 * Exported softdep operations. 209 */ 210 struct bio_ops bioops = { 211 softdep_disk_io_initiation, /* io_start */ 212 softdep_disk_write_complete, /* io_complete */ 213 softdep_deallocate_dependencies, /* io_deallocate */ 214 softdep_fsync, /* io_fsync */ 215 softdep_process_worklist, /* io_sync */ 216 }; 217 218 /* 219 * Locking primitives. 220 * 221 * For a uniprocessor, all we need to do is protect against disk 222 * interrupts. For a multiprocessor, this lock would have to be 223 * a mutex. A single mutex is used throughout this file, though 224 * finer grain locking could be used if contention warranted it. 225 * 226 * For a multiprocessor, the sleep call would accept a lock and 227 * release it after the sleep processing was complete. In a uniprocessor 228 * implementation there is no such interlock, so we simple mark 229 * the places where it needs to be done with the `interlocked' form 230 * of the lock calls. Since the uniprocessor sleep already interlocks 231 * the spl, there is nothing that really needs to be done. 232 */ 233 #ifndef /* NOT */ DEBUG 234 static struct lockit { 235 int lkt_spl; 236 } lk = { 0 }; 237 #define ACQUIRE_LOCK(lk) (lk)->lkt_spl = splbio() 238 #define FREE_LOCK(lk) splx((lk)->lkt_spl) 239 #define ACQUIRE_LOCK_INTERLOCKED(lk) 240 #define FREE_LOCK_INTERLOCKED(lk) 241 242 #else /* DEBUG */ 243 static struct lockit { 244 int lkt_spl; 245 pid_t lkt_held; 246 } lk = { 0, -1 }; 247 static int lockcnt; 248 249 static void acquire_lock __P((struct lockit *)); 250 static void free_lock __P((struct lockit *)); 251 static void acquire_lock_interlocked __P((struct lockit *)); 252 static void free_lock_interlocked __P((struct lockit *)); 253 254 #define ACQUIRE_LOCK(lk) acquire_lock(lk) 255 #define FREE_LOCK(lk) free_lock(lk) 256 #define ACQUIRE_LOCK_INTERLOCKED(lk) acquire_lock_interlocked(lk) 257 #define FREE_LOCK_INTERLOCKED(lk) free_lock_interlocked(lk) 258 259 static void 260 acquire_lock(lk) 261 struct lockit *lk; 262 { 263 264 if (lk->lkt_held != -1) { 265 if (lk->lkt_held == CURPROC->p_pid) 266 panic("softdep_lock: locking against myself"); 267 else 268 panic("softdep_lock: lock held by %d", lk->lkt_held); 269 } 270 lk->lkt_spl = splbio(); 271 lk->lkt_held = CURPROC->p_pid; 272 lockcnt++; 273 } 274 275 static void 276 free_lock(lk) 277 struct lockit *lk; 278 { 279 280 if (lk->lkt_held == -1) 281 panic("softdep_unlock: lock not held"); 282 lk->lkt_held = -1; 283 splx(lk->lkt_spl); 284 } 285 286 static void 287 acquire_lock_interlocked(lk) 288 struct lockit *lk; 289 { 290 291 if (lk->lkt_held != -1) { 292 if (lk->lkt_held == CURPROC->p_pid) 293 panic("softdep_lock_interlocked: locking against self"); 294 else 295 panic("softdep_lock_interlocked: lock held by %d", 296 lk->lkt_held); 297 } 298 lk->lkt_held = CURPROC->p_pid; 299 lockcnt++; 300 } 301 302 static void 303 free_lock_interlocked(lk) 304 struct lockit *lk; 305 { 306 307 if (lk->lkt_held == -1) 308 panic("softdep_unlock_interlocked: lock not held"); 309 lk->lkt_held = -1; 310 } 311 #endif /* DEBUG */ 312 313 /* 314 * Place holder for real semaphores. 315 */ 316 struct sema { 317 int value; 318 pid_t holder; 319 char *name; 320 int prio; 321 int timo; 322 }; 323 static void sema_init __P((struct sema *, char *, int, int)); 324 static int sema_get __P((struct sema *, struct lockit *)); 325 static void sema_release __P((struct sema *)); 326 327 static void 328 sema_init(semap, name, prio, timo) 329 struct sema *semap; 330 char *name; 331 int prio, timo; 332 { 333 334 semap->holder = -1; 335 semap->value = 0; 336 semap->name = name; 337 semap->prio = prio; 338 semap->timo = timo; 339 } 340 341 static int 342 sema_get(semap, interlock) 343 struct sema *semap; 344 struct lockit *interlock; 345 { 346 347 if (semap->value++ > 0) { 348 if (interlock != NULL) 349 FREE_LOCK_INTERLOCKED(interlock); 350 tsleep((caddr_t)semap, semap->prio, semap->name, semap->timo); 351 if (interlock != NULL) { 352 ACQUIRE_LOCK_INTERLOCKED(interlock); 353 FREE_LOCK(interlock); 354 } 355 return (0); 356 } 357 semap->holder = CURPROC->p_pid; 358 if (interlock != NULL) 359 FREE_LOCK(interlock); 360 return (1); 361 } 362 363 static void 364 sema_release(semap) 365 struct sema *semap; 366 { 367 368 if (semap->value <= 0 || semap->holder != CURPROC->p_pid) 369 panic("sema_release: not held"); 370 if (--semap->value > 0) { 371 semap->value = 0; 372 wakeup(semap); 373 } 374 semap->holder = -1; 375 } 376 377 /* 378 * Worklist queue management. 379 * These routines require that the lock be held. 380 */ 381 #ifndef /* NOT */ DEBUG 382 #define WORKLIST_INSERT(head, item) do { \ 383 (item)->wk_state |= ONWORKLIST; \ 384 LIST_INSERT_HEAD(head, item, wk_list); \ 385 } while (0) 386 #define WORKLIST_REMOVE(item) do { \ 387 (item)->wk_state &= ~ONWORKLIST; \ 388 LIST_REMOVE(item, wk_list); \ 389 } while (0) 390 #define WORKITEM_FREE(item, type) FREE(item, DtoM(type)) 391 392 #else /* DEBUG */ 393 static void worklist_insert __P((struct workhead *, struct worklist *)); 394 static void worklist_remove __P((struct worklist *)); 395 static void workitem_free __P((struct worklist *, int)); 396 397 #define WORKLIST_INSERT(head, item) worklist_insert(head, item) 398 #define WORKLIST_REMOVE(item) worklist_remove(item) 399 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type) 400 401 static void 402 worklist_insert(head, item) 403 struct workhead *head; 404 struct worklist *item; 405 { 406 407 if (lk.lkt_held == -1) 408 panic("worklist_insert: lock not held"); 409 if (item->wk_state & ONWORKLIST) 410 panic("worklist_insert: already on list"); 411 item->wk_state |= ONWORKLIST; 412 LIST_INSERT_HEAD(head, item, wk_list); 413 } 414 415 static void 416 worklist_remove(item) 417 struct worklist *item; 418 { 419 420 if (lk.lkt_held == -1) 421 panic("worklist_remove: lock not held"); 422 if ((item->wk_state & ONWORKLIST) == 0) 423 panic("worklist_remove: not on list"); 424 item->wk_state &= ~ONWORKLIST; 425 LIST_REMOVE(item, wk_list); 426 } 427 428 static void 429 workitem_free(item, type) 430 struct worklist *item; 431 int type; 432 { 433 434 if (item->wk_state & ONWORKLIST) 435 panic("workitem_free: still on list"); 436 if (item->wk_type != type) 437 panic("workitem_free: type mismatch"); 438 FREE(item, DtoM(type)); 439 } 440 #endif /* DEBUG */ 441 442 /* 443 * Workitem queue management 444 */ 445 static struct workhead softdep_workitem_pending; 446 static int softdep_worklist_busy; 447 static int max_softdeps; /* maximum number of structs before slowdown */ 448 static int tickdelay = 2; /* number of ticks to pause during slowdown */ 449 static int proc_waiting; /* tracks whether we have a timeout posted */ 450 static struct proc *filesys_syncer; /* proc of filesystem syncer process */ 451 static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 452 #define FLUSH_INODES 1 453 static int req_clear_remove; /* syncer process flush some freeblks */ 454 #define FLUSH_REMOVE 2 455 /* 456 * runtime statistics 457 */ 458 static int stat_blk_limit_push; /* number of times block limit neared */ 459 static int stat_ino_limit_push; /* number of times inode limit neared */ 460 static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 461 static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 462 static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 463 static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 464 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 465 static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 466 #ifdef DEBUG 467 #include <vm/vm.h> 468 #include <sys/sysctl.h> 469 #if defined(__FreeBSD__) 470 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, ""); 471 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, ""); 472 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,""); 473 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,""); 474 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, ""); 475 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, ""); 476 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, ""); 477 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, ""); 478 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, ""); 479 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, ""); 480 #else /* !__FreeBSD__ */ 481 struct ctldebug debug20 = { "max_softdeps", &max_softdeps }; 482 struct ctldebug debug21 = { "tickdelay", &tickdelay }; 483 struct ctldebug debug23 = { "blk_limit_push", &stat_blk_limit_push }; 484 struct ctldebug debug24 = { "ino_limit_push", &stat_ino_limit_push }; 485 struct ctldebug debug25 = { "blk_limit_hit", &stat_blk_limit_hit }; 486 struct ctldebug debug26 = { "ino_limit_hit", &stat_ino_limit_hit }; 487 struct ctldebug debug27 = { "indir_blk_ptrs", &stat_indir_blk_ptrs }; 488 struct ctldebug debug28 = { "inode_bitmap", &stat_inode_bitmap }; 489 struct ctldebug debug29 = { "direct_blk_ptrs", &stat_direct_blk_ptrs }; 490 struct ctldebug debug30 = { "dir_entry", &stat_dir_entry }; 491 #endif /* !__FreeBSD__ */ 492 493 #endif /* DEBUG */ 494 495 /* 496 * Add an item to the end of the work queue. 497 * This routine requires that the lock be held. 498 * This is the only routine that adds items to the list. 499 * The following routine is the only one that removes items 500 * and does so in order from first to last. 501 */ 502 static void 503 add_to_worklist(wk) 504 struct worklist *wk; 505 { 506 static struct worklist *worklist_tail; 507 508 if (wk->wk_state & ONWORKLIST) 509 panic("add_to_worklist: already on list"); 510 wk->wk_state |= ONWORKLIST; 511 if (LIST_FIRST(&softdep_workitem_pending) == NULL) 512 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list); 513 else 514 LIST_INSERT_AFTER(worklist_tail, wk, wk_list); 515 worklist_tail = wk; 516 } 517 518 /* 519 * Process that runs once per second to handle items in the background queue. 520 * 521 * Note that we ensure that everything is done in the order in which they 522 * appear in the queue. The code below depends on this property to ensure 523 * that blocks of a file are freed before the inode itself is freed. This 524 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 525 * until all the old ones have been purged from the dependency lists. 526 */ 527 int 528 softdep_process_worklist(matchmnt) 529 struct mount *matchmnt; 530 { 531 struct proc *p = CURPROC; 532 struct worklist *wk; 533 struct fs *matchfs; 534 int matchcnt; 535 536 /* 537 * Record the process identifier of our caller so that we can give 538 * this process preferential treatment in request_cleanup below. 539 */ 540 filesys_syncer = p; 541 matchcnt = 0; 542 matchfs = NULL; 543 if (matchmnt != NULL) 544 matchfs = VFSTOUFS(matchmnt)->um_fs; 545 /* 546 * There is no danger of having multiple processes run this 547 * code. It is single threaded solely so that softdep_flushfiles 548 * (below) can get an accurate count of the number of items 549 * related to its mount point that are in the list. 550 */ 551 if (softdep_worklist_busy && matchmnt == NULL) 552 return (-1); 553 /* 554 * If requested, try removing inode or removal dependencies. 555 */ 556 if (req_clear_inodedeps) { 557 clear_inodedeps(p); 558 req_clear_inodedeps = 0; 559 wakeup(&proc_waiting); 560 } 561 if (req_clear_remove) { 562 clear_remove(p); 563 req_clear_remove = 0; 564 wakeup(&proc_waiting); 565 } 566 ACQUIRE_LOCK(&lk); 567 while ((wk = LIST_FIRST(&softdep_workitem_pending)) != 0) { 568 WORKLIST_REMOVE(wk); 569 FREE_LOCK(&lk); 570 switch (wk->wk_type) { 571 572 case D_DIRREM: 573 /* removal of a directory entry */ 574 if (WK_DIRREM(wk)->dm_mnt == matchmnt) 575 matchcnt += 1; 576 handle_workitem_remove(WK_DIRREM(wk)); 577 break; 578 579 case D_FREEBLKS: 580 /* releasing blocks and/or fragments from a file */ 581 if (WK_FREEBLKS(wk)->fb_fs == matchfs) 582 matchcnt += 1; 583 handle_workitem_freeblocks(WK_FREEBLKS(wk)); 584 break; 585 586 case D_FREEFRAG: 587 /* releasing a fragment when replaced as a file grows */ 588 if (WK_FREEFRAG(wk)->ff_fs == matchfs) 589 matchcnt += 1; 590 handle_workitem_freefrag(WK_FREEFRAG(wk)); 591 break; 592 593 case D_FREEFILE: 594 /* releasing an inode when its link count drops to 0 */ 595 if (WK_FREEFILE(wk)->fx_fs == matchfs) 596 matchcnt += 1; 597 handle_workitem_freefile(WK_FREEFILE(wk)); 598 break; 599 600 default: 601 panic("%s_process_worklist: Unknown type %s", 602 "softdep", TYPENAME(wk->wk_type)); 603 /* NOTREACHED */ 604 } 605 if (softdep_worklist_busy && matchmnt == NULL) 606 return (-1); 607 /* 608 * If requested, try removing inode or removal dependencies. 609 */ 610 if (req_clear_inodedeps) { 611 clear_inodedeps(p); 612 req_clear_inodedeps = 0; 613 wakeup(&proc_waiting); 614 } 615 if (req_clear_remove) { 616 clear_remove(p); 617 req_clear_remove = 0; 618 wakeup(&proc_waiting); 619 } 620 ACQUIRE_LOCK(&lk); 621 } 622 FREE_LOCK(&lk); 623 return (matchcnt); 624 } 625 626 /* 627 * Purge the work list of all items associated with a particular mount point. 628 */ 629 int 630 softdep_flushfiles(oldmnt, flags, p) 631 struct mount *oldmnt; 632 int flags; 633 struct proc *p; 634 { 635 struct vnode *devvp; 636 int error, loopcnt; 637 638 /* 639 * Await our turn to clear out the queue. 640 */ 641 while (softdep_worklist_busy) 642 tsleep(&lbolt, PRIBIO, "softflush", 0); 643 softdep_worklist_busy = 1; 644 if ((error = ffs_flushfiles(oldmnt, flags, p)) != 0) { 645 softdep_worklist_busy = 0; 646 return (error); 647 } 648 /* 649 * Alternately flush the block device associated with the mount 650 * point and process any dependencies that the flushing 651 * creates. In theory, this loop can happen at most twice, 652 * but we give it a few extra just to be sure. 653 */ 654 devvp = VFSTOUFS(oldmnt)->um_devvp; 655 for (loopcnt = 10; loopcnt > 0; loopcnt--) { 656 if (softdep_process_worklist(oldmnt) == 0) { 657 /* 658 * Do another flush in case any vnodes were brought in 659 * as part of the cleanup operations. 660 */ 661 if ((error = ffs_flushfiles(oldmnt, flags, p)) != 0) 662 break; 663 /* 664 * If we still found nothing to do, we are really done. 665 */ 666 if (softdep_process_worklist(oldmnt) == 0) 667 break; 668 } 669 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, p); 670 error = VOP_FSYNC(devvp, p->p_ucred, MNT_WAIT, p); 671 VOP_UNLOCK(devvp, 0, p); 672 if (error) 673 break; 674 } 675 softdep_worklist_busy = 0; 676 /* 677 * If we are unmounting then it is an error to fail. If we 678 * are simply trying to downgrade to read-only, then filesystem 679 * activity can keep us busy forever, so we just fail with EBUSY. 680 */ 681 if (loopcnt == 0) { 682 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 683 panic("softdep_flushfiles: looping"); 684 error = EBUSY; 685 } 686 return (error); 687 } 688 689 /* 690 * Structure hashing. 691 * 692 * There are three types of structures that can be looked up: 693 * 1) pagedep structures identified by mount point, inode number, 694 * and logical block. 695 * 2) inodedep structures identified by mount point and inode number. 696 * 3) newblk structures identified by mount point and 697 * physical block number. 698 * 699 * The "pagedep" and "inodedep" dependency structures are hashed 700 * separately from the file blocks and inodes to which they correspond. 701 * This separation helps when the in-memory copy of an inode or 702 * file block must be replaced. It also obviates the need to access 703 * an inode or file page when simply updating (or de-allocating) 704 * dependency structures. Lookup of newblk structures is needed to 705 * find newly allocated blocks when trying to associate them with 706 * their allocdirect or allocindir structure. 707 * 708 * The lookup routines optionally create and hash a new instance when 709 * an existing entry is not found. 710 */ 711 #define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 712 713 /* 714 * Structures and routines associated with pagedep caching. 715 */ 716 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl; 717 u_long pagedep_hash; /* size of hash table - 1 */ 718 #define PAGEDEP_HASH(mp, inum, lbn) \ 719 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \ 720 pagedep_hash]) 721 static struct sema pagedep_in_progress; 722 723 /* 724 * Look up a pagedep. Return 1 if found, 0 if not found. 725 * If not found, allocate if DEPALLOC flag is passed. 726 * Found or allocated entry is returned in pagedeppp. 727 * This routine must be called with splbio interrupts blocked. 728 */ 729 static int 730 pagedep_lookup(ip, lbn, flags, pagedeppp) 731 struct inode *ip; 732 ufs_lbn_t lbn; 733 int flags; 734 struct pagedep **pagedeppp; 735 { 736 struct pagedep *pagedep; 737 struct pagedep_hashhead *pagedephd; 738 struct mount *mp; 739 int i; 740 741 #ifdef DEBUG 742 if (lk.lkt_held == -1) 743 panic("pagedep_lookup: lock not held"); 744 #endif 745 mp = ITOV(ip)->v_mount; 746 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn); 747 top: 748 for (pagedep = LIST_FIRST(pagedephd); pagedep; 749 pagedep = LIST_NEXT(pagedep, pd_hash)) 750 if (ip->i_number == pagedep->pd_ino && 751 lbn == pagedep->pd_lbn && 752 mp == pagedep->pd_mnt) 753 break; 754 if (pagedep) { 755 *pagedeppp = pagedep; 756 return (1); 757 } 758 if ((flags & DEPALLOC) == 0) { 759 *pagedeppp = NULL; 760 return (0); 761 } 762 if (sema_get(&pagedep_in_progress, &lk) == 0) { 763 ACQUIRE_LOCK(&lk); 764 goto top; 765 } 766 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP, 767 M_WAITOK); 768 bzero(pagedep, sizeof(struct pagedep)); 769 pagedep->pd_list.wk_type = D_PAGEDEP; 770 pagedep->pd_mnt = mp; 771 pagedep->pd_ino = ip->i_number; 772 pagedep->pd_lbn = lbn; 773 LIST_INIT(&pagedep->pd_dirremhd); 774 LIST_INIT(&pagedep->pd_pendinghd); 775 for (i = 0; i < DAHASHSZ; i++) 776 LIST_INIT(&pagedep->pd_diraddhd[i]); 777 ACQUIRE_LOCK(&lk); 778 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 779 sema_release(&pagedep_in_progress); 780 *pagedeppp = pagedep; 781 return (0); 782 } 783 784 /* 785 * Structures and routines associated with inodedep caching. 786 */ 787 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl; 788 static u_long inodedep_hash; /* size of hash table - 1 */ 789 static long num_inodedep; /* number of inodedep allocated */ 790 #define INODEDEP_HASH(fs, inum) \ 791 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash]) 792 static struct sema inodedep_in_progress; 793 794 /* 795 * Look up a inodedep. Return 1 if found, 0 if not found. 796 * If not found, allocate if DEPALLOC flag is passed. 797 * Found or allocated entry is returned in inodedeppp. 798 * This routine must be called with splbio interrupts blocked. 799 */ 800 static int 801 inodedep_lookup(fs, inum, flags, inodedeppp) 802 struct fs *fs; 803 ino_t inum; 804 int flags; 805 struct inodedep **inodedeppp; 806 { 807 struct inodedep *inodedep; 808 struct inodedep_hashhead *inodedephd; 809 int firsttry; 810 811 #ifdef DEBUG 812 if (lk.lkt_held == -1) 813 panic("inodedep_lookup: lock not held"); 814 #endif 815 firsttry = 1; 816 inodedephd = INODEDEP_HASH(fs, inum); 817 top: 818 for (inodedep = LIST_FIRST(inodedephd); inodedep; 819 inodedep = LIST_NEXT(inodedep, id_hash)) 820 if (inum == inodedep->id_ino && fs == inodedep->id_fs) 821 break; 822 if (inodedep) { 823 *inodedeppp = inodedep; 824 return (1); 825 } 826 if ((flags & DEPALLOC) == 0) { 827 *inodedeppp = NULL; 828 return (0); 829 } 830 /* 831 * If we are over our limit, try to improve the situation. 832 */ 833 if (num_inodedep > max_softdeps && firsttry && speedup_syncer() == 0 && 834 request_cleanup(FLUSH_INODES, 1)) { 835 firsttry = 0; 836 goto top; 837 } 838 if (sema_get(&inodedep_in_progress, &lk) == 0) { 839 ACQUIRE_LOCK(&lk); 840 goto top; 841 } 842 num_inodedep += 1; 843 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep), 844 M_INODEDEP, M_WAITOK); 845 inodedep->id_list.wk_type = D_INODEDEP; 846 inodedep->id_fs = fs; 847 inodedep->id_ino = inum; 848 inodedep->id_state = ALLCOMPLETE; 849 inodedep->id_nlinkdelta = 0; 850 inodedep->id_savedino = NULL; 851 inodedep->id_savedsize = -1; 852 inodedep->id_buf = NULL; 853 LIST_INIT(&inodedep->id_pendinghd); 854 LIST_INIT(&inodedep->id_inowait); 855 LIST_INIT(&inodedep->id_bufwait); 856 TAILQ_INIT(&inodedep->id_inoupdt); 857 TAILQ_INIT(&inodedep->id_newinoupdt); 858 ACQUIRE_LOCK(&lk); 859 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 860 sema_release(&inodedep_in_progress); 861 *inodedeppp = inodedep; 862 return (0); 863 } 864 865 /* 866 * Structures and routines associated with newblk caching. 867 */ 868 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl; 869 u_long newblk_hash; /* size of hash table - 1 */ 870 #define NEWBLK_HASH(fs, inum) \ 871 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash]) 872 static struct sema newblk_in_progress; 873 874 /* 875 * Look up a newblk. Return 1 if found, 0 if not found. 876 * If not found, allocate if DEPALLOC flag is passed. 877 * Found or allocated entry is returned in newblkpp. 878 */ 879 static int 880 newblk_lookup(fs, newblkno, flags, newblkpp) 881 struct fs *fs; 882 ufs_daddr_t newblkno; 883 int flags; 884 struct newblk **newblkpp; 885 { 886 struct newblk *newblk; 887 struct newblk_hashhead *newblkhd; 888 889 newblkhd = NEWBLK_HASH(fs, newblkno); 890 top: 891 for (newblk = LIST_FIRST(newblkhd); newblk; 892 newblk = LIST_NEXT(newblk, nb_hash)) 893 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs) 894 break; 895 if (newblk) { 896 *newblkpp = newblk; 897 return (1); 898 } 899 if ((flags & DEPALLOC) == 0) { 900 *newblkpp = NULL; 901 return (0); 902 } 903 if (sema_get(&newblk_in_progress, 0) == 0) 904 goto top; 905 MALLOC(newblk, struct newblk *, sizeof(struct newblk), 906 M_NEWBLK, M_WAITOK); 907 newblk->nb_state = 0; 908 newblk->nb_fs = fs; 909 newblk->nb_newblkno = newblkno; 910 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 911 sema_release(&newblk_in_progress); 912 *newblkpp = newblk; 913 return (0); 914 } 915 916 /* 917 * Executed during filesystem system initialization before 918 * mounting any file systems. 919 */ 920 void 921 softdep_initialize() 922 { 923 924 LIST_INIT(&mkdirlisthd); 925 LIST_INIT(&softdep_workitem_pending); 926 max_softdeps = desiredvnodes * 8; 927 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 928 &pagedep_hash); 929 sema_init(&pagedep_in_progress, "pagedep", PRIBIO, 0); 930 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash); 931 sema_init(&inodedep_in_progress, "inodedep", PRIBIO, 0); 932 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash); 933 sema_init(&newblk_in_progress, "newblk", PRIBIO, 0); 934 } 935 936 /* 937 * Called at mount time to notify the dependency code that a 938 * filesystem wishes to use it. 939 */ 940 int 941 softdep_mount(devvp, mp, fs, cred) 942 struct vnode *devvp; 943 struct mount *mp; 944 struct fs *fs; 945 struct ucred *cred; 946 { 947 struct csum cstotal; 948 struct cg *cgp; 949 struct buf *bp; 950 int error, cyl; 951 952 mp->mnt_flag &= ~MNT_ASYNC; 953 mp->mnt_flag |= MNT_SOFTDEP; 954 /* 955 * When doing soft updates, the counters in the 956 * superblock may have gotten out of sync, so we have 957 * to scan the cylinder groups and recalculate them. 958 */ 959 if (fs->fs_clean != 0) 960 return (0); 961 bzero(&cstotal, sizeof cstotal); 962 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 963 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 964 fs->fs_cgsize, cred, &bp)) != 0) { 965 brelse(bp); 966 return (error); 967 } 968 cgp = (struct cg *)bp->b_data; 969 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 970 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 971 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 972 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 973 fs->fs_cs(fs, cyl) = cgp->cg_cs; 974 brelse(bp); 975 } 976 #ifdef DEBUG 977 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 978 printf("ffs_mountfs: superblock updated for soft updates\n"); 979 #endif 980 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 981 return (0); 982 } 983 984 /* 985 * Protecting the freemaps (or bitmaps). 986 * 987 * To eliminate the need to execute fsck before mounting a file system 988 * after a power failure, one must (conservatively) guarantee that the 989 * on-disk copy of the bitmaps never indicate that a live inode or block is 990 * free. So, when a block or inode is allocated, the bitmap should be 991 * updated (on disk) before any new pointers. When a block or inode is 992 * freed, the bitmap should not be updated until all pointers have been 993 * reset. The latter dependency is handled by the delayed de-allocation 994 * approach described below for block and inode de-allocation. The former 995 * dependency is handled by calling the following procedure when a block or 996 * inode is allocated. When an inode is allocated an "inodedep" is created 997 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 998 * Each "inodedep" is also inserted into the hash indexing structure so 999 * that any additional link additions can be made dependent on the inode 1000 * allocation. 1001 * 1002 * The ufs file system maintains a number of free block counts (e.g., per 1003 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 1004 * in addition to the bitmaps. These counts are used to improve efficiency 1005 * during allocation and therefore must be consistent with the bitmaps. 1006 * There is no convenient way to guarantee post-crash consistency of these 1007 * counts with simple update ordering, for two main reasons: (1) The counts 1008 * and bitmaps for a single cylinder group block are not in the same disk 1009 * sector. If a disk write is interrupted (e.g., by power failure), one may 1010 * be written and the other not. (2) Some of the counts are located in the 1011 * superblock rather than the cylinder group block. So, we focus our soft 1012 * updates implementation on protecting the bitmaps. When mounting a 1013 * filesystem, we recompute the auxiliary counts from the bitmaps. 1014 */ 1015 1016 /* 1017 * Called just after updating the cylinder group block to allocate an inode. 1018 */ 1019 void 1020 softdep_setup_inomapdep(bp, ip, newinum) 1021 struct buf *bp; /* buffer for cylgroup block with inode map */ 1022 struct inode *ip; /* inode related to allocation */ 1023 ino_t newinum; /* new inode number being allocated */ 1024 { 1025 struct inodedep *inodedep; 1026 struct bmsafemap *bmsafemap; 1027 1028 /* 1029 * Create a dependency for the newly allocated inode. 1030 * Panic if it already exists as something is seriously wrong. 1031 * Otherwise add it to the dependency list for the buffer holding 1032 * the cylinder group map from which it was allocated. 1033 */ 1034 ACQUIRE_LOCK(&lk); 1035 if (inodedep_lookup(ip->i_fs, newinum, DEPALLOC, &inodedep) != 0) 1036 panic("softdep_setup_inomapdep: found inode"); 1037 inodedep->id_buf = bp; 1038 inodedep->id_state &= ~DEPCOMPLETE; 1039 bmsafemap = bmsafemap_lookup(bp); 1040 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 1041 FREE_LOCK(&lk); 1042 } 1043 1044 /* 1045 * Called just after updating the cylinder group block to 1046 * allocate block or fragment. 1047 */ 1048 void 1049 softdep_setup_blkmapdep(bp, fs, newblkno) 1050 struct buf *bp; /* buffer for cylgroup block with block map */ 1051 struct fs *fs; /* filesystem doing allocation */ 1052 ufs_daddr_t newblkno; /* number of newly allocated block */ 1053 { 1054 struct newblk *newblk; 1055 struct bmsafemap *bmsafemap; 1056 1057 /* 1058 * Create a dependency for the newly allocated block. 1059 * Add it to the dependency list for the buffer holding 1060 * the cylinder group map from which it was allocated. 1061 */ 1062 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0) 1063 panic("softdep_setup_blkmapdep: found block"); 1064 ACQUIRE_LOCK(&lk); 1065 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp); 1066 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 1067 FREE_LOCK(&lk); 1068 } 1069 1070 /* 1071 * Find the bmsafemap associated with a cylinder group buffer. 1072 * If none exists, create one. The buffer must be locked when 1073 * this routine is called and this routine must be called with 1074 * splbio interrupts blocked. 1075 */ 1076 static struct bmsafemap * 1077 bmsafemap_lookup(bp) 1078 struct buf *bp; 1079 { 1080 struct bmsafemap *bmsafemap; 1081 struct worklist *wk; 1082 1083 #ifdef DEBUG 1084 if (lk.lkt_held == -1) 1085 panic("bmsafemap_lookup: lock not held"); 1086 #endif 1087 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = LIST_NEXT(wk, wk_list)) 1088 if (wk->wk_type == D_BMSAFEMAP) 1089 return (WK_BMSAFEMAP(wk)); 1090 FREE_LOCK(&lk); 1091 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap), 1092 M_BMSAFEMAP, M_WAITOK); 1093 bmsafemap->sm_list.wk_type = D_BMSAFEMAP; 1094 bmsafemap->sm_list.wk_state = 0; 1095 bmsafemap->sm_buf = bp; 1096 LIST_INIT(&bmsafemap->sm_allocdirecthd); 1097 LIST_INIT(&bmsafemap->sm_allocindirhd); 1098 LIST_INIT(&bmsafemap->sm_inodedephd); 1099 LIST_INIT(&bmsafemap->sm_newblkhd); 1100 ACQUIRE_LOCK(&lk); 1101 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 1102 return (bmsafemap); 1103 } 1104 1105 /* 1106 * Direct block allocation dependencies. 1107 * 1108 * When a new block is allocated, the corresponding disk locations must be 1109 * initialized (with zeros or new data) before the on-disk inode points to 1110 * them. Also, the freemap from which the block was allocated must be 1111 * updated (on disk) before the inode's pointer. These two dependencies are 1112 * independent of each other and are needed for all file blocks and indirect 1113 * blocks that are pointed to directly by the inode. Just before the 1114 * "in-core" version of the inode is updated with a newly allocated block 1115 * number, a procedure (below) is called to setup allocation dependency 1116 * structures. These structures are removed when the corresponding 1117 * dependencies are satisfied or when the block allocation becomes obsolete 1118 * (i.e., the file is deleted, the block is de-allocated, or the block is a 1119 * fragment that gets upgraded). All of these cases are handled in 1120 * procedures described later. 1121 * 1122 * When a file extension causes a fragment to be upgraded, either to a larger 1123 * fragment or to a full block, the on-disk location may change (if the 1124 * previous fragment could not simply be extended). In this case, the old 1125 * fragment must be de-allocated, but not until after the inode's pointer has 1126 * been updated. In most cases, this is handled by later procedures, which 1127 * will construct a "freefrag" structure to be added to the workitem queue 1128 * when the inode update is complete (or obsolete). The main exception to 1129 * this is when an allocation occurs while a pending allocation dependency 1130 * (for the same block pointer) remains. This case is handled in the main 1131 * allocation dependency setup procedure by immediately freeing the 1132 * unreferenced fragments. 1133 */ 1134 void 1135 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 1136 struct inode *ip; /* inode to which block is being added */ 1137 ufs_lbn_t lbn; /* block pointer within inode */ 1138 ufs_daddr_t newblkno; /* disk block number being added */ 1139 ufs_daddr_t oldblkno; /* previous block number, 0 unless frag */ 1140 long newsize; /* size of new block */ 1141 long oldsize; /* size of new block */ 1142 struct buf *bp; /* bp for allocated block */ 1143 { 1144 struct allocdirect *adp, *oldadp; 1145 struct allocdirectlst *adphead; 1146 struct bmsafemap *bmsafemap; 1147 struct inodedep *inodedep; 1148 struct pagedep *pagedep; 1149 struct newblk *newblk; 1150 1151 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect), 1152 M_ALLOCDIRECT, M_WAITOK); 1153 bzero(adp, sizeof(struct allocdirect)); 1154 adp->ad_list.wk_type = D_ALLOCDIRECT; 1155 adp->ad_lbn = lbn; 1156 adp->ad_newblkno = newblkno; 1157 adp->ad_oldblkno = oldblkno; 1158 adp->ad_newsize = newsize; 1159 adp->ad_oldsize = oldsize; 1160 adp->ad_state = ATTACHED; 1161 if (newblkno == oldblkno) 1162 adp->ad_freefrag = NULL; 1163 else 1164 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1165 1166 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1167 panic("softdep_setup_allocdirect: lost block"); 1168 1169 ACQUIRE_LOCK(&lk); 1170 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 1171 adp->ad_inodedep = inodedep; 1172 1173 if (newblk->nb_state == DEPCOMPLETE) { 1174 adp->ad_state |= DEPCOMPLETE; 1175 adp->ad_buf = NULL; 1176 } else { 1177 bmsafemap = newblk->nb_bmsafemap; 1178 adp->ad_buf = bmsafemap->sm_buf; 1179 LIST_REMOVE(newblk, nb_deps); 1180 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1181 } 1182 LIST_REMOVE(newblk, nb_hash); 1183 FREE(newblk, M_NEWBLK); 1184 1185 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list); 1186 if (lbn >= NDADDR) { 1187 /* allocating an indirect block */ 1188 if (oldblkno != 0) 1189 panic("softdep_setup_allocdirect: non-zero indir"); 1190 } else { 1191 /* 1192 * Allocating a direct block. 1193 * 1194 * If we are allocating a directory block, then we must 1195 * allocate an associated pagedep to track additions and 1196 * deletions. 1197 */ 1198 if ((ip->i_mode & IFMT) == IFDIR && 1199 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1200 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 1201 } 1202 /* 1203 * The list of allocdirects must be kept in sorted and ascending 1204 * order so that the rollback routines can quickly determine the 1205 * first uncommitted block (the size of the file stored on disk 1206 * ends at the end of the lowest committed fragment, or if there 1207 * are no fragments, at the end of the highest committed block). 1208 * Since files generally grow, the typical case is that the new 1209 * block is to be added at the end of the list. We speed this 1210 * special case by checking against the last allocdirect in the 1211 * list before laboriously traversing the list looking for the 1212 * insertion point. 1213 */ 1214 adphead = &inodedep->id_newinoupdt; 1215 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1216 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1217 /* insert at end of list */ 1218 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1219 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1220 allocdirect_merge(adphead, adp, oldadp); 1221 FREE_LOCK(&lk); 1222 return; 1223 } 1224 for (oldadp = TAILQ_FIRST(adphead); oldadp; 1225 oldadp = TAILQ_NEXT(oldadp, ad_next)) { 1226 if (oldadp->ad_lbn >= lbn) 1227 break; 1228 } 1229 if (oldadp == NULL) 1230 panic("softdep_setup_allocdirect: lost entry"); 1231 /* insert in middle of list */ 1232 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1233 if (oldadp->ad_lbn == lbn) 1234 allocdirect_merge(adphead, adp, oldadp); 1235 FREE_LOCK(&lk); 1236 } 1237 1238 /* 1239 * Replace an old allocdirect dependency with a newer one. 1240 * This routine must be called with splbio interrupts blocked. 1241 */ 1242 static void 1243 allocdirect_merge(adphead, newadp, oldadp) 1244 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 1245 struct allocdirect *newadp; /* allocdirect being added */ 1246 struct allocdirect *oldadp; /* existing allocdirect being checked */ 1247 { 1248 struct freefrag *freefrag; 1249 1250 #ifdef DEBUG 1251 if (lk.lkt_held == -1) 1252 panic("allocdirect_merge: lock not held"); 1253 #endif 1254 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 1255 newadp->ad_oldsize != oldadp->ad_newsize || 1256 newadp->ad_lbn >= NDADDR) 1257 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d", 1258 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn, 1259 NDADDR); 1260 newadp->ad_oldblkno = oldadp->ad_oldblkno; 1261 newadp->ad_oldsize = oldadp->ad_oldsize; 1262 /* 1263 * If the old dependency had a fragment to free or had never 1264 * previously had a block allocated, then the new dependency 1265 * can immediately post its freefrag and adopt the old freefrag. 1266 * This action is done by swapping the freefrag dependencies. 1267 * The new dependency gains the old one's freefrag, and the 1268 * old one gets the new one and then immediately puts it on 1269 * the worklist when it is freed by free_allocdirect. It is 1270 * not possible to do this swap when the old dependency had a 1271 * non-zero size but no previous fragment to free. This condition 1272 * arises when the new block is an extension of the old block. 1273 * Here, the first part of the fragment allocated to the new 1274 * dependency is part of the block currently claimed on disk by 1275 * the old dependency, so cannot legitimately be freed until the 1276 * conditions for the new dependency are fulfilled. 1277 */ 1278 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 1279 freefrag = newadp->ad_freefrag; 1280 newadp->ad_freefrag = oldadp->ad_freefrag; 1281 oldadp->ad_freefrag = freefrag; 1282 } 1283 free_allocdirect(adphead, oldadp, 0); 1284 } 1285 1286 /* 1287 * Allocate a new freefrag structure if needed. 1288 */ 1289 static struct freefrag * 1290 newfreefrag(ip, blkno, size) 1291 struct inode *ip; 1292 ufs_daddr_t blkno; 1293 long size; 1294 { 1295 struct freefrag *freefrag; 1296 struct fs *fs; 1297 1298 if (blkno == 0) 1299 return (NULL); 1300 fs = ip->i_fs; 1301 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 1302 panic("newfreefrag: frag size"); 1303 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag), 1304 M_FREEFRAG, M_WAITOK); 1305 freefrag->ff_list.wk_type = D_FREEFRAG; 1306 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */ 1307 freefrag->ff_inum = ip->i_number; 1308 freefrag->ff_fs = fs; 1309 freefrag->ff_devvp = ip->i_devvp; 1310 freefrag->ff_blkno = blkno; 1311 freefrag->ff_fragsize = size; 1312 return (freefrag); 1313 } 1314 1315 /* 1316 * This workitem de-allocates fragments that were replaced during 1317 * file block allocation. 1318 */ 1319 static void 1320 handle_workitem_freefrag(freefrag) 1321 struct freefrag *freefrag; 1322 { 1323 struct inode tip; 1324 1325 tip.i_fs = freefrag->ff_fs; 1326 tip.i_devvp = freefrag->ff_devvp; 1327 tip.i_dev = freefrag->ff_devvp->v_rdev; 1328 tip.i_number = freefrag->ff_inum; 1329 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */ 1330 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize); 1331 FREE(freefrag, M_FREEFRAG); 1332 } 1333 1334 /* 1335 * Indirect block allocation dependencies. 1336 * 1337 * The same dependencies that exist for a direct block also exist when 1338 * a new block is allocated and pointed to by an entry in a block of 1339 * indirect pointers. The undo/redo states described above are also 1340 * used here. Because an indirect block contains many pointers that 1341 * may have dependencies, a second copy of the entire in-memory indirect 1342 * block is kept. The buffer cache copy is always completely up-to-date. 1343 * The second copy, which is used only as a source for disk writes, 1344 * contains only the safe pointers (i.e., those that have no remaining 1345 * update dependencies). The second copy is freed when all pointers 1346 * are safe. The cache is not allowed to replace indirect blocks with 1347 * pending update dependencies. If a buffer containing an indirect 1348 * block with dependencies is written, these routines will mark it 1349 * dirty again. It can only be successfully written once all the 1350 * dependencies are removed. The ffs_fsync routine in conjunction with 1351 * softdep_sync_metadata work together to get all the dependencies 1352 * removed so that a file can be successfully written to disk. Three 1353 * procedures are used when setting up indirect block pointer 1354 * dependencies. The division is necessary because of the organization 1355 * of the "balloc" routine and because of the distinction between file 1356 * pages and file metadata blocks. 1357 */ 1358 1359 /* 1360 * Allocate a new allocindir structure. 1361 */ 1362 static struct allocindir * 1363 newallocindir(ip, ptrno, newblkno, oldblkno) 1364 struct inode *ip; /* inode for file being extended */ 1365 int ptrno; /* offset of pointer in indirect block */ 1366 ufs_daddr_t newblkno; /* disk block number being added */ 1367 ufs_daddr_t oldblkno; /* previous block number, 0 if none */ 1368 { 1369 struct allocindir *aip; 1370 1371 MALLOC(aip, struct allocindir *, sizeof(struct allocindir), 1372 M_ALLOCINDIR, M_WAITOK); 1373 bzero(aip, sizeof(struct allocindir)); 1374 aip->ai_list.wk_type = D_ALLOCINDIR; 1375 aip->ai_state = ATTACHED; 1376 aip->ai_offset = ptrno; 1377 aip->ai_newblkno = newblkno; 1378 aip->ai_oldblkno = oldblkno; 1379 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize); 1380 return (aip); 1381 } 1382 1383 /* 1384 * Called just before setting an indirect block pointer 1385 * to a newly allocated file page. 1386 */ 1387 void 1388 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 1389 struct inode *ip; /* inode for file being extended */ 1390 ufs_lbn_t lbn; /* allocated block number within file */ 1391 struct buf *bp; /* buffer with indirect blk referencing page */ 1392 int ptrno; /* offset of pointer in indirect block */ 1393 ufs_daddr_t newblkno; /* disk block number being added */ 1394 ufs_daddr_t oldblkno; /* previous block number, 0 if none */ 1395 struct buf *nbp; /* buffer holding allocated page */ 1396 { 1397 struct allocindir *aip; 1398 struct pagedep *pagedep; 1399 1400 aip = newallocindir(ip, ptrno, newblkno, oldblkno); 1401 ACQUIRE_LOCK(&lk); 1402 /* 1403 * If we are allocating a directory page, then we must 1404 * allocate an associated pagedep to track additions and 1405 * deletions. 1406 */ 1407 if ((ip->i_mode & IFMT) == IFDIR && 1408 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1409 WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list); 1410 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1411 FREE_LOCK(&lk); 1412 setup_allocindir_phase2(bp, ip, aip); 1413 } 1414 1415 /* 1416 * Called just before setting an indirect block pointer to a 1417 * newly allocated indirect block. 1418 */ 1419 void 1420 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 1421 struct buf *nbp; /* newly allocated indirect block */ 1422 struct inode *ip; /* inode for file being extended */ 1423 struct buf *bp; /* indirect block referencing allocated block */ 1424 int ptrno; /* offset of pointer in indirect block */ 1425 ufs_daddr_t newblkno; /* disk block number being added */ 1426 { 1427 struct allocindir *aip; 1428 1429 aip = newallocindir(ip, ptrno, newblkno, 0); 1430 ACQUIRE_LOCK(&lk); 1431 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1432 FREE_LOCK(&lk); 1433 setup_allocindir_phase2(bp, ip, aip); 1434 } 1435 1436 /* 1437 * Called to finish the allocation of the "aip" allocated 1438 * by one of the two routines above. 1439 */ 1440 static void 1441 setup_allocindir_phase2(bp, ip, aip) 1442 struct buf *bp; /* in-memory copy of the indirect block */ 1443 struct inode *ip; /* inode for file being extended */ 1444 struct allocindir *aip; /* allocindir allocated by the above routines */ 1445 { 1446 struct worklist *wk; 1447 struct indirdep *indirdep, *newindirdep; 1448 struct bmsafemap *bmsafemap; 1449 struct allocindir *oldaip; 1450 struct freefrag *freefrag; 1451 struct newblk *newblk; 1452 1453 if (bp->b_lblkno >= 0) 1454 panic("setup_allocindir_phase2: not indir blk"); 1455 for (indirdep = NULL, newindirdep = NULL; ; ) { 1456 ACQUIRE_LOCK(&lk); 1457 for (wk = LIST_FIRST(&bp->b_dep); wk; 1458 wk = LIST_NEXT(wk, wk_list)) { 1459 if (wk->wk_type != D_INDIRDEP) 1460 continue; 1461 indirdep = WK_INDIRDEP(wk); 1462 break; 1463 } 1464 if (indirdep == NULL && newindirdep) { 1465 indirdep = newindirdep; 1466 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 1467 newindirdep = NULL; 1468 } 1469 FREE_LOCK(&lk); 1470 if (indirdep) { 1471 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0, 1472 &newblk) == 0) 1473 panic("setup_allocindir: lost block"); 1474 ACQUIRE_LOCK(&lk); 1475 if (newblk->nb_state == DEPCOMPLETE) { 1476 aip->ai_state |= DEPCOMPLETE; 1477 aip->ai_buf = NULL; 1478 } else { 1479 bmsafemap = newblk->nb_bmsafemap; 1480 aip->ai_buf = bmsafemap->sm_buf; 1481 LIST_REMOVE(newblk, nb_deps); 1482 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd, 1483 aip, ai_deps); 1484 } 1485 LIST_REMOVE(newblk, nb_hash); 1486 FREE(newblk, M_NEWBLK); 1487 aip->ai_indirdep = indirdep; 1488 /* 1489 * Check to see if there is an existing dependency 1490 * for this block. If there is, merge the old 1491 * dependency into the new one. 1492 */ 1493 if (aip->ai_oldblkno == 0) 1494 oldaip = NULL; 1495 else 1496 for (oldaip=LIST_FIRST(&indirdep->ir_deplisthd); 1497 oldaip; oldaip = LIST_NEXT(oldaip, ai_next)) 1498 if (oldaip->ai_offset == aip->ai_offset) 1499 break; 1500 if (oldaip != NULL) { 1501 if (oldaip->ai_newblkno != aip->ai_oldblkno) 1502 panic("setup_allocindir_phase2: blkno"); 1503 aip->ai_oldblkno = oldaip->ai_oldblkno; 1504 freefrag = oldaip->ai_freefrag; 1505 oldaip->ai_freefrag = aip->ai_freefrag; 1506 aip->ai_freefrag = freefrag; 1507 free_allocindir(oldaip, NULL); 1508 } 1509 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 1510 ((ufs_daddr_t *)indirdep->ir_savebp->b_data) 1511 [aip->ai_offset] = aip->ai_oldblkno; 1512 FREE_LOCK(&lk); 1513 } 1514 if (newindirdep) { 1515 if (indirdep->ir_savebp != NULL) 1516 brelse(newindirdep->ir_savebp); 1517 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP); 1518 } 1519 if (indirdep) 1520 break; 1521 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep), 1522 M_INDIRDEP, M_WAITOK); 1523 newindirdep->ir_list.wk_type = D_INDIRDEP; 1524 newindirdep->ir_state = ATTACHED; 1525 LIST_INIT(&newindirdep->ir_deplisthd); 1526 LIST_INIT(&newindirdep->ir_donehd); 1527 if (bp->b_blkno == bp->b_lblkno) { 1528 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, 1529 NULL, NULL); 1530 } 1531 newindirdep->ir_savebp = 1532 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0); 1533 BUF_KERNPROC(newindirdep->ir_savebp); 1534 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 1535 } 1536 } 1537 1538 /* 1539 * Block de-allocation dependencies. 1540 * 1541 * When blocks are de-allocated, the on-disk pointers must be nullified before 1542 * the blocks are made available for use by other files. (The true 1543 * requirement is that old pointers must be nullified before new on-disk 1544 * pointers are set. We chose this slightly more stringent requirement to 1545 * reduce complexity.) Our implementation handles this dependency by updating 1546 * the inode (or indirect block) appropriately but delaying the actual block 1547 * de-allocation (i.e., freemap and free space count manipulation) until 1548 * after the updated versions reach stable storage. After the disk is 1549 * updated, the blocks can be safely de-allocated whenever it is convenient. 1550 * This implementation handles only the common case of reducing a file's 1551 * length to zero. Other cases are handled by the conventional synchronous 1552 * write approach. 1553 * 1554 * The ffs implementation with which we worked double-checks 1555 * the state of the block pointers and file size as it reduces 1556 * a file's length. Some of this code is replicated here in our 1557 * soft updates implementation. The freeblks->fb_chkcnt field is 1558 * used to transfer a part of this information to the procedure 1559 * that eventually de-allocates the blocks. 1560 * 1561 * This routine should be called from the routine that shortens 1562 * a file's length, before the inode's size or block pointers 1563 * are modified. It will save the block pointer information for 1564 * later release and zero the inode so that the calling routine 1565 * can release it. 1566 */ 1567 static long num_freeblks; /* number of freeblks allocated */ 1568 void 1569 softdep_setup_freeblocks(ip, length) 1570 struct inode *ip; /* The inode whose length is to be reduced */ 1571 off_t length; /* The new length for the file */ 1572 { 1573 struct freeblks *freeblks; 1574 struct inodedep *inodedep; 1575 struct allocdirect *adp; 1576 struct vnode *vp; 1577 struct buf *bp; 1578 struct fs *fs; 1579 int i, error; 1580 1581 fs = ip->i_fs; 1582 if (length != 0) 1583 panic("softde_setup_freeblocks: non-zero length"); 1584 /* 1585 * If we are over our limit, try to improve the situation. 1586 */ 1587 if (num_freeblks > max_softdeps / 2 && speedup_syncer() == 0) 1588 (void) request_cleanup(FLUSH_REMOVE, 0); 1589 num_freeblks += 1; 1590 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks), 1591 M_FREEBLKS, M_WAITOK); 1592 bzero(freeblks, sizeof(struct freeblks)); 1593 freeblks->fb_list.wk_type = D_FREEBLKS; 1594 freeblks->fb_uid = ip->i_uid; 1595 freeblks->fb_previousinum = ip->i_number; 1596 freeblks->fb_devvp = ip->i_devvp; 1597 freeblks->fb_fs = fs; 1598 freeblks->fb_oldsize = ip->i_size; 1599 freeblks->fb_newsize = length; 1600 freeblks->fb_chkcnt = ip->i_blocks; 1601 for (i = 0; i < NDADDR; i++) { 1602 freeblks->fb_dblks[i] = ip->i_db[i]; 1603 ip->i_db[i] = 0; 1604 } 1605 for (i = 0; i < NIADDR; i++) { 1606 freeblks->fb_iblks[i] = ip->i_ib[i]; 1607 ip->i_ib[i] = 0; 1608 } 1609 ip->i_blocks = 0; 1610 ip->i_size = 0; 1611 /* 1612 * Push the zero'ed inode to to its disk buffer so that we are free 1613 * to delete its dependencies below. Once the dependencies are gone 1614 * the buffer can be safely released. 1615 */ 1616 if ((error = bread(ip->i_devvp, 1617 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 1618 (int)fs->fs_bsize, NOCRED, &bp)) != 0) 1619 softdep_error("softdep_setup_freeblocks", error); 1620 *((struct dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) = 1621 ip->i_din; 1622 /* 1623 * Find and eliminate any inode dependencies. 1624 */ 1625 ACQUIRE_LOCK(&lk); 1626 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep); 1627 if ((inodedep->id_state & IOSTARTED) != 0) 1628 panic("softdep_setup_freeblocks: inode busy"); 1629 /* 1630 * Add the freeblks structure to the list of operations that 1631 * must await the zero'ed inode being written to disk. 1632 */ 1633 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list); 1634 /* 1635 * Because the file length has been truncated to zero, any 1636 * pending block allocation dependency structures associated 1637 * with this inode are obsolete and can simply be de-allocated. 1638 * We must first merge the two dependency lists to get rid of 1639 * any duplicate freefrag structures, then purge the merged list. 1640 */ 1641 merge_inode_lists(inodedep); 1642 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 1643 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 1644 FREE_LOCK(&lk); 1645 bdwrite(bp); 1646 /* 1647 * We must wait for any I/O in progress to finish so that 1648 * all potential buffers on the dirty list will be visible. 1649 * Once they are all there, walk the list and get rid of 1650 * any dependencies. 1651 */ 1652 vp = ITOV(ip); 1653 ACQUIRE_LOCK(&lk); 1654 drain_output(vp, 1); 1655 while (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT)) { 1656 bp = TAILQ_FIRST(&vp->v_dirtyblkhd); 1657 (void) inodedep_lookup(fs, ip->i_number, 0, &inodedep); 1658 deallocate_dependencies(bp, inodedep); 1659 bp->b_flags |= B_INVAL | B_NOCACHE; 1660 FREE_LOCK(&lk); 1661 brelse(bp); 1662 ACQUIRE_LOCK(&lk); 1663 } 1664 /* 1665 * Try freeing the inodedep in case that was the last dependency. 1666 */ 1667 if ((inodedep_lookup(fs, ip->i_number, 0, &inodedep)) != 0) 1668 (void) free_inodedep(inodedep); 1669 FREE_LOCK(&lk); 1670 } 1671 1672 /* 1673 * Reclaim any dependency structures from a buffer that is about to 1674 * be reallocated to a new vnode. The buffer must be locked, thus, 1675 * no I/O completion operations can occur while we are manipulating 1676 * its associated dependencies. The mutex is held so that other I/O's 1677 * associated with related dependencies do not occur. 1678 */ 1679 static void 1680 deallocate_dependencies(bp, inodedep) 1681 struct buf *bp; 1682 struct inodedep *inodedep; 1683 { 1684 struct worklist *wk; 1685 struct indirdep *indirdep; 1686 struct allocindir *aip; 1687 struct pagedep *pagedep; 1688 struct dirrem *dirrem; 1689 struct diradd *dap; 1690 int i; 1691 1692 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 1693 switch (wk->wk_type) { 1694 1695 case D_INDIRDEP: 1696 indirdep = WK_INDIRDEP(wk); 1697 /* 1698 * None of the indirect pointers will ever be visible, 1699 * so they can simply be tossed. GOINGAWAY ensures 1700 * that allocated pointers will be saved in the buffer 1701 * cache until they are freed. Note that they will 1702 * only be able to be found by their physical address 1703 * since the inode mapping the logical address will 1704 * be gone. The save buffer used for the safe copy 1705 * was allocated in setup_allocindir_phase2 using 1706 * the physical address so it could be used for this 1707 * purpose. Hence we swap the safe copy with the real 1708 * copy, allowing the safe copy to be freed and holding 1709 * on to the real copy for later use in indir_trunc. 1710 */ 1711 if (indirdep->ir_state & GOINGAWAY) 1712 panic("deallocate_dependencies: already gone"); 1713 indirdep->ir_state |= GOINGAWAY; 1714 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 1715 free_allocindir(aip, inodedep); 1716 if (bp->b_lblkno >= 0 || 1717 bp->b_blkno != indirdep->ir_savebp->b_lblkno) 1718 panic("deallocate_dependencies: not indir"); 1719 bcopy(bp->b_data, indirdep->ir_savebp->b_data, 1720 bp->b_bcount); 1721 WORKLIST_REMOVE(wk); 1722 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk); 1723 continue; 1724 1725 case D_PAGEDEP: 1726 pagedep = WK_PAGEDEP(wk); 1727 /* 1728 * None of the directory additions will ever be 1729 * visible, so they can simply be tossed. 1730 */ 1731 for (i = 0; i < DAHASHSZ; i++) 1732 while ((dap = 1733 LIST_FIRST(&pagedep->pd_diraddhd[i]))) 1734 free_diradd(dap); 1735 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0) 1736 free_diradd(dap); 1737 /* 1738 * Copy any directory remove dependencies to the list 1739 * to be processed after the zero'ed inode is written. 1740 * If the inode has already been written, then they 1741 * can be dumped directly onto the work list. 1742 */ 1743 for (dirrem = LIST_FIRST(&pagedep->pd_dirremhd); dirrem; 1744 dirrem = LIST_NEXT(dirrem, dm_next)) { 1745 LIST_REMOVE(dirrem, dm_next); 1746 dirrem->dm_dirinum = pagedep->pd_ino; 1747 if (inodedep == NULL) 1748 add_to_worklist(&dirrem->dm_list); 1749 else 1750 WORKLIST_INSERT(&inodedep->id_bufwait, 1751 &dirrem->dm_list); 1752 } 1753 WORKLIST_REMOVE(&pagedep->pd_list); 1754 LIST_REMOVE(pagedep, pd_hash); 1755 WORKITEM_FREE(pagedep, D_PAGEDEP); 1756 continue; 1757 1758 case D_ALLOCINDIR: 1759 free_allocindir(WK_ALLOCINDIR(wk), inodedep); 1760 continue; 1761 1762 case D_ALLOCDIRECT: 1763 case D_INODEDEP: 1764 panic("deallocate_dependencies: Unexpected type %s", 1765 TYPENAME(wk->wk_type)); 1766 /* NOTREACHED */ 1767 1768 default: 1769 panic("deallocate_dependencies: Unknown type %s", 1770 TYPENAME(wk->wk_type)); 1771 /* NOTREACHED */ 1772 } 1773 } 1774 } 1775 1776 /* 1777 * Free an allocdirect. Generate a new freefrag work request if appropriate. 1778 * This routine must be called with splbio interrupts blocked. 1779 */ 1780 static void 1781 free_allocdirect(adphead, adp, delay) 1782 struct allocdirectlst *adphead; 1783 struct allocdirect *adp; 1784 int delay; 1785 { 1786 1787 #ifdef DEBUG 1788 if (lk.lkt_held == -1) 1789 panic("free_allocdirect: lock not held"); 1790 #endif 1791 if ((adp->ad_state & DEPCOMPLETE) == 0) 1792 LIST_REMOVE(adp, ad_deps); 1793 TAILQ_REMOVE(adphead, adp, ad_next); 1794 if ((adp->ad_state & COMPLETE) == 0) 1795 WORKLIST_REMOVE(&adp->ad_list); 1796 if (adp->ad_freefrag != NULL) { 1797 if (delay) 1798 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 1799 &adp->ad_freefrag->ff_list); 1800 else 1801 add_to_worklist(&adp->ad_freefrag->ff_list); 1802 } 1803 WORKITEM_FREE(adp, D_ALLOCDIRECT); 1804 } 1805 1806 /* 1807 * Prepare an inode to be freed. The actual free operation is not 1808 * done until the zero'ed inode has been written to disk. 1809 */ 1810 static long num_freefile; /* number of freefile allocated */ 1811 void 1812 softdep_freefile(pvp, ino, mode) 1813 struct vnode *pvp; 1814 ino_t ino; 1815 int mode; 1816 { 1817 struct inode *ip = VTOI(pvp); 1818 struct inodedep *inodedep; 1819 struct freefile *freefile; 1820 1821 /* 1822 * If we are over our limit, try to improve the situation. 1823 */ 1824 if (num_freefile > max_softdeps / 2 && speedup_syncer() == 0) 1825 (void) request_cleanup(FLUSH_REMOVE, 0); 1826 /* 1827 * This sets up the inode de-allocation dependency. 1828 */ 1829 num_freefile += 1; 1830 MALLOC(freefile, struct freefile *, sizeof(struct freefile), 1831 M_FREEFILE, M_WAITOK); 1832 freefile->fx_list.wk_type = D_FREEFILE; 1833 freefile->fx_list.wk_state = 0; 1834 freefile->fx_mode = mode; 1835 freefile->fx_oldinum = ino; 1836 freefile->fx_devvp = ip->i_devvp; 1837 freefile->fx_fs = ip->i_fs; 1838 1839 /* 1840 * If the inodedep does not exist, then the zero'ed inode has 1841 * been written to disk and we can free the file immediately. 1842 */ 1843 ACQUIRE_LOCK(&lk); 1844 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0) { 1845 add_to_worklist(&freefile->fx_list); 1846 FREE_LOCK(&lk); 1847 return; 1848 } 1849 1850 /* 1851 * If we still have a bitmap dependency, then the inode has never 1852 * been written to disk. Drop the dependency as it is no longer 1853 * necessary since the inode is being deallocated. We could process 1854 * the freefile immediately, but then we would have to clear the 1855 * id_inowait dependencies here and it is easier just to let the 1856 * zero'ed inode be written and let them be cleaned up in the 1857 * normal followup actions that follow the inode write. 1858 */ 1859 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 1860 inodedep->id_state |= DEPCOMPLETE; 1861 LIST_REMOVE(inodedep, id_deps); 1862 inodedep->id_buf = NULL; 1863 } 1864 /* 1865 * If the inodedep has no dependencies associated with it, 1866 * then we must free it here and free the file immediately. 1867 * This case arises when an early allocation fails (for 1868 * example, the user is over their file quota). 1869 */ 1870 if (free_inodedep(inodedep) == 0) 1871 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 1872 else 1873 add_to_worklist(&freefile->fx_list); 1874 FREE_LOCK(&lk); 1875 } 1876 1877 /* 1878 * Try to free an inodedep structure. Return 1 if it could be freed. 1879 */ 1880 static int 1881 free_inodedep(inodedep) 1882 struct inodedep *inodedep; 1883 { 1884 1885 if ((inodedep->id_state & ONWORKLIST) != 0 || 1886 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 1887 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 1888 LIST_FIRST(&inodedep->id_bufwait) != NULL || 1889 LIST_FIRST(&inodedep->id_inowait) != NULL || 1890 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 1891 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 1892 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL) 1893 return (0); 1894 LIST_REMOVE(inodedep, id_hash); 1895 WORKITEM_FREE(inodedep, D_INODEDEP); 1896 num_inodedep -= 1; 1897 return (1); 1898 } 1899 1900 /* 1901 * This workitem routine performs the block de-allocation. 1902 * The workitem is added to the pending list after the updated 1903 * inode block has been written to disk. As mentioned above, 1904 * checks regarding the number of blocks de-allocated (compared 1905 * to the number of blocks allocated for the file) are also 1906 * performed in this function. 1907 */ 1908 static void 1909 handle_workitem_freeblocks(freeblks) 1910 struct freeblks *freeblks; 1911 { 1912 struct inode tip; 1913 ufs_daddr_t bn; 1914 struct fs *fs; 1915 int i, level, bsize; 1916 long nblocks, blocksreleased = 0; 1917 int error, allerror = 0; 1918 ufs_lbn_t baselbns[NIADDR], tmpval; 1919 1920 tip.i_number = freeblks->fb_previousinum; 1921 tip.i_devvp = freeblks->fb_devvp; 1922 tip.i_dev = freeblks->fb_devvp->v_rdev; 1923 tip.i_fs = freeblks->fb_fs; 1924 tip.i_size = freeblks->fb_oldsize; 1925 tip.i_uid = freeblks->fb_uid; 1926 fs = freeblks->fb_fs; 1927 tmpval = 1; 1928 baselbns[0] = NDADDR; 1929 for (i = 1; i < NIADDR; i++) { 1930 tmpval *= NINDIR(fs); 1931 baselbns[i] = baselbns[i - 1] + tmpval; 1932 } 1933 nblocks = btodb(fs->fs_bsize); 1934 blocksreleased = 0; 1935 /* 1936 * Indirect blocks first. 1937 */ 1938 for (level = (NIADDR - 1); level >= 0; level--) { 1939 if ((bn = freeblks->fb_iblks[level]) == 0) 1940 continue; 1941 if ((error = indir_trunc(&tip, fsbtodb(fs, bn), level, 1942 baselbns[level], &blocksreleased)) == 0) 1943 allerror = error; 1944 ffs_blkfree(&tip, bn, fs->fs_bsize); 1945 blocksreleased += nblocks; 1946 } 1947 /* 1948 * All direct blocks or frags. 1949 */ 1950 for (i = (NDADDR - 1); i >= 0; i--) { 1951 if ((bn = freeblks->fb_dblks[i]) == 0) 1952 continue; 1953 bsize = blksize(fs, &tip, i); 1954 ffs_blkfree(&tip, bn, bsize); 1955 blocksreleased += btodb(bsize); 1956 } 1957 1958 #ifdef DIAGNOSTIC 1959 if (freeblks->fb_chkcnt != blocksreleased) 1960 panic("handle_workitem_freeblocks: block count"); 1961 if (allerror) 1962 softdep_error("handle_workitem_freeblks", allerror); 1963 #endif /* DIAGNOSTIC */ 1964 WORKITEM_FREE(freeblks, D_FREEBLKS); 1965 num_freeblks -= 1; 1966 } 1967 1968 /* 1969 * Release blocks associated with the inode ip and stored in the indirect 1970 * block dbn. If level is greater than SINGLE, the block is an indirect block 1971 * and recursive calls to indirtrunc must be used to cleanse other indirect 1972 * blocks. 1973 */ 1974 static int 1975 indir_trunc(ip, dbn, level, lbn, countp) 1976 struct inode *ip; 1977 ufs_daddr_t dbn; 1978 int level; 1979 ufs_lbn_t lbn; 1980 long *countp; 1981 { 1982 struct buf *bp; 1983 ufs_daddr_t *bap; 1984 ufs_daddr_t nb; 1985 struct fs *fs; 1986 struct worklist *wk; 1987 struct indirdep *indirdep; 1988 int i, lbnadd, nblocks; 1989 int error, allerror = 0; 1990 1991 fs = ip->i_fs; 1992 lbnadd = 1; 1993 for (i = level; i > 0; i--) 1994 lbnadd *= NINDIR(fs); 1995 /* 1996 * Get buffer of block pointers to be freed. This routine is not 1997 * called until the zero'ed inode has been written, so it is safe 1998 * to free blocks as they are encountered. Because the inode has 1999 * been zero'ed, calls to bmap on these blocks will fail. So, we 2000 * have to use the on-disk address and the block device for the 2001 * filesystem to look them up. If the file was deleted before its 2002 * indirect blocks were all written to disk, the routine that set 2003 * us up (deallocate_dependencies) will have arranged to leave 2004 * a complete copy of the indirect block in memory for our use. 2005 * Otherwise we have to read the blocks in from the disk. 2006 */ 2007 ACQUIRE_LOCK(&lk); 2008 if ((bp = incore(ip->i_devvp, dbn)) != NULL && 2009 (wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2010 if (wk->wk_type != D_INDIRDEP || 2011 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp || 2012 (indirdep->ir_state & GOINGAWAY) == 0) 2013 panic("indir_trunc: lost indirdep"); 2014 WORKLIST_REMOVE(wk); 2015 WORKITEM_FREE(indirdep, D_INDIRDEP); 2016 if (LIST_FIRST(&bp->b_dep) != NULL) 2017 panic("indir_trunc: dangling dep"); 2018 FREE_LOCK(&lk); 2019 } else { 2020 FREE_LOCK(&lk); 2021 error = bread(ip->i_devvp, dbn, (int)fs->fs_bsize, NOCRED, &bp); 2022 if (error) 2023 return (error); 2024 } 2025 /* 2026 * Recursively free indirect blocks. 2027 */ 2028 bap = (ufs_daddr_t *)bp->b_data; 2029 nblocks = btodb(fs->fs_bsize); 2030 for (i = NINDIR(fs) - 1; i >= 0; i--) { 2031 if ((nb = bap[i]) == 0) 2032 continue; 2033 if (level != 0) { 2034 if ((error = indir_trunc(ip, fsbtodb(fs, nb), 2035 level - 1, lbn + (i * lbnadd), countp)) != 0) 2036 allerror = error; 2037 } 2038 ffs_blkfree(ip, nb, fs->fs_bsize); 2039 *countp += nblocks; 2040 } 2041 bp->b_flags |= B_INVAL | B_NOCACHE; 2042 brelse(bp); 2043 return (allerror); 2044 } 2045 2046 /* 2047 * Free an allocindir. 2048 * This routine must be called with splbio interrupts blocked. 2049 */ 2050 static void 2051 free_allocindir(aip, inodedep) 2052 struct allocindir *aip; 2053 struct inodedep *inodedep; 2054 { 2055 struct freefrag *freefrag; 2056 2057 #ifdef DEBUG 2058 if (lk.lkt_held == -1) 2059 panic("free_allocindir: lock not held"); 2060 #endif 2061 if ((aip->ai_state & DEPCOMPLETE) == 0) 2062 LIST_REMOVE(aip, ai_deps); 2063 if (aip->ai_state & ONWORKLIST) 2064 WORKLIST_REMOVE(&aip->ai_list); 2065 LIST_REMOVE(aip, ai_next); 2066 if ((freefrag = aip->ai_freefrag) != NULL) { 2067 if (inodedep == NULL) 2068 add_to_worklist(&freefrag->ff_list); 2069 else 2070 WORKLIST_INSERT(&inodedep->id_bufwait, 2071 &freefrag->ff_list); 2072 } 2073 WORKITEM_FREE(aip, D_ALLOCINDIR); 2074 } 2075 2076 /* 2077 * Directory entry addition dependencies. 2078 * 2079 * When adding a new directory entry, the inode (with its incremented link 2080 * count) must be written to disk before the directory entry's pointer to it. 2081 * Also, if the inode is newly allocated, the corresponding freemap must be 2082 * updated (on disk) before the directory entry's pointer. These requirements 2083 * are met via undo/redo on the directory entry's pointer, which consists 2084 * simply of the inode number. 2085 * 2086 * As directory entries are added and deleted, the free space within a 2087 * directory block can become fragmented. The ufs file system will compact 2088 * a fragmented directory block to make space for a new entry. When this 2089 * occurs, the offsets of previously added entries change. Any "diradd" 2090 * dependency structures corresponding to these entries must be updated with 2091 * the new offsets. 2092 */ 2093 2094 /* 2095 * This routine is called after the in-memory inode's link 2096 * count has been incremented, but before the directory entry's 2097 * pointer to the inode has been set. 2098 */ 2099 void 2100 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp) 2101 struct buf *bp; /* buffer containing directory block */ 2102 struct inode *dp; /* inode for directory */ 2103 off_t diroffset; /* offset of new entry in directory */ 2104 long newinum; /* inode referenced by new directory entry */ 2105 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 2106 { 2107 int offset; /* offset of new entry within directory block */ 2108 ufs_lbn_t lbn; /* block in directory containing new entry */ 2109 struct fs *fs; 2110 struct diradd *dap; 2111 struct pagedep *pagedep; 2112 struct inodedep *inodedep; 2113 struct mkdir *mkdir1, *mkdir2; 2114 2115 /* 2116 * Whiteouts have no dependencies. 2117 */ 2118 if (newinum == WINO) { 2119 if (newdirbp != NULL) 2120 bdwrite(newdirbp); 2121 return; 2122 } 2123 2124 fs = dp->i_fs; 2125 lbn = lblkno(fs, diroffset); 2126 offset = blkoff(fs, diroffset); 2127 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, M_WAITOK); 2128 bzero(dap, sizeof(struct diradd)); 2129 dap->da_list.wk_type = D_DIRADD; 2130 dap->da_offset = offset; 2131 dap->da_newinum = newinum; 2132 dap->da_state = ATTACHED; 2133 if (newdirbp == NULL) { 2134 dap->da_state |= DEPCOMPLETE; 2135 ACQUIRE_LOCK(&lk); 2136 } else { 2137 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 2138 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2139 M_WAITOK); 2140 mkdir1->md_list.wk_type = D_MKDIR; 2141 mkdir1->md_state = MKDIR_BODY; 2142 mkdir1->md_diradd = dap; 2143 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2144 M_WAITOK); 2145 mkdir2->md_list.wk_type = D_MKDIR; 2146 mkdir2->md_state = MKDIR_PARENT; 2147 mkdir2->md_diradd = dap; 2148 /* 2149 * Dependency on "." and ".." being written to disk. 2150 */ 2151 mkdir1->md_buf = newdirbp; 2152 ACQUIRE_LOCK(&lk); 2153 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 2154 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list); 2155 FREE_LOCK(&lk); 2156 bdwrite(newdirbp); 2157 /* 2158 * Dependency on link count increase for parent directory 2159 */ 2160 ACQUIRE_LOCK(&lk); 2161 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0 2162 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2163 dap->da_state &= ~MKDIR_PARENT; 2164 WORKITEM_FREE(mkdir2, D_MKDIR); 2165 } else { 2166 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 2167 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 2168 } 2169 } 2170 /* 2171 * Link into parent directory pagedep to await its being written. 2172 */ 2173 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2174 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2175 dap->da_pagedep = pagedep; 2176 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 2177 da_pdlist); 2178 /* 2179 * Link into its inodedep. Put it on the id_bufwait list if the inode 2180 * is not yet written. If it is written, do the post-inode write 2181 * processing to put it on the id_pendinghd list. 2182 */ 2183 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep); 2184 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 2185 diradd_inode_written(dap, inodedep); 2186 else 2187 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2188 FREE_LOCK(&lk); 2189 } 2190 2191 /* 2192 * This procedure is called to change the offset of a directory 2193 * entry when compacting a directory block which must be owned 2194 * exclusively by the caller. Note that the actual entry movement 2195 * must be done in this procedure to ensure that no I/O completions 2196 * occur while the move is in progress. 2197 */ 2198 void 2199 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize) 2200 struct inode *dp; /* inode for directory */ 2201 caddr_t base; /* address of dp->i_offset */ 2202 caddr_t oldloc; /* address of old directory location */ 2203 caddr_t newloc; /* address of new directory location */ 2204 int entrysize; /* size of directory entry */ 2205 { 2206 int offset, oldoffset, newoffset; 2207 struct pagedep *pagedep; 2208 struct diradd *dap; 2209 ufs_lbn_t lbn; 2210 2211 ACQUIRE_LOCK(&lk); 2212 lbn = lblkno(dp->i_fs, dp->i_offset); 2213 offset = blkoff(dp->i_fs, dp->i_offset); 2214 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) 2215 goto done; 2216 oldoffset = offset + (oldloc - base); 2217 newoffset = offset + (newloc - base); 2218 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[DIRADDHASH(oldoffset)]); 2219 dap; dap = LIST_NEXT(dap, da_pdlist)) { 2220 if (dap->da_offset != oldoffset) 2221 continue; 2222 dap->da_offset = newoffset; 2223 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) 2224 break; 2225 LIST_REMOVE(dap, da_pdlist); 2226 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], 2227 dap, da_pdlist); 2228 break; 2229 } 2230 if (dap == NULL) { 2231 for (dap = LIST_FIRST(&pagedep->pd_pendinghd); 2232 dap; dap = LIST_NEXT(dap, da_pdlist)) { 2233 if (dap->da_offset == oldoffset) { 2234 dap->da_offset = newoffset; 2235 break; 2236 } 2237 } 2238 } 2239 done: 2240 bcopy(oldloc, newloc, entrysize); 2241 FREE_LOCK(&lk); 2242 } 2243 2244 /* 2245 * Free a diradd dependency structure. This routine must be called 2246 * with splbio interrupts blocked. 2247 */ 2248 static void 2249 free_diradd(dap) 2250 struct diradd *dap; 2251 { 2252 struct dirrem *dirrem; 2253 struct pagedep *pagedep; 2254 struct inodedep *inodedep; 2255 struct mkdir *mkdir, *nextmd; 2256 2257 #ifdef DEBUG 2258 if (lk.lkt_held == -1) 2259 panic("free_diradd: lock not held"); 2260 #endif 2261 WORKLIST_REMOVE(&dap->da_list); 2262 LIST_REMOVE(dap, da_pdlist); 2263 if ((dap->da_state & DIRCHG) == 0) { 2264 pagedep = dap->da_pagedep; 2265 } else { 2266 dirrem = dap->da_previous; 2267 pagedep = dirrem->dm_pagedep; 2268 dirrem->dm_dirinum = pagedep->pd_ino; 2269 add_to_worklist(&dirrem->dm_list); 2270 } 2271 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum, 2272 0, &inodedep) != 0) 2273 (void) free_inodedep(inodedep); 2274 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2275 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 2276 nextmd = LIST_NEXT(mkdir, md_mkdirs); 2277 if (mkdir->md_diradd != dap) 2278 continue; 2279 dap->da_state &= ~mkdir->md_state; 2280 WORKLIST_REMOVE(&mkdir->md_list); 2281 LIST_REMOVE(mkdir, md_mkdirs); 2282 WORKITEM_FREE(mkdir, D_MKDIR); 2283 } 2284 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 2285 panic("free_diradd: unfound ref"); 2286 } 2287 WORKITEM_FREE(dap, D_DIRADD); 2288 } 2289 2290 /* 2291 * Directory entry removal dependencies. 2292 * 2293 * When removing a directory entry, the entry's inode pointer must be 2294 * zero'ed on disk before the corresponding inode's link count is decremented 2295 * (possibly freeing the inode for re-use). This dependency is handled by 2296 * updating the directory entry but delaying the inode count reduction until 2297 * after the directory block has been written to disk. After this point, the 2298 * inode count can be decremented whenever it is convenient. 2299 */ 2300 2301 /* 2302 * This routine should be called immediately after removing 2303 * a directory entry. The inode's link count should not be 2304 * decremented by the calling procedure -- the soft updates 2305 * code will do this task when it is safe. 2306 */ 2307 void 2308 softdep_setup_remove(bp, dp, ip, isrmdir) 2309 struct buf *bp; /* buffer containing directory block */ 2310 struct inode *dp; /* inode for the directory being modified */ 2311 struct inode *ip; /* inode for directory entry being removed */ 2312 int isrmdir; /* indicates if doing RMDIR */ 2313 { 2314 struct dirrem *dirrem; 2315 2316 /* 2317 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 2318 */ 2319 dirrem = newdirrem(bp, dp, ip, isrmdir); 2320 if ((dirrem->dm_state & COMPLETE) == 0) { 2321 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 2322 dm_next); 2323 } else { 2324 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 2325 add_to_worklist(&dirrem->dm_list); 2326 } 2327 FREE_LOCK(&lk); 2328 } 2329 2330 /* 2331 * Allocate a new dirrem if appropriate and return it along with 2332 * its associated pagedep. Called without a lock, returns with lock. 2333 */ 2334 static struct dirrem * 2335 newdirrem(bp, dp, ip, isrmdir) 2336 struct buf *bp; /* buffer containing directory block */ 2337 struct inode *dp; /* inode for the directory being modified */ 2338 struct inode *ip; /* inode for directory entry being removed */ 2339 int isrmdir; /* indicates if doing RMDIR */ 2340 { 2341 int offset; 2342 ufs_lbn_t lbn; 2343 struct diradd *dap; 2344 struct dirrem *dirrem; 2345 struct pagedep *pagedep; 2346 2347 /* 2348 * Whiteouts have no deletion dependencies. 2349 */ 2350 if (ip == NULL) 2351 panic("newdirrem: whiteout"); 2352 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem), 2353 M_DIRREM, M_WAITOK); 2354 bzero(dirrem, sizeof(struct dirrem)); 2355 dirrem->dm_list.wk_type = D_DIRREM; 2356 dirrem->dm_state = isrmdir ? RMDIR : 0; 2357 dirrem->dm_mnt = ITOV(ip)->v_mount; 2358 dirrem->dm_oldinum = ip->i_number; 2359 2360 ACQUIRE_LOCK(&lk); 2361 lbn = lblkno(dp->i_fs, dp->i_offset); 2362 offset = blkoff(dp->i_fs, dp->i_offset); 2363 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2364 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2365 dirrem->dm_pagedep = pagedep; 2366 /* 2367 * Check for a diradd dependency for the same directory entry. 2368 * If present, then both dependencies become obsolete and can 2369 * be de-allocated. Check for an entry on both the pd_dirraddhd 2370 * list and the pd_pendinghd list. 2371 */ 2372 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[DIRADDHASH(offset)]); 2373 dap; dap = LIST_NEXT(dap, da_pdlist)) 2374 if (dap->da_offset == offset) 2375 break; 2376 if (dap == NULL) { 2377 for (dap = LIST_FIRST(&pagedep->pd_pendinghd); 2378 dap; dap = LIST_NEXT(dap, da_pdlist)) 2379 if (dap->da_offset == offset) 2380 break; 2381 if (dap == NULL) 2382 return (dirrem); 2383 } 2384 /* 2385 * Must be ATTACHED at this point, so just delete it. 2386 */ 2387 if ((dap->da_state & ATTACHED) == 0) 2388 panic("newdirrem: not ATTACHED"); 2389 if (dap->da_newinum != ip->i_number) 2390 panic("newdirrem: inum %d should be %d", 2391 ip->i_number, dap->da_newinum); 2392 free_diradd(dap); 2393 dirrem->dm_state |= COMPLETE; 2394 return (dirrem); 2395 } 2396 2397 /* 2398 * Directory entry change dependencies. 2399 * 2400 * Changing an existing directory entry requires that an add operation 2401 * be completed first followed by a deletion. The semantics for the addition 2402 * are identical to the description of adding a new entry above except 2403 * that the rollback is to the old inode number rather than zero. Once 2404 * the addition dependency is completed, the removal is done as described 2405 * in the removal routine above. 2406 */ 2407 2408 /* 2409 * This routine should be called immediately after changing 2410 * a directory entry. The inode's link count should not be 2411 * decremented by the calling procedure -- the soft updates 2412 * code will perform this task when it is safe. 2413 */ 2414 void 2415 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 2416 struct buf *bp; /* buffer containing directory block */ 2417 struct inode *dp; /* inode for the directory being modified */ 2418 struct inode *ip; /* inode for directory entry being removed */ 2419 long newinum; /* new inode number for changed entry */ 2420 int isrmdir; /* indicates if doing RMDIR */ 2421 { 2422 int offset; 2423 struct diradd *dap = NULL; 2424 struct dirrem *dirrem; 2425 struct pagedep *pagedep; 2426 struct inodedep *inodedep; 2427 2428 offset = blkoff(dp->i_fs, dp->i_offset); 2429 2430 /* 2431 * Whiteouts do not need diradd dependencies. 2432 */ 2433 if (newinum != WINO) { 2434 MALLOC(dap, struct diradd *, sizeof(struct diradd), 2435 M_DIRADD, M_WAITOK); 2436 bzero(dap, sizeof(struct diradd)); 2437 dap->da_list.wk_type = D_DIRADD; 2438 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 2439 dap->da_offset = offset; 2440 dap->da_newinum = newinum; 2441 } 2442 2443 /* 2444 * Allocate a new dirrem and ACQUIRE_LOCK. 2445 */ 2446 dirrem = newdirrem(bp, dp, ip, isrmdir); 2447 pagedep = dirrem->dm_pagedep; 2448 /* 2449 * The possible values for isrmdir: 2450 * 0 - non-directory file rename 2451 * 1 - directory rename within same directory 2452 * inum - directory rename to new directory of given inode number 2453 * When renaming to a new directory, we are both deleting and 2454 * creating a new directory entry, so the link count on the new 2455 * directory should not change. Thus we do not need the followup 2456 * dirrem which is usually done in handle_workitem_remove. We set 2457 * the DIRCHG flag to tell handle_workitem_remove to skip the 2458 * followup dirrem. 2459 */ 2460 if (isrmdir > 1) 2461 dirrem->dm_state |= DIRCHG; 2462 2463 /* 2464 * Whiteouts have no additional dependencies, 2465 * so just put the dirrem on the correct list. 2466 */ 2467 if (newinum == WINO) { 2468 if ((dirrem->dm_state & COMPLETE) == 0) { 2469 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 2470 dm_next); 2471 } else { 2472 dirrem->dm_dirinum = pagedep->pd_ino; 2473 add_to_worklist(&dirrem->dm_list); 2474 } 2475 FREE_LOCK(&lk); 2476 return; 2477 } 2478 2479 /* 2480 * Link into its inodedep. Put it on the id_bufwait list if the inode 2481 * is not yet written. If it is written, do the post-inode write 2482 * processing to put it on the id_pendinghd list. 2483 */ 2484 dap->da_previous = dirrem; 2485 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 || 2486 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2487 dap->da_state |= COMPLETE; 2488 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 2489 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 2490 } else { 2491 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 2492 dap, da_pdlist); 2493 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2494 } 2495 /* 2496 * If the previous inode was never written or its previous directory 2497 * entry was never written, then we do not want to roll back to this 2498 * previous value. Instead we want to roll back to zero and immediately 2499 * free the unwritten or unreferenced inode. 2500 */ 2501 if (dirrem->dm_state & COMPLETE) { 2502 dap->da_state &= ~DIRCHG; 2503 dap->da_pagedep = pagedep; 2504 dirrem->dm_dirinum = pagedep->pd_ino; 2505 add_to_worklist(&dirrem->dm_list); 2506 } 2507 FREE_LOCK(&lk); 2508 } 2509 2510 /* 2511 * Called whenever the link count on an inode is increased. 2512 * It creates an inode dependency so that the new reference(s) 2513 * to the inode cannot be committed to disk until the updated 2514 * inode has been written. 2515 */ 2516 void 2517 softdep_increase_linkcnt(ip) 2518 struct inode *ip; /* the inode with the increased link count */ 2519 { 2520 struct inodedep *inodedep; 2521 2522 ACQUIRE_LOCK(&lk); 2523 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 2524 FREE_LOCK(&lk); 2525 } 2526 2527 /* 2528 * This workitem decrements the inode's link count. 2529 * If the link count reaches zero, the file is removed. 2530 */ 2531 static void 2532 handle_workitem_remove(dirrem) 2533 struct dirrem *dirrem; 2534 { 2535 struct proc *p = CURPROC; /* XXX */ 2536 struct inodedep *inodedep; 2537 struct vnode *vp; 2538 struct inode *ip; 2539 int error; 2540 2541 if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) { 2542 softdep_error("handle_workitem_remove: vget", error); 2543 return; 2544 } 2545 ip = VTOI(vp); 2546 /* 2547 * Normal file deletion. 2548 */ 2549 if ((dirrem->dm_state & RMDIR) == 0) { 2550 ip->i_nlink--; 2551 if (ip->i_nlink < ip->i_effnlink) 2552 panic("handle_workitem_remove: bad file delta"); 2553 ip->i_flag |= IN_CHANGE; 2554 vput(vp); 2555 WORKITEM_FREE(dirrem, D_DIRREM); 2556 return; 2557 } 2558 /* 2559 * Directory deletion. Decrement reference count for both the 2560 * just deleted parent directory entry and the reference for ".". 2561 * Next truncate the directory to length zero. When the 2562 * truncation completes, arrange to have the reference count on 2563 * the parent decremented to account for the loss of "..". 2564 */ 2565 ip->i_nlink -= 2; 2566 if (ip->i_nlink < ip->i_effnlink) 2567 panic("handle_workitem_remove: bad dir delta"); 2568 ip->i_flag |= IN_CHANGE; 2569 if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, p->p_ucred, p)) != 0) 2570 softdep_error("handle_workitem_remove: truncate", error); 2571 /* 2572 * Rename a directory to a new parent. Since, we are both deleting 2573 * and creating a new directory entry, the link count on the new 2574 * directory should not change. Thus we skip the followup dirrem. 2575 */ 2576 if (dirrem->dm_state & DIRCHG) { 2577 vput(vp); 2578 WORKITEM_FREE(dirrem, D_DIRREM); 2579 return; 2580 } 2581 ACQUIRE_LOCK(&lk); 2582 (void) inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, DEPALLOC, 2583 &inodedep); 2584 dirrem->dm_state = 0; 2585 dirrem->dm_oldinum = dirrem->dm_dirinum; 2586 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 2587 FREE_LOCK(&lk); 2588 vput(vp); 2589 } 2590 2591 /* 2592 * Inode de-allocation dependencies. 2593 * 2594 * When an inode's link count is reduced to zero, it can be de-allocated. We 2595 * found it convenient to postpone de-allocation until after the inode is 2596 * written to disk with its new link count (zero). At this point, all of the 2597 * on-disk inode's block pointers are nullified and, with careful dependency 2598 * list ordering, all dependencies related to the inode will be satisfied and 2599 * the corresponding dependency structures de-allocated. So, if/when the 2600 * inode is reused, there will be no mixing of old dependencies with new 2601 * ones. This artificial dependency is set up by the block de-allocation 2602 * procedure above (softdep_setup_freeblocks) and completed by the 2603 * following procedure. 2604 */ 2605 static void 2606 handle_workitem_freefile(freefile) 2607 struct freefile *freefile; 2608 { 2609 struct vnode vp; 2610 struct inode tip; 2611 struct inodedep *idp; 2612 int error; 2613 2614 #ifdef DEBUG 2615 ACQUIRE_LOCK(&lk); 2616 if (inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp)) 2617 panic("handle_workitem_freefile: inodedep survived"); 2618 FREE_LOCK(&lk); 2619 #endif 2620 tip.i_devvp = freefile->fx_devvp; 2621 tip.i_dev = freefile->fx_devvp->v_rdev; 2622 tip.i_fs = freefile->fx_fs; 2623 vp.v_data = &tip; 2624 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0) 2625 softdep_error("handle_workitem_freefile", error); 2626 WORKITEM_FREE(freefile, D_FREEFILE); 2627 num_freefile -= 1; 2628 } 2629 2630 /* 2631 * Disk writes. 2632 * 2633 * The dependency structures constructed above are most actively used when file 2634 * system blocks are written to disk. No constraints are placed on when a 2635 * block can be written, but unsatisfied update dependencies are made safe by 2636 * modifying (or replacing) the source memory for the duration of the disk 2637 * write. When the disk write completes, the memory block is again brought 2638 * up-to-date. 2639 * 2640 * In-core inode structure reclamation. 2641 * 2642 * Because there are a finite number of "in-core" inode structures, they are 2643 * reused regularly. By transferring all inode-related dependencies to the 2644 * in-memory inode block and indexing them separately (via "inodedep"s), we 2645 * can allow "in-core" inode structures to be reused at any time and avoid 2646 * any increase in contention. 2647 * 2648 * Called just before entering the device driver to initiate a new disk I/O. 2649 * The buffer must be locked, thus, no I/O completion operations can occur 2650 * while we are manipulating its associated dependencies. 2651 */ 2652 void 2653 softdep_disk_io_initiation(bp) 2654 struct buf *bp; /* structure describing disk write to occur */ 2655 { 2656 struct worklist *wk, *nextwk; 2657 struct indirdep *indirdep; 2658 2659 /* 2660 * We only care about write operations. There should never 2661 * be dependencies for reads. 2662 */ 2663 if (bp->b_flags & B_READ) 2664 panic("softdep_disk_io_initiation: read"); 2665 /* 2666 * Do any necessary pre-I/O processing. 2667 */ 2668 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = nextwk) { 2669 nextwk = LIST_NEXT(wk, wk_list); 2670 switch (wk->wk_type) { 2671 2672 case D_PAGEDEP: 2673 initiate_write_filepage(WK_PAGEDEP(wk), bp); 2674 continue; 2675 2676 case D_INODEDEP: 2677 initiate_write_inodeblock(WK_INODEDEP(wk), bp); 2678 continue; 2679 2680 case D_INDIRDEP: 2681 indirdep = WK_INDIRDEP(wk); 2682 if (indirdep->ir_state & GOINGAWAY) 2683 panic("disk_io_initiation: indirdep gone"); 2684 /* 2685 * If there are no remaining dependencies, this 2686 * will be writing the real pointers, so the 2687 * dependency can be freed. 2688 */ 2689 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) { 2690 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 2691 brelse(indirdep->ir_savebp); 2692 /* inline expand WORKLIST_REMOVE(wk); */ 2693 wk->wk_state &= ~ONWORKLIST; 2694 LIST_REMOVE(wk, wk_list); 2695 WORKITEM_FREE(indirdep, D_INDIRDEP); 2696 continue; 2697 } 2698 /* 2699 * Replace up-to-date version with safe version. 2700 */ 2701 ACQUIRE_LOCK(&lk); 2702 indirdep->ir_state &= ~ATTACHED; 2703 indirdep->ir_state |= UNDONE; 2704 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount, 2705 M_INDIRDEP, M_WAITOK); 2706 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 2707 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 2708 bp->b_bcount); 2709 FREE_LOCK(&lk); 2710 continue; 2711 2712 case D_MKDIR: 2713 case D_BMSAFEMAP: 2714 case D_ALLOCDIRECT: 2715 case D_ALLOCINDIR: 2716 continue; 2717 2718 default: 2719 panic("handle_disk_io_initiation: Unexpected type %s", 2720 TYPENAME(wk->wk_type)); 2721 /* NOTREACHED */ 2722 } 2723 } 2724 } 2725 2726 /* 2727 * Called from within the procedure above to deal with unsatisfied 2728 * allocation dependencies in a directory. The buffer must be locked, 2729 * thus, no I/O completion operations can occur while we are 2730 * manipulating its associated dependencies. 2731 */ 2732 static void 2733 initiate_write_filepage(pagedep, bp) 2734 struct pagedep *pagedep; 2735 struct buf *bp; 2736 { 2737 struct diradd *dap; 2738 struct direct *ep; 2739 int i; 2740 2741 if (pagedep->pd_state & IOSTARTED) { 2742 /* 2743 * This can only happen if there is a driver that does not 2744 * understand chaining. Here biodone will reissue the call 2745 * to strategy for the incomplete buffers. 2746 */ 2747 printf("initiate_write_filepage: already started\n"); 2748 return; 2749 } 2750 pagedep->pd_state |= IOSTARTED; 2751 ACQUIRE_LOCK(&lk); 2752 for (i = 0; i < DAHASHSZ; i++) { 2753 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 2754 dap = LIST_NEXT(dap, da_pdlist)) { 2755 ep = (struct direct *) 2756 ((char *)bp->b_data + dap->da_offset); 2757 if (ep->d_ino != dap->da_newinum) 2758 panic("%s: dir inum %d != new %d", 2759 "initiate_write_filepage", 2760 ep->d_ino, dap->da_newinum); 2761 if (dap->da_state & DIRCHG) 2762 ep->d_ino = dap->da_previous->dm_oldinum; 2763 else 2764 ep->d_ino = 0; 2765 dap->da_state &= ~ATTACHED; 2766 dap->da_state |= UNDONE; 2767 } 2768 } 2769 FREE_LOCK(&lk); 2770 } 2771 2772 /* 2773 * Called from within the procedure above to deal with unsatisfied 2774 * allocation dependencies in an inodeblock. The buffer must be 2775 * locked, thus, no I/O completion operations can occur while we 2776 * are manipulating its associated dependencies. 2777 */ 2778 static void 2779 initiate_write_inodeblock(inodedep, bp) 2780 struct inodedep *inodedep; 2781 struct buf *bp; /* The inode block */ 2782 { 2783 struct allocdirect *adp, *lastadp; 2784 struct dinode *dp; 2785 struct fs *fs; 2786 ufs_lbn_t prevlbn = 0; 2787 int i, deplist; 2788 2789 if (inodedep->id_state & IOSTARTED) 2790 panic("initiate_write_inodeblock: already started"); 2791 inodedep->id_state |= IOSTARTED; 2792 fs = inodedep->id_fs; 2793 dp = (struct dinode *)bp->b_data + 2794 ino_to_fsbo(fs, inodedep->id_ino); 2795 /* 2796 * If the bitmap is not yet written, then the allocated 2797 * inode cannot be written to disk. 2798 */ 2799 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 2800 if (inodedep->id_savedino != NULL) 2801 panic("initiate_write_inodeblock: already doing I/O"); 2802 MALLOC(inodedep->id_savedino, struct dinode *, 2803 sizeof(struct dinode), M_INODEDEP, M_WAITOK); 2804 *inodedep->id_savedino = *dp; 2805 bzero((caddr_t)dp, sizeof(struct dinode)); 2806 return; 2807 } 2808 /* 2809 * If no dependencies, then there is nothing to roll back. 2810 */ 2811 inodedep->id_savedsize = dp->di_size; 2812 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 2813 return; 2814 /* 2815 * Set the dependencies to busy. 2816 */ 2817 ACQUIRE_LOCK(&lk); 2818 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 2819 adp = TAILQ_NEXT(adp, ad_next)) { 2820 #ifdef DIAGNOSTIC 2821 if (deplist != 0 && prevlbn >= adp->ad_lbn) 2822 panic("softdep_write_inodeblock: lbn order"); 2823 prevlbn = adp->ad_lbn; 2824 if (adp->ad_lbn < NDADDR && 2825 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) 2826 panic("%s: direct pointer #%ld mismatch %d != %d", 2827 "softdep_write_inodeblock", adp->ad_lbn, 2828 dp->di_db[adp->ad_lbn], adp->ad_newblkno); 2829 if (adp->ad_lbn >= NDADDR && 2830 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) 2831 panic("%s: indirect pointer #%ld mismatch %d != %d", 2832 "softdep_write_inodeblock", adp->ad_lbn - NDADDR, 2833 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno); 2834 deplist |= 1 << adp->ad_lbn; 2835 if ((adp->ad_state & ATTACHED) == 0) 2836 panic("softdep_write_inodeblock: Unknown state 0x%x", 2837 adp->ad_state); 2838 #endif /* DIAGNOSTIC */ 2839 adp->ad_state &= ~ATTACHED; 2840 adp->ad_state |= UNDONE; 2841 } 2842 /* 2843 * The on-disk inode cannot claim to be any larger than the last 2844 * fragment that has been written. Otherwise, the on-disk inode 2845 * might have fragments that were not the last block in the file 2846 * which would corrupt the filesystem. 2847 */ 2848 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 2849 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 2850 if (adp->ad_lbn >= NDADDR) 2851 break; 2852 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 2853 /* keep going until hitting a rollback to a frag */ 2854 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 2855 continue; 2856 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 2857 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 2858 #ifdef DIAGNOSTIC 2859 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 2860 panic("softdep_write_inodeblock: lost dep1"); 2861 #endif /* DIAGNOSTIC */ 2862 dp->di_db[i] = 0; 2863 } 2864 for (i = 0; i < NIADDR; i++) { 2865 #ifdef DIAGNOSTIC 2866 if (dp->di_ib[i] != 0 && 2867 (deplist & ((1 << NDADDR) << i)) == 0) 2868 panic("softdep_write_inodeblock: lost dep2"); 2869 #endif /* DIAGNOSTIC */ 2870 dp->di_ib[i] = 0; 2871 } 2872 FREE_LOCK(&lk); 2873 return; 2874 } 2875 /* 2876 * If we have zero'ed out the last allocated block of the file, 2877 * roll back the size to the last currently allocated block. 2878 * We know that this last allocated block is a full-sized as 2879 * we already checked for fragments in the loop above. 2880 */ 2881 if (lastadp != NULL && 2882 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 2883 for (i = lastadp->ad_lbn; i >= 0; i--) 2884 if (dp->di_db[i] != 0) 2885 break; 2886 dp->di_size = (i + 1) * fs->fs_bsize; 2887 } 2888 /* 2889 * The only dependencies are for indirect blocks. 2890 * 2891 * The file size for indirect block additions is not guaranteed. 2892 * Such a guarantee would be non-trivial to achieve. The conventional 2893 * synchronous write implementation also does not make this guarantee. 2894 * Fsck should catch and fix discrepancies. Arguably, the file size 2895 * can be over-estimated without destroying integrity when the file 2896 * moves into the indirect blocks (i.e., is large). If we want to 2897 * postpone fsck, we are stuck with this argument. 2898 */ 2899 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 2900 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 2901 FREE_LOCK(&lk); 2902 } 2903 2904 /* 2905 * This routine is called during the completion interrupt 2906 * service routine for a disk write (from the procedure called 2907 * by the device driver to inform the file system caches of 2908 * a request completion). It should be called early in this 2909 * procedure, before the block is made available to other 2910 * processes or other routines are called. 2911 */ 2912 void 2913 softdep_disk_write_complete(bp) 2914 struct buf *bp; /* describes the completed disk write */ 2915 { 2916 struct worklist *wk; 2917 struct workhead reattach; 2918 struct newblk *newblk; 2919 struct allocindir *aip; 2920 struct allocdirect *adp; 2921 struct indirdep *indirdep; 2922 struct inodedep *inodedep; 2923 struct bmsafemap *bmsafemap; 2924 2925 #ifdef DEBUG 2926 if (lk.lkt_held != -1) 2927 panic("softdep_disk_write_complete: lock is held"); 2928 lk.lkt_held = -2; 2929 #endif 2930 LIST_INIT(&reattach); 2931 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2932 WORKLIST_REMOVE(wk); 2933 switch (wk->wk_type) { 2934 2935 case D_PAGEDEP: 2936 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 2937 WORKLIST_INSERT(&reattach, wk); 2938 continue; 2939 2940 case D_INODEDEP: 2941 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 2942 WORKLIST_INSERT(&reattach, wk); 2943 continue; 2944 2945 case D_BMSAFEMAP: 2946 bmsafemap = WK_BMSAFEMAP(wk); 2947 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) { 2948 newblk->nb_state |= DEPCOMPLETE; 2949 newblk->nb_bmsafemap = NULL; 2950 LIST_REMOVE(newblk, nb_deps); 2951 } 2952 while ((adp = 2953 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) { 2954 adp->ad_state |= DEPCOMPLETE; 2955 adp->ad_buf = NULL; 2956 LIST_REMOVE(adp, ad_deps); 2957 handle_allocdirect_partdone(adp); 2958 } 2959 while ((aip = 2960 LIST_FIRST(&bmsafemap->sm_allocindirhd))) { 2961 aip->ai_state |= DEPCOMPLETE; 2962 aip->ai_buf = NULL; 2963 LIST_REMOVE(aip, ai_deps); 2964 handle_allocindir_partdone(aip); 2965 } 2966 while ((inodedep = 2967 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { 2968 inodedep->id_state |= DEPCOMPLETE; 2969 LIST_REMOVE(inodedep, id_deps); 2970 inodedep->id_buf = NULL; 2971 } 2972 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 2973 continue; 2974 2975 case D_MKDIR: 2976 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 2977 continue; 2978 2979 case D_ALLOCDIRECT: 2980 adp = WK_ALLOCDIRECT(wk); 2981 adp->ad_state |= COMPLETE; 2982 handle_allocdirect_partdone(adp); 2983 continue; 2984 2985 case D_ALLOCINDIR: 2986 aip = WK_ALLOCINDIR(wk); 2987 aip->ai_state |= COMPLETE; 2988 handle_allocindir_partdone(aip); 2989 continue; 2990 2991 case D_INDIRDEP: 2992 indirdep = WK_INDIRDEP(wk); 2993 if (indirdep->ir_state & GOINGAWAY) 2994 panic("disk_write_complete: indirdep gone"); 2995 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 2996 FREE(indirdep->ir_saveddata, M_INDIRDEP); 2997 indirdep->ir_saveddata = 0; 2998 indirdep->ir_state &= ~UNDONE; 2999 indirdep->ir_state |= ATTACHED; 3000 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 3001 handle_allocindir_partdone(aip); 3002 if (aip == LIST_FIRST(&indirdep->ir_donehd)) 3003 panic("disk_write_complete: not gone"); 3004 } 3005 WORKLIST_INSERT(&reattach, wk); 3006 if ((bp->b_flags & B_DELWRI) == 0) 3007 stat_indir_blk_ptrs++; 3008 bdirty(bp); 3009 continue; 3010 3011 default: 3012 panic("handle_disk_write_complete: Unknown type %s", 3013 TYPENAME(wk->wk_type)); 3014 /* NOTREACHED */ 3015 } 3016 } 3017 /* 3018 * Reattach any requests that must be redone. 3019 */ 3020 while ((wk = LIST_FIRST(&reattach)) != NULL) { 3021 WORKLIST_REMOVE(wk); 3022 WORKLIST_INSERT(&bp->b_dep, wk); 3023 } 3024 #ifdef DEBUG 3025 if (lk.lkt_held != -2) 3026 panic("softdep_disk_write_complete: lock lost"); 3027 lk.lkt_held = -1; 3028 #endif 3029 } 3030 3031 /* 3032 * Called from within softdep_disk_write_complete above. Note that 3033 * this routine is always called from interrupt level with further 3034 * splbio interrupts blocked. 3035 */ 3036 static void 3037 handle_allocdirect_partdone(adp) 3038 struct allocdirect *adp; /* the completed allocdirect */ 3039 { 3040 struct allocdirect *listadp; 3041 struct inodedep *inodedep; 3042 long bsize; 3043 3044 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3045 return; 3046 if (adp->ad_buf != NULL) 3047 panic("handle_allocdirect_partdone: dangling dep"); 3048 /* 3049 * The on-disk inode cannot claim to be any larger than the last 3050 * fragment that has been written. Otherwise, the on-disk inode 3051 * might have fragments that were not the last block in the file 3052 * which would corrupt the filesystem. Thus, we cannot free any 3053 * allocdirects after one whose ad_oldblkno claims a fragment as 3054 * these blocks must be rolled back to zero before writing the inode. 3055 * We check the currently active set of allocdirects in id_inoupdt. 3056 */ 3057 inodedep = adp->ad_inodedep; 3058 bsize = inodedep->id_fs->fs_bsize; 3059 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp; 3060 listadp = TAILQ_NEXT(listadp, ad_next)) { 3061 /* found our block */ 3062 if (listadp == adp) 3063 break; 3064 /* continue if ad_oldlbn is not a fragment */ 3065 if (listadp->ad_oldsize == 0 || 3066 listadp->ad_oldsize == bsize) 3067 continue; 3068 /* hit a fragment */ 3069 return; 3070 } 3071 /* 3072 * If we have reached the end of the current list without 3073 * finding the just finished dependency, then it must be 3074 * on the future dependency list. Future dependencies cannot 3075 * be freed until they are moved to the current list. 3076 */ 3077 if (listadp == NULL) { 3078 #ifdef DEBUG 3079 for (listadp = TAILQ_FIRST(&inodedep->id_newinoupdt); listadp; 3080 listadp = TAILQ_NEXT(listadp, ad_next)) 3081 /* found our block */ 3082 if (listadp == adp) 3083 break; 3084 if (listadp == NULL) 3085 panic("handle_allocdirect_partdone: lost dep"); 3086 #endif /* DEBUG */ 3087 return; 3088 } 3089 /* 3090 * If we have found the just finished dependency, then free 3091 * it along with anything that follows it that is complete. 3092 */ 3093 for (; adp; adp = listadp) { 3094 listadp = TAILQ_NEXT(adp, ad_next); 3095 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3096 return; 3097 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 3098 } 3099 } 3100 3101 /* 3102 * Called from within softdep_disk_write_complete above. Note that 3103 * this routine is always called from interrupt level with further 3104 * splbio interrupts blocked. 3105 */ 3106 static void 3107 handle_allocindir_partdone(aip) 3108 struct allocindir *aip; /* the completed allocindir */ 3109 { 3110 struct indirdep *indirdep; 3111 3112 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 3113 return; 3114 if (aip->ai_buf != NULL) 3115 panic("handle_allocindir_partdone: dangling dependency"); 3116 indirdep = aip->ai_indirdep; 3117 if (indirdep->ir_state & UNDONE) { 3118 LIST_REMOVE(aip, ai_next); 3119 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 3120 return; 3121 } 3122 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 3123 aip->ai_newblkno; 3124 LIST_REMOVE(aip, ai_next); 3125 if (aip->ai_freefrag != NULL) 3126 add_to_worklist(&aip->ai_freefrag->ff_list); 3127 WORKITEM_FREE(aip, D_ALLOCINDIR); 3128 } 3129 3130 /* 3131 * Called from within softdep_disk_write_complete above to restore 3132 * in-memory inode block contents to their most up-to-date state. Note 3133 * that this routine is always called from interrupt level with further 3134 * splbio interrupts blocked. 3135 */ 3136 static int 3137 handle_written_inodeblock(inodedep, bp) 3138 struct inodedep *inodedep; 3139 struct buf *bp; /* buffer containing the inode block */ 3140 { 3141 struct worklist *wk, *filefree; 3142 struct allocdirect *adp, *nextadp; 3143 struct dinode *dp; 3144 int hadchanges; 3145 3146 if ((inodedep->id_state & IOSTARTED) == 0) 3147 panic("handle_written_inodeblock: not started"); 3148 inodedep->id_state &= ~IOSTARTED; 3149 inodedep->id_state |= COMPLETE; 3150 dp = (struct dinode *)bp->b_data + 3151 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 3152 /* 3153 * If we had to rollback the inode allocation because of 3154 * bitmaps being incomplete, then simply restore it. 3155 * Keep the block dirty so that it will not be reclaimed until 3156 * all associated dependencies have been cleared and the 3157 * corresponding updates written to disk. 3158 */ 3159 if (inodedep->id_savedino != NULL) { 3160 *dp = *inodedep->id_savedino; 3161 FREE(inodedep->id_savedino, M_INODEDEP); 3162 inodedep->id_savedino = NULL; 3163 if ((bp->b_flags & B_DELWRI) == 0) 3164 stat_inode_bitmap++; 3165 bdirty(bp); 3166 return (1); 3167 } 3168 /* 3169 * Roll forward anything that had to be rolled back before 3170 * the inode could be updated. 3171 */ 3172 hadchanges = 0; 3173 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 3174 nextadp = TAILQ_NEXT(adp, ad_next); 3175 if (adp->ad_state & ATTACHED) 3176 panic("handle_written_inodeblock: new entry"); 3177 if (adp->ad_lbn < NDADDR) { 3178 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) 3179 panic("%s: %s #%ld mismatch %d != %d", 3180 "handle_written_inodeblock", 3181 "direct pointer", adp->ad_lbn, 3182 dp->di_db[adp->ad_lbn], adp->ad_oldblkno); 3183 dp->di_db[adp->ad_lbn] = adp->ad_newblkno; 3184 } else { 3185 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) 3186 panic("%s: %s #%ld allocated as %d", 3187 "handle_written_inodeblock", 3188 "indirect pointer", adp->ad_lbn - NDADDR, 3189 dp->di_ib[adp->ad_lbn - NDADDR]); 3190 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno; 3191 } 3192 adp->ad_state &= ~UNDONE; 3193 adp->ad_state |= ATTACHED; 3194 hadchanges = 1; 3195 } 3196 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 3197 stat_direct_blk_ptrs++; 3198 /* 3199 * Reset the file size to its most up-to-date value. 3200 */ 3201 if (inodedep->id_savedsize == -1) 3202 panic("handle_written_inodeblock: bad size"); 3203 if (dp->di_size != inodedep->id_savedsize) { 3204 dp->di_size = inodedep->id_savedsize; 3205 hadchanges = 1; 3206 } 3207 inodedep->id_savedsize = -1; 3208 /* 3209 * If there were any rollbacks in the inode block, then it must be 3210 * marked dirty so that its will eventually get written back in 3211 * its correct form. 3212 */ 3213 if (hadchanges) 3214 bdirty(bp); 3215 /* 3216 * Process any allocdirects that completed during the update. 3217 */ 3218 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 3219 handle_allocdirect_partdone(adp); 3220 /* 3221 * Process deallocations that were held pending until the 3222 * inode had been written to disk. Freeing of the inode 3223 * is delayed until after all blocks have been freed to 3224 * avoid creation of new <vfsid, inum, lbn> triples 3225 * before the old ones have been deleted. 3226 */ 3227 filefree = NULL; 3228 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 3229 WORKLIST_REMOVE(wk); 3230 switch (wk->wk_type) { 3231 3232 case D_FREEFILE: 3233 /* 3234 * We defer adding filefree to the worklist until 3235 * all other additions have been made to ensure 3236 * that it will be done after all the old blocks 3237 * have been freed. 3238 */ 3239 if (filefree != NULL) 3240 panic("handle_written_inodeblock: filefree"); 3241 filefree = wk; 3242 continue; 3243 3244 case D_MKDIR: 3245 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 3246 continue; 3247 3248 case D_DIRADD: 3249 diradd_inode_written(WK_DIRADD(wk), inodedep); 3250 continue; 3251 3252 case D_FREEBLKS: 3253 case D_FREEFRAG: 3254 case D_DIRREM: 3255 add_to_worklist(wk); 3256 continue; 3257 3258 default: 3259 panic("handle_written_inodeblock: Unknown type %s", 3260 TYPENAME(wk->wk_type)); 3261 /* NOTREACHED */ 3262 } 3263 } 3264 if (filefree != NULL) { 3265 if (free_inodedep(inodedep) == 0) 3266 panic("handle_written_inodeblock: live inodedep"); 3267 add_to_worklist(filefree); 3268 return (0); 3269 } 3270 3271 /* 3272 * If no outstanding dependencies, free it. 3273 */ 3274 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0) 3275 return (0); 3276 return (hadchanges); 3277 } 3278 3279 /* 3280 * Process a diradd entry after its dependent inode has been written. 3281 * This routine must be called with splbio interrupts blocked. 3282 */ 3283 static void 3284 diradd_inode_written(dap, inodedep) 3285 struct diradd *dap; 3286 struct inodedep *inodedep; 3287 { 3288 struct pagedep *pagedep; 3289 3290 dap->da_state |= COMPLETE; 3291 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3292 if (dap->da_state & DIRCHG) 3293 pagedep = dap->da_previous->dm_pagedep; 3294 else 3295 pagedep = dap->da_pagedep; 3296 LIST_REMOVE(dap, da_pdlist); 3297 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3298 } 3299 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 3300 } 3301 3302 /* 3303 * Handle the completion of a mkdir dependency. 3304 */ 3305 static void 3306 handle_written_mkdir(mkdir, type) 3307 struct mkdir *mkdir; 3308 int type; 3309 { 3310 struct diradd *dap; 3311 struct pagedep *pagedep; 3312 3313 if (mkdir->md_state != type) 3314 panic("handle_written_mkdir: bad type"); 3315 dap = mkdir->md_diradd; 3316 dap->da_state &= ~type; 3317 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 3318 dap->da_state |= DEPCOMPLETE; 3319 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3320 if (dap->da_state & DIRCHG) 3321 pagedep = dap->da_previous->dm_pagedep; 3322 else 3323 pagedep = dap->da_pagedep; 3324 LIST_REMOVE(dap, da_pdlist); 3325 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3326 } 3327 LIST_REMOVE(mkdir, md_mkdirs); 3328 WORKITEM_FREE(mkdir, D_MKDIR); 3329 } 3330 3331 /* 3332 * Called from within softdep_disk_write_complete above. 3333 * A write operation was just completed. Removed inodes can 3334 * now be freed and associated block pointers may be committed. 3335 * Note that this routine is always called from interrupt level 3336 * with further splbio interrupts blocked. 3337 */ 3338 static int 3339 handle_written_filepage(pagedep, bp) 3340 struct pagedep *pagedep; 3341 struct buf *bp; /* buffer containing the written page */ 3342 { 3343 struct dirrem *dirrem; 3344 struct diradd *dap, *nextdap; 3345 struct direct *ep; 3346 int i, chgs; 3347 3348 if ((pagedep->pd_state & IOSTARTED) == 0) 3349 panic("handle_written_filepage: not started"); 3350 pagedep->pd_state &= ~IOSTARTED; 3351 /* 3352 * Process any directory removals that have been committed. 3353 */ 3354 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 3355 LIST_REMOVE(dirrem, dm_next); 3356 dirrem->dm_dirinum = pagedep->pd_ino; 3357 add_to_worklist(&dirrem->dm_list); 3358 } 3359 /* 3360 * Free any directory additions that have been committed. 3361 */ 3362 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 3363 free_diradd(dap); 3364 /* 3365 * Uncommitted directory entries must be restored. 3366 */ 3367 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 3368 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 3369 dap = nextdap) { 3370 nextdap = LIST_NEXT(dap, da_pdlist); 3371 if (dap->da_state & ATTACHED) 3372 panic("handle_written_filepage: attached"); 3373 ep = (struct direct *) 3374 ((char *)bp->b_data + dap->da_offset); 3375 ep->d_ino = dap->da_newinum; 3376 dap->da_state &= ~UNDONE; 3377 dap->da_state |= ATTACHED; 3378 chgs = 1; 3379 /* 3380 * If the inode referenced by the directory has 3381 * been written out, then the dependency can be 3382 * moved to the pending list. 3383 */ 3384 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3385 LIST_REMOVE(dap, da_pdlist); 3386 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 3387 da_pdlist); 3388 } 3389 } 3390 } 3391 /* 3392 * If there were any rollbacks in the directory, then it must be 3393 * marked dirty so that its will eventually get written back in 3394 * its correct form. 3395 */ 3396 if (chgs) { 3397 if ((bp->b_flags & B_DELWRI) == 0) 3398 stat_dir_entry++; 3399 bdirty(bp); 3400 } 3401 /* 3402 * If no dependencies remain, the pagedep will be freed. 3403 * Otherwise it will remain to update the page before it 3404 * is written back to disk. 3405 */ 3406 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) { 3407 for (i = 0; i < DAHASHSZ; i++) 3408 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL) 3409 break; 3410 if (i == DAHASHSZ) { 3411 LIST_REMOVE(pagedep, pd_hash); 3412 WORKITEM_FREE(pagedep, D_PAGEDEP); 3413 return (0); 3414 } 3415 } 3416 return (1); 3417 } 3418 3419 /* 3420 * Writing back in-core inode structures. 3421 * 3422 * The file system only accesses an inode's contents when it occupies an 3423 * "in-core" inode structure. These "in-core" structures are separate from 3424 * the page frames used to cache inode blocks. Only the latter are 3425 * transferred to/from the disk. So, when the updated contents of the 3426 * "in-core" inode structure are copied to the corresponding in-memory inode 3427 * block, the dependencies are also transferred. The following procedure is 3428 * called when copying a dirty "in-core" inode to a cached inode block. 3429 */ 3430 3431 /* 3432 * Called when an inode is loaded from disk. If the effective link count 3433 * differed from the actual link count when it was last flushed, then we 3434 * need to ensure that the correct effective link count is put back. 3435 */ 3436 void 3437 softdep_load_inodeblock(ip) 3438 struct inode *ip; /* the "in_core" copy of the inode */ 3439 { 3440 struct inodedep *inodedep; 3441 3442 /* 3443 * Check for alternate nlink count. 3444 */ 3445 ip->i_effnlink = ip->i_nlink; 3446 ACQUIRE_LOCK(&lk); 3447 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3448 FREE_LOCK(&lk); 3449 return; 3450 } 3451 if (inodedep->id_nlinkdelta != 0) { 3452 ip->i_effnlink -= inodedep->id_nlinkdelta; 3453 ip->i_flag |= IN_MODIFIED; 3454 inodedep->id_nlinkdelta = 0; 3455 (void) free_inodedep(inodedep); 3456 } 3457 FREE_LOCK(&lk); 3458 } 3459 3460 /* 3461 * This routine is called just before the "in-core" inode 3462 * information is to be copied to the in-memory inode block. 3463 * Recall that an inode block contains several inodes. If 3464 * the force flag is set, then the dependencies will be 3465 * cleared so that the update can always be made. Note that 3466 * the buffer is locked when this routine is called, so we 3467 * will never be in the middle of writing the inode block 3468 * to disk. 3469 */ 3470 void 3471 softdep_update_inodeblock(ip, bp, waitfor) 3472 struct inode *ip; /* the "in_core" copy of the inode */ 3473 struct buf *bp; /* the buffer containing the inode block */ 3474 int waitfor; /* nonzero => update must be allowed */ 3475 { 3476 struct inodedep *inodedep; 3477 struct worklist *wk; 3478 int error, gotit; 3479 3480 /* 3481 * If the effective link count is not equal to the actual link 3482 * count, then we must track the difference in an inodedep while 3483 * the inode is (potentially) tossed out of the cache. Otherwise, 3484 * if there is no existing inodedep, then there are no dependencies 3485 * to track. 3486 */ 3487 ACQUIRE_LOCK(&lk); 3488 if (ip->i_effnlink != ip->i_nlink) { 3489 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, 3490 &inodedep); 3491 } else if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3492 FREE_LOCK(&lk); 3493 return; 3494 } 3495 if (ip->i_nlink < ip->i_effnlink) 3496 panic("softdep_update_inodeblock: bad delta"); 3497 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3498 /* 3499 * Changes have been initiated. Anything depending on these 3500 * changes cannot occur until this inode has been written. 3501 */ 3502 inodedep->id_state &= ~COMPLETE; 3503 if ((inodedep->id_state & ONWORKLIST) == 0) 3504 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 3505 /* 3506 * Any new dependencies associated with the incore inode must 3507 * now be moved to the list associated with the buffer holding 3508 * the in-memory copy of the inode. Once merged process any 3509 * allocdirects that are completed by the merger. 3510 */ 3511 merge_inode_lists(inodedep); 3512 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL) 3513 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); 3514 /* 3515 * Now that the inode has been pushed into the buffer, the 3516 * operations dependent on the inode being written to disk 3517 * can be moved to the id_bufwait so that they will be 3518 * processed when the buffer I/O completes. 3519 */ 3520 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 3521 WORKLIST_REMOVE(wk); 3522 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 3523 } 3524 /* 3525 * Newly allocated inodes cannot be written until the bitmap 3526 * that allocates them have been written (indicated by 3527 * DEPCOMPLETE being set in id_state). If we are doing a 3528 * forced sync (e.g., an fsync on a file), we force the bitmap 3529 * to be written so that the update can be done. 3530 */ 3531 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) { 3532 FREE_LOCK(&lk); 3533 return; 3534 } 3535 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 3536 FREE_LOCK(&lk); 3537 if (gotit && 3538 (error = VOP_BWRITE(inodedep->id_buf->b_vp, inodedep->id_buf)) != 0) 3539 softdep_error("softdep_update_inodeblock: bwrite", error); 3540 if ((inodedep->id_state & DEPCOMPLETE) == 0) 3541 panic("softdep_update_inodeblock: update failed"); 3542 } 3543 3544 /* 3545 * Merge the new inode dependency list (id_newinoupdt) into the old 3546 * inode dependency list (id_inoupdt). This routine must be called 3547 * with splbio interrupts blocked. 3548 */ 3549 static void 3550 merge_inode_lists(inodedep) 3551 struct inodedep *inodedep; 3552 { 3553 struct allocdirect *listadp, *newadp; 3554 3555 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 3556 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) { 3557 if (listadp->ad_lbn < newadp->ad_lbn) { 3558 listadp = TAILQ_NEXT(listadp, ad_next); 3559 continue; 3560 } 3561 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 3562 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 3563 if (listadp->ad_lbn == newadp->ad_lbn) { 3564 allocdirect_merge(&inodedep->id_inoupdt, newadp, 3565 listadp); 3566 listadp = newadp; 3567 } 3568 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 3569 } 3570 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) { 3571 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 3572 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next); 3573 } 3574 } 3575 3576 /* 3577 * If we are doing an fsync, then we must ensure that any directory 3578 * entries for the inode have been written after the inode gets to disk. 3579 */ 3580 int 3581 softdep_fsync(vp) 3582 struct vnode *vp; /* the "in_core" copy of the inode */ 3583 { 3584 struct diradd *dap, *olddap; 3585 struct inodedep *inodedep; 3586 struct pagedep *pagedep; 3587 struct worklist *wk; 3588 struct mount *mnt; 3589 struct vnode *pvp; 3590 struct inode *ip; 3591 struct buf *bp; 3592 struct fs *fs; 3593 struct proc *p = CURPROC; /* XXX */ 3594 int error, ret, flushparent; 3595 ino_t parentino; 3596 ufs_lbn_t lbn; 3597 3598 ip = VTOI(vp); 3599 fs = ip->i_fs; 3600 for (error = 0, flushparent = 0, olddap = NULL; ; ) { 3601 ACQUIRE_LOCK(&lk); 3602 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) 3603 break; 3604 if (LIST_FIRST(&inodedep->id_inowait) != NULL || 3605 LIST_FIRST(&inodedep->id_bufwait) != NULL || 3606 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 3607 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) 3608 panic("softdep_fsync: pending ops"); 3609 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 3610 break; 3611 if (wk->wk_type != D_DIRADD) 3612 panic("softdep_fsync: Unexpected type %s", 3613 TYPENAME(wk->wk_type)); 3614 dap = WK_DIRADD(wk); 3615 /* 3616 * If we have failed to get rid of all the dependencies 3617 * then something is seriously wrong. 3618 */ 3619 if (dap == olddap) 3620 panic("softdep_fsync: flush failed"); 3621 olddap = dap; 3622 /* 3623 * Flush our parent if this directory entry 3624 * has a MKDIR_PARENT dependency. 3625 */ 3626 if (dap->da_state & DIRCHG) 3627 pagedep = dap->da_previous->dm_pagedep; 3628 else 3629 pagedep = dap->da_pagedep; 3630 mnt = pagedep->pd_mnt; 3631 parentino = pagedep->pd_ino; 3632 lbn = pagedep->pd_lbn; 3633 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) 3634 panic("softdep_fsync: dirty"); 3635 flushparent = dap->da_state & MKDIR_PARENT; 3636 /* 3637 * If we are being fsync'ed as part of vgone'ing this vnode, 3638 * then we will not be able to release and recover the 3639 * vnode below, so we just have to give up on writing its 3640 * directory entry out. It will eventually be written, just 3641 * not now, but then the user was not asking to have it 3642 * written, so we are not breaking any promises. 3643 */ 3644 if (vp->v_flag & VXLOCK) 3645 break; 3646 /* 3647 * We prevent deadlock by always fetching inodes from the 3648 * root, moving down the directory tree. Thus, when fetching 3649 * our parent directory, we must unlock ourselves before 3650 * requesting the lock on our parent. See the comment in 3651 * ufs_lookup for details on possible races. 3652 */ 3653 FREE_LOCK(&lk); 3654 VOP_UNLOCK(vp, 0, p); 3655 if ((error = VFS_VGET(mnt, parentino, &pvp)) != 0) { 3656 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, p); 3657 return (error); 3658 } 3659 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, p); 3660 if (flushparent) { 3661 if ((error = UFS_UPDATE(pvp, 1)) != 0) { 3662 vput(pvp); 3663 return (error); 3664 } 3665 } 3666 /* 3667 * Flush directory page containing the inode's name. 3668 */ 3669 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), p->p_ucred, 3670 &bp); 3671 ret = VOP_BWRITE(bp->b_vp, bp); 3672 vput(pvp); 3673 if (error != 0) 3674 return (error); 3675 if (ret != 0) 3676 return (ret); 3677 } 3678 FREE_LOCK(&lk); 3679 return (0); 3680 } 3681 3682 /* 3683 * Flush all the dirty bitmaps associated with the block device 3684 * before flushing the rest of the dirty blocks so as to reduce 3685 * the number of dependencies that will have to be rolled back. 3686 */ 3687 void 3688 softdep_fsync_mountdev(vp) 3689 struct vnode *vp; 3690 { 3691 struct buf *bp, *nbp; 3692 struct worklist *wk; 3693 3694 if (vp->v_type != VBLK) 3695 panic("softdep_fsync_mountdev: vnode not VBLK"); 3696 ACQUIRE_LOCK(&lk); 3697 for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 3698 nbp = TAILQ_NEXT(bp, b_vnbufs); 3699 /* 3700 * If it is already scheduled, skip to the next buffer. 3701 */ 3702 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) 3703 continue; 3704 if ((bp->b_flags & B_DELWRI) == 0) 3705 panic("softdep_fsync_mountdev: not dirty"); 3706 /* 3707 * We are only interested in bitmaps with outstanding 3708 * dependencies. 3709 */ 3710 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 3711 wk->wk_type != D_BMSAFEMAP) { 3712 BUF_UNLOCK(bp); 3713 continue; 3714 } 3715 bremfree(bp); 3716 FREE_LOCK(&lk); 3717 (void) bawrite(bp); 3718 ACQUIRE_LOCK(&lk); 3719 /* 3720 * Since we may have slept during the I/O, we need 3721 * to start from a known point. 3722 */ 3723 nbp = TAILQ_FIRST(&vp->v_dirtyblkhd); 3724 } 3725 drain_output(vp, 1); 3726 FREE_LOCK(&lk); 3727 } 3728 3729 /* 3730 * This routine is called when we are trying to synchronously flush a 3731 * file. This routine must eliminate any filesystem metadata dependencies 3732 * so that the syncing routine can succeed by pushing the dirty blocks 3733 * associated with the file. If any I/O errors occur, they are returned. 3734 */ 3735 int 3736 softdep_sync_metadata(ap) 3737 struct vop_fsync_args /* { 3738 struct vnode *a_vp; 3739 struct ucred *a_cred; 3740 int a_waitfor; 3741 struct proc *a_p; 3742 } */ *ap; 3743 { 3744 struct vnode *vp = ap->a_vp; 3745 struct pagedep *pagedep; 3746 struct allocdirect *adp; 3747 struct allocindir *aip; 3748 struct buf *bp, *nbp; 3749 struct worklist *wk; 3750 int i, error, waitfor; 3751 3752 /* 3753 * Check whether this vnode is involved in a filesystem 3754 * that is doing soft dependency processing. 3755 */ 3756 if (vp->v_type != VBLK) { 3757 if (!DOINGSOFTDEP(vp)) 3758 return (0); 3759 } else 3760 if (vp->v_specmountpoint == NULL || 3761 (vp->v_specmountpoint->mnt_flag & MNT_SOFTDEP) == 0) 3762 return (0); 3763 /* 3764 * Ensure that any direct block dependencies have been cleared. 3765 */ 3766 ACQUIRE_LOCK(&lk); 3767 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) { 3768 FREE_LOCK(&lk); 3769 return (error); 3770 } 3771 /* 3772 * For most files, the only metadata dependencies are the 3773 * cylinder group maps that allocate their inode or blocks. 3774 * The block allocation dependencies can be found by traversing 3775 * the dependency lists for any buffers that remain on their 3776 * dirty buffer list. The inode allocation dependency will 3777 * be resolved when the inode is updated with MNT_WAIT. 3778 * This work is done in two passes. The first pass grabs most 3779 * of the buffers and begins asynchronously writing them. The 3780 * only way to wait for these asynchronous writes is to sleep 3781 * on the filesystem vnode which may stay busy for a long time 3782 * if the filesystem is active. So, instead, we make a second 3783 * pass over the dependencies blocking on each write. In the 3784 * usual case we will be blocking against a write that we 3785 * initiated, so when it is done the dependency will have been 3786 * resolved. Thus the second pass is expected to end quickly. 3787 */ 3788 waitfor = MNT_NOWAIT; 3789 top: 3790 if (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT) == 0) { 3791 FREE_LOCK(&lk); 3792 return (0); 3793 } 3794 bp = TAILQ_FIRST(&vp->v_dirtyblkhd); 3795 loop: 3796 /* 3797 * As we hold the buffer locked, none of its dependencies 3798 * will disappear. 3799 */ 3800 for (wk = LIST_FIRST(&bp->b_dep); wk; 3801 wk = LIST_NEXT(wk, wk_list)) { 3802 switch (wk->wk_type) { 3803 3804 case D_ALLOCDIRECT: 3805 adp = WK_ALLOCDIRECT(wk); 3806 if (adp->ad_state & DEPCOMPLETE) 3807 break; 3808 nbp = adp->ad_buf; 3809 if (getdirtybuf(&nbp, waitfor) == 0) 3810 break; 3811 FREE_LOCK(&lk); 3812 if (waitfor == MNT_NOWAIT) { 3813 bawrite(nbp); 3814 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 3815 bawrite(bp); 3816 return (error); 3817 } 3818 ACQUIRE_LOCK(&lk); 3819 break; 3820 3821 case D_ALLOCINDIR: 3822 aip = WK_ALLOCINDIR(wk); 3823 if (aip->ai_state & DEPCOMPLETE) 3824 break; 3825 nbp = aip->ai_buf; 3826 if (getdirtybuf(&nbp, waitfor) == 0) 3827 break; 3828 FREE_LOCK(&lk); 3829 if (waitfor == MNT_NOWAIT) { 3830 bawrite(nbp); 3831 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 3832 bawrite(bp); 3833 return (error); 3834 } 3835 ACQUIRE_LOCK(&lk); 3836 break; 3837 3838 case D_INDIRDEP: 3839 restart: 3840 for (aip = LIST_FIRST(&WK_INDIRDEP(wk)->ir_deplisthd); 3841 aip; aip = LIST_NEXT(aip, ai_next)) { 3842 if (aip->ai_state & DEPCOMPLETE) 3843 continue; 3844 nbp = aip->ai_buf; 3845 if (getdirtybuf(&nbp, MNT_WAIT) == 0) 3846 goto restart; 3847 FREE_LOCK(&lk); 3848 if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 3849 bawrite(bp); 3850 return (error); 3851 } 3852 ACQUIRE_LOCK(&lk); 3853 goto restart; 3854 } 3855 break; 3856 3857 case D_INODEDEP: 3858 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs, 3859 WK_INODEDEP(wk)->id_ino)) != 0) { 3860 FREE_LOCK(&lk); 3861 bawrite(bp); 3862 return (error); 3863 } 3864 break; 3865 3866 case D_PAGEDEP: 3867 /* 3868 * We are trying to sync a directory that may 3869 * have dependencies on both its own metadata 3870 * and/or dependencies on the inodes of any 3871 * recently allocated files. We walk its diradd 3872 * lists pushing out the associated inode. 3873 */ 3874 pagedep = WK_PAGEDEP(wk); 3875 for (i = 0; i < DAHASHSZ; i++) { 3876 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 3877 continue; 3878 if ((error = 3879 flush_pagedep_deps(vp, pagedep->pd_mnt, 3880 &pagedep->pd_diraddhd[i]))) { 3881 FREE_LOCK(&lk); 3882 bawrite(bp); 3883 return (error); 3884 } 3885 } 3886 break; 3887 3888 case D_MKDIR: 3889 /* 3890 * This case should never happen if the vnode has 3891 * been properly sync'ed. However, if this function 3892 * is used at a place where the vnode has not yet 3893 * been sync'ed, this dependency can show up. So, 3894 * rather than panic, just flush it. 3895 */ 3896 nbp = WK_MKDIR(wk)->md_buf; 3897 if (getdirtybuf(&nbp, waitfor) == 0) 3898 break; 3899 FREE_LOCK(&lk); 3900 if (waitfor == MNT_NOWAIT) { 3901 bawrite(nbp); 3902 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 3903 bawrite(bp); 3904 return (error); 3905 } 3906 ACQUIRE_LOCK(&lk); 3907 break; 3908 3909 case D_BMSAFEMAP: 3910 /* 3911 * This case should never happen if the vnode has 3912 * been properly sync'ed. However, if this function 3913 * is used at a place where the vnode has not yet 3914 * been sync'ed, this dependency can show up. So, 3915 * rather than panic, just flush it. 3916 */ 3917 nbp = WK_BMSAFEMAP(wk)->sm_buf; 3918 if (getdirtybuf(&nbp, waitfor) == 0) 3919 break; 3920 FREE_LOCK(&lk); 3921 if (waitfor == MNT_NOWAIT) { 3922 bawrite(nbp); 3923 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 3924 bawrite(bp); 3925 return (error); 3926 } 3927 ACQUIRE_LOCK(&lk); 3928 break; 3929 3930 default: 3931 panic("softdep_sync_metadata: Unknown type %s", 3932 TYPENAME(wk->wk_type)); 3933 /* NOTREACHED */ 3934 } 3935 } 3936 (void) getdirtybuf(&TAILQ_NEXT(bp, b_vnbufs), MNT_WAIT); 3937 nbp = TAILQ_NEXT(bp, b_vnbufs); 3938 FREE_LOCK(&lk); 3939 bawrite(bp); 3940 ACQUIRE_LOCK(&lk); 3941 if (nbp != NULL) { 3942 bp = nbp; 3943 goto loop; 3944 } 3945 /* 3946 * We must wait for any I/O in progress to finish so that 3947 * all potential buffers on the dirty list will be visible. 3948 * Once they are all there, proceed with the second pass 3949 * which will wait for the I/O as per above. 3950 */ 3951 drain_output(vp, 1); 3952 /* 3953 * The brief unlock is to allow any pent up dependency 3954 * processing to be done. 3955 */ 3956 if (waitfor == MNT_NOWAIT) { 3957 waitfor = MNT_WAIT; 3958 FREE_LOCK(&lk); 3959 ACQUIRE_LOCK(&lk); 3960 goto top; 3961 } 3962 3963 /* 3964 * If we have managed to get rid of all the dirty buffers, 3965 * then we are done. For certain directories and block 3966 * devices, we may need to do further work. 3967 */ 3968 if (TAILQ_FIRST(&vp->v_dirtyblkhd) == NULL) { 3969 FREE_LOCK(&lk); 3970 return (0); 3971 } 3972 3973 FREE_LOCK(&lk); 3974 /* 3975 * If we are trying to sync a block device, some of its buffers may 3976 * contain metadata that cannot be written until the contents of some 3977 * partially written files have been written to disk. The only easy 3978 * way to accomplish this is to sync the entire filesystem (luckily 3979 * this happens rarely). 3980 */ 3981 if (vp->v_type == VBLK && vp->v_specmountpoint && !VOP_ISLOCKED(vp) && 3982 (error = VFS_SYNC(vp->v_specmountpoint, MNT_WAIT, ap->a_cred, 3983 ap->a_p)) != 0) 3984 return (error); 3985 return (0); 3986 } 3987 3988 /* 3989 * Flush the dependencies associated with an inodedep. 3990 * Called with splbio blocked. 3991 */ 3992 static int 3993 flush_inodedep_deps(fs, ino) 3994 struct fs *fs; 3995 ino_t ino; 3996 { 3997 struct inodedep *inodedep; 3998 struct allocdirect *adp; 3999 int error, waitfor; 4000 struct buf *bp; 4001 4002 /* 4003 * This work is done in two passes. The first pass grabs most 4004 * of the buffers and begins asynchronously writing them. The 4005 * only way to wait for these asynchronous writes is to sleep 4006 * on the filesystem vnode which may stay busy for a long time 4007 * if the filesystem is active. So, instead, we make a second 4008 * pass over the dependencies blocking on each write. In the 4009 * usual case we will be blocking against a write that we 4010 * initiated, so when it is done the dependency will have been 4011 * resolved. Thus the second pass is expected to end quickly. 4012 * We give a brief window at the top of the loop to allow 4013 * any pending I/O to complete. 4014 */ 4015 for (waitfor = MNT_NOWAIT; ; ) { 4016 FREE_LOCK(&lk); 4017 ACQUIRE_LOCK(&lk); 4018 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4019 return (0); 4020 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 4021 adp = TAILQ_NEXT(adp, ad_next)) { 4022 if (adp->ad_state & DEPCOMPLETE) 4023 continue; 4024 bp = adp->ad_buf; 4025 if (getdirtybuf(&bp, waitfor) == 0) { 4026 if (waitfor == MNT_NOWAIT) 4027 continue; 4028 break; 4029 } 4030 FREE_LOCK(&lk); 4031 if (waitfor == MNT_NOWAIT) { 4032 bawrite(bp); 4033 } else if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) { 4034 ACQUIRE_LOCK(&lk); 4035 return (error); 4036 } 4037 ACQUIRE_LOCK(&lk); 4038 break; 4039 } 4040 if (adp != NULL) 4041 continue; 4042 for (adp = TAILQ_FIRST(&inodedep->id_newinoupdt); adp; 4043 adp = TAILQ_NEXT(adp, ad_next)) { 4044 if (adp->ad_state & DEPCOMPLETE) 4045 continue; 4046 bp = adp->ad_buf; 4047 if (getdirtybuf(&bp, waitfor) == 0) { 4048 if (waitfor == MNT_NOWAIT) 4049 continue; 4050 break; 4051 } 4052 FREE_LOCK(&lk); 4053 if (waitfor == MNT_NOWAIT) { 4054 bawrite(bp); 4055 } else if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) { 4056 ACQUIRE_LOCK(&lk); 4057 return (error); 4058 } 4059 ACQUIRE_LOCK(&lk); 4060 break; 4061 } 4062 if (adp != NULL) 4063 continue; 4064 /* 4065 * If pass2, we are done, otherwise do pass 2. 4066 */ 4067 if (waitfor == MNT_WAIT) 4068 break; 4069 waitfor = MNT_WAIT; 4070 } 4071 /* 4072 * Try freeing inodedep in case all dependencies have been removed. 4073 */ 4074 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0) 4075 (void) free_inodedep(inodedep); 4076 return (0); 4077 } 4078 4079 /* 4080 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 4081 * Called with splbio blocked. 4082 */ 4083 static int 4084 flush_pagedep_deps(pvp, mp, diraddhdp) 4085 struct vnode *pvp; 4086 struct mount *mp; 4087 struct diraddhd *diraddhdp; 4088 { 4089 struct proc *p = CURPROC; /* XXX */ 4090 struct inodedep *inodedep; 4091 struct ufsmount *ump; 4092 struct diradd *dap; 4093 struct vnode *vp; 4094 int gotit, error = 0; 4095 struct buf *bp; 4096 ino_t inum; 4097 4098 ump = VFSTOUFS(mp); 4099 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 4100 /* 4101 * Flush ourselves if this directory entry 4102 * has a MKDIR_PARENT dependency. 4103 */ 4104 if (dap->da_state & MKDIR_PARENT) { 4105 FREE_LOCK(&lk); 4106 if ((error = UFS_UPDATE(pvp, 1)) != 0) 4107 break; 4108 ACQUIRE_LOCK(&lk); 4109 /* 4110 * If that cleared dependencies, go on to next. 4111 */ 4112 if (dap != LIST_FIRST(diraddhdp)) 4113 continue; 4114 if (dap->da_state & MKDIR_PARENT) 4115 panic("flush_pagedep_deps: MKDIR"); 4116 } 4117 /* 4118 * Flush the file on which the directory entry depends. 4119 * If the inode has already been pushed out of the cache, 4120 * then all the block dependencies will have been flushed 4121 * leaving only inode dependencies (e.g., bitmaps). Thus, 4122 * we do a ufs_ihashget to check for the vnode in the cache. 4123 * If it is there, we do a full flush. If it is no longer 4124 * there we need only dispose of any remaining bitmap 4125 * dependencies and write the inode to disk. 4126 */ 4127 inum = dap->da_newinum; 4128 FREE_LOCK(&lk); 4129 if ((vp = ufs_ihashget(ump->um_dev, inum)) == NULL) { 4130 ACQUIRE_LOCK(&lk); 4131 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0 4132 && dap == LIST_FIRST(diraddhdp)) 4133 panic("flush_pagedep_deps: flush 1 failed"); 4134 /* 4135 * If the inode still has bitmap dependencies, 4136 * push them to disk. 4137 */ 4138 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4139 gotit = getdirtybuf(&inodedep->id_buf,MNT_WAIT); 4140 FREE_LOCK(&lk); 4141 if (gotit && 4142 (error = VOP_BWRITE(inodedep->id_buf->b_vp, 4143 inodedep->id_buf)) != 0) 4144 break; 4145 ACQUIRE_LOCK(&lk); 4146 } 4147 if (dap != LIST_FIRST(diraddhdp)) 4148 continue; 4149 /* 4150 * If the inode is still sitting in a buffer waiting 4151 * to be written, push it to disk. 4152 */ 4153 FREE_LOCK(&lk); 4154 if ((error = bread(ump->um_devvp, 4155 fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), 4156 (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0) 4157 break; 4158 if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) 4159 break; 4160 ACQUIRE_LOCK(&lk); 4161 if (dap == LIST_FIRST(diraddhdp)) 4162 panic("flush_pagedep_deps: flush 2 failed"); 4163 continue; 4164 } 4165 if (vp->v_type == VDIR) { 4166 /* 4167 * A newly allocated directory must have its "." and 4168 * ".." entries written out before its name can be 4169 * committed in its parent. We do not want or need 4170 * the full semantics of a synchronous VOP_FSYNC as 4171 * that may end up here again, once for each directory 4172 * level in the filesystem. Instead, we push the blocks 4173 * and wait for them to clear. 4174 */ 4175 if ((error = 4176 VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p))) { 4177 vput(vp); 4178 break; 4179 } 4180 drain_output(vp, 0); 4181 } 4182 error = UFS_UPDATE(vp, 1); 4183 vput(vp); 4184 if (error) 4185 break; 4186 /* 4187 * If we have failed to get rid of all the dependencies 4188 * then something is seriously wrong. 4189 */ 4190 if (dap == LIST_FIRST(diraddhdp)) 4191 panic("flush_pagedep_deps: flush 3 failed"); 4192 ACQUIRE_LOCK(&lk); 4193 } 4194 if (error) 4195 ACQUIRE_LOCK(&lk); 4196 return (error); 4197 } 4198 4199 /* 4200 * A large burst of file addition or deletion activity can drive the 4201 * memory load excessively high. Therefore we deliberately slow things 4202 * down and speed up the I/O processing if we find ourselves with too 4203 * many dependencies in progress. 4204 */ 4205 static int 4206 request_cleanup(resource, islocked) 4207 int resource; 4208 int islocked; 4209 { 4210 struct callout_handle handle; 4211 struct proc *p = CURPROC; 4212 4213 /* 4214 * We never hold up the filesystem syncer process. 4215 */ 4216 if (p == filesys_syncer) 4217 return (0); 4218 /* 4219 * If we are resource constrained on inode dependencies, try 4220 * flushing some dirty inodes. Otherwise, we are constrained 4221 * by file deletions, so try accelerating flushes of directories 4222 * with removal dependencies. We would like to do the cleanup 4223 * here, but we probably hold an inode locked at this point and 4224 * that might deadlock against one that we try to clean. So, 4225 * the best that we can do is request the syncer daemon to do 4226 * the cleanup for us. 4227 */ 4228 switch (resource) { 4229 4230 case FLUSH_INODES: 4231 stat_ino_limit_push += 1; 4232 req_clear_inodedeps = 1; 4233 break; 4234 4235 case FLUSH_REMOVE: 4236 stat_blk_limit_push += 1; 4237 req_clear_remove = 1; 4238 break; 4239 4240 default: 4241 panic("request_cleanup: unknown type"); 4242 } 4243 /* 4244 * Hopefully the syncer daemon will catch up and awaken us. 4245 * We wait at most tickdelay before proceeding in any case. 4246 */ 4247 if (islocked == 0) 4248 ACQUIRE_LOCK(&lk); 4249 if (proc_waiting == 0) { 4250 proc_waiting = 1; 4251 handle = timeout(pause_timer, NULL, 4252 tickdelay > 2 ? tickdelay : 2); 4253 } 4254 FREE_LOCK_INTERLOCKED(&lk); 4255 (void) tsleep((caddr_t)&proc_waiting, PPAUSE | PCATCH, "softupdate", 0); 4256 ACQUIRE_LOCK_INTERLOCKED(&lk); 4257 if (proc_waiting) { 4258 untimeout(pause_timer, NULL, handle); 4259 proc_waiting = 0; 4260 } else { 4261 switch (resource) { 4262 4263 case FLUSH_INODES: 4264 stat_ino_limit_hit += 1; 4265 break; 4266 4267 case FLUSH_REMOVE: 4268 stat_blk_limit_hit += 1; 4269 break; 4270 } 4271 } 4272 if (islocked == 0) 4273 FREE_LOCK(&lk); 4274 return (1); 4275 } 4276 4277 /* 4278 * Awaken processes pausing in request_cleanup and clear proc_waiting 4279 * to indicate that there is no longer a timer running. 4280 */ 4281 void 4282 pause_timer(arg) 4283 void *arg; 4284 { 4285 4286 proc_waiting = 0; 4287 wakeup(&proc_waiting); 4288 } 4289 4290 /* 4291 * Flush out a directory with at least one removal dependency in an effort 4292 * to reduce the number of freefile and freeblks dependency structures. 4293 */ 4294 static void 4295 clear_remove(p) 4296 struct proc *p; 4297 { 4298 struct pagedep_hashhead *pagedephd; 4299 struct pagedep *pagedep; 4300 static int next = 0; 4301 struct mount *mp; 4302 struct vnode *vp; 4303 int error, cnt; 4304 ino_t ino; 4305 4306 ACQUIRE_LOCK(&lk); 4307 for (cnt = 0; cnt < pagedep_hash; cnt++) { 4308 pagedephd = &pagedep_hashtbl[next++]; 4309 if (next >= pagedep_hash) 4310 next = 0; 4311 for (pagedep = LIST_FIRST(pagedephd); pagedep; 4312 pagedep = LIST_NEXT(pagedep, pd_hash)) { 4313 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL) 4314 continue; 4315 mp = pagedep->pd_mnt; 4316 ino = pagedep->pd_ino; 4317 FREE_LOCK(&lk); 4318 if ((error = VFS_VGET(mp, ino, &vp)) != 0) { 4319 softdep_error("clear_remove: vget", error); 4320 return; 4321 } 4322 if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p))) 4323 softdep_error("clear_remove: fsync", error); 4324 drain_output(vp, 0); 4325 vput(vp); 4326 return; 4327 } 4328 } 4329 FREE_LOCK(&lk); 4330 } 4331 4332 /* 4333 * Clear out a block of dirty inodes in an effort to reduce 4334 * the number of inodedep dependency structures. 4335 */ 4336 static void 4337 clear_inodedeps(p) 4338 struct proc *p; 4339 { 4340 struct inodedep_hashhead *inodedephd; 4341 struct inodedep *inodedep; 4342 static int next = 0; 4343 struct mount *mp; 4344 struct vnode *vp; 4345 struct fs *fs; 4346 int error, cnt; 4347 ino_t firstino, lastino, ino; 4348 4349 ACQUIRE_LOCK(&lk); 4350 /* 4351 * Pick a random inode dependency to be cleared. 4352 * We will then gather up all the inodes in its block 4353 * that have dependencies and flush them out. 4354 */ 4355 for (cnt = 0; cnt < inodedep_hash; cnt++) { 4356 inodedephd = &inodedep_hashtbl[next++]; 4357 if (next >= inodedep_hash) 4358 next = 0; 4359 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 4360 break; 4361 } 4362 /* 4363 * Ugly code to find mount point given pointer to superblock. 4364 */ 4365 fs = inodedep->id_fs; 4366 for (mp = CIRCLEQ_FIRST(&mountlist); mp != (void *)&mountlist; 4367 mp = CIRCLEQ_NEXT(mp, mnt_list)) 4368 if ((mp->mnt_flag & MNT_SOFTDEP) && fs == VFSTOUFS(mp)->um_fs) 4369 break; 4370 /* 4371 * Find the last inode in the block with dependencies. 4372 */ 4373 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 4374 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 4375 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0) 4376 break; 4377 /* 4378 * Asynchronously push all but the last inode with dependencies. 4379 * Synchronously push the last inode with dependencies to ensure 4380 * that the inode block gets written to free up the inodedeps. 4381 */ 4382 for (ino = firstino; ino <= lastino; ino++) { 4383 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4384 continue; 4385 FREE_LOCK(&lk); 4386 if ((error = VFS_VGET(mp, ino, &vp)) != 0) { 4387 softdep_error("clear_inodedeps: vget", error); 4388 return; 4389 } 4390 if (ino == lastino) { 4391 if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_WAIT, p))) 4392 softdep_error("clear_inodedeps: fsync1", error); 4393 } else { 4394 if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p))) 4395 softdep_error("clear_inodedeps: fsync2", error); 4396 drain_output(vp, 0); 4397 } 4398 vput(vp); 4399 ACQUIRE_LOCK(&lk); 4400 } 4401 FREE_LOCK(&lk); 4402 } 4403 4404 /* 4405 * Acquire exclusive access to a buffer. 4406 * Must be called with splbio blocked. 4407 * Return 1 if buffer was acquired. 4408 */ 4409 static int 4410 getdirtybuf(bpp, waitfor) 4411 struct buf **bpp; 4412 int waitfor; 4413 { 4414 struct buf *bp; 4415 4416 for (;;) { 4417 if ((bp = *bpp) == NULL) 4418 return (0); 4419 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) 4420 break; 4421 if (waitfor != MNT_WAIT) 4422 return (0); 4423 FREE_LOCK_INTERLOCKED(&lk); 4424 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL) != ENOLCK) 4425 panic("getdirtybuf: inconsistent lock"); 4426 ACQUIRE_LOCK_INTERLOCKED(&lk); 4427 } 4428 if ((bp->b_flags & B_DELWRI) == 0) { 4429 BUF_UNLOCK(bp); 4430 return (0); 4431 } 4432 bremfree(bp); 4433 return (1); 4434 } 4435 4436 /* 4437 * Wait for pending output on a vnode to complete. 4438 * Must be called with vnode locked. 4439 */ 4440 static void 4441 drain_output(vp, islocked) 4442 struct vnode *vp; 4443 int islocked; 4444 { 4445 4446 if (!islocked) 4447 ACQUIRE_LOCK(&lk); 4448 while (vp->v_numoutput) { 4449 vp->v_flag |= VBWAIT; 4450 FREE_LOCK_INTERLOCKED(&lk); 4451 tsleep((caddr_t)&vp->v_numoutput, PRIBIO + 1, "drainvp", 0); 4452 ACQUIRE_LOCK_INTERLOCKED(&lk); 4453 } 4454 if (!islocked) 4455 FREE_LOCK(&lk); 4456 } 4457 4458 /* 4459 * Called whenever a buffer that is being invalidated or reallocated 4460 * contains dependencies. This should only happen if an I/O error has 4461 * occurred. The routine is called with the buffer locked. 4462 */ 4463 void 4464 softdep_deallocate_dependencies(bp) 4465 struct buf *bp; 4466 { 4467 4468 if ((bp->b_flags & B_ERROR) == 0) 4469 panic("softdep_deallocate_dependencies: dangling deps"); 4470 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); 4471 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 4472 } 4473 4474 /* 4475 * Function to handle asynchronous write errors in the filesystem. 4476 */ 4477 void 4478 softdep_error(func, error) 4479 char *func; 4480 int error; 4481 { 4482 4483 /* XXX should do something better! */ 4484 printf("%s: got error %d while accessing filesystem\n", func, error); 4485 } 4486