ffs_softdep.c revision 71820
1/* 2 * Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved. 3 * 4 * The soft updates code is derived from the appendix of a University 5 * of Michigan technical report (Gregory R. Ganger and Yale N. Patt, 6 * "Soft Updates: A Solution to the Metadata Update Problem in File 7 * Systems", CSE-TR-254-95, August 1995). 8 * 9 * Further information about soft updates can be obtained from: 10 * 11 * Marshall Kirk McKusick http://www.mckusick.com/softdep/ 12 * 1614 Oxford Street mckusick@mckusick.com 13 * Berkeley, CA 94709-1608 +1-510-843-9542 14 * USA 15 * 16 * Redistribution and use in source and binary forms, with or without 17 * modification, are permitted provided that the following conditions 18 * are met: 19 * 20 * 1. Redistributions of source code must retain the above copyright 21 * notice, this list of conditions and the following disclaimer. 22 * 2. Redistributions in binary form must reproduce the above copyright 23 * notice, this list of conditions and the following disclaimer in the 24 * documentation and/or other materials provided with the distribution. 25 * 26 * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY 27 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 28 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 29 * DISCLAIMED. IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR 30 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * from: @(#)ffs_softdep.c 9.59 (McKusick) 6/21/00 39 * $FreeBSD: head/sys/ufs/ffs/ffs_softdep.c 71820 2001-01-30 06:31:59Z dillon $ 40 */ 41 42/* 43 * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide. 44 */ 45#ifndef DIAGNOSTIC 46#define DIAGNOSTIC 47#endif 48#ifndef DEBUG 49#define DEBUG 50#endif 51 52#include <sys/param.h> 53#include <sys/kernel.h> 54#include <sys/systm.h> 55#include <sys/bio.h> 56#include <sys/buf.h> 57#include <sys/malloc.h> 58#include <sys/mount.h> 59#include <sys/proc.h> 60#include <sys/syslog.h> 61#include <sys/vnode.h> 62#include <sys/conf.h> 63#include <ufs/ufs/dir.h> 64#include <ufs/ufs/extattr.h> 65#include <ufs/ufs/quota.h> 66#include <ufs/ufs/inode.h> 67#include <ufs/ufs/ufsmount.h> 68#include <ufs/ffs/fs.h> 69#include <ufs/ffs/softdep.h> 70#include <ufs/ffs/ffs_extern.h> 71#include <ufs/ufs/ufs_extern.h> 72 73/* 74 * These definitions need to be adapted to the system to which 75 * this file is being ported. 76 */ 77/* 78 * malloc types defined for the softdep system. 79 */ 80static MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies"); 81static MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies"); 82static MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation"); 83static MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map"); 84static MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode"); 85static MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies"); 86static MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block"); 87static MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode"); 88static MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode"); 89static MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated"); 90static MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry"); 91static MALLOC_DEFINE(M_MKDIR, "mkdir","New directory"); 92static MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted"); 93 94#define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE) 95 96#define D_PAGEDEP 0 97#define D_INODEDEP 1 98#define D_NEWBLK 2 99#define D_BMSAFEMAP 3 100#define D_ALLOCDIRECT 4 101#define D_INDIRDEP 5 102#define D_ALLOCINDIR 6 103#define D_FREEFRAG 7 104#define D_FREEBLKS 8 105#define D_FREEFILE 9 106#define D_DIRADD 10 107#define D_MKDIR 11 108#define D_DIRREM 12 109#define D_LAST D_DIRREM 110 111/* 112 * translate from workitem type to memory type 113 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 114 */ 115static struct malloc_type *memtype[] = { 116 M_PAGEDEP, 117 M_INODEDEP, 118 M_NEWBLK, 119 M_BMSAFEMAP, 120 M_ALLOCDIRECT, 121 M_INDIRDEP, 122 M_ALLOCINDIR, 123 M_FREEFRAG, 124 M_FREEBLKS, 125 M_FREEFILE, 126 M_DIRADD, 127 M_MKDIR, 128 M_DIRREM 129}; 130 131#define DtoM(type) (memtype[type]) 132 133/* 134 * Names of malloc types. 135 */ 136#define TYPENAME(type) \ 137 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???") 138/* 139 * End system adaptaion definitions. 140 */ 141 142/* 143 * Internal function prototypes. 144 */ 145static void softdep_error __P((char *, int)); 146static void drain_output __P((struct vnode *, int)); 147static int getdirtybuf __P((struct buf **, int)); 148static void clear_remove __P((struct proc *)); 149static void clear_inodedeps __P((struct proc *)); 150static int flush_pagedep_deps __P((struct vnode *, struct mount *, 151 struct diraddhd *)); 152static int flush_inodedep_deps __P((struct fs *, ino_t)); 153static int handle_written_filepage __P((struct pagedep *, struct buf *)); 154static void diradd_inode_written __P((struct diradd *, struct inodedep *)); 155static int handle_written_inodeblock __P((struct inodedep *, struct buf *)); 156static void handle_allocdirect_partdone __P((struct allocdirect *)); 157static void handle_allocindir_partdone __P((struct allocindir *)); 158static void initiate_write_filepage __P((struct pagedep *, struct buf *)); 159static void handle_written_mkdir __P((struct mkdir *, int)); 160static void initiate_write_inodeblock __P((struct inodedep *, struct buf *)); 161static void handle_workitem_freefile __P((struct freefile *)); 162static void handle_workitem_remove __P((struct dirrem *)); 163static struct dirrem *newdirrem __P((struct buf *, struct inode *, 164 struct inode *, int, struct dirrem **)); 165static void free_diradd __P((struct diradd *)); 166static void free_allocindir __P((struct allocindir *, struct inodedep *)); 167static int indir_trunc __P((struct inode *, ufs_daddr_t, int, ufs_lbn_t, 168 long *)); 169static void deallocate_dependencies __P((struct buf *, struct inodedep *)); 170static void free_allocdirect __P((struct allocdirectlst *, 171 struct allocdirect *, int)); 172static int check_inode_unwritten __P((struct inodedep *)); 173static int free_inodedep __P((struct inodedep *)); 174static void handle_workitem_freeblocks __P((struct freeblks *)); 175static void merge_inode_lists __P((struct inodedep *)); 176static void setup_allocindir_phase2 __P((struct buf *, struct inode *, 177 struct allocindir *)); 178static struct allocindir *newallocindir __P((struct inode *, int, ufs_daddr_t, 179 ufs_daddr_t)); 180static void handle_workitem_freefrag __P((struct freefrag *)); 181static struct freefrag *newfreefrag __P((struct inode *, ufs_daddr_t, long)); 182static void allocdirect_merge __P((struct allocdirectlst *, 183 struct allocdirect *, struct allocdirect *)); 184static struct bmsafemap *bmsafemap_lookup __P((struct buf *)); 185static int newblk_lookup __P((struct fs *, ufs_daddr_t, int, 186 struct newblk **)); 187static int inodedep_lookup __P((struct fs *, ino_t, int, struct inodedep **)); 188static int pagedep_lookup __P((struct inode *, ufs_lbn_t, int, 189 struct pagedep **)); 190static void pause_timer __P((void *)); 191static int request_cleanup __P((int, int)); 192static int process_worklist_item __P((struct mount *, int)); 193static void add_to_worklist __P((struct worklist *)); 194 195/* 196 * Exported softdep operations. 197 */ 198static void softdep_disk_io_initiation __P((struct buf *)); 199static void softdep_disk_write_complete __P((struct buf *)); 200static void softdep_deallocate_dependencies __P((struct buf *)); 201static void softdep_move_dependencies __P((struct buf *, struct buf *)); 202static int softdep_count_dependencies __P((struct buf *bp, int)); 203 204struct bio_ops bioops = { 205 softdep_disk_io_initiation, /* io_start */ 206 softdep_disk_write_complete, /* io_complete */ 207 softdep_deallocate_dependencies, /* io_deallocate */ 208 softdep_move_dependencies, /* io_movedeps */ 209 softdep_count_dependencies, /* io_countdeps */ 210}; 211 212/* 213 * Locking primitives. 214 * 215 * For a uniprocessor, all we need to do is protect against disk 216 * interrupts. For a multiprocessor, this lock would have to be 217 * a mutex. A single mutex is used throughout this file, though 218 * finer grain locking could be used if contention warranted it. 219 * 220 * For a multiprocessor, the sleep call would accept a lock and 221 * release it after the sleep processing was complete. In a uniprocessor 222 * implementation there is no such interlock, so we simple mark 223 * the places where it needs to be done with the `interlocked' form 224 * of the lock calls. Since the uniprocessor sleep already interlocks 225 * the spl, there is nothing that really needs to be done. 226 */ 227#ifndef /* NOT */ DEBUG 228static struct lockit { 229 int lkt_spl; 230} lk = { 0 }; 231#define ACQUIRE_LOCK(lk) (lk)->lkt_spl = splbio() 232#define FREE_LOCK(lk) splx((lk)->lkt_spl) 233#define ACQUIRE_LOCK_INTERLOCKED(lk) 234#define FREE_LOCK_INTERLOCKED(lk) 235 236#else /* DEBUG */ 237static struct lockit { 238 int lkt_spl; 239 pid_t lkt_held; 240} lk = { 0, -1 }; 241static int lockcnt; 242 243static void acquire_lock __P((struct lockit *)); 244static void free_lock __P((struct lockit *)); 245static void acquire_lock_interlocked __P((struct lockit *)); 246static void free_lock_interlocked __P((struct lockit *)); 247 248#define ACQUIRE_LOCK(lk) acquire_lock(lk) 249#define FREE_LOCK(lk) free_lock(lk) 250#define ACQUIRE_LOCK_INTERLOCKED(lk) acquire_lock_interlocked(lk) 251#define FREE_LOCK_INTERLOCKED(lk) free_lock_interlocked(lk) 252 253static void 254acquire_lock(lk) 255 struct lockit *lk; 256{ 257 258 if (lk->lkt_held != -1) { 259 if (lk->lkt_held == CURPROC->p_pid) 260 panic("softdep_lock: locking against myself"); 261 else 262 panic("softdep_lock: lock held by %d", lk->lkt_held); 263 } 264 lk->lkt_spl = splbio(); 265 lk->lkt_held = CURPROC->p_pid; 266 lockcnt++; 267} 268 269static void 270free_lock(lk) 271 struct lockit *lk; 272{ 273 274 if (lk->lkt_held == -1) 275 panic("softdep_unlock: lock not held"); 276 lk->lkt_held = -1; 277 splx(lk->lkt_spl); 278} 279 280static void 281acquire_lock_interlocked(lk) 282 struct lockit *lk; 283{ 284 285 if (lk->lkt_held != -1) { 286 if (lk->lkt_held == CURPROC->p_pid) 287 panic("softdep_lock_interlocked: locking against self"); 288 else 289 panic("softdep_lock_interlocked: lock held by %d", 290 lk->lkt_held); 291 } 292 lk->lkt_held = CURPROC->p_pid; 293 lockcnt++; 294} 295 296static void 297free_lock_interlocked(lk) 298 struct lockit *lk; 299{ 300 301 if (lk->lkt_held == -1) 302 panic("softdep_unlock_interlocked: lock not held"); 303 lk->lkt_held = -1; 304} 305#endif /* DEBUG */ 306 307/* 308 * Place holder for real semaphores. 309 */ 310struct sema { 311 int value; 312 pid_t holder; 313 char *name; 314 int prio; 315 int timo; 316}; 317static void sema_init __P((struct sema *, char *, int, int)); 318static int sema_get __P((struct sema *, struct lockit *)); 319static void sema_release __P((struct sema *)); 320 321static void 322sema_init(semap, name, prio, timo) 323 struct sema *semap; 324 char *name; 325 int prio, timo; 326{ 327 328 semap->holder = -1; 329 semap->value = 0; 330 semap->name = name; 331 semap->prio = prio; 332 semap->timo = timo; 333} 334 335static int 336sema_get(semap, interlock) 337 struct sema *semap; 338 struct lockit *interlock; 339{ 340 341 if (semap->value++ > 0) { 342 if (interlock != NULL) 343 FREE_LOCK_INTERLOCKED(interlock); 344 tsleep((caddr_t)semap, semap->prio, semap->name, semap->timo); 345 if (interlock != NULL) { 346 ACQUIRE_LOCK_INTERLOCKED(interlock); 347 FREE_LOCK(interlock); 348 } 349 return (0); 350 } 351 semap->holder = CURPROC->p_pid; 352 if (interlock != NULL) 353 FREE_LOCK(interlock); 354 return (1); 355} 356 357static void 358sema_release(semap) 359 struct sema *semap; 360{ 361 362 if (semap->value <= 0 || semap->holder != CURPROC->p_pid) 363 panic("sema_release: not held"); 364 if (--semap->value > 0) { 365 semap->value = 0; 366 wakeup(semap); 367 } 368 semap->holder = -1; 369} 370 371/* 372 * Worklist queue management. 373 * These routines require that the lock be held. 374 */ 375#ifndef /* NOT */ DEBUG 376#define WORKLIST_INSERT(head, item) do { \ 377 (item)->wk_state |= ONWORKLIST; \ 378 LIST_INSERT_HEAD(head, item, wk_list); \ 379} while (0) 380#define WORKLIST_REMOVE(item) do { \ 381 (item)->wk_state &= ~ONWORKLIST; \ 382 LIST_REMOVE(item, wk_list); \ 383} while (0) 384#define WORKITEM_FREE(item, type) FREE(item, DtoM(type)) 385 386#else /* DEBUG */ 387static void worklist_insert __P((struct workhead *, struct worklist *)); 388static void worklist_remove __P((struct worklist *)); 389static void workitem_free __P((struct worklist *, int)); 390 391#define WORKLIST_INSERT(head, item) worklist_insert(head, item) 392#define WORKLIST_REMOVE(item) worklist_remove(item) 393#define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type) 394 395static void 396worklist_insert(head, item) 397 struct workhead *head; 398 struct worklist *item; 399{ 400 401 if (lk.lkt_held == -1) 402 panic("worklist_insert: lock not held"); 403 if (item->wk_state & ONWORKLIST) 404 panic("worklist_insert: already on list"); 405 item->wk_state |= ONWORKLIST; 406 LIST_INSERT_HEAD(head, item, wk_list); 407} 408 409static void 410worklist_remove(item) 411 struct worklist *item; 412{ 413 414 if (lk.lkt_held == -1) 415 panic("worklist_remove: lock not held"); 416 if ((item->wk_state & ONWORKLIST) == 0) 417 panic("worklist_remove: not on list"); 418 item->wk_state &= ~ONWORKLIST; 419 LIST_REMOVE(item, wk_list); 420} 421 422static void 423workitem_free(item, type) 424 struct worklist *item; 425 int type; 426{ 427 428 if (item->wk_state & ONWORKLIST) 429 panic("workitem_free: still on list"); 430 if (item->wk_type != type) 431 panic("workitem_free: type mismatch"); 432 FREE(item, DtoM(type)); 433} 434#endif /* DEBUG */ 435 436/* 437 * Workitem queue management 438 */ 439static struct workhead softdep_workitem_pending; 440static int num_on_worklist; /* number of worklist items to be processed */ 441static int softdep_worklist_busy; /* 1 => trying to do unmount */ 442static int softdep_worklist_req; /* serialized waiters */ 443static int max_softdeps; /* maximum number of structs before slowdown */ 444static int tickdelay = 2; /* number of ticks to pause during slowdown */ 445static int proc_waiting; /* tracks whether we have a timeout posted */ 446static int *stat_countp; /* statistic to count in proc_waiting timeout */ 447static struct callout_handle handle; /* handle on posted proc_waiting timeout */ 448static struct proc *filesys_syncer; /* proc of filesystem syncer process */ 449static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 450#define FLUSH_INODES 1 451static int req_clear_remove; /* syncer process flush some freeblks */ 452#define FLUSH_REMOVE 2 453/* 454 * runtime statistics 455 */ 456static int stat_worklist_push; /* number of worklist cleanups */ 457static int stat_blk_limit_push; /* number of times block limit neared */ 458static int stat_ino_limit_push; /* number of times inode limit neared */ 459static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 460static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 461static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 462static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 463static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 464static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 465static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 466#ifdef DEBUG 467#include <vm/vm.h> 468#include <sys/sysctl.h> 469SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, ""); 470SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, ""); 471SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,""); 472SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,""); 473SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,""); 474SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, ""); 475SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, ""); 476SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, ""); 477SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, ""); 478SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, ""); 479SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, ""); 480SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, ""); 481#endif /* DEBUG */ 482 483/* 484 * Add an item to the end of the work queue. 485 * This routine requires that the lock be held. 486 * This is the only routine that adds items to the list. 487 * The following routine is the only one that removes items 488 * and does so in order from first to last. 489 */ 490static void 491add_to_worklist(wk) 492 struct worklist *wk; 493{ 494 static struct worklist *worklist_tail; 495 496 if (wk->wk_state & ONWORKLIST) 497 panic("add_to_worklist: already on list"); 498 wk->wk_state |= ONWORKLIST; 499 if (LIST_FIRST(&softdep_workitem_pending) == NULL) 500 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list); 501 else 502 LIST_INSERT_AFTER(worklist_tail, wk, wk_list); 503 worklist_tail = wk; 504 num_on_worklist += 1; 505} 506 507/* 508 * Process that runs once per second to handle items in the background queue. 509 * 510 * Note that we ensure that everything is done in the order in which they 511 * appear in the queue. The code below depends on this property to ensure 512 * that blocks of a file are freed before the inode itself is freed. This 513 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 514 * until all the old ones have been purged from the dependency lists. 515 */ 516int 517softdep_process_worklist(matchmnt) 518 struct mount *matchmnt; 519{ 520 struct proc *p = CURPROC; 521 int matchcnt, loopcount; 522 long starttime; 523 524 /* 525 * Record the process identifier of our caller so that we can give 526 * this process preferential treatment in request_cleanup below. 527 */ 528 filesys_syncer = p; 529 matchcnt = 0; 530 531 /* 532 * There is no danger of having multiple processes run this 533 * code, but we have to single-thread it when softdep_flushfiles() 534 * is in operation to get an accurate count of the number of items 535 * related to its mount point that are in the list. 536 */ 537 if (matchmnt == NULL) { 538 if (softdep_worklist_busy < 0) 539 return(-1); 540 softdep_worklist_busy += 1; 541 } 542 543 /* 544 * If requested, try removing inode or removal dependencies. 545 */ 546 if (req_clear_inodedeps) { 547 clear_inodedeps(p); 548 req_clear_inodedeps -= 1; 549 wakeup_one(&proc_waiting); 550 } 551 if (req_clear_remove) { 552 clear_remove(p); 553 req_clear_remove -= 1; 554 wakeup_one(&proc_waiting); 555 } 556 loopcount = 1; 557 starttime = time_second; 558 while (num_on_worklist > 0) { 559 matchcnt += process_worklist_item(matchmnt, 0); 560 561 /* 562 * If a umount operation wants to run the worklist 563 * accurately, abort. 564 */ 565 if (softdep_worklist_req && matchmnt == NULL) { 566 matchcnt = -1; 567 break; 568 } 569 570 /* 571 * If requested, try removing inode or removal dependencies. 572 */ 573 if (req_clear_inodedeps) { 574 clear_inodedeps(p); 575 req_clear_inodedeps -= 1; 576 wakeup_one(&proc_waiting); 577 } 578 if (req_clear_remove) { 579 clear_remove(p); 580 req_clear_remove -= 1; 581 wakeup_one(&proc_waiting); 582 } 583 /* 584 * We do not generally want to stop for buffer space, but if 585 * we are really being a buffer hog, we will stop and wait. 586 */ 587 if (loopcount++ % 128 == 0) 588 bwillwrite(); 589 /* 590 * Never allow processing to run for more than one 591 * second. Otherwise the other syncer tasks may get 592 * excessively backlogged. 593 */ 594 if (starttime != time_second && matchmnt == NULL) { 595 matchcnt = -1; 596 break; 597 } 598 } 599 if (matchmnt == NULL) { 600 softdep_worklist_busy -= 1; 601 if (softdep_worklist_req && softdep_worklist_busy == 0) 602 wakeup(&softdep_worklist_req); 603 } 604 return (matchcnt); 605} 606 607/* 608 * Process one item on the worklist. 609 */ 610static int 611process_worklist_item(matchmnt, flags) 612 struct mount *matchmnt; 613 int flags; 614{ 615 struct worklist *wk; 616 struct dirrem *dirrem; 617 struct mount *mp; 618 struct vnode *vp; 619 int matchcnt = 0; 620 621 ACQUIRE_LOCK(&lk); 622 /* 623 * Normally we just process each item on the worklist in order. 624 * However, if we are in a situation where we cannot lock any 625 * inodes, we have to skip over any dirrem requests whose 626 * vnodes are resident and locked. 627 */ 628 LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) { 629 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM) 630 break; 631 dirrem = WK_DIRREM(wk); 632 vp = ufs_ihashlookup(VFSTOUFS(dirrem->dm_mnt)->um_dev, 633 dirrem->dm_oldinum); 634 if (vp == NULL || !VOP_ISLOCKED(vp, CURPROC)) 635 break; 636 } 637 if (wk == 0) 638 return (0); 639 WORKLIST_REMOVE(wk); 640 num_on_worklist -= 1; 641 FREE_LOCK(&lk); 642 switch (wk->wk_type) { 643 644 case D_DIRREM: 645 /* removal of a directory entry */ 646 mp = WK_DIRREM(wk)->dm_mnt; 647 if (vn_write_suspend_wait(NULL, mp, V_NOWAIT)) 648 panic("%s: dirrem on suspended filesystem", 649 "process_worklist_item"); 650 if (mp == matchmnt) 651 matchcnt += 1; 652 handle_workitem_remove(WK_DIRREM(wk)); 653 break; 654 655 case D_FREEBLKS: 656 /* releasing blocks and/or fragments from a file */ 657 mp = WK_FREEBLKS(wk)->fb_mnt; 658 if (vn_write_suspend_wait(NULL, mp, V_NOWAIT)) 659 panic("%s: freeblks on suspended filesystem", 660 "process_worklist_item"); 661 if (mp == matchmnt) 662 matchcnt += 1; 663 handle_workitem_freeblocks(WK_FREEBLKS(wk)); 664 break; 665 666 case D_FREEFRAG: 667 /* releasing a fragment when replaced as a file grows */ 668 mp = WK_FREEFRAG(wk)->ff_mnt; 669 if (vn_write_suspend_wait(NULL, mp, V_NOWAIT)) 670 panic("%s: freefrag on suspended filesystem", 671 "process_worklist_item"); 672 if (mp == matchmnt) 673 matchcnt += 1; 674 handle_workitem_freefrag(WK_FREEFRAG(wk)); 675 break; 676 677 case D_FREEFILE: 678 /* releasing an inode when its link count drops to 0 */ 679 mp = WK_FREEFILE(wk)->fx_mnt; 680 if (vn_write_suspend_wait(NULL, mp, V_NOWAIT)) 681 panic("%s: freefile on suspended filesystem", 682 "process_worklist_item"); 683 if (mp == matchmnt) 684 matchcnt += 1; 685 handle_workitem_freefile(WK_FREEFILE(wk)); 686 break; 687 688 default: 689 panic("%s_process_worklist: Unknown type %s", 690 "softdep", TYPENAME(wk->wk_type)); 691 /* NOTREACHED */ 692 } 693 return (matchcnt); 694} 695 696/* 697 * Move dependencies from one buffer to another. 698 */ 699static void 700softdep_move_dependencies(oldbp, newbp) 701 struct buf *oldbp; 702 struct buf *newbp; 703{ 704 struct worklist *wk, *wktail; 705 706 if (LIST_FIRST(&newbp->b_dep) != NULL) 707 panic("softdep_move_dependencies: need merge code"); 708 wktail = 0; 709 ACQUIRE_LOCK(&lk); 710 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 711 LIST_REMOVE(wk, wk_list); 712 if (wktail == 0) 713 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 714 else 715 LIST_INSERT_AFTER(wktail, wk, wk_list); 716 wktail = wk; 717 } 718 FREE_LOCK(&lk); 719} 720 721/* 722 * Purge the work list of all items associated with a particular mount point. 723 */ 724int 725softdep_flushworklist(oldmnt, countp, p) 726 struct mount *oldmnt; 727 int *countp; 728 struct proc *p; 729{ 730 struct vnode *devvp; 731 int count, error = 0; 732 733 /* 734 * Await our turn to clear out the queue, then serialize access. 735 */ 736 while (softdep_worklist_busy) { 737 softdep_worklist_req += 1; 738 tsleep(&softdep_worklist_req, PRIBIO, "softflush", 0); 739 softdep_worklist_req -= 1; 740 } 741 softdep_worklist_busy = -1; 742 /* 743 * Alternately flush the block device associated with the mount 744 * point and process any dependencies that the flushing 745 * creates. We continue until no more worklist dependencies 746 * are found. 747 */ 748 *countp = 0; 749 devvp = VFSTOUFS(oldmnt)->um_devvp; 750 while ((count = softdep_process_worklist(oldmnt)) > 0) { 751 *countp += count; 752 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, p); 753 error = VOP_FSYNC(devvp, p->p_ucred, MNT_WAIT, p); 754 VOP_UNLOCK(devvp, 0, p); 755 if (error) 756 break; 757 } 758 softdep_worklist_busy = 0; 759 if (softdep_worklist_req) 760 wakeup(&softdep_worklist_req); 761 return (error); 762} 763 764/* 765 * Flush all vnodes and worklist items associated with a specified mount point. 766 */ 767int 768softdep_flushfiles(oldmnt, flags, p) 769 struct mount *oldmnt; 770 int flags; 771 struct proc *p; 772{ 773 int error, count, loopcnt; 774 775 /* 776 * Alternately flush the vnodes associated with the mount 777 * point and process any dependencies that the flushing 778 * creates. In theory, this loop can happen at most twice, 779 * but we give it a few extra just to be sure. 780 */ 781 for (loopcnt = 10; loopcnt > 0; loopcnt--) { 782 /* 783 * Do another flush in case any vnodes were brought in 784 * as part of the cleanup operations. 785 */ 786 if ((error = ffs_flushfiles(oldmnt, flags, p)) != 0) 787 break; 788 if ((error = softdep_flushworklist(oldmnt, &count, p)) != 0 || 789 count == 0) 790 break; 791 } 792 /* 793 * If we are unmounting then it is an error to fail. If we 794 * are simply trying to downgrade to read-only, then filesystem 795 * activity can keep us busy forever, so we just fail with EBUSY. 796 */ 797 if (loopcnt == 0) { 798 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 799 panic("softdep_flushfiles: looping"); 800 error = EBUSY; 801 } 802 return (error); 803} 804 805/* 806 * Structure hashing. 807 * 808 * There are three types of structures that can be looked up: 809 * 1) pagedep structures identified by mount point, inode number, 810 * and logical block. 811 * 2) inodedep structures identified by mount point and inode number. 812 * 3) newblk structures identified by mount point and 813 * physical block number. 814 * 815 * The "pagedep" and "inodedep" dependency structures are hashed 816 * separately from the file blocks and inodes to which they correspond. 817 * This separation helps when the in-memory copy of an inode or 818 * file block must be replaced. It also obviates the need to access 819 * an inode or file page when simply updating (or de-allocating) 820 * dependency structures. Lookup of newblk structures is needed to 821 * find newly allocated blocks when trying to associate them with 822 * their allocdirect or allocindir structure. 823 * 824 * The lookup routines optionally create and hash a new instance when 825 * an existing entry is not found. 826 */ 827#define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 828 829/* 830 * Structures and routines associated with pagedep caching. 831 */ 832LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl; 833u_long pagedep_hash; /* size of hash table - 1 */ 834#define PAGEDEP_HASH(mp, inum, lbn) \ 835 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \ 836 pagedep_hash]) 837static struct sema pagedep_in_progress; 838 839/* 840 * Look up a pagedep. Return 1 if found, 0 if not found. 841 * If not found, allocate if DEPALLOC flag is passed. 842 * Found or allocated entry is returned in pagedeppp. 843 * This routine must be called with splbio interrupts blocked. 844 */ 845static int 846pagedep_lookup(ip, lbn, flags, pagedeppp) 847 struct inode *ip; 848 ufs_lbn_t lbn; 849 int flags; 850 struct pagedep **pagedeppp; 851{ 852 struct pagedep *pagedep; 853 struct pagedep_hashhead *pagedephd; 854 struct mount *mp; 855 int i; 856 857#ifdef DEBUG 858 if (lk.lkt_held == -1) 859 panic("pagedep_lookup: lock not held"); 860#endif 861 mp = ITOV(ip)->v_mount; 862 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn); 863top: 864 for (pagedep = LIST_FIRST(pagedephd); pagedep; 865 pagedep = LIST_NEXT(pagedep, pd_hash)) 866 if (ip->i_number == pagedep->pd_ino && 867 lbn == pagedep->pd_lbn && 868 mp == pagedep->pd_mnt) 869 break; 870 if (pagedep) { 871 *pagedeppp = pagedep; 872 return (1); 873 } 874 if ((flags & DEPALLOC) == 0) { 875 *pagedeppp = NULL; 876 return (0); 877 } 878 if (sema_get(&pagedep_in_progress, &lk) == 0) { 879 ACQUIRE_LOCK(&lk); 880 goto top; 881 } 882 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP, 883 M_SOFTDEP_FLAGS|M_ZERO); 884 pagedep->pd_list.wk_type = D_PAGEDEP; 885 pagedep->pd_mnt = mp; 886 pagedep->pd_ino = ip->i_number; 887 pagedep->pd_lbn = lbn; 888 LIST_INIT(&pagedep->pd_dirremhd); 889 LIST_INIT(&pagedep->pd_pendinghd); 890 for (i = 0; i < DAHASHSZ; i++) 891 LIST_INIT(&pagedep->pd_diraddhd[i]); 892 ACQUIRE_LOCK(&lk); 893 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 894 sema_release(&pagedep_in_progress); 895 *pagedeppp = pagedep; 896 return (0); 897} 898 899/* 900 * Structures and routines associated with inodedep caching. 901 */ 902LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl; 903static u_long inodedep_hash; /* size of hash table - 1 */ 904static long num_inodedep; /* number of inodedep allocated */ 905#define INODEDEP_HASH(fs, inum) \ 906 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash]) 907static struct sema inodedep_in_progress; 908 909/* 910 * Look up a inodedep. Return 1 if found, 0 if not found. 911 * If not found, allocate if DEPALLOC flag is passed. 912 * Found or allocated entry is returned in inodedeppp. 913 * This routine must be called with splbio interrupts blocked. 914 */ 915static int 916inodedep_lookup(fs, inum, flags, inodedeppp) 917 struct fs *fs; 918 ino_t inum; 919 int flags; 920 struct inodedep **inodedeppp; 921{ 922 struct inodedep *inodedep; 923 struct inodedep_hashhead *inodedephd; 924 int firsttry; 925 926#ifdef DEBUG 927 if (lk.lkt_held == -1) 928 panic("inodedep_lookup: lock not held"); 929#endif 930 firsttry = 1; 931 inodedephd = INODEDEP_HASH(fs, inum); 932top: 933 for (inodedep = LIST_FIRST(inodedephd); inodedep; 934 inodedep = LIST_NEXT(inodedep, id_hash)) 935 if (inum == inodedep->id_ino && fs == inodedep->id_fs) 936 break; 937 if (inodedep) { 938 *inodedeppp = inodedep; 939 return (1); 940 } 941 if ((flags & DEPALLOC) == 0) { 942 *inodedeppp = NULL; 943 return (0); 944 } 945 /* 946 * If we are over our limit, try to improve the situation. 947 */ 948 if (num_inodedep > max_softdeps && firsttry && 949 request_cleanup(FLUSH_INODES, 1)) { 950 firsttry = 0; 951 goto top; 952 } 953 if (sema_get(&inodedep_in_progress, &lk) == 0) { 954 ACQUIRE_LOCK(&lk); 955 goto top; 956 } 957 num_inodedep += 1; 958 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep), 959 M_INODEDEP, M_SOFTDEP_FLAGS); 960 inodedep->id_list.wk_type = D_INODEDEP; 961 inodedep->id_fs = fs; 962 inodedep->id_ino = inum; 963 inodedep->id_state = ALLCOMPLETE; 964 inodedep->id_nlinkdelta = 0; 965 inodedep->id_savedino = NULL; 966 inodedep->id_savedsize = -1; 967 inodedep->id_buf = NULL; 968 LIST_INIT(&inodedep->id_pendinghd); 969 LIST_INIT(&inodedep->id_inowait); 970 LIST_INIT(&inodedep->id_bufwait); 971 TAILQ_INIT(&inodedep->id_inoupdt); 972 TAILQ_INIT(&inodedep->id_newinoupdt); 973 ACQUIRE_LOCK(&lk); 974 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 975 sema_release(&inodedep_in_progress); 976 *inodedeppp = inodedep; 977 return (0); 978} 979 980/* 981 * Structures and routines associated with newblk caching. 982 */ 983LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl; 984u_long newblk_hash; /* size of hash table - 1 */ 985#define NEWBLK_HASH(fs, inum) \ 986 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash]) 987static struct sema newblk_in_progress; 988 989/* 990 * Look up a newblk. Return 1 if found, 0 if not found. 991 * If not found, allocate if DEPALLOC flag is passed. 992 * Found or allocated entry is returned in newblkpp. 993 */ 994static int 995newblk_lookup(fs, newblkno, flags, newblkpp) 996 struct fs *fs; 997 ufs_daddr_t newblkno; 998 int flags; 999 struct newblk **newblkpp; 1000{ 1001 struct newblk *newblk; 1002 struct newblk_hashhead *newblkhd; 1003 1004 newblkhd = NEWBLK_HASH(fs, newblkno); 1005top: 1006 for (newblk = LIST_FIRST(newblkhd); newblk; 1007 newblk = LIST_NEXT(newblk, nb_hash)) 1008 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs) 1009 break; 1010 if (newblk) { 1011 *newblkpp = newblk; 1012 return (1); 1013 } 1014 if ((flags & DEPALLOC) == 0) { 1015 *newblkpp = NULL; 1016 return (0); 1017 } 1018 if (sema_get(&newblk_in_progress, 0) == 0) 1019 goto top; 1020 MALLOC(newblk, struct newblk *, sizeof(struct newblk), 1021 M_NEWBLK, M_SOFTDEP_FLAGS); 1022 newblk->nb_state = 0; 1023 newblk->nb_fs = fs; 1024 newblk->nb_newblkno = newblkno; 1025 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 1026 sema_release(&newblk_in_progress); 1027 *newblkpp = newblk; 1028 return (0); 1029} 1030 1031/* 1032 * Executed during filesystem system initialization before 1033 * mounting any file systems. 1034 */ 1035void 1036softdep_initialize() 1037{ 1038 1039 LIST_INIT(&mkdirlisthd); 1040 LIST_INIT(&softdep_workitem_pending); 1041 max_softdeps = min(desiredvnodes * 8, 1042 M_INODEDEP->ks_limit / (2 * sizeof(struct inodedep))); 1043 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 1044 &pagedep_hash); 1045 sema_init(&pagedep_in_progress, "pagedep", PRIBIO, 0); 1046 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash); 1047 sema_init(&inodedep_in_progress, "inodedep", PRIBIO, 0); 1048 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash); 1049 sema_init(&newblk_in_progress, "newblk", PRIBIO, 0); 1050} 1051 1052/* 1053 * Called at mount time to notify the dependency code that a 1054 * filesystem wishes to use it. 1055 */ 1056int 1057softdep_mount(devvp, mp, fs, cred) 1058 struct vnode *devvp; 1059 struct mount *mp; 1060 struct fs *fs; 1061 struct ucred *cred; 1062{ 1063 struct csum cstotal; 1064 struct cg *cgp; 1065 struct buf *bp; 1066 int error, cyl; 1067 1068 mp->mnt_flag &= ~MNT_ASYNC; 1069 mp->mnt_flag |= MNT_SOFTDEP; 1070 /* 1071 * When doing soft updates, the counters in the 1072 * superblock may have gotten out of sync, so we have 1073 * to scan the cylinder groups and recalculate them. 1074 */ 1075 if (fs->fs_clean != 0) 1076 return (0); 1077 bzero(&cstotal, sizeof cstotal); 1078 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 1079 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 1080 fs->fs_cgsize, cred, &bp)) != 0) { 1081 brelse(bp); 1082 return (error); 1083 } 1084 cgp = (struct cg *)bp->b_data; 1085 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 1086 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 1087 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 1088 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 1089 fs->fs_cs(fs, cyl) = cgp->cg_cs; 1090 brelse(bp); 1091 } 1092#ifdef DEBUG 1093 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 1094 printf("ffs_mountfs: superblock updated for soft updates\n"); 1095#endif 1096 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 1097 return (0); 1098} 1099 1100/* 1101 * Protecting the freemaps (or bitmaps). 1102 * 1103 * To eliminate the need to execute fsck before mounting a file system 1104 * after a power failure, one must (conservatively) guarantee that the 1105 * on-disk copy of the bitmaps never indicate that a live inode or block is 1106 * free. So, when a block or inode is allocated, the bitmap should be 1107 * updated (on disk) before any new pointers. When a block or inode is 1108 * freed, the bitmap should not be updated until all pointers have been 1109 * reset. The latter dependency is handled by the delayed de-allocation 1110 * approach described below for block and inode de-allocation. The former 1111 * dependency is handled by calling the following procedure when a block or 1112 * inode is allocated. When an inode is allocated an "inodedep" is created 1113 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 1114 * Each "inodedep" is also inserted into the hash indexing structure so 1115 * that any additional link additions can be made dependent on the inode 1116 * allocation. 1117 * 1118 * The ufs file system maintains a number of free block counts (e.g., per 1119 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 1120 * in addition to the bitmaps. These counts are used to improve efficiency 1121 * during allocation and therefore must be consistent with the bitmaps. 1122 * There is no convenient way to guarantee post-crash consistency of these 1123 * counts with simple update ordering, for two main reasons: (1) The counts 1124 * and bitmaps for a single cylinder group block are not in the same disk 1125 * sector. If a disk write is interrupted (e.g., by power failure), one may 1126 * be written and the other not. (2) Some of the counts are located in the 1127 * superblock rather than the cylinder group block. So, we focus our soft 1128 * updates implementation on protecting the bitmaps. When mounting a 1129 * filesystem, we recompute the auxiliary counts from the bitmaps. 1130 */ 1131 1132/* 1133 * Called just after updating the cylinder group block to allocate an inode. 1134 */ 1135void 1136softdep_setup_inomapdep(bp, ip, newinum) 1137 struct buf *bp; /* buffer for cylgroup block with inode map */ 1138 struct inode *ip; /* inode related to allocation */ 1139 ino_t newinum; /* new inode number being allocated */ 1140{ 1141 struct inodedep *inodedep; 1142 struct bmsafemap *bmsafemap; 1143 1144 /* 1145 * Create a dependency for the newly allocated inode. 1146 * Panic if it already exists as something is seriously wrong. 1147 * Otherwise add it to the dependency list for the buffer holding 1148 * the cylinder group map from which it was allocated. 1149 */ 1150 ACQUIRE_LOCK(&lk); 1151 if (inodedep_lookup(ip->i_fs, newinum, DEPALLOC, &inodedep) != 0) 1152 panic("softdep_setup_inomapdep: found inode"); 1153 inodedep->id_buf = bp; 1154 inodedep->id_state &= ~DEPCOMPLETE; 1155 bmsafemap = bmsafemap_lookup(bp); 1156 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 1157 FREE_LOCK(&lk); 1158} 1159 1160/* 1161 * Called just after updating the cylinder group block to 1162 * allocate block or fragment. 1163 */ 1164void 1165softdep_setup_blkmapdep(bp, fs, newblkno) 1166 struct buf *bp; /* buffer for cylgroup block with block map */ 1167 struct fs *fs; /* filesystem doing allocation */ 1168 ufs_daddr_t newblkno; /* number of newly allocated block */ 1169{ 1170 struct newblk *newblk; 1171 struct bmsafemap *bmsafemap; 1172 1173 /* 1174 * Create a dependency for the newly allocated block. 1175 * Add it to the dependency list for the buffer holding 1176 * the cylinder group map from which it was allocated. 1177 */ 1178 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0) 1179 panic("softdep_setup_blkmapdep: found block"); 1180 ACQUIRE_LOCK(&lk); 1181 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp); 1182 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 1183 FREE_LOCK(&lk); 1184} 1185 1186/* 1187 * Find the bmsafemap associated with a cylinder group buffer. 1188 * If none exists, create one. The buffer must be locked when 1189 * this routine is called and this routine must be called with 1190 * splbio interrupts blocked. 1191 */ 1192static struct bmsafemap * 1193bmsafemap_lookup(bp) 1194 struct buf *bp; 1195{ 1196 struct bmsafemap *bmsafemap; 1197 struct worklist *wk; 1198 1199#ifdef DEBUG 1200 if (lk.lkt_held == -1) 1201 panic("bmsafemap_lookup: lock not held"); 1202#endif 1203 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = LIST_NEXT(wk, wk_list)) 1204 if (wk->wk_type == D_BMSAFEMAP) 1205 return (WK_BMSAFEMAP(wk)); 1206 FREE_LOCK(&lk); 1207 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap), 1208 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 1209 bmsafemap->sm_list.wk_type = D_BMSAFEMAP; 1210 bmsafemap->sm_list.wk_state = 0; 1211 bmsafemap->sm_buf = bp; 1212 LIST_INIT(&bmsafemap->sm_allocdirecthd); 1213 LIST_INIT(&bmsafemap->sm_allocindirhd); 1214 LIST_INIT(&bmsafemap->sm_inodedephd); 1215 LIST_INIT(&bmsafemap->sm_newblkhd); 1216 ACQUIRE_LOCK(&lk); 1217 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 1218 return (bmsafemap); 1219} 1220 1221/* 1222 * Direct block allocation dependencies. 1223 * 1224 * When a new block is allocated, the corresponding disk locations must be 1225 * initialized (with zeros or new data) before the on-disk inode points to 1226 * them. Also, the freemap from which the block was allocated must be 1227 * updated (on disk) before the inode's pointer. These two dependencies are 1228 * independent of each other and are needed for all file blocks and indirect 1229 * blocks that are pointed to directly by the inode. Just before the 1230 * "in-core" version of the inode is updated with a newly allocated block 1231 * number, a procedure (below) is called to setup allocation dependency 1232 * structures. These structures are removed when the corresponding 1233 * dependencies are satisfied or when the block allocation becomes obsolete 1234 * (i.e., the file is deleted, the block is de-allocated, or the block is a 1235 * fragment that gets upgraded). All of these cases are handled in 1236 * procedures described later. 1237 * 1238 * When a file extension causes a fragment to be upgraded, either to a larger 1239 * fragment or to a full block, the on-disk location may change (if the 1240 * previous fragment could not simply be extended). In this case, the old 1241 * fragment must be de-allocated, but not until after the inode's pointer has 1242 * been updated. In most cases, this is handled by later procedures, which 1243 * will construct a "freefrag" structure to be added to the workitem queue 1244 * when the inode update is complete (or obsolete). The main exception to 1245 * this is when an allocation occurs while a pending allocation dependency 1246 * (for the same block pointer) remains. This case is handled in the main 1247 * allocation dependency setup procedure by immediately freeing the 1248 * unreferenced fragments. 1249 */ 1250void 1251softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 1252 struct inode *ip; /* inode to which block is being added */ 1253 ufs_lbn_t lbn; /* block pointer within inode */ 1254 ufs_daddr_t newblkno; /* disk block number being added */ 1255 ufs_daddr_t oldblkno; /* previous block number, 0 unless frag */ 1256 long newsize; /* size of new block */ 1257 long oldsize; /* size of new block */ 1258 struct buf *bp; /* bp for allocated block */ 1259{ 1260 struct allocdirect *adp, *oldadp; 1261 struct allocdirectlst *adphead; 1262 struct bmsafemap *bmsafemap; 1263 struct inodedep *inodedep; 1264 struct pagedep *pagedep; 1265 struct newblk *newblk; 1266 1267 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect), 1268 M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO); 1269 adp->ad_list.wk_type = D_ALLOCDIRECT; 1270 adp->ad_lbn = lbn; 1271 adp->ad_newblkno = newblkno; 1272 adp->ad_oldblkno = oldblkno; 1273 adp->ad_newsize = newsize; 1274 adp->ad_oldsize = oldsize; 1275 adp->ad_state = ATTACHED; 1276 if (newblkno == oldblkno) 1277 adp->ad_freefrag = NULL; 1278 else 1279 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1280 1281 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1282 panic("softdep_setup_allocdirect: lost block"); 1283 1284 ACQUIRE_LOCK(&lk); 1285 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 1286 adp->ad_inodedep = inodedep; 1287 1288 if (newblk->nb_state == DEPCOMPLETE) { 1289 adp->ad_state |= DEPCOMPLETE; 1290 adp->ad_buf = NULL; 1291 } else { 1292 bmsafemap = newblk->nb_bmsafemap; 1293 adp->ad_buf = bmsafemap->sm_buf; 1294 LIST_REMOVE(newblk, nb_deps); 1295 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1296 } 1297 LIST_REMOVE(newblk, nb_hash); 1298 FREE(newblk, M_NEWBLK); 1299 1300 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list); 1301 if (lbn >= NDADDR) { 1302 /* allocating an indirect block */ 1303 if (oldblkno != 0) 1304 panic("softdep_setup_allocdirect: non-zero indir"); 1305 } else { 1306 /* 1307 * Allocating a direct block. 1308 * 1309 * If we are allocating a directory block, then we must 1310 * allocate an associated pagedep to track additions and 1311 * deletions. 1312 */ 1313 if ((ip->i_mode & IFMT) == IFDIR && 1314 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1315 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 1316 } 1317 /* 1318 * The list of allocdirects must be kept in sorted and ascending 1319 * order so that the rollback routines can quickly determine the 1320 * first uncommitted block (the size of the file stored on disk 1321 * ends at the end of the lowest committed fragment, or if there 1322 * are no fragments, at the end of the highest committed block). 1323 * Since files generally grow, the typical case is that the new 1324 * block is to be added at the end of the list. We speed this 1325 * special case by checking against the last allocdirect in the 1326 * list before laboriously traversing the list looking for the 1327 * insertion point. 1328 */ 1329 adphead = &inodedep->id_newinoupdt; 1330 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1331 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1332 /* insert at end of list */ 1333 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1334 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1335 allocdirect_merge(adphead, adp, oldadp); 1336 FREE_LOCK(&lk); 1337 return; 1338 } 1339 for (oldadp = TAILQ_FIRST(adphead); oldadp; 1340 oldadp = TAILQ_NEXT(oldadp, ad_next)) { 1341 if (oldadp->ad_lbn >= lbn) 1342 break; 1343 } 1344 if (oldadp == NULL) 1345 panic("softdep_setup_allocdirect: lost entry"); 1346 /* insert in middle of list */ 1347 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1348 if (oldadp->ad_lbn == lbn) 1349 allocdirect_merge(adphead, adp, oldadp); 1350 FREE_LOCK(&lk); 1351} 1352 1353/* 1354 * Replace an old allocdirect dependency with a newer one. 1355 * This routine must be called with splbio interrupts blocked. 1356 */ 1357static void 1358allocdirect_merge(adphead, newadp, oldadp) 1359 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 1360 struct allocdirect *newadp; /* allocdirect being added */ 1361 struct allocdirect *oldadp; /* existing allocdirect being checked */ 1362{ 1363 struct freefrag *freefrag; 1364 1365#ifdef DEBUG 1366 if (lk.lkt_held == -1) 1367 panic("allocdirect_merge: lock not held"); 1368#endif 1369 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 1370 newadp->ad_oldsize != oldadp->ad_newsize || 1371 newadp->ad_lbn >= NDADDR) 1372 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d", 1373 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn, 1374 NDADDR); 1375 newadp->ad_oldblkno = oldadp->ad_oldblkno; 1376 newadp->ad_oldsize = oldadp->ad_oldsize; 1377 /* 1378 * If the old dependency had a fragment to free or had never 1379 * previously had a block allocated, then the new dependency 1380 * can immediately post its freefrag and adopt the old freefrag. 1381 * This action is done by swapping the freefrag dependencies. 1382 * The new dependency gains the old one's freefrag, and the 1383 * old one gets the new one and then immediately puts it on 1384 * the worklist when it is freed by free_allocdirect. It is 1385 * not possible to do this swap when the old dependency had a 1386 * non-zero size but no previous fragment to free. This condition 1387 * arises when the new block is an extension of the old block. 1388 * Here, the first part of the fragment allocated to the new 1389 * dependency is part of the block currently claimed on disk by 1390 * the old dependency, so cannot legitimately be freed until the 1391 * conditions for the new dependency are fulfilled. 1392 */ 1393 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 1394 freefrag = newadp->ad_freefrag; 1395 newadp->ad_freefrag = oldadp->ad_freefrag; 1396 oldadp->ad_freefrag = freefrag; 1397 } 1398 free_allocdirect(adphead, oldadp, 0); 1399} 1400 1401/* 1402 * Allocate a new freefrag structure if needed. 1403 */ 1404static struct freefrag * 1405newfreefrag(ip, blkno, size) 1406 struct inode *ip; 1407 ufs_daddr_t blkno; 1408 long size; 1409{ 1410 struct freefrag *freefrag; 1411 struct fs *fs; 1412 1413 if (blkno == 0) 1414 return (NULL); 1415 fs = ip->i_fs; 1416 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 1417 panic("newfreefrag: frag size"); 1418 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag), 1419 M_FREEFRAG, M_SOFTDEP_FLAGS); 1420 freefrag->ff_list.wk_type = D_FREEFRAG; 1421 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */ 1422 freefrag->ff_inum = ip->i_number; 1423 freefrag->ff_mnt = ITOV(ip)->v_mount; 1424 freefrag->ff_devvp = ip->i_devvp; 1425 freefrag->ff_blkno = blkno; 1426 freefrag->ff_fragsize = size; 1427 return (freefrag); 1428} 1429 1430/* 1431 * This workitem de-allocates fragments that were replaced during 1432 * file block allocation. 1433 */ 1434static void 1435handle_workitem_freefrag(freefrag) 1436 struct freefrag *freefrag; 1437{ 1438 struct inode tip; 1439 1440 tip.i_vnode = NULL; 1441 tip.i_fs = VFSTOUFS(freefrag->ff_mnt)->um_fs; 1442 tip.i_devvp = freefrag->ff_devvp; 1443 tip.i_dev = freefrag->ff_devvp->v_rdev; 1444 tip.i_number = freefrag->ff_inum; 1445 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */ 1446 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize); 1447 FREE(freefrag, M_FREEFRAG); 1448} 1449 1450/* 1451 * Indirect block allocation dependencies. 1452 * 1453 * The same dependencies that exist for a direct block also exist when 1454 * a new block is allocated and pointed to by an entry in a block of 1455 * indirect pointers. The undo/redo states described above are also 1456 * used here. Because an indirect block contains many pointers that 1457 * may have dependencies, a second copy of the entire in-memory indirect 1458 * block is kept. The buffer cache copy is always completely up-to-date. 1459 * The second copy, which is used only as a source for disk writes, 1460 * contains only the safe pointers (i.e., those that have no remaining 1461 * update dependencies). The second copy is freed when all pointers 1462 * are safe. The cache is not allowed to replace indirect blocks with 1463 * pending update dependencies. If a buffer containing an indirect 1464 * block with dependencies is written, these routines will mark it 1465 * dirty again. It can only be successfully written once all the 1466 * dependencies are removed. The ffs_fsync routine in conjunction with 1467 * softdep_sync_metadata work together to get all the dependencies 1468 * removed so that a file can be successfully written to disk. Three 1469 * procedures are used when setting up indirect block pointer 1470 * dependencies. The division is necessary because of the organization 1471 * of the "balloc" routine and because of the distinction between file 1472 * pages and file metadata blocks. 1473 */ 1474 1475/* 1476 * Allocate a new allocindir structure. 1477 */ 1478static struct allocindir * 1479newallocindir(ip, ptrno, newblkno, oldblkno) 1480 struct inode *ip; /* inode for file being extended */ 1481 int ptrno; /* offset of pointer in indirect block */ 1482 ufs_daddr_t newblkno; /* disk block number being added */ 1483 ufs_daddr_t oldblkno; /* previous block number, 0 if none */ 1484{ 1485 struct allocindir *aip; 1486 1487 MALLOC(aip, struct allocindir *, sizeof(struct allocindir), 1488 M_ALLOCINDIR, M_SOFTDEP_FLAGS|M_ZERO); 1489 aip->ai_list.wk_type = D_ALLOCINDIR; 1490 aip->ai_state = ATTACHED; 1491 aip->ai_offset = ptrno; 1492 aip->ai_newblkno = newblkno; 1493 aip->ai_oldblkno = oldblkno; 1494 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize); 1495 return (aip); 1496} 1497 1498/* 1499 * Called just before setting an indirect block pointer 1500 * to a newly allocated file page. 1501 */ 1502void 1503softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 1504 struct inode *ip; /* inode for file being extended */ 1505 ufs_lbn_t lbn; /* allocated block number within file */ 1506 struct buf *bp; /* buffer with indirect blk referencing page */ 1507 int ptrno; /* offset of pointer in indirect block */ 1508 ufs_daddr_t newblkno; /* disk block number being added */ 1509 ufs_daddr_t oldblkno; /* previous block number, 0 if none */ 1510 struct buf *nbp; /* buffer holding allocated page */ 1511{ 1512 struct allocindir *aip; 1513 struct pagedep *pagedep; 1514 1515 aip = newallocindir(ip, ptrno, newblkno, oldblkno); 1516 ACQUIRE_LOCK(&lk); 1517 /* 1518 * If we are allocating a directory page, then we must 1519 * allocate an associated pagedep to track additions and 1520 * deletions. 1521 */ 1522 if ((ip->i_mode & IFMT) == IFDIR && 1523 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1524 WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list); 1525 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1526 FREE_LOCK(&lk); 1527 setup_allocindir_phase2(bp, ip, aip); 1528} 1529 1530/* 1531 * Called just before setting an indirect block pointer to a 1532 * newly allocated indirect block. 1533 */ 1534void 1535softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 1536 struct buf *nbp; /* newly allocated indirect block */ 1537 struct inode *ip; /* inode for file being extended */ 1538 struct buf *bp; /* indirect block referencing allocated block */ 1539 int ptrno; /* offset of pointer in indirect block */ 1540 ufs_daddr_t newblkno; /* disk block number being added */ 1541{ 1542 struct allocindir *aip; 1543 1544 aip = newallocindir(ip, ptrno, newblkno, 0); 1545 ACQUIRE_LOCK(&lk); 1546 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1547 FREE_LOCK(&lk); 1548 setup_allocindir_phase2(bp, ip, aip); 1549} 1550 1551/* 1552 * Called to finish the allocation of the "aip" allocated 1553 * by one of the two routines above. 1554 */ 1555static void 1556setup_allocindir_phase2(bp, ip, aip) 1557 struct buf *bp; /* in-memory copy of the indirect block */ 1558 struct inode *ip; /* inode for file being extended */ 1559 struct allocindir *aip; /* allocindir allocated by the above routines */ 1560{ 1561 struct worklist *wk; 1562 struct indirdep *indirdep, *newindirdep; 1563 struct bmsafemap *bmsafemap; 1564 struct allocindir *oldaip; 1565 struct freefrag *freefrag; 1566 struct newblk *newblk; 1567 1568 if (bp->b_lblkno >= 0) 1569 panic("setup_allocindir_phase2: not indir blk"); 1570 for (indirdep = NULL, newindirdep = NULL; ; ) { 1571 ACQUIRE_LOCK(&lk); 1572 for (wk = LIST_FIRST(&bp->b_dep); wk; 1573 wk = LIST_NEXT(wk, wk_list)) { 1574 if (wk->wk_type != D_INDIRDEP) 1575 continue; 1576 indirdep = WK_INDIRDEP(wk); 1577 break; 1578 } 1579 if (indirdep == NULL && newindirdep) { 1580 indirdep = newindirdep; 1581 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 1582 newindirdep = NULL; 1583 } 1584 FREE_LOCK(&lk); 1585 if (indirdep) { 1586 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0, 1587 &newblk) == 0) 1588 panic("setup_allocindir: lost block"); 1589 ACQUIRE_LOCK(&lk); 1590 if (newblk->nb_state == DEPCOMPLETE) { 1591 aip->ai_state |= DEPCOMPLETE; 1592 aip->ai_buf = NULL; 1593 } else { 1594 bmsafemap = newblk->nb_bmsafemap; 1595 aip->ai_buf = bmsafemap->sm_buf; 1596 LIST_REMOVE(newblk, nb_deps); 1597 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd, 1598 aip, ai_deps); 1599 } 1600 LIST_REMOVE(newblk, nb_hash); 1601 FREE(newblk, M_NEWBLK); 1602 aip->ai_indirdep = indirdep; 1603 /* 1604 * Check to see if there is an existing dependency 1605 * for this block. If there is, merge the old 1606 * dependency into the new one. 1607 */ 1608 if (aip->ai_oldblkno == 0) 1609 oldaip = NULL; 1610 else 1611 for (oldaip=LIST_FIRST(&indirdep->ir_deplisthd); 1612 oldaip; oldaip = LIST_NEXT(oldaip, ai_next)) 1613 if (oldaip->ai_offset == aip->ai_offset) 1614 break; 1615 freefrag = NULL; 1616 if (oldaip != NULL) { 1617 if (oldaip->ai_newblkno != aip->ai_oldblkno) 1618 panic("setup_allocindir_phase2: blkno"); 1619 aip->ai_oldblkno = oldaip->ai_oldblkno; 1620 freefrag = aip->ai_freefrag; 1621 aip->ai_freefrag = oldaip->ai_freefrag; 1622 oldaip->ai_freefrag = NULL; 1623 free_allocindir(oldaip, NULL); 1624 } 1625 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 1626 ((ufs_daddr_t *)indirdep->ir_savebp->b_data) 1627 [aip->ai_offset] = aip->ai_oldblkno; 1628 FREE_LOCK(&lk); 1629 if (freefrag != NULL) 1630 handle_workitem_freefrag(freefrag); 1631 } 1632 if (newindirdep) { 1633 if (indirdep->ir_savebp != NULL) 1634 brelse(newindirdep->ir_savebp); 1635 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP); 1636 } 1637 if (indirdep) 1638 break; 1639 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep), 1640 M_INDIRDEP, M_SOFTDEP_FLAGS); 1641 newindirdep->ir_list.wk_type = D_INDIRDEP; 1642 newindirdep->ir_state = ATTACHED; 1643 LIST_INIT(&newindirdep->ir_deplisthd); 1644 LIST_INIT(&newindirdep->ir_donehd); 1645 if (bp->b_blkno == bp->b_lblkno) { 1646 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, 1647 NULL, NULL); 1648 } 1649 newindirdep->ir_savebp = 1650 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0); 1651 BUF_KERNPROC(newindirdep->ir_savebp); 1652 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 1653 } 1654} 1655 1656/* 1657 * Block de-allocation dependencies. 1658 * 1659 * When blocks are de-allocated, the on-disk pointers must be nullified before 1660 * the blocks are made available for use by other files. (The true 1661 * requirement is that old pointers must be nullified before new on-disk 1662 * pointers are set. We chose this slightly more stringent requirement to 1663 * reduce complexity.) Our implementation handles this dependency by updating 1664 * the inode (or indirect block) appropriately but delaying the actual block 1665 * de-allocation (i.e., freemap and free space count manipulation) until 1666 * after the updated versions reach stable storage. After the disk is 1667 * updated, the blocks can be safely de-allocated whenever it is convenient. 1668 * This implementation handles only the common case of reducing a file's 1669 * length to zero. Other cases are handled by the conventional synchronous 1670 * write approach. 1671 * 1672 * The ffs implementation with which we worked double-checks 1673 * the state of the block pointers and file size as it reduces 1674 * a file's length. Some of this code is replicated here in our 1675 * soft updates implementation. The freeblks->fb_chkcnt field is 1676 * used to transfer a part of this information to the procedure 1677 * that eventually de-allocates the blocks. 1678 * 1679 * This routine should be called from the routine that shortens 1680 * a file's length, before the inode's size or block pointers 1681 * are modified. It will save the block pointer information for 1682 * later release and zero the inode so that the calling routine 1683 * can release it. 1684 */ 1685void 1686softdep_setup_freeblocks(ip, length) 1687 struct inode *ip; /* The inode whose length is to be reduced */ 1688 off_t length; /* The new length for the file */ 1689{ 1690 struct freeblks *freeblks; 1691 struct inodedep *inodedep; 1692 struct allocdirect *adp; 1693 struct vnode *vp; 1694 struct buf *bp; 1695 struct fs *fs; 1696 int i, delay, error; 1697 1698 fs = ip->i_fs; 1699 if (length != 0) 1700 panic("softde_setup_freeblocks: non-zero length"); 1701 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks), 1702 M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO); 1703 freeblks->fb_list.wk_type = D_FREEBLKS; 1704 freeblks->fb_uid = ip->i_uid; 1705 freeblks->fb_previousinum = ip->i_number; 1706 freeblks->fb_devvp = ip->i_devvp; 1707 freeblks->fb_mnt = ITOV(ip)->v_mount; 1708 freeblks->fb_oldsize = ip->i_size; 1709 freeblks->fb_newsize = length; 1710 freeblks->fb_chkcnt = ip->i_blocks; 1711 for (i = 0; i < NDADDR; i++) { 1712 freeblks->fb_dblks[i] = ip->i_db[i]; 1713 ip->i_db[i] = 0; 1714 } 1715 for (i = 0; i < NIADDR; i++) { 1716 freeblks->fb_iblks[i] = ip->i_ib[i]; 1717 ip->i_ib[i] = 0; 1718 } 1719 ip->i_blocks = 0; 1720 ip->i_size = 0; 1721 /* 1722 * Push the zero'ed inode to to its disk buffer so that we are free 1723 * to delete its dependencies below. Once the dependencies are gone 1724 * the buffer can be safely released. 1725 */ 1726 if ((error = bread(ip->i_devvp, 1727 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 1728 (int)fs->fs_bsize, NOCRED, &bp)) != 0) 1729 softdep_error("softdep_setup_freeblocks", error); 1730 *((struct dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) = 1731 ip->i_din; 1732 /* 1733 * Find and eliminate any inode dependencies. 1734 */ 1735 ACQUIRE_LOCK(&lk); 1736 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep); 1737 if ((inodedep->id_state & IOSTARTED) != 0) 1738 panic("softdep_setup_freeblocks: inode busy"); 1739 /* 1740 * Add the freeblks structure to the list of operations that 1741 * must await the zero'ed inode being written to disk. If we 1742 * still have a bitmap dependency (delay == 0), then the inode 1743 * has never been written to disk, so we can process the 1744 * freeblks below once we have deleted the dependencies. 1745 */ 1746 delay = (inodedep->id_state & DEPCOMPLETE); 1747 if (delay) 1748 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list); 1749 /* 1750 * Because the file length has been truncated to zero, any 1751 * pending block allocation dependency structures associated 1752 * with this inode are obsolete and can simply be de-allocated. 1753 * We must first merge the two dependency lists to get rid of 1754 * any duplicate freefrag structures, then purge the merged list. 1755 * If we still have a bitmap dependency, then the inode has never 1756 * been written to disk, so we can free any fragments without delay. 1757 */ 1758 merge_inode_lists(inodedep); 1759 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 1760 free_allocdirect(&inodedep->id_inoupdt, adp, delay); 1761 FREE_LOCK(&lk); 1762 bdwrite(bp); 1763 /* 1764 * We must wait for any I/O in progress to finish so that 1765 * all potential buffers on the dirty list will be visible. 1766 * Once they are all there, walk the list and get rid of 1767 * any dependencies. 1768 */ 1769 vp = ITOV(ip); 1770 ACQUIRE_LOCK(&lk); 1771 drain_output(vp, 1); 1772 while (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT)) { 1773 bp = TAILQ_FIRST(&vp->v_dirtyblkhd); 1774 (void) inodedep_lookup(fs, ip->i_number, 0, &inodedep); 1775 deallocate_dependencies(bp, inodedep); 1776 bp->b_flags |= B_INVAL | B_NOCACHE; 1777 FREE_LOCK(&lk); 1778 brelse(bp); 1779 ACQUIRE_LOCK(&lk); 1780 } 1781 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0) 1782 (void) free_inodedep(inodedep); 1783 FREE_LOCK(&lk); 1784 /* 1785 * If the inode has never been written to disk (delay == 0), 1786 * then we can process the freeblks now that we have deleted 1787 * the dependencies. 1788 */ 1789 if (!delay) 1790 handle_workitem_freeblocks(freeblks); 1791} 1792 1793/* 1794 * Reclaim any dependency structures from a buffer that is about to 1795 * be reallocated to a new vnode. The buffer must be locked, thus, 1796 * no I/O completion operations can occur while we are manipulating 1797 * its associated dependencies. The mutex is held so that other I/O's 1798 * associated with related dependencies do not occur. 1799 */ 1800static void 1801deallocate_dependencies(bp, inodedep) 1802 struct buf *bp; 1803 struct inodedep *inodedep; 1804{ 1805 struct worklist *wk; 1806 struct indirdep *indirdep; 1807 struct allocindir *aip; 1808 struct pagedep *pagedep; 1809 struct dirrem *dirrem; 1810 struct diradd *dap; 1811 int i; 1812 1813 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 1814 switch (wk->wk_type) { 1815 1816 case D_INDIRDEP: 1817 indirdep = WK_INDIRDEP(wk); 1818 /* 1819 * None of the indirect pointers will ever be visible, 1820 * so they can simply be tossed. GOINGAWAY ensures 1821 * that allocated pointers will be saved in the buffer 1822 * cache until they are freed. Note that they will 1823 * only be able to be found by their physical address 1824 * since the inode mapping the logical address will 1825 * be gone. The save buffer used for the safe copy 1826 * was allocated in setup_allocindir_phase2 using 1827 * the physical address so it could be used for this 1828 * purpose. Hence we swap the safe copy with the real 1829 * copy, allowing the safe copy to be freed and holding 1830 * on to the real copy for later use in indir_trunc. 1831 */ 1832 if (indirdep->ir_state & GOINGAWAY) 1833 panic("deallocate_dependencies: already gone"); 1834 indirdep->ir_state |= GOINGAWAY; 1835 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 1836 free_allocindir(aip, inodedep); 1837 if (bp->b_lblkno >= 0 || 1838 bp->b_blkno != indirdep->ir_savebp->b_lblkno) 1839 panic("deallocate_dependencies: not indir"); 1840 bcopy(bp->b_data, indirdep->ir_savebp->b_data, 1841 bp->b_bcount); 1842 WORKLIST_REMOVE(wk); 1843 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk); 1844 continue; 1845 1846 case D_PAGEDEP: 1847 pagedep = WK_PAGEDEP(wk); 1848 /* 1849 * None of the directory additions will ever be 1850 * visible, so they can simply be tossed. 1851 */ 1852 for (i = 0; i < DAHASHSZ; i++) 1853 while ((dap = 1854 LIST_FIRST(&pagedep->pd_diraddhd[i]))) 1855 free_diradd(dap); 1856 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0) 1857 free_diradd(dap); 1858 /* 1859 * Copy any directory remove dependencies to the list 1860 * to be processed after the zero'ed inode is written. 1861 * If the inode has already been written, then they 1862 * can be dumped directly onto the work list. 1863 */ 1864 for (dirrem = LIST_FIRST(&pagedep->pd_dirremhd); dirrem; 1865 dirrem = LIST_NEXT(dirrem, dm_next)) { 1866 LIST_REMOVE(dirrem, dm_next); 1867 dirrem->dm_dirinum = pagedep->pd_ino; 1868 if (inodedep == NULL || 1869 (inodedep->id_state & ALLCOMPLETE) == 1870 ALLCOMPLETE) 1871 add_to_worklist(&dirrem->dm_list); 1872 else 1873 WORKLIST_INSERT(&inodedep->id_bufwait, 1874 &dirrem->dm_list); 1875 } 1876 WORKLIST_REMOVE(&pagedep->pd_list); 1877 LIST_REMOVE(pagedep, pd_hash); 1878 WORKITEM_FREE(pagedep, D_PAGEDEP); 1879 continue; 1880 1881 case D_ALLOCINDIR: 1882 free_allocindir(WK_ALLOCINDIR(wk), inodedep); 1883 continue; 1884 1885 case D_ALLOCDIRECT: 1886 case D_INODEDEP: 1887 panic("deallocate_dependencies: Unexpected type %s", 1888 TYPENAME(wk->wk_type)); 1889 /* NOTREACHED */ 1890 1891 default: 1892 panic("deallocate_dependencies: Unknown type %s", 1893 TYPENAME(wk->wk_type)); 1894 /* NOTREACHED */ 1895 } 1896 } 1897} 1898 1899/* 1900 * Free an allocdirect. Generate a new freefrag work request if appropriate. 1901 * This routine must be called with splbio interrupts blocked. 1902 */ 1903static void 1904free_allocdirect(adphead, adp, delay) 1905 struct allocdirectlst *adphead; 1906 struct allocdirect *adp; 1907 int delay; 1908{ 1909 1910#ifdef DEBUG 1911 if (lk.lkt_held == -1) 1912 panic("free_allocdirect: lock not held"); 1913#endif 1914 if ((adp->ad_state & DEPCOMPLETE) == 0) 1915 LIST_REMOVE(adp, ad_deps); 1916 TAILQ_REMOVE(adphead, adp, ad_next); 1917 if ((adp->ad_state & COMPLETE) == 0) 1918 WORKLIST_REMOVE(&adp->ad_list); 1919 if (adp->ad_freefrag != NULL) { 1920 if (delay) 1921 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 1922 &adp->ad_freefrag->ff_list); 1923 else 1924 add_to_worklist(&adp->ad_freefrag->ff_list); 1925 } 1926 WORKITEM_FREE(adp, D_ALLOCDIRECT); 1927} 1928 1929/* 1930 * Prepare an inode to be freed. The actual free operation is not 1931 * done until the zero'ed inode has been written to disk. 1932 */ 1933void 1934softdep_freefile(pvp, ino, mode) 1935 struct vnode *pvp; 1936 ino_t ino; 1937 int mode; 1938{ 1939 struct inode *ip = VTOI(pvp); 1940 struct inodedep *inodedep; 1941 struct freefile *freefile; 1942 1943 /* 1944 * This sets up the inode de-allocation dependency. 1945 */ 1946 MALLOC(freefile, struct freefile *, sizeof(struct freefile), 1947 M_FREEFILE, M_SOFTDEP_FLAGS); 1948 freefile->fx_list.wk_type = D_FREEFILE; 1949 freefile->fx_list.wk_state = 0; 1950 freefile->fx_mode = mode; 1951 freefile->fx_oldinum = ino; 1952 freefile->fx_devvp = ip->i_devvp; 1953 freefile->fx_mnt = ITOV(ip)->v_mount; 1954 1955 /* 1956 * If the inodedep does not exist, then the zero'ed inode has 1957 * been written to disk. If the allocated inode has never been 1958 * written to disk, then the on-disk inode is zero'ed. In either 1959 * case we can free the file immediately. 1960 */ 1961 ACQUIRE_LOCK(&lk); 1962 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 || 1963 check_inode_unwritten(inodedep)) { 1964 FREE_LOCK(&lk); 1965 handle_workitem_freefile(freefile); 1966 return; 1967 } 1968 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 1969 FREE_LOCK(&lk); 1970} 1971 1972/* 1973 * Check to see if an inode has never been written to disk. If 1974 * so free the inodedep and return success, otherwise return failure. 1975 * This routine must be called with splbio interrupts blocked. 1976 * 1977 * If we still have a bitmap dependency, then the inode has never 1978 * been written to disk. Drop the dependency as it is no longer 1979 * necessary since the inode is being deallocated. We set the 1980 * ALLCOMPLETE flags since the bitmap now properly shows that the 1981 * inode is not allocated. Even if the inode is actively being 1982 * written, it has been rolled back to its zero'ed state, so we 1983 * are ensured that a zero inode is what is on the disk. For short 1984 * lived files, this change will usually result in removing all the 1985 * dependencies from the inode so that it can be freed immediately. 1986 */ 1987static int 1988check_inode_unwritten(inodedep) 1989 struct inodedep *inodedep; 1990{ 1991 1992 if ((inodedep->id_state & DEPCOMPLETE) != 0 || 1993 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 1994 LIST_FIRST(&inodedep->id_bufwait) != NULL || 1995 LIST_FIRST(&inodedep->id_inowait) != NULL || 1996 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 1997 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 1998 inodedep->id_nlinkdelta != 0) 1999 return (0); 2000 inodedep->id_state |= ALLCOMPLETE; 2001 LIST_REMOVE(inodedep, id_deps); 2002 inodedep->id_buf = NULL; 2003 if (inodedep->id_state & ONWORKLIST) 2004 WORKLIST_REMOVE(&inodedep->id_list); 2005 if (inodedep->id_savedino != NULL) { 2006 FREE(inodedep->id_savedino, M_INODEDEP); 2007 inodedep->id_savedino = NULL; 2008 } 2009 if (free_inodedep(inodedep) == 0) 2010 panic("check_inode_unwritten: busy inode"); 2011 return (1); 2012} 2013 2014/* 2015 * Try to free an inodedep structure. Return 1 if it could be freed. 2016 */ 2017static int 2018free_inodedep(inodedep) 2019 struct inodedep *inodedep; 2020{ 2021 2022 if ((inodedep->id_state & ONWORKLIST) != 0 || 2023 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 2024 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2025 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2026 LIST_FIRST(&inodedep->id_inowait) != NULL || 2027 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2028 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2029 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL) 2030 return (0); 2031 LIST_REMOVE(inodedep, id_hash); 2032 WORKITEM_FREE(inodedep, D_INODEDEP); 2033 num_inodedep -= 1; 2034 return (1); 2035} 2036 2037/* 2038 * This workitem routine performs the block de-allocation. 2039 * The workitem is added to the pending list after the updated 2040 * inode block has been written to disk. As mentioned above, 2041 * checks regarding the number of blocks de-allocated (compared 2042 * to the number of blocks allocated for the file) are also 2043 * performed in this function. 2044 */ 2045static void 2046handle_workitem_freeblocks(freeblks) 2047 struct freeblks *freeblks; 2048{ 2049 struct inode tip; 2050 ufs_daddr_t bn; 2051 struct fs *fs; 2052 int i, level, bsize; 2053 long nblocks, blocksreleased = 0; 2054 int error, allerror = 0; 2055 ufs_lbn_t baselbns[NIADDR], tmpval; 2056 2057 tip.i_fs = fs = VFSTOUFS(freeblks->fb_mnt)->um_fs; 2058 tip.i_number = freeblks->fb_previousinum; 2059 tip.i_devvp = freeblks->fb_devvp; 2060 tip.i_dev = freeblks->fb_devvp->v_rdev; 2061 tip.i_size = freeblks->fb_oldsize; 2062 tip.i_uid = freeblks->fb_uid; 2063 tip.i_vnode = NULL; 2064 tmpval = 1; 2065 baselbns[0] = NDADDR; 2066 for (i = 1; i < NIADDR; i++) { 2067 tmpval *= NINDIR(fs); 2068 baselbns[i] = baselbns[i - 1] + tmpval; 2069 } 2070 nblocks = btodb(fs->fs_bsize); 2071 blocksreleased = 0; 2072 /* 2073 * Indirect blocks first. 2074 */ 2075 for (level = (NIADDR - 1); level >= 0; level--) { 2076 if ((bn = freeblks->fb_iblks[level]) == 0) 2077 continue; 2078 if ((error = indir_trunc(&tip, fsbtodb(fs, bn), level, 2079 baselbns[level], &blocksreleased)) == 0) 2080 allerror = error; 2081 ffs_blkfree(&tip, bn, fs->fs_bsize); 2082 blocksreleased += nblocks; 2083 } 2084 /* 2085 * All direct blocks or frags. 2086 */ 2087 for (i = (NDADDR - 1); i >= 0; i--) { 2088 if ((bn = freeblks->fb_dblks[i]) == 0) 2089 continue; 2090 bsize = blksize(fs, &tip, i); 2091 ffs_blkfree(&tip, bn, bsize); 2092 blocksreleased += btodb(bsize); 2093 } 2094 2095#ifdef DIAGNOSTIC 2096 if (freeblks->fb_chkcnt != blocksreleased) 2097 printf("handle_workitem_freeblocks: block count"); 2098 if (allerror) 2099 softdep_error("handle_workitem_freeblks", allerror); 2100#endif /* DIAGNOSTIC */ 2101 WORKITEM_FREE(freeblks, D_FREEBLKS); 2102} 2103 2104/* 2105 * Release blocks associated with the inode ip and stored in the indirect 2106 * block dbn. If level is greater than SINGLE, the block is an indirect block 2107 * and recursive calls to indirtrunc must be used to cleanse other indirect 2108 * blocks. 2109 */ 2110static int 2111indir_trunc(ip, dbn, level, lbn, countp) 2112 struct inode *ip; 2113 ufs_daddr_t dbn; 2114 int level; 2115 ufs_lbn_t lbn; 2116 long *countp; 2117{ 2118 struct buf *bp; 2119 ufs_daddr_t *bap; 2120 ufs_daddr_t nb; 2121 struct fs *fs; 2122 struct worklist *wk; 2123 struct indirdep *indirdep; 2124 int i, lbnadd, nblocks; 2125 int error, allerror = 0; 2126 2127 fs = ip->i_fs; 2128 lbnadd = 1; 2129 for (i = level; i > 0; i--) 2130 lbnadd *= NINDIR(fs); 2131 /* 2132 * Get buffer of block pointers to be freed. This routine is not 2133 * called until the zero'ed inode has been written, so it is safe 2134 * to free blocks as they are encountered. Because the inode has 2135 * been zero'ed, calls to bmap on these blocks will fail. So, we 2136 * have to use the on-disk address and the block device for the 2137 * filesystem to look them up. If the file was deleted before its 2138 * indirect blocks were all written to disk, the routine that set 2139 * us up (deallocate_dependencies) will have arranged to leave 2140 * a complete copy of the indirect block in memory for our use. 2141 * Otherwise we have to read the blocks in from the disk. 2142 */ 2143 ACQUIRE_LOCK(&lk); 2144 if ((bp = incore(ip->i_devvp, dbn)) != NULL && 2145 (wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2146 if (wk->wk_type != D_INDIRDEP || 2147 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp || 2148 (indirdep->ir_state & GOINGAWAY) == 0) 2149 panic("indir_trunc: lost indirdep"); 2150 WORKLIST_REMOVE(wk); 2151 WORKITEM_FREE(indirdep, D_INDIRDEP); 2152 if (LIST_FIRST(&bp->b_dep) != NULL) 2153 panic("indir_trunc: dangling dep"); 2154 FREE_LOCK(&lk); 2155 } else { 2156 FREE_LOCK(&lk); 2157 error = bread(ip->i_devvp, dbn, (int)fs->fs_bsize, NOCRED, &bp); 2158 if (error) 2159 return (error); 2160 } 2161 /* 2162 * Recursively free indirect blocks. 2163 */ 2164 bap = (ufs_daddr_t *)bp->b_data; 2165 nblocks = btodb(fs->fs_bsize); 2166 for (i = NINDIR(fs) - 1; i >= 0; i--) { 2167 if ((nb = bap[i]) == 0) 2168 continue; 2169 if (level != 0) { 2170 if ((error = indir_trunc(ip, fsbtodb(fs, nb), 2171 level - 1, lbn + (i * lbnadd), countp)) != 0) 2172 allerror = error; 2173 } 2174 ffs_blkfree(ip, nb, fs->fs_bsize); 2175 *countp += nblocks; 2176 } 2177 bp->b_flags |= B_INVAL | B_NOCACHE; 2178 brelse(bp); 2179 return (allerror); 2180} 2181 2182/* 2183 * Free an allocindir. 2184 * This routine must be called with splbio interrupts blocked. 2185 */ 2186static void 2187free_allocindir(aip, inodedep) 2188 struct allocindir *aip; 2189 struct inodedep *inodedep; 2190{ 2191 struct freefrag *freefrag; 2192 2193#ifdef DEBUG 2194 if (lk.lkt_held == -1) 2195 panic("free_allocindir: lock not held"); 2196#endif 2197 if ((aip->ai_state & DEPCOMPLETE) == 0) 2198 LIST_REMOVE(aip, ai_deps); 2199 if (aip->ai_state & ONWORKLIST) 2200 WORKLIST_REMOVE(&aip->ai_list); 2201 LIST_REMOVE(aip, ai_next); 2202 if ((freefrag = aip->ai_freefrag) != NULL) { 2203 if (inodedep == NULL) 2204 add_to_worklist(&freefrag->ff_list); 2205 else 2206 WORKLIST_INSERT(&inodedep->id_bufwait, 2207 &freefrag->ff_list); 2208 } 2209 WORKITEM_FREE(aip, D_ALLOCINDIR); 2210} 2211 2212/* 2213 * Directory entry addition dependencies. 2214 * 2215 * When adding a new directory entry, the inode (with its incremented link 2216 * count) must be written to disk before the directory entry's pointer to it. 2217 * Also, if the inode is newly allocated, the corresponding freemap must be 2218 * updated (on disk) before the directory entry's pointer. These requirements 2219 * are met via undo/redo on the directory entry's pointer, which consists 2220 * simply of the inode number. 2221 * 2222 * As directory entries are added and deleted, the free space within a 2223 * directory block can become fragmented. The ufs file system will compact 2224 * a fragmented directory block to make space for a new entry. When this 2225 * occurs, the offsets of previously added entries change. Any "diradd" 2226 * dependency structures corresponding to these entries must be updated with 2227 * the new offsets. 2228 */ 2229 2230/* 2231 * This routine is called after the in-memory inode's link 2232 * count has been incremented, but before the directory entry's 2233 * pointer to the inode has been set. 2234 */ 2235void 2236softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp) 2237 struct buf *bp; /* buffer containing directory block */ 2238 struct inode *dp; /* inode for directory */ 2239 off_t diroffset; /* offset of new entry in directory */ 2240 long newinum; /* inode referenced by new directory entry */ 2241 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 2242{ 2243 int offset; /* offset of new entry within directory block */ 2244 ufs_lbn_t lbn; /* block in directory containing new entry */ 2245 struct fs *fs; 2246 struct diradd *dap; 2247 struct pagedep *pagedep; 2248 struct inodedep *inodedep; 2249 struct mkdir *mkdir1, *mkdir2; 2250 2251 /* 2252 * Whiteouts have no dependencies. 2253 */ 2254 if (newinum == WINO) { 2255 if (newdirbp != NULL) 2256 bdwrite(newdirbp); 2257 return; 2258 } 2259 2260 fs = dp->i_fs; 2261 lbn = lblkno(fs, diroffset); 2262 offset = blkoff(fs, diroffset); 2263 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, 2264 M_SOFTDEP_FLAGS|M_ZERO); 2265 dap->da_list.wk_type = D_DIRADD; 2266 dap->da_offset = offset; 2267 dap->da_newinum = newinum; 2268 dap->da_state = ATTACHED; 2269 if (newdirbp == NULL) { 2270 dap->da_state |= DEPCOMPLETE; 2271 ACQUIRE_LOCK(&lk); 2272 } else { 2273 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 2274 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2275 M_SOFTDEP_FLAGS); 2276 mkdir1->md_list.wk_type = D_MKDIR; 2277 mkdir1->md_state = MKDIR_BODY; 2278 mkdir1->md_diradd = dap; 2279 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2280 M_SOFTDEP_FLAGS); 2281 mkdir2->md_list.wk_type = D_MKDIR; 2282 mkdir2->md_state = MKDIR_PARENT; 2283 mkdir2->md_diradd = dap; 2284 /* 2285 * Dependency on "." and ".." being written to disk. 2286 */ 2287 mkdir1->md_buf = newdirbp; 2288 ACQUIRE_LOCK(&lk); 2289 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 2290 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list); 2291 FREE_LOCK(&lk); 2292 bdwrite(newdirbp); 2293 /* 2294 * Dependency on link count increase for parent directory 2295 */ 2296 ACQUIRE_LOCK(&lk); 2297 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0 2298 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2299 dap->da_state &= ~MKDIR_PARENT; 2300 WORKITEM_FREE(mkdir2, D_MKDIR); 2301 } else { 2302 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 2303 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 2304 } 2305 } 2306 /* 2307 * Link into parent directory pagedep to await its being written. 2308 */ 2309 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2310 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2311 dap->da_pagedep = pagedep; 2312 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 2313 da_pdlist); 2314 /* 2315 * Link into its inodedep. Put it on the id_bufwait list if the inode 2316 * is not yet written. If it is written, do the post-inode write 2317 * processing to put it on the id_pendinghd list. 2318 */ 2319 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep); 2320 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 2321 diradd_inode_written(dap, inodedep); 2322 else 2323 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2324 FREE_LOCK(&lk); 2325} 2326 2327/* 2328 * This procedure is called to change the offset of a directory 2329 * entry when compacting a directory block which must be owned 2330 * exclusively by the caller. Note that the actual entry movement 2331 * must be done in this procedure to ensure that no I/O completions 2332 * occur while the move is in progress. 2333 */ 2334void 2335softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize) 2336 struct inode *dp; /* inode for directory */ 2337 caddr_t base; /* address of dp->i_offset */ 2338 caddr_t oldloc; /* address of old directory location */ 2339 caddr_t newloc; /* address of new directory location */ 2340 int entrysize; /* size of directory entry */ 2341{ 2342 int offset, oldoffset, newoffset; 2343 struct pagedep *pagedep; 2344 struct diradd *dap; 2345 ufs_lbn_t lbn; 2346 2347 ACQUIRE_LOCK(&lk); 2348 lbn = lblkno(dp->i_fs, dp->i_offset); 2349 offset = blkoff(dp->i_fs, dp->i_offset); 2350 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) 2351 goto done; 2352 oldoffset = offset + (oldloc - base); 2353 newoffset = offset + (newloc - base); 2354 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[DIRADDHASH(oldoffset)]); 2355 dap; dap = LIST_NEXT(dap, da_pdlist)) { 2356 if (dap->da_offset != oldoffset) 2357 continue; 2358 dap->da_offset = newoffset; 2359 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) 2360 break; 2361 LIST_REMOVE(dap, da_pdlist); 2362 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], 2363 dap, da_pdlist); 2364 break; 2365 } 2366 if (dap == NULL) { 2367 for (dap = LIST_FIRST(&pagedep->pd_pendinghd); 2368 dap; dap = LIST_NEXT(dap, da_pdlist)) { 2369 if (dap->da_offset == oldoffset) { 2370 dap->da_offset = newoffset; 2371 break; 2372 } 2373 } 2374 } 2375done: 2376 bcopy(oldloc, newloc, entrysize); 2377 FREE_LOCK(&lk); 2378} 2379 2380/* 2381 * Free a diradd dependency structure. This routine must be called 2382 * with splbio interrupts blocked. 2383 */ 2384static void 2385free_diradd(dap) 2386 struct diradd *dap; 2387{ 2388 struct dirrem *dirrem; 2389 struct pagedep *pagedep; 2390 struct inodedep *inodedep; 2391 struct mkdir *mkdir, *nextmd; 2392 2393#ifdef DEBUG 2394 if (lk.lkt_held == -1) 2395 panic("free_diradd: lock not held"); 2396#endif 2397 WORKLIST_REMOVE(&dap->da_list); 2398 LIST_REMOVE(dap, da_pdlist); 2399 if ((dap->da_state & DIRCHG) == 0) { 2400 pagedep = dap->da_pagedep; 2401 } else { 2402 dirrem = dap->da_previous; 2403 pagedep = dirrem->dm_pagedep; 2404 dirrem->dm_dirinum = pagedep->pd_ino; 2405 add_to_worklist(&dirrem->dm_list); 2406 } 2407 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum, 2408 0, &inodedep) != 0) 2409 (void) free_inodedep(inodedep); 2410 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2411 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 2412 nextmd = LIST_NEXT(mkdir, md_mkdirs); 2413 if (mkdir->md_diradd != dap) 2414 continue; 2415 dap->da_state &= ~mkdir->md_state; 2416 WORKLIST_REMOVE(&mkdir->md_list); 2417 LIST_REMOVE(mkdir, md_mkdirs); 2418 WORKITEM_FREE(mkdir, D_MKDIR); 2419 } 2420 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 2421 panic("free_diradd: unfound ref"); 2422 } 2423 WORKITEM_FREE(dap, D_DIRADD); 2424} 2425 2426/* 2427 * Directory entry removal dependencies. 2428 * 2429 * When removing a directory entry, the entry's inode pointer must be 2430 * zero'ed on disk before the corresponding inode's link count is decremented 2431 * (possibly freeing the inode for re-use). This dependency is handled by 2432 * updating the directory entry but delaying the inode count reduction until 2433 * after the directory block has been written to disk. After this point, the 2434 * inode count can be decremented whenever it is convenient. 2435 */ 2436 2437/* 2438 * This routine should be called immediately after removing 2439 * a directory entry. The inode's link count should not be 2440 * decremented by the calling procedure -- the soft updates 2441 * code will do this task when it is safe. 2442 */ 2443void 2444softdep_setup_remove(bp, dp, ip, isrmdir) 2445 struct buf *bp; /* buffer containing directory block */ 2446 struct inode *dp; /* inode for the directory being modified */ 2447 struct inode *ip; /* inode for directory entry being removed */ 2448 int isrmdir; /* indicates if doing RMDIR */ 2449{ 2450 struct dirrem *dirrem, *prevdirrem; 2451 2452 /* 2453 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 2454 */ 2455 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2456 2457 /* 2458 * If the COMPLETE flag is clear, then there were no active 2459 * entries and we want to roll back to a zeroed entry until 2460 * the new inode is committed to disk. If the COMPLETE flag is 2461 * set then we have deleted an entry that never made it to 2462 * disk. If the entry we deleted resulted from a name change, 2463 * then the old name still resides on disk. We cannot delete 2464 * its inode (returned to us in prevdirrem) until the zeroed 2465 * directory entry gets to disk. The new inode has never been 2466 * referenced on the disk, so can be deleted immediately. 2467 */ 2468 if ((dirrem->dm_state & COMPLETE) == 0) { 2469 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 2470 dm_next); 2471 FREE_LOCK(&lk); 2472 } else { 2473 if (prevdirrem != NULL) 2474 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 2475 prevdirrem, dm_next); 2476 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 2477 FREE_LOCK(&lk); 2478 handle_workitem_remove(dirrem); 2479 } 2480} 2481 2482/* 2483 * Allocate a new dirrem if appropriate and return it along with 2484 * its associated pagedep. Called without a lock, returns with lock. 2485 */ 2486static long num_dirrem; /* number of dirrem allocated */ 2487static struct dirrem * 2488newdirrem(bp, dp, ip, isrmdir, prevdirremp) 2489 struct buf *bp; /* buffer containing directory block */ 2490 struct inode *dp; /* inode for the directory being modified */ 2491 struct inode *ip; /* inode for directory entry being removed */ 2492 int isrmdir; /* indicates if doing RMDIR */ 2493 struct dirrem **prevdirremp; /* previously referenced inode, if any */ 2494{ 2495 int offset; 2496 ufs_lbn_t lbn; 2497 struct diradd *dap; 2498 struct dirrem *dirrem; 2499 struct pagedep *pagedep; 2500 2501 /* 2502 * Whiteouts have no deletion dependencies. 2503 */ 2504 if (ip == NULL) 2505 panic("newdirrem: whiteout"); 2506 /* 2507 * If we are over our limit, try to improve the situation. 2508 * Limiting the number of dirrem structures will also limit 2509 * the number of freefile and freeblks structures. 2510 */ 2511 if (num_dirrem > max_softdeps / 2) 2512 (void) request_cleanup(FLUSH_REMOVE, 0); 2513 num_dirrem += 1; 2514 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem), 2515 M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO); 2516 dirrem->dm_list.wk_type = D_DIRREM; 2517 dirrem->dm_state = isrmdir ? RMDIR : 0; 2518 dirrem->dm_mnt = ITOV(ip)->v_mount; 2519 dirrem->dm_oldinum = ip->i_number; 2520 *prevdirremp = NULL; 2521 2522 ACQUIRE_LOCK(&lk); 2523 lbn = lblkno(dp->i_fs, dp->i_offset); 2524 offset = blkoff(dp->i_fs, dp->i_offset); 2525 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2526 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2527 dirrem->dm_pagedep = pagedep; 2528 /* 2529 * Check for a diradd dependency for the same directory entry. 2530 * If present, then both dependencies become obsolete and can 2531 * be de-allocated. Check for an entry on both the pd_dirraddhd 2532 * list and the pd_pendinghd list. 2533 */ 2534 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[DIRADDHASH(offset)]); 2535 dap; dap = LIST_NEXT(dap, da_pdlist)) 2536 if (dap->da_offset == offset) 2537 break; 2538 if (dap == NULL) { 2539 for (dap = LIST_FIRST(&pagedep->pd_pendinghd); 2540 dap; dap = LIST_NEXT(dap, da_pdlist)) 2541 if (dap->da_offset == offset) 2542 break; 2543 if (dap == NULL) 2544 return (dirrem); 2545 } 2546 /* 2547 * Must be ATTACHED at this point. 2548 */ 2549 if ((dap->da_state & ATTACHED) == 0) 2550 panic("newdirrem: not ATTACHED"); 2551 if (dap->da_newinum != ip->i_number) 2552 panic("newdirrem: inum %d should be %d", 2553 ip->i_number, dap->da_newinum); 2554 /* 2555 * If we are deleting a changed name that never made it to disk, 2556 * then return the dirrem describing the previous inode (which 2557 * represents the inode currently referenced from this entry on disk). 2558 */ 2559 if ((dap->da_state & DIRCHG) != 0) { 2560 *prevdirremp = dap->da_previous; 2561 dap->da_state &= ~DIRCHG; 2562 dap->da_pagedep = pagedep; 2563 } 2564 /* 2565 * We are deleting an entry that never made it to disk. 2566 * Mark it COMPLETE so we can delete its inode immediately. 2567 */ 2568 dirrem->dm_state |= COMPLETE; 2569 free_diradd(dap); 2570 return (dirrem); 2571} 2572 2573/* 2574 * Directory entry change dependencies. 2575 * 2576 * Changing an existing directory entry requires that an add operation 2577 * be completed first followed by a deletion. The semantics for the addition 2578 * are identical to the description of adding a new entry above except 2579 * that the rollback is to the old inode number rather than zero. Once 2580 * the addition dependency is completed, the removal is done as described 2581 * in the removal routine above. 2582 */ 2583 2584/* 2585 * This routine should be called immediately after changing 2586 * a directory entry. The inode's link count should not be 2587 * decremented by the calling procedure -- the soft updates 2588 * code will perform this task when it is safe. 2589 */ 2590void 2591softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 2592 struct buf *bp; /* buffer containing directory block */ 2593 struct inode *dp; /* inode for the directory being modified */ 2594 struct inode *ip; /* inode for directory entry being removed */ 2595 long newinum; /* new inode number for changed entry */ 2596 int isrmdir; /* indicates if doing RMDIR */ 2597{ 2598 int offset; 2599 struct diradd *dap = NULL; 2600 struct dirrem *dirrem, *prevdirrem; 2601 struct pagedep *pagedep; 2602 struct inodedep *inodedep; 2603 2604 offset = blkoff(dp->i_fs, dp->i_offset); 2605 2606 /* 2607 * Whiteouts do not need diradd dependencies. 2608 */ 2609 if (newinum != WINO) { 2610 MALLOC(dap, struct diradd *, sizeof(struct diradd), 2611 M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO); 2612 dap->da_list.wk_type = D_DIRADD; 2613 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 2614 dap->da_offset = offset; 2615 dap->da_newinum = newinum; 2616 } 2617 2618 /* 2619 * Allocate a new dirrem and ACQUIRE_LOCK. 2620 */ 2621 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2622 pagedep = dirrem->dm_pagedep; 2623 /* 2624 * The possible values for isrmdir: 2625 * 0 - non-directory file rename 2626 * 1 - directory rename within same directory 2627 * inum - directory rename to new directory of given inode number 2628 * When renaming to a new directory, we are both deleting and 2629 * creating a new directory entry, so the link count on the new 2630 * directory should not change. Thus we do not need the followup 2631 * dirrem which is usually done in handle_workitem_remove. We set 2632 * the DIRCHG flag to tell handle_workitem_remove to skip the 2633 * followup dirrem. 2634 */ 2635 if (isrmdir > 1) 2636 dirrem->dm_state |= DIRCHG; 2637 2638 /* 2639 * Whiteouts have no additional dependencies, 2640 * so just put the dirrem on the correct list. 2641 */ 2642 if (newinum == WINO) { 2643 if ((dirrem->dm_state & COMPLETE) == 0) { 2644 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 2645 dm_next); 2646 } else { 2647 dirrem->dm_dirinum = pagedep->pd_ino; 2648 add_to_worklist(&dirrem->dm_list); 2649 } 2650 FREE_LOCK(&lk); 2651 return; 2652 } 2653 2654 /* 2655 * If the COMPLETE flag is clear, then there were no active 2656 * entries and we want to roll back to the previous inode until 2657 * the new inode is committed to disk. If the COMPLETE flag is 2658 * set, then we have deleted an entry that never made it to disk. 2659 * If the entry we deleted resulted from a name change, then the old 2660 * inode reference still resides on disk. Any rollback that we do 2661 * needs to be to that old inode (returned to us in prevdirrem). If 2662 * the entry we deleted resulted from a create, then there is 2663 * no entry on the disk, so we want to roll back to zero rather 2664 * than the uncommitted inode. In either of the COMPLETE cases we 2665 * want to immediately free the unwritten and unreferenced inode. 2666 */ 2667 if ((dirrem->dm_state & COMPLETE) == 0) { 2668 dap->da_previous = dirrem; 2669 } else { 2670 if (prevdirrem != NULL) { 2671 dap->da_previous = prevdirrem; 2672 } else { 2673 dap->da_state &= ~DIRCHG; 2674 dap->da_pagedep = pagedep; 2675 } 2676 dirrem->dm_dirinum = pagedep->pd_ino; 2677 add_to_worklist(&dirrem->dm_list); 2678 } 2679 /* 2680 * Link into its inodedep. Put it on the id_bufwait list if the inode 2681 * is not yet written. If it is written, do the post-inode write 2682 * processing to put it on the id_pendinghd list. 2683 */ 2684 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 || 2685 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2686 dap->da_state |= COMPLETE; 2687 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 2688 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 2689 } else { 2690 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 2691 dap, da_pdlist); 2692 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2693 } 2694 FREE_LOCK(&lk); 2695} 2696 2697/* 2698 * Called whenever the link count on an inode is changed. 2699 * It creates an inode dependency so that the new reference(s) 2700 * to the inode cannot be committed to disk until the updated 2701 * inode has been written. 2702 */ 2703void 2704softdep_change_linkcnt(ip) 2705 struct inode *ip; /* the inode with the increased link count */ 2706{ 2707 struct inodedep *inodedep; 2708 2709 ACQUIRE_LOCK(&lk); 2710 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 2711 if (ip->i_nlink < ip->i_effnlink) 2712 panic("softdep_change_linkcnt: bad delta"); 2713 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2714 FREE_LOCK(&lk); 2715} 2716 2717/* 2718 * This workitem decrements the inode's link count. 2719 * If the link count reaches zero, the file is removed. 2720 */ 2721static void 2722handle_workitem_remove(dirrem) 2723 struct dirrem *dirrem; 2724{ 2725 struct proc *p = CURPROC; /* XXX */ 2726 struct inodedep *inodedep; 2727 struct vnode *vp; 2728 struct inode *ip; 2729 ino_t oldinum; 2730 int error; 2731 2732 if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) { 2733 softdep_error("handle_workitem_remove: vget", error); 2734 return; 2735 } 2736 ip = VTOI(vp); 2737 ACQUIRE_LOCK(&lk); 2738 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0) 2739 panic("handle_workitem_remove: lost inodedep"); 2740 /* 2741 * Normal file deletion. 2742 */ 2743 if ((dirrem->dm_state & RMDIR) == 0) { 2744 ip->i_nlink--; 2745 ip->i_flag |= IN_CHANGE; 2746 if (ip->i_nlink < ip->i_effnlink) 2747 panic("handle_workitem_remove: bad file delta"); 2748 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2749 FREE_LOCK(&lk); 2750 vput(vp); 2751 num_dirrem -= 1; 2752 WORKITEM_FREE(dirrem, D_DIRREM); 2753 return; 2754 } 2755 /* 2756 * Directory deletion. Decrement reference count for both the 2757 * just deleted parent directory entry and the reference for ".". 2758 * Next truncate the directory to length zero. When the 2759 * truncation completes, arrange to have the reference count on 2760 * the parent decremented to account for the loss of "..". 2761 */ 2762 ip->i_nlink -= 2; 2763 ip->i_flag |= IN_CHANGE; 2764 if (ip->i_nlink < ip->i_effnlink) 2765 panic("handle_workitem_remove: bad dir delta"); 2766 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2767 FREE_LOCK(&lk); 2768 if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, p->p_ucred, p)) != 0) 2769 softdep_error("handle_workitem_remove: truncate", error); 2770 /* 2771 * Rename a directory to a new parent. Since, we are both deleting 2772 * and creating a new directory entry, the link count on the new 2773 * directory should not change. Thus we skip the followup dirrem. 2774 */ 2775 if (dirrem->dm_state & DIRCHG) { 2776 vput(vp); 2777 num_dirrem -= 1; 2778 WORKITEM_FREE(dirrem, D_DIRREM); 2779 return; 2780 } 2781 /* 2782 * If the inodedep does not exist, then the zero'ed inode has 2783 * been written to disk. If the allocated inode has never been 2784 * written to disk, then the on-disk inode is zero'ed. In either 2785 * case we can remove the file immediately. 2786 */ 2787 ACQUIRE_LOCK(&lk); 2788 dirrem->dm_state = 0; 2789 oldinum = dirrem->dm_oldinum; 2790 dirrem->dm_oldinum = dirrem->dm_dirinum; 2791 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 || 2792 check_inode_unwritten(inodedep)) { 2793 FREE_LOCK(&lk); 2794 vput(vp); 2795 handle_workitem_remove(dirrem); 2796 return; 2797 } 2798 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 2799 FREE_LOCK(&lk); 2800 vput(vp); 2801} 2802 2803/* 2804 * Inode de-allocation dependencies. 2805 * 2806 * When an inode's link count is reduced to zero, it can be de-allocated. We 2807 * found it convenient to postpone de-allocation until after the inode is 2808 * written to disk with its new link count (zero). At this point, all of the 2809 * on-disk inode's block pointers are nullified and, with careful dependency 2810 * list ordering, all dependencies related to the inode will be satisfied and 2811 * the corresponding dependency structures de-allocated. So, if/when the 2812 * inode is reused, there will be no mixing of old dependencies with new 2813 * ones. This artificial dependency is set up by the block de-allocation 2814 * procedure above (softdep_setup_freeblocks) and completed by the 2815 * following procedure. 2816 */ 2817static void 2818handle_workitem_freefile(freefile) 2819 struct freefile *freefile; 2820{ 2821 struct fs *fs; 2822 struct vnode vp; 2823 struct inode tip; 2824 struct inodedep *idp; 2825 int error; 2826 2827 fs = VFSTOUFS(freefile->fx_mnt)->um_fs; 2828#ifdef DEBUG 2829 ACQUIRE_LOCK(&lk); 2830 if (inodedep_lookup(fs, freefile->fx_oldinum, 0, &idp)) 2831 panic("handle_workitem_freefile: inodedep survived"); 2832 FREE_LOCK(&lk); 2833#endif 2834 tip.i_devvp = freefile->fx_devvp; 2835 tip.i_dev = freefile->fx_devvp->v_rdev; 2836 tip.i_fs = fs; 2837 tip.i_vnode = &vp; 2838 vp.v_data = &tip; 2839 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0) 2840 softdep_error("handle_workitem_freefile", error); 2841 WORKITEM_FREE(freefile, D_FREEFILE); 2842} 2843 2844/* 2845 * Disk writes. 2846 * 2847 * The dependency structures constructed above are most actively used when file 2848 * system blocks are written to disk. No constraints are placed on when a 2849 * block can be written, but unsatisfied update dependencies are made safe by 2850 * modifying (or replacing) the source memory for the duration of the disk 2851 * write. When the disk write completes, the memory block is again brought 2852 * up-to-date. 2853 * 2854 * In-core inode structure reclamation. 2855 * 2856 * Because there are a finite number of "in-core" inode structures, they are 2857 * reused regularly. By transferring all inode-related dependencies to the 2858 * in-memory inode block and indexing them separately (via "inodedep"s), we 2859 * can allow "in-core" inode structures to be reused at any time and avoid 2860 * any increase in contention. 2861 * 2862 * Called just before entering the device driver to initiate a new disk I/O. 2863 * The buffer must be locked, thus, no I/O completion operations can occur 2864 * while we are manipulating its associated dependencies. 2865 */ 2866static void 2867softdep_disk_io_initiation(bp) 2868 struct buf *bp; /* structure describing disk write to occur */ 2869{ 2870 struct worklist *wk, *nextwk; 2871 struct indirdep *indirdep; 2872 2873 /* 2874 * We only care about write operations. There should never 2875 * be dependencies for reads. 2876 */ 2877 if (bp->b_iocmd == BIO_READ) 2878 panic("softdep_disk_io_initiation: read"); 2879 /* 2880 * Do any necessary pre-I/O processing. 2881 */ 2882 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = nextwk) { 2883 nextwk = LIST_NEXT(wk, wk_list); 2884 switch (wk->wk_type) { 2885 2886 case D_PAGEDEP: 2887 initiate_write_filepage(WK_PAGEDEP(wk), bp); 2888 continue; 2889 2890 case D_INODEDEP: 2891 initiate_write_inodeblock(WK_INODEDEP(wk), bp); 2892 continue; 2893 2894 case D_INDIRDEP: 2895 indirdep = WK_INDIRDEP(wk); 2896 if (indirdep->ir_state & GOINGAWAY) 2897 panic("disk_io_initiation: indirdep gone"); 2898 /* 2899 * If there are no remaining dependencies, this 2900 * will be writing the real pointers, so the 2901 * dependency can be freed. 2902 */ 2903 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) { 2904 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 2905 brelse(indirdep->ir_savebp); 2906 /* inline expand WORKLIST_REMOVE(wk); */ 2907 wk->wk_state &= ~ONWORKLIST; 2908 LIST_REMOVE(wk, wk_list); 2909 WORKITEM_FREE(indirdep, D_INDIRDEP); 2910 continue; 2911 } 2912 /* 2913 * Replace up-to-date version with safe version. 2914 */ 2915 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount, 2916 M_INDIRDEP, M_SOFTDEP_FLAGS); 2917 ACQUIRE_LOCK(&lk); 2918 indirdep->ir_state &= ~ATTACHED; 2919 indirdep->ir_state |= UNDONE; 2920 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 2921 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 2922 bp->b_bcount); 2923 FREE_LOCK(&lk); 2924 continue; 2925 2926 case D_MKDIR: 2927 case D_BMSAFEMAP: 2928 case D_ALLOCDIRECT: 2929 case D_ALLOCINDIR: 2930 continue; 2931 2932 default: 2933 panic("handle_disk_io_initiation: Unexpected type %s", 2934 TYPENAME(wk->wk_type)); 2935 /* NOTREACHED */ 2936 } 2937 } 2938} 2939 2940/* 2941 * Called from within the procedure above to deal with unsatisfied 2942 * allocation dependencies in a directory. The buffer must be locked, 2943 * thus, no I/O completion operations can occur while we are 2944 * manipulating its associated dependencies. 2945 */ 2946static void 2947initiate_write_filepage(pagedep, bp) 2948 struct pagedep *pagedep; 2949 struct buf *bp; 2950{ 2951 struct diradd *dap; 2952 struct direct *ep; 2953 int i; 2954 2955 if (pagedep->pd_state & IOSTARTED) { 2956 /* 2957 * This can only happen if there is a driver that does not 2958 * understand chaining. Here biodone will reissue the call 2959 * to strategy for the incomplete buffers. 2960 */ 2961 printf("initiate_write_filepage: already started\n"); 2962 return; 2963 } 2964 pagedep->pd_state |= IOSTARTED; 2965 ACQUIRE_LOCK(&lk); 2966 for (i = 0; i < DAHASHSZ; i++) { 2967 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 2968 dap = LIST_NEXT(dap, da_pdlist)) { 2969 ep = (struct direct *) 2970 ((char *)bp->b_data + dap->da_offset); 2971 if (ep->d_ino != dap->da_newinum) 2972 panic("%s: dir inum %d != new %d", 2973 "initiate_write_filepage", 2974 ep->d_ino, dap->da_newinum); 2975 if (dap->da_state & DIRCHG) 2976 ep->d_ino = dap->da_previous->dm_oldinum; 2977 else 2978 ep->d_ino = 0; 2979 dap->da_state &= ~ATTACHED; 2980 dap->da_state |= UNDONE; 2981 } 2982 } 2983 FREE_LOCK(&lk); 2984} 2985 2986/* 2987 * Called from within the procedure above to deal with unsatisfied 2988 * allocation dependencies in an inodeblock. The buffer must be 2989 * locked, thus, no I/O completion operations can occur while we 2990 * are manipulating its associated dependencies. 2991 */ 2992static void 2993initiate_write_inodeblock(inodedep, bp) 2994 struct inodedep *inodedep; 2995 struct buf *bp; /* The inode block */ 2996{ 2997 struct allocdirect *adp, *lastadp; 2998 struct dinode *dp; 2999 struct fs *fs; 3000 ufs_lbn_t prevlbn = 0; 3001 int i, deplist; 3002 3003 if (inodedep->id_state & IOSTARTED) 3004 panic("initiate_write_inodeblock: already started"); 3005 inodedep->id_state |= IOSTARTED; 3006 fs = inodedep->id_fs; 3007 dp = (struct dinode *)bp->b_data + 3008 ino_to_fsbo(fs, inodedep->id_ino); 3009 /* 3010 * If the bitmap is not yet written, then the allocated 3011 * inode cannot be written to disk. 3012 */ 3013 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 3014 if (inodedep->id_savedino != NULL) 3015 panic("initiate_write_inodeblock: already doing I/O"); 3016 MALLOC(inodedep->id_savedino, struct dinode *, 3017 sizeof(struct dinode), M_INODEDEP, M_SOFTDEP_FLAGS); 3018 *inodedep->id_savedino = *dp; 3019 bzero((caddr_t)dp, sizeof(struct dinode)); 3020 return; 3021 } 3022 /* 3023 * If no dependencies, then there is nothing to roll back. 3024 */ 3025 inodedep->id_savedsize = dp->di_size; 3026 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 3027 return; 3028 /* 3029 * Set the dependencies to busy. 3030 */ 3031 ACQUIRE_LOCK(&lk); 3032 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3033 adp = TAILQ_NEXT(adp, ad_next)) { 3034#ifdef DIAGNOSTIC 3035 if (deplist != 0 && prevlbn >= adp->ad_lbn) 3036 panic("softdep_write_inodeblock: lbn order"); 3037 prevlbn = adp->ad_lbn; 3038 if (adp->ad_lbn < NDADDR && 3039 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) 3040 panic("%s: direct pointer #%ld mismatch %d != %d", 3041 "softdep_write_inodeblock", adp->ad_lbn, 3042 dp->di_db[adp->ad_lbn], adp->ad_newblkno); 3043 if (adp->ad_lbn >= NDADDR && 3044 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) 3045 panic("%s: indirect pointer #%ld mismatch %d != %d", 3046 "softdep_write_inodeblock", adp->ad_lbn - NDADDR, 3047 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno); 3048 deplist |= 1 << adp->ad_lbn; 3049 if ((adp->ad_state & ATTACHED) == 0) 3050 panic("softdep_write_inodeblock: Unknown state 0x%x", 3051 adp->ad_state); 3052#endif /* DIAGNOSTIC */ 3053 adp->ad_state &= ~ATTACHED; 3054 adp->ad_state |= UNDONE; 3055 } 3056 /* 3057 * The on-disk inode cannot claim to be any larger than the last 3058 * fragment that has been written. Otherwise, the on-disk inode 3059 * might have fragments that were not the last block in the file 3060 * which would corrupt the filesystem. 3061 */ 3062 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3063 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 3064 if (adp->ad_lbn >= NDADDR) 3065 break; 3066 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 3067 /* keep going until hitting a rollback to a frag */ 3068 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 3069 continue; 3070 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 3071 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 3072#ifdef DIAGNOSTIC 3073 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 3074 panic("softdep_write_inodeblock: lost dep1"); 3075#endif /* DIAGNOSTIC */ 3076 dp->di_db[i] = 0; 3077 } 3078 for (i = 0; i < NIADDR; i++) { 3079#ifdef DIAGNOSTIC 3080 if (dp->di_ib[i] != 0 && 3081 (deplist & ((1 << NDADDR) << i)) == 0) 3082 panic("softdep_write_inodeblock: lost dep2"); 3083#endif /* DIAGNOSTIC */ 3084 dp->di_ib[i] = 0; 3085 } 3086 FREE_LOCK(&lk); 3087 return; 3088 } 3089 /* 3090 * If we have zero'ed out the last allocated block of the file, 3091 * roll back the size to the last currently allocated block. 3092 * We know that this last allocated block is a full-sized as 3093 * we already checked for fragments in the loop above. 3094 */ 3095 if (lastadp != NULL && 3096 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 3097 for (i = lastadp->ad_lbn; i >= 0; i--) 3098 if (dp->di_db[i] != 0) 3099 break; 3100 dp->di_size = (i + 1) * fs->fs_bsize; 3101 } 3102 /* 3103 * The only dependencies are for indirect blocks. 3104 * 3105 * The file size for indirect block additions is not guaranteed. 3106 * Such a guarantee would be non-trivial to achieve. The conventional 3107 * synchronous write implementation also does not make this guarantee. 3108 * Fsck should catch and fix discrepancies. Arguably, the file size 3109 * can be over-estimated without destroying integrity when the file 3110 * moves into the indirect blocks (i.e., is large). If we want to 3111 * postpone fsck, we are stuck with this argument. 3112 */ 3113 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 3114 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 3115 FREE_LOCK(&lk); 3116} 3117 3118/* 3119 * This routine is called during the completion interrupt 3120 * service routine for a disk write (from the procedure called 3121 * by the device driver to inform the file system caches of 3122 * a request completion). It should be called early in this 3123 * procedure, before the block is made available to other 3124 * processes or other routines are called. 3125 */ 3126static void 3127softdep_disk_write_complete(bp) 3128 struct buf *bp; /* describes the completed disk write */ 3129{ 3130 struct worklist *wk; 3131 struct workhead reattach; 3132 struct newblk *newblk; 3133 struct allocindir *aip; 3134 struct allocdirect *adp; 3135 struct indirdep *indirdep; 3136 struct inodedep *inodedep; 3137 struct bmsafemap *bmsafemap; 3138 3139#ifdef DEBUG 3140 if (lk.lkt_held != -1) 3141 panic("softdep_disk_write_complete: lock is held"); 3142 lk.lkt_held = -2; 3143#endif 3144 LIST_INIT(&reattach); 3145 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 3146 WORKLIST_REMOVE(wk); 3147 switch (wk->wk_type) { 3148 3149 case D_PAGEDEP: 3150 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 3151 WORKLIST_INSERT(&reattach, wk); 3152 continue; 3153 3154 case D_INODEDEP: 3155 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 3156 WORKLIST_INSERT(&reattach, wk); 3157 continue; 3158 3159 case D_BMSAFEMAP: 3160 bmsafemap = WK_BMSAFEMAP(wk); 3161 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) { 3162 newblk->nb_state |= DEPCOMPLETE; 3163 newblk->nb_bmsafemap = NULL; 3164 LIST_REMOVE(newblk, nb_deps); 3165 } 3166 while ((adp = 3167 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) { 3168 adp->ad_state |= DEPCOMPLETE; 3169 adp->ad_buf = NULL; 3170 LIST_REMOVE(adp, ad_deps); 3171 handle_allocdirect_partdone(adp); 3172 } 3173 while ((aip = 3174 LIST_FIRST(&bmsafemap->sm_allocindirhd))) { 3175 aip->ai_state |= DEPCOMPLETE; 3176 aip->ai_buf = NULL; 3177 LIST_REMOVE(aip, ai_deps); 3178 handle_allocindir_partdone(aip); 3179 } 3180 while ((inodedep = 3181 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { 3182 inodedep->id_state |= DEPCOMPLETE; 3183 LIST_REMOVE(inodedep, id_deps); 3184 inodedep->id_buf = NULL; 3185 } 3186 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 3187 continue; 3188 3189 case D_MKDIR: 3190 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 3191 continue; 3192 3193 case D_ALLOCDIRECT: 3194 adp = WK_ALLOCDIRECT(wk); 3195 adp->ad_state |= COMPLETE; 3196 handle_allocdirect_partdone(adp); 3197 continue; 3198 3199 case D_ALLOCINDIR: 3200 aip = WK_ALLOCINDIR(wk); 3201 aip->ai_state |= COMPLETE; 3202 handle_allocindir_partdone(aip); 3203 continue; 3204 3205 case D_INDIRDEP: 3206 indirdep = WK_INDIRDEP(wk); 3207 if (indirdep->ir_state & GOINGAWAY) 3208 panic("disk_write_complete: indirdep gone"); 3209 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 3210 FREE(indirdep->ir_saveddata, M_INDIRDEP); 3211 indirdep->ir_saveddata = 0; 3212 indirdep->ir_state &= ~UNDONE; 3213 indirdep->ir_state |= ATTACHED; 3214 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 3215 handle_allocindir_partdone(aip); 3216 if (aip == LIST_FIRST(&indirdep->ir_donehd)) 3217 panic("disk_write_complete: not gone"); 3218 } 3219 WORKLIST_INSERT(&reattach, wk); 3220 if ((bp->b_flags & B_DELWRI) == 0) 3221 stat_indir_blk_ptrs++; 3222 bdirty(bp); 3223 continue; 3224 3225 default: 3226 panic("handle_disk_write_complete: Unknown type %s", 3227 TYPENAME(wk->wk_type)); 3228 /* NOTREACHED */ 3229 } 3230 } 3231 /* 3232 * Reattach any requests that must be redone. 3233 */ 3234 while ((wk = LIST_FIRST(&reattach)) != NULL) { 3235 WORKLIST_REMOVE(wk); 3236 WORKLIST_INSERT(&bp->b_dep, wk); 3237 } 3238#ifdef DEBUG 3239 if (lk.lkt_held != -2) 3240 panic("softdep_disk_write_complete: lock lost"); 3241 lk.lkt_held = -1; 3242#endif 3243} 3244 3245/* 3246 * Called from within softdep_disk_write_complete above. Note that 3247 * this routine is always called from interrupt level with further 3248 * splbio interrupts blocked. 3249 */ 3250static void 3251handle_allocdirect_partdone(adp) 3252 struct allocdirect *adp; /* the completed allocdirect */ 3253{ 3254 struct allocdirect *listadp; 3255 struct inodedep *inodedep; 3256 long bsize, delay; 3257 3258 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3259 return; 3260 if (adp->ad_buf != NULL) 3261 panic("handle_allocdirect_partdone: dangling dep"); 3262 /* 3263 * The on-disk inode cannot claim to be any larger than the last 3264 * fragment that has been written. Otherwise, the on-disk inode 3265 * might have fragments that were not the last block in the file 3266 * which would corrupt the filesystem. Thus, we cannot free any 3267 * allocdirects after one whose ad_oldblkno claims a fragment as 3268 * these blocks must be rolled back to zero before writing the inode. 3269 * We check the currently active set of allocdirects in id_inoupdt. 3270 */ 3271 inodedep = adp->ad_inodedep; 3272 bsize = inodedep->id_fs->fs_bsize; 3273 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp; 3274 listadp = TAILQ_NEXT(listadp, ad_next)) { 3275 /* found our block */ 3276 if (listadp == adp) 3277 break; 3278 /* continue if ad_oldlbn is not a fragment */ 3279 if (listadp->ad_oldsize == 0 || 3280 listadp->ad_oldsize == bsize) 3281 continue; 3282 /* hit a fragment */ 3283 return; 3284 } 3285 /* 3286 * If we have reached the end of the current list without 3287 * finding the just finished dependency, then it must be 3288 * on the future dependency list. Future dependencies cannot 3289 * be freed until they are moved to the current list. 3290 */ 3291 if (listadp == NULL) { 3292#ifdef DEBUG 3293 for (listadp = TAILQ_FIRST(&inodedep->id_newinoupdt); listadp; 3294 listadp = TAILQ_NEXT(listadp, ad_next)) 3295 /* found our block */ 3296 if (listadp == adp) 3297 break; 3298 if (listadp == NULL) 3299 panic("handle_allocdirect_partdone: lost dep"); 3300#endif /* DEBUG */ 3301 return; 3302 } 3303 /* 3304 * If we have found the just finished dependency, then free 3305 * it along with anything that follows it that is complete. 3306 * If the inode still has a bitmap dependency, then it has 3307 * never been written to disk, hence the on-disk inode cannot 3308 * reference the old fragment so we can free it without delay. 3309 */ 3310 delay = (inodedep->id_state & DEPCOMPLETE); 3311 for (; adp; adp = listadp) { 3312 listadp = TAILQ_NEXT(adp, ad_next); 3313 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3314 return; 3315 free_allocdirect(&inodedep->id_inoupdt, adp, delay); 3316 } 3317} 3318 3319/* 3320 * Called from within softdep_disk_write_complete above. Note that 3321 * this routine is always called from interrupt level with further 3322 * splbio interrupts blocked. 3323 */ 3324static void 3325handle_allocindir_partdone(aip) 3326 struct allocindir *aip; /* the completed allocindir */ 3327{ 3328 struct indirdep *indirdep; 3329 3330 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 3331 return; 3332 if (aip->ai_buf != NULL) 3333 panic("handle_allocindir_partdone: dangling dependency"); 3334 indirdep = aip->ai_indirdep; 3335 if (indirdep->ir_state & UNDONE) { 3336 LIST_REMOVE(aip, ai_next); 3337 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 3338 return; 3339 } 3340 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 3341 aip->ai_newblkno; 3342 LIST_REMOVE(aip, ai_next); 3343 if (aip->ai_freefrag != NULL) 3344 add_to_worklist(&aip->ai_freefrag->ff_list); 3345 WORKITEM_FREE(aip, D_ALLOCINDIR); 3346} 3347 3348/* 3349 * Called from within softdep_disk_write_complete above to restore 3350 * in-memory inode block contents to their most up-to-date state. Note 3351 * that this routine is always called from interrupt level with further 3352 * splbio interrupts blocked. 3353 */ 3354static int 3355handle_written_inodeblock(inodedep, bp) 3356 struct inodedep *inodedep; 3357 struct buf *bp; /* buffer containing the inode block */ 3358{ 3359 struct worklist *wk, *filefree; 3360 struct allocdirect *adp, *nextadp; 3361 struct dinode *dp; 3362 int hadchanges; 3363 3364 if ((inodedep->id_state & IOSTARTED) == 0) 3365 panic("handle_written_inodeblock: not started"); 3366 inodedep->id_state &= ~IOSTARTED; 3367 inodedep->id_state |= COMPLETE; 3368 dp = (struct dinode *)bp->b_data + 3369 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 3370 /* 3371 * If we had to rollback the inode allocation because of 3372 * bitmaps being incomplete, then simply restore it. 3373 * Keep the block dirty so that it will not be reclaimed until 3374 * all associated dependencies have been cleared and the 3375 * corresponding updates written to disk. 3376 */ 3377 if (inodedep->id_savedino != NULL) { 3378 *dp = *inodedep->id_savedino; 3379 FREE(inodedep->id_savedino, M_INODEDEP); 3380 inodedep->id_savedino = NULL; 3381 if ((bp->b_flags & B_DELWRI) == 0) 3382 stat_inode_bitmap++; 3383 bdirty(bp); 3384 return (1); 3385 } 3386 /* 3387 * Roll forward anything that had to be rolled back before 3388 * the inode could be updated. 3389 */ 3390 hadchanges = 0; 3391 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 3392 nextadp = TAILQ_NEXT(adp, ad_next); 3393 if (adp->ad_state & ATTACHED) 3394 panic("handle_written_inodeblock: new entry"); 3395 if (adp->ad_lbn < NDADDR) { 3396 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) 3397 panic("%s: %s #%ld mismatch %d != %d", 3398 "handle_written_inodeblock", 3399 "direct pointer", adp->ad_lbn, 3400 dp->di_db[adp->ad_lbn], adp->ad_oldblkno); 3401 dp->di_db[adp->ad_lbn] = adp->ad_newblkno; 3402 } else { 3403 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) 3404 panic("%s: %s #%ld allocated as %d", 3405 "handle_written_inodeblock", 3406 "indirect pointer", adp->ad_lbn - NDADDR, 3407 dp->di_ib[adp->ad_lbn - NDADDR]); 3408 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno; 3409 } 3410 adp->ad_state &= ~UNDONE; 3411 adp->ad_state |= ATTACHED; 3412 hadchanges = 1; 3413 } 3414 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 3415 stat_direct_blk_ptrs++; 3416 /* 3417 * Reset the file size to its most up-to-date value. 3418 */ 3419 if (inodedep->id_savedsize == -1) 3420 panic("handle_written_inodeblock: bad size"); 3421 if (dp->di_size != inodedep->id_savedsize) { 3422 dp->di_size = inodedep->id_savedsize; 3423 hadchanges = 1; 3424 } 3425 inodedep->id_savedsize = -1; 3426 /* 3427 * If there were any rollbacks in the inode block, then it must be 3428 * marked dirty so that its will eventually get written back in 3429 * its correct form. 3430 */ 3431 if (hadchanges) 3432 bdirty(bp); 3433 /* 3434 * Process any allocdirects that completed during the update. 3435 */ 3436 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 3437 handle_allocdirect_partdone(adp); 3438 /* 3439 * Process deallocations that were held pending until the 3440 * inode had been written to disk. Freeing of the inode 3441 * is delayed until after all blocks have been freed to 3442 * avoid creation of new <vfsid, inum, lbn> triples 3443 * before the old ones have been deleted. 3444 */ 3445 filefree = NULL; 3446 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 3447 WORKLIST_REMOVE(wk); 3448 switch (wk->wk_type) { 3449 3450 case D_FREEFILE: 3451 /* 3452 * We defer adding filefree to the worklist until 3453 * all other additions have been made to ensure 3454 * that it will be done after all the old blocks 3455 * have been freed. 3456 */ 3457 if (filefree != NULL) 3458 panic("handle_written_inodeblock: filefree"); 3459 filefree = wk; 3460 continue; 3461 3462 case D_MKDIR: 3463 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 3464 continue; 3465 3466 case D_DIRADD: 3467 diradd_inode_written(WK_DIRADD(wk), inodedep); 3468 continue; 3469 3470 case D_FREEBLKS: 3471 case D_FREEFRAG: 3472 case D_DIRREM: 3473 add_to_worklist(wk); 3474 continue; 3475 3476 default: 3477 panic("handle_written_inodeblock: Unknown type %s", 3478 TYPENAME(wk->wk_type)); 3479 /* NOTREACHED */ 3480 } 3481 } 3482 if (filefree != NULL) { 3483 if (free_inodedep(inodedep) == 0) 3484 panic("handle_written_inodeblock: live inodedep"); 3485 add_to_worklist(filefree); 3486 return (0); 3487 } 3488 3489 /* 3490 * If no outstanding dependencies, free it. 3491 */ 3492 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0) 3493 return (0); 3494 return (hadchanges); 3495} 3496 3497/* 3498 * Process a diradd entry after its dependent inode has been written. 3499 * This routine must be called with splbio interrupts blocked. 3500 */ 3501static void 3502diradd_inode_written(dap, inodedep) 3503 struct diradd *dap; 3504 struct inodedep *inodedep; 3505{ 3506 struct pagedep *pagedep; 3507 3508 dap->da_state |= COMPLETE; 3509 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3510 if (dap->da_state & DIRCHG) 3511 pagedep = dap->da_previous->dm_pagedep; 3512 else 3513 pagedep = dap->da_pagedep; 3514 LIST_REMOVE(dap, da_pdlist); 3515 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3516 } 3517 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 3518} 3519 3520/* 3521 * Handle the completion of a mkdir dependency. 3522 */ 3523static void 3524handle_written_mkdir(mkdir, type) 3525 struct mkdir *mkdir; 3526 int type; 3527{ 3528 struct diradd *dap; 3529 struct pagedep *pagedep; 3530 3531 if (mkdir->md_state != type) 3532 panic("handle_written_mkdir: bad type"); 3533 dap = mkdir->md_diradd; 3534 dap->da_state &= ~type; 3535 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 3536 dap->da_state |= DEPCOMPLETE; 3537 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3538 if (dap->da_state & DIRCHG) 3539 pagedep = dap->da_previous->dm_pagedep; 3540 else 3541 pagedep = dap->da_pagedep; 3542 LIST_REMOVE(dap, da_pdlist); 3543 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3544 } 3545 LIST_REMOVE(mkdir, md_mkdirs); 3546 WORKITEM_FREE(mkdir, D_MKDIR); 3547} 3548 3549/* 3550 * Called from within softdep_disk_write_complete above. 3551 * A write operation was just completed. Removed inodes can 3552 * now be freed and associated block pointers may be committed. 3553 * Note that this routine is always called from interrupt level 3554 * with further splbio interrupts blocked. 3555 */ 3556static int 3557handle_written_filepage(pagedep, bp) 3558 struct pagedep *pagedep; 3559 struct buf *bp; /* buffer containing the written page */ 3560{ 3561 struct dirrem *dirrem; 3562 struct diradd *dap, *nextdap; 3563 struct direct *ep; 3564 int i, chgs; 3565 3566 if ((pagedep->pd_state & IOSTARTED) == 0) 3567 panic("handle_written_filepage: not started"); 3568 pagedep->pd_state &= ~IOSTARTED; 3569 /* 3570 * Process any directory removals that have been committed. 3571 */ 3572 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 3573 LIST_REMOVE(dirrem, dm_next); 3574 dirrem->dm_dirinum = pagedep->pd_ino; 3575 add_to_worklist(&dirrem->dm_list); 3576 } 3577 /* 3578 * Free any directory additions that have been committed. 3579 */ 3580 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 3581 free_diradd(dap); 3582 /* 3583 * Uncommitted directory entries must be restored. 3584 */ 3585 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 3586 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 3587 dap = nextdap) { 3588 nextdap = LIST_NEXT(dap, da_pdlist); 3589 if (dap->da_state & ATTACHED) 3590 panic("handle_written_filepage: attached"); 3591 ep = (struct direct *) 3592 ((char *)bp->b_data + dap->da_offset); 3593 ep->d_ino = dap->da_newinum; 3594 dap->da_state &= ~UNDONE; 3595 dap->da_state |= ATTACHED; 3596 chgs = 1; 3597 /* 3598 * If the inode referenced by the directory has 3599 * been written out, then the dependency can be 3600 * moved to the pending list. 3601 */ 3602 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3603 LIST_REMOVE(dap, da_pdlist); 3604 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 3605 da_pdlist); 3606 } 3607 } 3608 } 3609 /* 3610 * If there were any rollbacks in the directory, then it must be 3611 * marked dirty so that its will eventually get written back in 3612 * its correct form. 3613 */ 3614 if (chgs) { 3615 if ((bp->b_flags & B_DELWRI) == 0) 3616 stat_dir_entry++; 3617 bdirty(bp); 3618 } 3619 /* 3620 * If no dependencies remain, the pagedep will be freed. 3621 * Otherwise it will remain to update the page before it 3622 * is written back to disk. 3623 */ 3624 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) { 3625 for (i = 0; i < DAHASHSZ; i++) 3626 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL) 3627 break; 3628 if (i == DAHASHSZ) { 3629 LIST_REMOVE(pagedep, pd_hash); 3630 WORKITEM_FREE(pagedep, D_PAGEDEP); 3631 return (0); 3632 } 3633 } 3634 return (1); 3635} 3636 3637/* 3638 * Writing back in-core inode structures. 3639 * 3640 * The file system only accesses an inode's contents when it occupies an 3641 * "in-core" inode structure. These "in-core" structures are separate from 3642 * the page frames used to cache inode blocks. Only the latter are 3643 * transferred to/from the disk. So, when the updated contents of the 3644 * "in-core" inode structure are copied to the corresponding in-memory inode 3645 * block, the dependencies are also transferred. The following procedure is 3646 * called when copying a dirty "in-core" inode to a cached inode block. 3647 */ 3648 3649/* 3650 * Called when an inode is loaded from disk. If the effective link count 3651 * differed from the actual link count when it was last flushed, then we 3652 * need to ensure that the correct effective link count is put back. 3653 */ 3654void 3655softdep_load_inodeblock(ip) 3656 struct inode *ip; /* the "in_core" copy of the inode */ 3657{ 3658 struct inodedep *inodedep; 3659 3660 /* 3661 * Check for alternate nlink count. 3662 */ 3663 ip->i_effnlink = ip->i_nlink; 3664 ACQUIRE_LOCK(&lk); 3665 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3666 FREE_LOCK(&lk); 3667 return; 3668 } 3669 ip->i_effnlink -= inodedep->id_nlinkdelta; 3670 FREE_LOCK(&lk); 3671} 3672 3673/* 3674 * This routine is called just before the "in-core" inode 3675 * information is to be copied to the in-memory inode block. 3676 * Recall that an inode block contains several inodes. If 3677 * the force flag is set, then the dependencies will be 3678 * cleared so that the update can always be made. Note that 3679 * the buffer is locked when this routine is called, so we 3680 * will never be in the middle of writing the inode block 3681 * to disk. 3682 */ 3683void 3684softdep_update_inodeblock(ip, bp, waitfor) 3685 struct inode *ip; /* the "in_core" copy of the inode */ 3686 struct buf *bp; /* the buffer containing the inode block */ 3687 int waitfor; /* nonzero => update must be allowed */ 3688{ 3689 struct inodedep *inodedep; 3690 struct worklist *wk; 3691 int error, gotit; 3692 3693 /* 3694 * If the effective link count is not equal to the actual link 3695 * count, then we must track the difference in an inodedep while 3696 * the inode is (potentially) tossed out of the cache. Otherwise, 3697 * if there is no existing inodedep, then there are no dependencies 3698 * to track. 3699 */ 3700 ACQUIRE_LOCK(&lk); 3701 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3702 if (ip->i_effnlink != ip->i_nlink) 3703 panic("softdep_update_inodeblock: bad link count"); 3704 FREE_LOCK(&lk); 3705 return; 3706 } 3707 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) 3708 panic("softdep_update_inodeblock: bad delta"); 3709 /* 3710 * Changes have been initiated. Anything depending on these 3711 * changes cannot occur until this inode has been written. 3712 */ 3713 inodedep->id_state &= ~COMPLETE; 3714 if ((inodedep->id_state & ONWORKLIST) == 0) 3715 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 3716 /* 3717 * Any new dependencies associated with the incore inode must 3718 * now be moved to the list associated with the buffer holding 3719 * the in-memory copy of the inode. Once merged process any 3720 * allocdirects that are completed by the merger. 3721 */ 3722 merge_inode_lists(inodedep); 3723 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL) 3724 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); 3725 /* 3726 * Now that the inode has been pushed into the buffer, the 3727 * operations dependent on the inode being written to disk 3728 * can be moved to the id_bufwait so that they will be 3729 * processed when the buffer I/O completes. 3730 */ 3731 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 3732 WORKLIST_REMOVE(wk); 3733 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 3734 } 3735 /* 3736 * Newly allocated inodes cannot be written until the bitmap 3737 * that allocates them have been written (indicated by 3738 * DEPCOMPLETE being set in id_state). If we are doing a 3739 * forced sync (e.g., an fsync on a file), we force the bitmap 3740 * to be written so that the update can be done. 3741 */ 3742 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) { 3743 FREE_LOCK(&lk); 3744 return; 3745 } 3746 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 3747 FREE_LOCK(&lk); 3748 if (gotit && 3749 (error = BUF_WRITE(inodedep->id_buf)) != 0) 3750 softdep_error("softdep_update_inodeblock: bwrite", error); 3751 if ((inodedep->id_state & DEPCOMPLETE) == 0) 3752 panic("softdep_update_inodeblock: update failed"); 3753} 3754 3755/* 3756 * Merge the new inode dependency list (id_newinoupdt) into the old 3757 * inode dependency list (id_inoupdt). This routine must be called 3758 * with splbio interrupts blocked. 3759 */ 3760static void 3761merge_inode_lists(inodedep) 3762 struct inodedep *inodedep; 3763{ 3764 struct allocdirect *listadp, *newadp; 3765 3766 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 3767 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) { 3768 if (listadp->ad_lbn < newadp->ad_lbn) { 3769 listadp = TAILQ_NEXT(listadp, ad_next); 3770 continue; 3771 } 3772 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 3773 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 3774 if (listadp->ad_lbn == newadp->ad_lbn) { 3775 allocdirect_merge(&inodedep->id_inoupdt, newadp, 3776 listadp); 3777 listadp = newadp; 3778 } 3779 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 3780 } 3781 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) { 3782 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 3783 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next); 3784 } 3785} 3786 3787/* 3788 * If we are doing an fsync, then we must ensure that any directory 3789 * entries for the inode have been written after the inode gets to disk. 3790 */ 3791int 3792softdep_fsync(vp) 3793 struct vnode *vp; /* the "in_core" copy of the inode */ 3794{ 3795 struct inodedep *inodedep; 3796 struct pagedep *pagedep; 3797 struct worklist *wk; 3798 struct diradd *dap; 3799 struct mount *mnt; 3800 struct vnode *pvp; 3801 struct inode *ip; 3802 struct buf *bp; 3803 struct fs *fs; 3804 struct proc *p = CURPROC; /* XXX */ 3805 int error, flushparent; 3806 ino_t parentino; 3807 ufs_lbn_t lbn; 3808 3809 ip = VTOI(vp); 3810 fs = ip->i_fs; 3811 ACQUIRE_LOCK(&lk); 3812 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) { 3813 FREE_LOCK(&lk); 3814 return (0); 3815 } 3816 if (LIST_FIRST(&inodedep->id_inowait) != NULL || 3817 LIST_FIRST(&inodedep->id_bufwait) != NULL || 3818 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 3819 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) 3820 panic("softdep_fsync: pending ops"); 3821 for (error = 0, flushparent = 0; ; ) { 3822 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 3823 break; 3824 if (wk->wk_type != D_DIRADD) 3825 panic("softdep_fsync: Unexpected type %s", 3826 TYPENAME(wk->wk_type)); 3827 dap = WK_DIRADD(wk); 3828 /* 3829 * Flush our parent if this directory entry 3830 * has a MKDIR_PARENT dependency. 3831 */ 3832 if (dap->da_state & DIRCHG) 3833 pagedep = dap->da_previous->dm_pagedep; 3834 else 3835 pagedep = dap->da_pagedep; 3836 mnt = pagedep->pd_mnt; 3837 parentino = pagedep->pd_ino; 3838 lbn = pagedep->pd_lbn; 3839 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) 3840 panic("softdep_fsync: dirty"); 3841 flushparent = dap->da_state & MKDIR_PARENT; 3842 /* 3843 * If we are being fsync'ed as part of vgone'ing this vnode, 3844 * then we will not be able to release and recover the 3845 * vnode below, so we just have to give up on writing its 3846 * directory entry out. It will eventually be written, just 3847 * not now, but then the user was not asking to have it 3848 * written, so we are not breaking any promises. 3849 */ 3850 if (vp->v_flag & VXLOCK) 3851 break; 3852 /* 3853 * We prevent deadlock by always fetching inodes from the 3854 * root, moving down the directory tree. Thus, when fetching 3855 * our parent directory, we must unlock ourselves before 3856 * requesting the lock on our parent. See the comment in 3857 * ufs_lookup for details on possible races. 3858 */ 3859 FREE_LOCK(&lk); 3860 VOP_UNLOCK(vp, 0, p); 3861 error = VFS_VGET(mnt, parentino, &pvp); 3862 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, p); 3863 if (error != 0) 3864 return (error); 3865 if (flushparent) { 3866 if ((error = UFS_UPDATE(pvp, 1)) != 0) { 3867 vput(pvp); 3868 return (error); 3869 } 3870 } 3871 /* 3872 * Flush directory page containing the inode's name. 3873 */ 3874 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), p->p_ucred, 3875 &bp); 3876 if (error == 0) 3877 error = BUF_WRITE(bp); 3878 vput(pvp); 3879 if (error != 0) 3880 return (error); 3881 ACQUIRE_LOCK(&lk); 3882 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) 3883 break; 3884 } 3885 FREE_LOCK(&lk); 3886 return (0); 3887} 3888 3889/* 3890 * Flush all the dirty bitmaps associated with the block device 3891 * before flushing the rest of the dirty blocks so as to reduce 3892 * the number of dependencies that will have to be rolled back. 3893 */ 3894void 3895softdep_fsync_mountdev(vp) 3896 struct vnode *vp; 3897{ 3898 struct buf *bp, *nbp; 3899 struct worklist *wk; 3900 3901 if (!vn_isdisk(vp, NULL)) 3902 panic("softdep_fsync_mountdev: vnode not a disk"); 3903 ACQUIRE_LOCK(&lk); 3904 for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 3905 nbp = TAILQ_NEXT(bp, b_vnbufs); 3906 /* 3907 * If it is already scheduled, skip to the next buffer. 3908 */ 3909 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) 3910 continue; 3911 if ((bp->b_flags & B_DELWRI) == 0) 3912 panic("softdep_fsync_mountdev: not dirty"); 3913 /* 3914 * We are only interested in bitmaps with outstanding 3915 * dependencies. 3916 */ 3917 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 3918 wk->wk_type != D_BMSAFEMAP || 3919 (bp->b_xflags & BX_BKGRDINPROG)) { 3920 BUF_UNLOCK(bp); 3921 continue; 3922 } 3923 bremfree(bp); 3924 FREE_LOCK(&lk); 3925 (void) bawrite(bp); 3926 ACQUIRE_LOCK(&lk); 3927 /* 3928 * Since we may have slept during the I/O, we need 3929 * to start from a known point. 3930 */ 3931 nbp = TAILQ_FIRST(&vp->v_dirtyblkhd); 3932 } 3933 drain_output(vp, 1); 3934 FREE_LOCK(&lk); 3935} 3936 3937/* 3938 * This routine is called when we are trying to synchronously flush a 3939 * file. This routine must eliminate any filesystem metadata dependencies 3940 * so that the syncing routine can succeed by pushing the dirty blocks 3941 * associated with the file. If any I/O errors occur, they are returned. 3942 */ 3943int 3944softdep_sync_metadata(ap) 3945 struct vop_fsync_args /* { 3946 struct vnode *a_vp; 3947 struct ucred *a_cred; 3948 int a_waitfor; 3949 struct proc *a_p; 3950 } */ *ap; 3951{ 3952 struct vnode *vp = ap->a_vp; 3953 struct pagedep *pagedep; 3954 struct allocdirect *adp; 3955 struct allocindir *aip; 3956 struct buf *bp, *nbp; 3957 struct worklist *wk; 3958 int i, error, waitfor; 3959 3960 /* 3961 * Check whether this vnode is involved in a filesystem 3962 * that is doing soft dependency processing. 3963 */ 3964 if (!vn_isdisk(vp, NULL)) { 3965 if (!DOINGSOFTDEP(vp)) 3966 return (0); 3967 } else 3968 if (vp->v_rdev->si_mountpoint == NULL || 3969 (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0) 3970 return (0); 3971 /* 3972 * Ensure that any direct block dependencies have been cleared. 3973 */ 3974 ACQUIRE_LOCK(&lk); 3975 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) { 3976 FREE_LOCK(&lk); 3977 return (error); 3978 } 3979 /* 3980 * For most files, the only metadata dependencies are the 3981 * cylinder group maps that allocate their inode or blocks. 3982 * The block allocation dependencies can be found by traversing 3983 * the dependency lists for any buffers that remain on their 3984 * dirty buffer list. The inode allocation dependency will 3985 * be resolved when the inode is updated with MNT_WAIT. 3986 * This work is done in two passes. The first pass grabs most 3987 * of the buffers and begins asynchronously writing them. The 3988 * only way to wait for these asynchronous writes is to sleep 3989 * on the filesystem vnode which may stay busy for a long time 3990 * if the filesystem is active. So, instead, we make a second 3991 * pass over the dependencies blocking on each write. In the 3992 * usual case we will be blocking against a write that we 3993 * initiated, so when it is done the dependency will have been 3994 * resolved. Thus the second pass is expected to end quickly. 3995 */ 3996 waitfor = MNT_NOWAIT; 3997top: 3998 if (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT) == 0) { 3999 FREE_LOCK(&lk); 4000 return (0); 4001 } 4002 bp = TAILQ_FIRST(&vp->v_dirtyblkhd); 4003loop: 4004 /* 4005 * As we hold the buffer locked, none of its dependencies 4006 * will disappear. 4007 */ 4008 for (wk = LIST_FIRST(&bp->b_dep); wk; 4009 wk = LIST_NEXT(wk, wk_list)) { 4010 switch (wk->wk_type) { 4011 4012 case D_ALLOCDIRECT: 4013 adp = WK_ALLOCDIRECT(wk); 4014 if (adp->ad_state & DEPCOMPLETE) 4015 break; 4016 nbp = adp->ad_buf; 4017 if (getdirtybuf(&nbp, waitfor) == 0) 4018 break; 4019 FREE_LOCK(&lk); 4020 if (waitfor == MNT_NOWAIT) { 4021 bawrite(nbp); 4022 } else if ((error = BUF_WRITE(nbp)) != 0) { 4023 bawrite(bp); 4024 return (error); 4025 } 4026 ACQUIRE_LOCK(&lk); 4027 break; 4028 4029 case D_ALLOCINDIR: 4030 aip = WK_ALLOCINDIR(wk); 4031 if (aip->ai_state & DEPCOMPLETE) 4032 break; 4033 nbp = aip->ai_buf; 4034 if (getdirtybuf(&nbp, waitfor) == 0) 4035 break; 4036 FREE_LOCK(&lk); 4037 if (waitfor == MNT_NOWAIT) { 4038 bawrite(nbp); 4039 } else if ((error = BUF_WRITE(nbp)) != 0) { 4040 bawrite(bp); 4041 return (error); 4042 } 4043 ACQUIRE_LOCK(&lk); 4044 break; 4045 4046 case D_INDIRDEP: 4047 restart: 4048 for (aip = LIST_FIRST(&WK_INDIRDEP(wk)->ir_deplisthd); 4049 aip; aip = LIST_NEXT(aip, ai_next)) { 4050 if (aip->ai_state & DEPCOMPLETE) 4051 continue; 4052 nbp = aip->ai_buf; 4053 if (getdirtybuf(&nbp, MNT_WAIT) == 0) 4054 goto restart; 4055 FREE_LOCK(&lk); 4056 if ((error = BUF_WRITE(nbp)) != 0) { 4057 bawrite(bp); 4058 return (error); 4059 } 4060 ACQUIRE_LOCK(&lk); 4061 goto restart; 4062 } 4063 break; 4064 4065 case D_INODEDEP: 4066 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs, 4067 WK_INODEDEP(wk)->id_ino)) != 0) { 4068 FREE_LOCK(&lk); 4069 bawrite(bp); 4070 return (error); 4071 } 4072 break; 4073 4074 case D_PAGEDEP: 4075 /* 4076 * We are trying to sync a directory that may 4077 * have dependencies on both its own metadata 4078 * and/or dependencies on the inodes of any 4079 * recently allocated files. We walk its diradd 4080 * lists pushing out the associated inode. 4081 */ 4082 pagedep = WK_PAGEDEP(wk); 4083 for (i = 0; i < DAHASHSZ; i++) { 4084 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 4085 continue; 4086 if ((error = 4087 flush_pagedep_deps(vp, pagedep->pd_mnt, 4088 &pagedep->pd_diraddhd[i]))) { 4089 FREE_LOCK(&lk); 4090 bawrite(bp); 4091 return (error); 4092 } 4093 } 4094 break; 4095 4096 case D_MKDIR: 4097 /* 4098 * This case should never happen if the vnode has 4099 * been properly sync'ed. However, if this function 4100 * is used at a place where the vnode has not yet 4101 * been sync'ed, this dependency can show up. So, 4102 * rather than panic, just flush it. 4103 */ 4104 nbp = WK_MKDIR(wk)->md_buf; 4105 if (getdirtybuf(&nbp, waitfor) == 0) 4106 break; 4107 FREE_LOCK(&lk); 4108 if (waitfor == MNT_NOWAIT) { 4109 bawrite(nbp); 4110 } else if ((error = BUF_WRITE(nbp)) != 0) { 4111 bawrite(bp); 4112 return (error); 4113 } 4114 ACQUIRE_LOCK(&lk); 4115 break; 4116 4117 case D_BMSAFEMAP: 4118 /* 4119 * This case should never happen if the vnode has 4120 * been properly sync'ed. However, if this function 4121 * is used at a place where the vnode has not yet 4122 * been sync'ed, this dependency can show up. So, 4123 * rather than panic, just flush it. 4124 */ 4125 nbp = WK_BMSAFEMAP(wk)->sm_buf; 4126 if (getdirtybuf(&nbp, waitfor) == 0) 4127 break; 4128 FREE_LOCK(&lk); 4129 if (waitfor == MNT_NOWAIT) { 4130 bawrite(nbp); 4131 } else if ((error = BUF_WRITE(nbp)) != 0) { 4132 bawrite(bp); 4133 return (error); 4134 } 4135 ACQUIRE_LOCK(&lk); 4136 break; 4137 4138 default: 4139 panic("softdep_sync_metadata: Unknown type %s", 4140 TYPENAME(wk->wk_type)); 4141 /* NOTREACHED */ 4142 } 4143 } 4144 (void) getdirtybuf(&TAILQ_NEXT(bp, b_vnbufs), MNT_WAIT); 4145 nbp = TAILQ_NEXT(bp, b_vnbufs); 4146 FREE_LOCK(&lk); 4147 bawrite(bp); 4148 ACQUIRE_LOCK(&lk); 4149 if (nbp != NULL) { 4150 bp = nbp; 4151 goto loop; 4152 } 4153 /* 4154 * We must wait for any I/O in progress to finish so that 4155 * all potential buffers on the dirty list will be visible. 4156 * Once they are all there, proceed with the second pass 4157 * which will wait for the I/O as per above. 4158 */ 4159 drain_output(vp, 1); 4160 /* 4161 * The brief unlock is to allow any pent up dependency 4162 * processing to be done. 4163 */ 4164 if (waitfor == MNT_NOWAIT) { 4165 waitfor = MNT_WAIT; 4166 FREE_LOCK(&lk); 4167 ACQUIRE_LOCK(&lk); 4168 goto top; 4169 } 4170 4171 /* 4172 * If we have managed to get rid of all the dirty buffers, 4173 * then we are done. For certain directories and block 4174 * devices, we may need to do further work. 4175 */ 4176 if (TAILQ_FIRST(&vp->v_dirtyblkhd) == NULL) { 4177 FREE_LOCK(&lk); 4178 return (0); 4179 } 4180 4181 FREE_LOCK(&lk); 4182 /* 4183 * If we are trying to sync a block device, some of its buffers may 4184 * contain metadata that cannot be written until the contents of some 4185 * partially written files have been written to disk. The only easy 4186 * way to accomplish this is to sync the entire filesystem (luckily 4187 * this happens rarely). 4188 */ 4189 if (vn_isdisk(vp, NULL) && 4190 vp->v_rdev->si_mountpoint && !VOP_ISLOCKED(vp, NULL) && 4191 (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT, ap->a_cred, 4192 ap->a_p)) != 0) 4193 return (error); 4194 return (0); 4195} 4196 4197/* 4198 * Flush the dependencies associated with an inodedep. 4199 * Called with splbio blocked. 4200 */ 4201static int 4202flush_inodedep_deps(fs, ino) 4203 struct fs *fs; 4204 ino_t ino; 4205{ 4206 struct inodedep *inodedep; 4207 struct allocdirect *adp; 4208 int error, waitfor; 4209 struct buf *bp; 4210 4211 /* 4212 * This work is done in two passes. The first pass grabs most 4213 * of the buffers and begins asynchronously writing them. The 4214 * only way to wait for these asynchronous writes is to sleep 4215 * on the filesystem vnode which may stay busy for a long time 4216 * if the filesystem is active. So, instead, we make a second 4217 * pass over the dependencies blocking on each write. In the 4218 * usual case we will be blocking against a write that we 4219 * initiated, so when it is done the dependency will have been 4220 * resolved. Thus the second pass is expected to end quickly. 4221 * We give a brief window at the top of the loop to allow 4222 * any pending I/O to complete. 4223 */ 4224 for (waitfor = MNT_NOWAIT; ; ) { 4225 FREE_LOCK(&lk); 4226 ACQUIRE_LOCK(&lk); 4227 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4228 return (0); 4229 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 4230 adp = TAILQ_NEXT(adp, ad_next)) { 4231 if (adp->ad_state & DEPCOMPLETE) 4232 continue; 4233 bp = adp->ad_buf; 4234 if (getdirtybuf(&bp, waitfor) == 0) { 4235 if (waitfor == MNT_NOWAIT) 4236 continue; 4237 break; 4238 } 4239 FREE_LOCK(&lk); 4240 if (waitfor == MNT_NOWAIT) { 4241 bawrite(bp); 4242 } else if ((error = BUF_WRITE(bp)) != 0) { 4243 ACQUIRE_LOCK(&lk); 4244 return (error); 4245 } 4246 ACQUIRE_LOCK(&lk); 4247 break; 4248 } 4249 if (adp != NULL) 4250 continue; 4251 for (adp = TAILQ_FIRST(&inodedep->id_newinoupdt); adp; 4252 adp = TAILQ_NEXT(adp, ad_next)) { 4253 if (adp->ad_state & DEPCOMPLETE) 4254 continue; 4255 bp = adp->ad_buf; 4256 if (getdirtybuf(&bp, waitfor) == 0) { 4257 if (waitfor == MNT_NOWAIT) 4258 continue; 4259 break; 4260 } 4261 FREE_LOCK(&lk); 4262 if (waitfor == MNT_NOWAIT) { 4263 bawrite(bp); 4264 } else if ((error = BUF_WRITE(bp)) != 0) { 4265 ACQUIRE_LOCK(&lk); 4266 return (error); 4267 } 4268 ACQUIRE_LOCK(&lk); 4269 break; 4270 } 4271 if (adp != NULL) 4272 continue; 4273 /* 4274 * If pass2, we are done, otherwise do pass 2. 4275 */ 4276 if (waitfor == MNT_WAIT) 4277 break; 4278 waitfor = MNT_WAIT; 4279 } 4280 /* 4281 * Try freeing inodedep in case all dependencies have been removed. 4282 */ 4283 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0) 4284 (void) free_inodedep(inodedep); 4285 return (0); 4286} 4287 4288/* 4289 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 4290 * Called with splbio blocked. 4291 */ 4292static int 4293flush_pagedep_deps(pvp, mp, diraddhdp) 4294 struct vnode *pvp; 4295 struct mount *mp; 4296 struct diraddhd *diraddhdp; 4297{ 4298 struct proc *p = CURPROC; /* XXX */ 4299 struct inodedep *inodedep; 4300 struct ufsmount *ump; 4301 struct diradd *dap; 4302 struct vnode *vp; 4303 int gotit, error = 0; 4304 struct buf *bp; 4305 ino_t inum; 4306 4307 ump = VFSTOUFS(mp); 4308 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 4309 /* 4310 * Flush ourselves if this directory entry 4311 * has a MKDIR_PARENT dependency. 4312 */ 4313 if (dap->da_state & MKDIR_PARENT) { 4314 FREE_LOCK(&lk); 4315 if ((error = UFS_UPDATE(pvp, 1)) != 0) 4316 break; 4317 ACQUIRE_LOCK(&lk); 4318 /* 4319 * If that cleared dependencies, go on to next. 4320 */ 4321 if (dap != LIST_FIRST(diraddhdp)) 4322 continue; 4323 if (dap->da_state & MKDIR_PARENT) 4324 panic("flush_pagedep_deps: MKDIR_PARENT"); 4325 } 4326 /* 4327 * A newly allocated directory must have its "." and 4328 * ".." entries written out before its name can be 4329 * committed in its parent. We do not want or need 4330 * the full semantics of a synchronous VOP_FSYNC as 4331 * that may end up here again, once for each directory 4332 * level in the filesystem. Instead, we push the blocks 4333 * and wait for them to clear. We have to fsync twice 4334 * because the first call may choose to defer blocks 4335 * that still have dependencies, but deferral will 4336 * happen at most once. 4337 */ 4338 inum = dap->da_newinum; 4339 if (dap->da_state & MKDIR_BODY) { 4340 FREE_LOCK(&lk); 4341 if ((error = VFS_VGET(mp, inum, &vp)) != 0) 4342 break; 4343 if ((error=VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p)) || 4344 (error=VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p))) { 4345 vput(vp); 4346 break; 4347 } 4348 drain_output(vp, 0); 4349 vput(vp); 4350 ACQUIRE_LOCK(&lk); 4351 /* 4352 * If that cleared dependencies, go on to next. 4353 */ 4354 if (dap != LIST_FIRST(diraddhdp)) 4355 continue; 4356 if (dap->da_state & MKDIR_BODY) 4357 panic("flush_pagedep_deps: MKDIR_BODY"); 4358 } 4359 /* 4360 * Flush the inode on which the directory entry depends. 4361 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 4362 * the only remaining dependency is that the updated inode 4363 * count must get pushed to disk. The inode has already 4364 * been pushed into its inode buffer (via VOP_UPDATE) at 4365 * the time of the reference count change. So we need only 4366 * locate that buffer, ensure that there will be no rollback 4367 * caused by a bitmap dependency, then write the inode buffer. 4368 */ 4369 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) 4370 panic("flush_pagedep_deps: lost inode"); 4371 /* 4372 * If the inode still has bitmap dependencies, 4373 * push them to disk. 4374 */ 4375 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4376 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 4377 FREE_LOCK(&lk); 4378 if (gotit && 4379 (error = BUF_WRITE(inodedep->id_buf)) != 0) 4380 break; 4381 ACQUIRE_LOCK(&lk); 4382 if (dap != LIST_FIRST(diraddhdp)) 4383 continue; 4384 } 4385 /* 4386 * If the inode is still sitting in a buffer waiting 4387 * to be written, push it to disk. 4388 */ 4389 FREE_LOCK(&lk); 4390 if ((error = bread(ump->um_devvp, 4391 fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), 4392 (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0) 4393 break; 4394 if ((error = BUF_WRITE(bp)) != 0) 4395 break; 4396 ACQUIRE_LOCK(&lk); 4397 /* 4398 * If we have failed to get rid of all the dependencies 4399 * then something is seriously wrong. 4400 */ 4401 if (dap == LIST_FIRST(diraddhdp)) 4402 panic("flush_pagedep_deps: flush failed"); 4403 } 4404 if (error) 4405 ACQUIRE_LOCK(&lk); 4406 return (error); 4407} 4408 4409/* 4410 * A large burst of file addition or deletion activity can drive the 4411 * memory load excessively high. First attempt to slow things down 4412 * using the techniques below. If that fails, this routine requests 4413 * the offending operations to fall back to running synchronously 4414 * until the memory load returns to a reasonable level. 4415 */ 4416int 4417softdep_slowdown(vp) 4418 struct vnode *vp; 4419{ 4420 int max_softdeps_hard; 4421 4422 max_softdeps_hard = max_softdeps * 11 / 10; 4423 if (num_dirrem < max_softdeps_hard / 2 && 4424 num_inodedep < max_softdeps_hard) 4425 return (0); 4426 stat_sync_limit_hit += 1; 4427 return (1); 4428} 4429 4430/* 4431 * If memory utilization has gotten too high, deliberately slow things 4432 * down and speed up the I/O processing. 4433 */ 4434static int 4435request_cleanup(resource, islocked) 4436 int resource; 4437 int islocked; 4438{ 4439 struct proc *p = CURPROC; 4440 4441 /* 4442 * We never hold up the filesystem syncer process. 4443 */ 4444 if (p == filesys_syncer) 4445 return (0); 4446 /* 4447 * First check to see if the work list has gotten backlogged. 4448 * If it has, co-opt this process to help clean up two entries. 4449 * Because this process may hold inodes locked, we cannot 4450 * handle any remove requests that might block on a locked 4451 * inode as that could lead to deadlock. 4452 */ 4453 if (num_on_worklist > max_softdeps / 10) { 4454 process_worklist_item(NULL, LK_NOWAIT); 4455 process_worklist_item(NULL, LK_NOWAIT); 4456 stat_worklist_push += 2; 4457 return(0); 4458 } 4459 /* 4460 * Next, we attempt to speed up the syncer process. If that 4461 * is successful, then we allow the process to continue. 4462 */ 4463 if (speedup_syncer()) 4464 return(0); 4465 /* 4466 * If we are resource constrained on inode dependencies, try 4467 * flushing some dirty inodes. Otherwise, we are constrained 4468 * by file deletions, so try accelerating flushes of directories 4469 * with removal dependencies. We would like to do the cleanup 4470 * here, but we probably hold an inode locked at this point and 4471 * that might deadlock against one that we try to clean. So, 4472 * the best that we can do is request the syncer daemon to do 4473 * the cleanup for us. 4474 */ 4475 switch (resource) { 4476 4477 case FLUSH_INODES: 4478 stat_ino_limit_push += 1; 4479 req_clear_inodedeps += 1; 4480 stat_countp = &stat_ino_limit_hit; 4481 break; 4482 4483 case FLUSH_REMOVE: 4484 stat_blk_limit_push += 1; 4485 req_clear_remove += 1; 4486 stat_countp = &stat_blk_limit_hit; 4487 break; 4488 4489 default: 4490 panic("request_cleanup: unknown type"); 4491 } 4492 /* 4493 * Hopefully the syncer daemon will catch up and awaken us. 4494 * We wait at most tickdelay before proceeding in any case. 4495 */ 4496 if (islocked == 0) 4497 ACQUIRE_LOCK(&lk); 4498 proc_waiting += 1; 4499 if (handle.callout == NULL) 4500 handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2); 4501 FREE_LOCK_INTERLOCKED(&lk); 4502 (void) tsleep((caddr_t)&proc_waiting, PPAUSE, "softupdate", 0); 4503 ACQUIRE_LOCK_INTERLOCKED(&lk); 4504 proc_waiting -= 1; 4505 if (islocked == 0) 4506 FREE_LOCK(&lk); 4507 return (1); 4508} 4509 4510/* 4511 * Awaken processes pausing in request_cleanup and clear proc_waiting 4512 * to indicate that there is no longer a timer running. 4513 */ 4514void 4515pause_timer(arg) 4516 void *arg; 4517{ 4518 4519 *stat_countp += 1; 4520 wakeup_one(&proc_waiting); 4521 if (proc_waiting > 0) 4522 handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2); 4523 else 4524 handle.callout = NULL; 4525} 4526 4527/* 4528 * Flush out a directory with at least one removal dependency in an effort to 4529 * reduce the number of dirrem, freefile, and freeblks dependency structures. 4530 */ 4531static void 4532clear_remove(p) 4533 struct proc *p; 4534{ 4535 struct pagedep_hashhead *pagedephd; 4536 struct pagedep *pagedep; 4537 static int next = 0; 4538 struct mount *mp; 4539 struct vnode *vp; 4540 int error, cnt; 4541 ino_t ino; 4542 4543 ACQUIRE_LOCK(&lk); 4544 for (cnt = 0; cnt < pagedep_hash; cnt++) { 4545 pagedephd = &pagedep_hashtbl[next++]; 4546 if (next >= pagedep_hash) 4547 next = 0; 4548 for (pagedep = LIST_FIRST(pagedephd); pagedep; 4549 pagedep = LIST_NEXT(pagedep, pd_hash)) { 4550 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL) 4551 continue; 4552 mp = pagedep->pd_mnt; 4553 ino = pagedep->pd_ino; 4554 FREE_LOCK(&lk); 4555 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 4556 continue; 4557 if ((error = VFS_VGET(mp, ino, &vp)) != 0) { 4558 softdep_error("clear_remove: vget", error); 4559 vn_finished_write(mp); 4560 return; 4561 } 4562 if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p))) 4563 softdep_error("clear_remove: fsync", error); 4564 drain_output(vp, 0); 4565 vput(vp); 4566 vn_finished_write(mp); 4567 return; 4568 } 4569 } 4570 FREE_LOCK(&lk); 4571} 4572 4573/* 4574 * Clear out a block of dirty inodes in an effort to reduce 4575 * the number of inodedep dependency structures. 4576 */ 4577static void 4578clear_inodedeps(p) 4579 struct proc *p; 4580{ 4581 struct inodedep_hashhead *inodedephd; 4582 struct inodedep *inodedep; 4583 static int next = 0; 4584 struct mount *mp; 4585 struct vnode *vp; 4586 struct fs *fs; 4587 int error, cnt; 4588 ino_t firstino, lastino, ino; 4589 4590 ACQUIRE_LOCK(&lk); 4591 /* 4592 * Pick a random inode dependency to be cleared. 4593 * We will then gather up all the inodes in its block 4594 * that have dependencies and flush them out. 4595 */ 4596 for (cnt = 0; cnt < inodedep_hash; cnt++) { 4597 inodedephd = &inodedep_hashtbl[next++]; 4598 if (next >= inodedep_hash) 4599 next = 0; 4600 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 4601 break; 4602 } 4603 /* 4604 * Ugly code to find mount point given pointer to superblock. 4605 */ 4606 fs = inodedep->id_fs; 4607 TAILQ_FOREACH(mp, &mountlist, mnt_list) 4608 if ((mp->mnt_flag & MNT_SOFTDEP) && fs == VFSTOUFS(mp)->um_fs) 4609 break; 4610 /* 4611 * Find the last inode in the block with dependencies. 4612 */ 4613 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 4614 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 4615 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0) 4616 break; 4617 /* 4618 * Asynchronously push all but the last inode with dependencies. 4619 * Synchronously push the last inode with dependencies to ensure 4620 * that the inode block gets written to free up the inodedeps. 4621 */ 4622 for (ino = firstino; ino <= lastino; ino++) { 4623 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4624 continue; 4625 FREE_LOCK(&lk); 4626 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 4627 continue; 4628 if ((error = VFS_VGET(mp, ino, &vp)) != 0) { 4629 softdep_error("clear_inodedeps: vget", error); 4630 vn_finished_write(mp); 4631 return; 4632 } 4633 if (ino == lastino) { 4634 if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_WAIT, p))) 4635 softdep_error("clear_inodedeps: fsync1", error); 4636 } else { 4637 if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p))) 4638 softdep_error("clear_inodedeps: fsync2", error); 4639 drain_output(vp, 0); 4640 } 4641 vput(vp); 4642 vn_finished_write(mp); 4643 ACQUIRE_LOCK(&lk); 4644 } 4645 FREE_LOCK(&lk); 4646} 4647 4648/* 4649 * Function to determine if the buffer has outstanding dependencies 4650 * that will cause a roll-back if the buffer is written. If wantcount 4651 * is set, return number of dependencies, otherwise just yes or no. 4652 */ 4653static int 4654softdep_count_dependencies(bp, wantcount) 4655 struct buf *bp; 4656 int wantcount; 4657{ 4658 struct worklist *wk; 4659 struct inodedep *inodedep; 4660 struct indirdep *indirdep; 4661 struct allocindir *aip; 4662 struct pagedep *pagedep; 4663 struct diradd *dap; 4664 int i, retval; 4665 4666 retval = 0; 4667 ACQUIRE_LOCK(&lk); 4668 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = LIST_NEXT(wk, wk_list)) { 4669 switch (wk->wk_type) { 4670 4671 case D_INODEDEP: 4672 inodedep = WK_INODEDEP(wk); 4673 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4674 /* bitmap allocation dependency */ 4675 retval += 1; 4676 if (!wantcount) 4677 goto out; 4678 } 4679 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 4680 /* direct block pointer dependency */ 4681 retval += 1; 4682 if (!wantcount) 4683 goto out; 4684 } 4685 continue; 4686 4687 case D_INDIRDEP: 4688 indirdep = WK_INDIRDEP(wk); 4689 for (aip = LIST_FIRST(&indirdep->ir_deplisthd); 4690 aip; aip = LIST_NEXT(aip, ai_next)) { 4691 /* indirect block pointer dependency */ 4692 retval += 1; 4693 if (!wantcount) 4694 goto out; 4695 } 4696 continue; 4697 4698 case D_PAGEDEP: 4699 pagedep = WK_PAGEDEP(wk); 4700 for (i = 0; i < DAHASHSZ; i++) { 4701 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); 4702 dap; dap = LIST_NEXT(dap, da_pdlist)) { 4703 /* directory entry dependency */ 4704 retval += 1; 4705 if (!wantcount) 4706 goto out; 4707 } 4708 } 4709 continue; 4710 4711 case D_BMSAFEMAP: 4712 case D_ALLOCDIRECT: 4713 case D_ALLOCINDIR: 4714 case D_MKDIR: 4715 /* never a dependency on these blocks */ 4716 continue; 4717 4718 default: 4719 panic("softdep_check_for_rollback: Unexpected type %s", 4720 TYPENAME(wk->wk_type)); 4721 /* NOTREACHED */ 4722 } 4723 } 4724out: 4725 FREE_LOCK(&lk); 4726 return retval; 4727} 4728 4729/* 4730 * Acquire exclusive access to a buffer. 4731 * Must be called with splbio blocked. 4732 * Return 1 if buffer was acquired. 4733 */ 4734static int 4735getdirtybuf(bpp, waitfor) 4736 struct buf **bpp; 4737 int waitfor; 4738{ 4739 struct buf *bp; 4740 4741 for (;;) { 4742 if ((bp = *bpp) == NULL) 4743 return (0); 4744 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 4745 if ((bp->b_xflags & BX_BKGRDINPROG) == 0) 4746 break; 4747 BUF_UNLOCK(bp); 4748 if (waitfor != MNT_WAIT) 4749 return (0); 4750 bp->b_xflags |= BX_BKGRDWAIT; 4751 FREE_LOCK_INTERLOCKED(&lk); 4752 tsleep(&bp->b_xflags, PRIBIO, "getbuf", 0); 4753 ACQUIRE_LOCK_INTERLOCKED(&lk); 4754 continue; 4755 } 4756 if (waitfor != MNT_WAIT) 4757 return (0); 4758 FREE_LOCK_INTERLOCKED(&lk); 4759 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL) != ENOLCK) 4760 panic("getdirtybuf: inconsistent lock"); 4761 ACQUIRE_LOCK_INTERLOCKED(&lk); 4762 } 4763 if ((bp->b_flags & B_DELWRI) == 0) { 4764 BUF_UNLOCK(bp); 4765 return (0); 4766 } 4767 bremfree(bp); 4768 return (1); 4769} 4770 4771/* 4772 * Wait for pending output on a vnode to complete. 4773 * Must be called with vnode locked. 4774 */ 4775static void 4776drain_output(vp, islocked) 4777 struct vnode *vp; 4778 int islocked; 4779{ 4780 4781 if (!islocked) 4782 ACQUIRE_LOCK(&lk); 4783 while (vp->v_numoutput) { 4784 vp->v_flag |= VBWAIT; 4785 FREE_LOCK_INTERLOCKED(&lk); 4786 tsleep((caddr_t)&vp->v_numoutput, PRIBIO + 1, "drainvp", 0); 4787 ACQUIRE_LOCK_INTERLOCKED(&lk); 4788 } 4789 if (!islocked) 4790 FREE_LOCK(&lk); 4791} 4792 4793/* 4794 * Called whenever a buffer that is being invalidated or reallocated 4795 * contains dependencies. This should only happen if an I/O error has 4796 * occurred. The routine is called with the buffer locked. 4797 */ 4798static void 4799softdep_deallocate_dependencies(bp) 4800 struct buf *bp; 4801{ 4802 4803 if ((bp->b_ioflags & BIO_ERROR) == 0) 4804 panic("softdep_deallocate_dependencies: dangling deps"); 4805 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); 4806 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 4807} 4808 4809/* 4810 * Function to handle asynchronous write errors in the filesystem. 4811 */ 4812void 4813softdep_error(func, error) 4814 char *func; 4815 int error; 4816{ 4817 4818 /* XXX should do something better! */ 4819 printf("%s: got error %d while accessing filesystem\n", func, error); 4820} 4821