ffs_softdep.c revision 126853
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 */ 40 41#include <sys/cdefs.h> 42__FBSDID("$FreeBSD: head/sys/ufs/ffs/ffs_softdep.c 126853 2004-03-11 18:02:36Z phk $"); 43 44/* 45 * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide. 46 */ 47#ifndef DIAGNOSTIC 48#define DIAGNOSTIC 49#endif 50#ifndef DEBUG 51#define DEBUG 52#endif 53 54#include <sys/param.h> 55#include <sys/kernel.h> 56#include <sys/systm.h> 57#include <sys/bio.h> 58#include <sys/buf.h> 59#include <sys/malloc.h> 60#include <sys/mount.h> 61#include <sys/proc.h> 62#include <sys/stat.h> 63#include <sys/syslog.h> 64#include <sys/vnode.h> 65#include <sys/conf.h> 66#include <ufs/ufs/dir.h> 67#include <ufs/ufs/extattr.h> 68#include <ufs/ufs/quota.h> 69#include <ufs/ufs/inode.h> 70#include <ufs/ufs/ufsmount.h> 71#include <ufs/ffs/fs.h> 72#include <ufs/ffs/softdep.h> 73#include <ufs/ffs/ffs_extern.h> 74#include <ufs/ufs/ufs_extern.h> 75 76/* 77 * These definitions need to be adapted to the system to which 78 * this file is being ported. 79 */ 80/* 81 * malloc types defined for the softdep system. 82 */ 83static MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies"); 84static MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies"); 85static MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation"); 86static MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map"); 87static MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode"); 88static MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies"); 89static MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block"); 90static MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode"); 91static MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode"); 92static MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated"); 93static MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry"); 94static MALLOC_DEFINE(M_MKDIR, "mkdir","New directory"); 95static MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted"); 96static MALLOC_DEFINE(M_NEWDIRBLK, "newdirblk","Unclaimed new directory block"); 97 98#define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE) 99 100#define D_PAGEDEP 0 101#define D_INODEDEP 1 102#define D_NEWBLK 2 103#define D_BMSAFEMAP 3 104#define D_ALLOCDIRECT 4 105#define D_INDIRDEP 5 106#define D_ALLOCINDIR 6 107#define D_FREEFRAG 7 108#define D_FREEBLKS 8 109#define D_FREEFILE 9 110#define D_DIRADD 10 111#define D_MKDIR 11 112#define D_DIRREM 12 113#define D_NEWDIRBLK 13 114#define D_LAST D_NEWDIRBLK 115 116/* 117 * translate from workitem type to memory type 118 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 119 */ 120static struct malloc_type *memtype[] = { 121 M_PAGEDEP, 122 M_INODEDEP, 123 M_NEWBLK, 124 M_BMSAFEMAP, 125 M_ALLOCDIRECT, 126 M_INDIRDEP, 127 M_ALLOCINDIR, 128 M_FREEFRAG, 129 M_FREEBLKS, 130 M_FREEFILE, 131 M_DIRADD, 132 M_MKDIR, 133 M_DIRREM, 134 M_NEWDIRBLK 135}; 136 137#define DtoM(type) (memtype[type]) 138 139/* 140 * Names of malloc types. 141 */ 142#define TYPENAME(type) \ 143 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???") 144/* 145 * End system adaptaion definitions. 146 */ 147 148/* 149 * Internal function prototypes. 150 */ 151static void softdep_error(char *, int); 152static void drain_output(struct vnode *, int); 153static struct buf *getdirtybuf(struct buf **, struct mtx *, int); 154static void clear_remove(struct thread *); 155static void clear_inodedeps(struct thread *); 156static int flush_pagedep_deps(struct vnode *, struct mount *, 157 struct diraddhd *); 158static int flush_inodedep_deps(struct fs *, ino_t); 159static int flush_deplist(struct allocdirectlst *, int, int *); 160static int handle_written_filepage(struct pagedep *, struct buf *); 161static void diradd_inode_written(struct diradd *, struct inodedep *); 162static int handle_written_inodeblock(struct inodedep *, struct buf *); 163static void handle_allocdirect_partdone(struct allocdirect *); 164static void handle_allocindir_partdone(struct allocindir *); 165static void initiate_write_filepage(struct pagedep *, struct buf *); 166static void handle_written_mkdir(struct mkdir *, int); 167static void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *); 168static void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *); 169static void handle_workitem_freefile(struct freefile *); 170static void handle_workitem_remove(struct dirrem *, struct vnode *); 171static struct dirrem *newdirrem(struct buf *, struct inode *, 172 struct inode *, int, struct dirrem **); 173static void free_diradd(struct diradd *); 174static void free_allocindir(struct allocindir *, struct inodedep *); 175static void free_newdirblk(struct newdirblk *); 176static int indir_trunc(struct freeblks *, ufs2_daddr_t, int, ufs_lbn_t, 177 ufs2_daddr_t *); 178static void deallocate_dependencies(struct buf *, struct inodedep *); 179static void free_allocdirect(struct allocdirectlst *, 180 struct allocdirect *, int); 181static int check_inode_unwritten(struct inodedep *); 182static int free_inodedep(struct inodedep *); 183static void handle_workitem_freeblocks(struct freeblks *, int); 184static void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *); 185static void setup_allocindir_phase2(struct buf *, struct inode *, 186 struct allocindir *); 187static struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t, 188 ufs2_daddr_t); 189static void handle_workitem_freefrag(struct freefrag *); 190static struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long); 191static void allocdirect_merge(struct allocdirectlst *, 192 struct allocdirect *, struct allocdirect *); 193static struct bmsafemap *bmsafemap_lookup(struct buf *); 194static int newblk_lookup(struct fs *, ufs2_daddr_t, int, struct newblk **); 195static int inodedep_lookup(struct fs *, ino_t, int, struct inodedep **); 196static int pagedep_lookup(struct inode *, ufs_lbn_t, int, struct pagedep **); 197static void pause_timer(void *); 198static int request_cleanup(int, int); 199static int process_worklist_item(struct mount *, int); 200static void add_to_worklist(struct worklist *); 201 202/* 203 * Exported softdep operations. 204 */ 205static void softdep_disk_io_initiation(struct buf *); 206static void softdep_disk_write_complete(struct buf *); 207static void softdep_deallocate_dependencies(struct buf *); 208static void softdep_move_dependencies(struct buf *, struct buf *); 209static int softdep_count_dependencies(struct buf *bp, int); 210 211/* 212 * Locking primitives. 213 * 214 * For a uniprocessor, all we need to do is protect against disk 215 * interrupts. For a multiprocessor, this lock would have to be 216 * a mutex. A single mutex is used throughout this file, though 217 * finer grain locking could be used if contention warranted it. 218 * 219 * For a multiprocessor, the sleep call would accept a lock and 220 * release it after the sleep processing was complete. In a uniprocessor 221 * implementation there is no such interlock, so we simple mark 222 * the places where it needs to be done with the `interlocked' form 223 * of the lock calls. Since the uniprocessor sleep already interlocks 224 * the spl, there is nothing that really needs to be done. 225 */ 226#ifndef /* NOT */ DEBUG 227static struct lockit { 228 int lkt_spl; 229} lk = { 0 }; 230#define ACQUIRE_LOCK(lk) (lk)->lkt_spl = splbio() 231#define FREE_LOCK(lk) splx((lk)->lkt_spl) 232 233#else /* DEBUG */ 234#define NOHOLDER ((struct thread *)-1) 235#define SPECIAL_FLAG ((struct thread *)-2) 236static struct lockit { 237 int lkt_spl; 238 struct thread *lkt_held; 239} lk = { 0, NOHOLDER }; 240 241static void acquire_lock(struct lockit *); 242static void free_lock(struct lockit *); 243void softdep_panic(char *); 244 245#define ACQUIRE_LOCK(lk) acquire_lock(lk) 246#define FREE_LOCK(lk) free_lock(lk) 247 248static void 249acquire_lock(lk) 250 struct lockit *lk; 251{ 252 struct thread *holder; 253 254 if (lk->lkt_held != NOHOLDER) { 255 holder = lk->lkt_held; 256 FREE_LOCK(lk); 257 if (holder == curthread) 258 panic("softdep_lock: locking against myself"); 259 else 260 panic("softdep_lock: lock held by %p", holder); 261 } 262 lk->lkt_spl = splbio(); 263 lk->lkt_held = curthread; 264} 265 266static void 267free_lock(lk) 268 struct lockit *lk; 269{ 270 271 if (lk->lkt_held == NOHOLDER) 272 panic("softdep_unlock: lock not held"); 273 lk->lkt_held = NOHOLDER; 274 splx(lk->lkt_spl); 275} 276 277/* 278 * Function to release soft updates lock and panic. 279 */ 280void 281softdep_panic(msg) 282 char *msg; 283{ 284 285 if (lk.lkt_held != NOHOLDER) 286 FREE_LOCK(&lk); 287 panic(msg); 288} 289#endif /* DEBUG */ 290 291static int interlocked_sleep(struct lockit *, int, void *, struct mtx *, int, 292 const char *, int); 293 294/* 295 * When going to sleep, we must save our SPL so that it does 296 * not get lost if some other process uses the lock while we 297 * are sleeping. We restore it after we have slept. This routine 298 * wraps the interlocking with functions that sleep. The list 299 * below enumerates the available set of operations. 300 */ 301#define UNKNOWN 0 302#define SLEEP 1 303#define LOCKBUF 2 304 305static int 306interlocked_sleep(lk, op, ident, mtx, flags, wmesg, timo) 307 struct lockit *lk; 308 int op; 309 void *ident; 310 struct mtx *mtx; 311 int flags; 312 const char *wmesg; 313 int timo; 314{ 315 struct thread *holder; 316 int s, retval; 317 318 s = lk->lkt_spl; 319# ifdef DEBUG 320 if (lk->lkt_held == NOHOLDER) 321 panic("interlocked_sleep: lock not held"); 322 lk->lkt_held = NOHOLDER; 323# endif /* DEBUG */ 324 switch (op) { 325 case SLEEP: 326 retval = msleep(ident, mtx, flags, wmesg, timo); 327 break; 328 case LOCKBUF: 329 retval = BUF_LOCK((struct buf *)ident, flags, mtx); 330 break; 331 default: 332 panic("interlocked_sleep: unknown operation"); 333 } 334# ifdef DEBUG 335 if (lk->lkt_held != NOHOLDER) { 336 holder = lk->lkt_held; 337 FREE_LOCK(lk); 338 if (holder == curthread) 339 panic("interlocked_sleep: locking against self"); 340 else 341 panic("interlocked_sleep: lock held by %p", holder); 342 } 343 lk->lkt_held = curthread; 344# endif /* DEBUG */ 345 lk->lkt_spl = s; 346 return (retval); 347} 348 349/* 350 * Place holder for real semaphores. 351 */ 352struct sema { 353 int value; 354 struct thread *holder; 355 char *name; 356 int prio; 357 int timo; 358}; 359static void sema_init(struct sema *, char *, int, int); 360static int sema_get(struct sema *, struct lockit *); 361static void sema_release(struct sema *); 362 363static void 364sema_init(semap, name, prio, timo) 365 struct sema *semap; 366 char *name; 367 int prio, timo; 368{ 369 370 semap->holder = NOHOLDER; 371 semap->value = 0; 372 semap->name = name; 373 semap->prio = prio; 374 semap->timo = timo; 375} 376 377static int 378sema_get(semap, interlock) 379 struct sema *semap; 380 struct lockit *interlock; 381{ 382 383 if (semap->value++ > 0) { 384 if (interlock != NULL) { 385 interlocked_sleep(interlock, SLEEP, (caddr_t)semap, 386 NULL, semap->prio, semap->name, 387 semap->timo); 388 FREE_LOCK(interlock); 389 } else { 390 tsleep(semap, semap->prio, semap->name, 391 semap->timo); 392 } 393 return (0); 394 } 395 semap->holder = curthread; 396 if (interlock != NULL) 397 FREE_LOCK(interlock); 398 return (1); 399} 400 401static void 402sema_release(semap) 403 struct sema *semap; 404{ 405 406 if (semap->value <= 0 || semap->holder != curthread) { 407 if (lk.lkt_held != NOHOLDER) 408 FREE_LOCK(&lk); 409 panic("sema_release: not held"); 410 } 411 if (--semap->value > 0) { 412 semap->value = 0; 413 wakeup(semap); 414 } 415 semap->holder = NOHOLDER; 416} 417 418/* 419 * Worklist queue management. 420 * These routines require that the lock be held. 421 */ 422#ifndef /* NOT */ DEBUG 423#define WORKLIST_INSERT(head, item) do { \ 424 (item)->wk_state |= ONWORKLIST; \ 425 LIST_INSERT_HEAD(head, item, wk_list); \ 426} while (0) 427#define WORKLIST_REMOVE(item) do { \ 428 (item)->wk_state &= ~ONWORKLIST; \ 429 LIST_REMOVE(item, wk_list); \ 430} while (0) 431#define WORKITEM_FREE(item, type) FREE(item, DtoM(type)) 432 433#else /* DEBUG */ 434static void worklist_insert(struct workhead *, struct worklist *); 435static void worklist_remove(struct worklist *); 436static void workitem_free(struct worklist *, int); 437 438#define WORKLIST_INSERT(head, item) worklist_insert(head, item) 439#define WORKLIST_REMOVE(item) worklist_remove(item) 440#define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type) 441 442static void 443worklist_insert(head, item) 444 struct workhead *head; 445 struct worklist *item; 446{ 447 448 if (lk.lkt_held == NOHOLDER) 449 panic("worklist_insert: lock not held"); 450 if (item->wk_state & ONWORKLIST) { 451 FREE_LOCK(&lk); 452 panic("worklist_insert: already on list"); 453 } 454 item->wk_state |= ONWORKLIST; 455 LIST_INSERT_HEAD(head, item, wk_list); 456} 457 458static void 459worklist_remove(item) 460 struct worklist *item; 461{ 462 463 if (lk.lkt_held == NOHOLDER) 464 panic("worklist_remove: lock not held"); 465 if ((item->wk_state & ONWORKLIST) == 0) { 466 FREE_LOCK(&lk); 467 panic("worklist_remove: not on list"); 468 } 469 item->wk_state &= ~ONWORKLIST; 470 LIST_REMOVE(item, wk_list); 471} 472 473static void 474workitem_free(item, type) 475 struct worklist *item; 476 int type; 477{ 478 479 if (item->wk_state & ONWORKLIST) { 480 if (lk.lkt_held != NOHOLDER) 481 FREE_LOCK(&lk); 482 panic("workitem_free: still on list"); 483 } 484 if (item->wk_type != type) { 485 if (lk.lkt_held != NOHOLDER) 486 FREE_LOCK(&lk); 487 panic("workitem_free: type mismatch"); 488 } 489 FREE(item, DtoM(type)); 490} 491#endif /* DEBUG */ 492 493/* 494 * Workitem queue management 495 */ 496static struct workhead softdep_workitem_pending; 497static struct worklist *worklist_tail; 498static int num_on_worklist; /* number of worklist items to be processed */ 499static int softdep_worklist_busy; /* 1 => trying to do unmount */ 500static int softdep_worklist_req; /* serialized waiters */ 501static int max_softdeps; /* maximum number of structs before slowdown */ 502static int maxindirdeps = 50; /* max number of indirdeps before slowdown */ 503static int tickdelay = 2; /* number of ticks to pause during slowdown */ 504static int proc_waiting; /* tracks whether we have a timeout posted */ 505static int *stat_countp; /* statistic to count in proc_waiting timeout */ 506static struct callout_handle handle; /* handle on posted proc_waiting timeout */ 507static struct thread *filesys_syncer; /* proc of filesystem syncer process */ 508static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 509#define FLUSH_INODES 1 510static int req_clear_remove; /* syncer process flush some freeblks */ 511#define FLUSH_REMOVE 2 512#define FLUSH_REMOVE_WAIT 3 513/* 514 * runtime statistics 515 */ 516static int stat_worklist_push; /* number of worklist cleanups */ 517static int stat_blk_limit_push; /* number of times block limit neared */ 518static int stat_ino_limit_push; /* number of times inode limit neared */ 519static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 520static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 521static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 522static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 523static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 524static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 525static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 526#ifdef DEBUG 527#include <vm/vm.h> 528#include <sys/sysctl.h> 529SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, ""); 530SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, ""); 531SYSCTL_INT(_debug, OID_AUTO, maxindirdeps, CTLFLAG_RW, &maxindirdeps, 0, ""); 532SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,""); 533SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,""); 534SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,""); 535SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, ""); 536SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, ""); 537SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, ""); 538SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, ""); 539SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, ""); 540SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, ""); 541SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, ""); 542#endif /* DEBUG */ 543 544/* 545 * Add an item to the end of the work queue. 546 * This routine requires that the lock be held. 547 * This is the only routine that adds items to the list. 548 * The following routine is the only one that removes items 549 * and does so in order from first to last. 550 */ 551static void 552add_to_worklist(wk) 553 struct worklist *wk; 554{ 555 556 if (wk->wk_state & ONWORKLIST) { 557 if (lk.lkt_held != NOHOLDER) 558 FREE_LOCK(&lk); 559 panic("add_to_worklist: already on list"); 560 } 561 wk->wk_state |= ONWORKLIST; 562 if (LIST_FIRST(&softdep_workitem_pending) == NULL) 563 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list); 564 else 565 LIST_INSERT_AFTER(worklist_tail, wk, wk_list); 566 worklist_tail = wk; 567 num_on_worklist += 1; 568} 569 570/* 571 * Process that runs once per second to handle items in the background queue. 572 * 573 * Note that we ensure that everything is done in the order in which they 574 * appear in the queue. The code below depends on this property to ensure 575 * that blocks of a file are freed before the inode itself is freed. This 576 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 577 * until all the old ones have been purged from the dependency lists. 578 */ 579int 580softdep_process_worklist(matchmnt) 581 struct mount *matchmnt; 582{ 583 struct thread *td = curthread; 584 int cnt, matchcnt, loopcount; 585 long starttime; 586 587 /* 588 * Record the process identifier of our caller so that we can give 589 * this process preferential treatment in request_cleanup below. 590 */ 591 filesys_syncer = td; 592 matchcnt = 0; 593 594 /* 595 * There is no danger of having multiple processes run this 596 * code, but we have to single-thread it when softdep_flushfiles() 597 * is in operation to get an accurate count of the number of items 598 * related to its mount point that are in the list. 599 */ 600 if (matchmnt == NULL) { 601 if (softdep_worklist_busy < 0) 602 return(-1); 603 softdep_worklist_busy += 1; 604 } 605 606 /* 607 * If requested, try removing inode or removal dependencies. 608 */ 609 if (req_clear_inodedeps) { 610 clear_inodedeps(td); 611 req_clear_inodedeps -= 1; 612 wakeup_one(&proc_waiting); 613 } 614 if (req_clear_remove) { 615 clear_remove(td); 616 req_clear_remove -= 1; 617 wakeup_one(&proc_waiting); 618 } 619 loopcount = 1; 620 starttime = time_second; 621 while (num_on_worklist > 0) { 622 if ((cnt = process_worklist_item(matchmnt, 0)) == -1) 623 break; 624 else 625 matchcnt += cnt; 626 627 /* 628 * If a umount operation wants to run the worklist 629 * accurately, abort. 630 */ 631 if (softdep_worklist_req && matchmnt == NULL) { 632 matchcnt = -1; 633 break; 634 } 635 636 /* 637 * If requested, try removing inode or removal dependencies. 638 */ 639 if (req_clear_inodedeps) { 640 clear_inodedeps(td); 641 req_clear_inodedeps -= 1; 642 wakeup_one(&proc_waiting); 643 } 644 if (req_clear_remove) { 645 clear_remove(td); 646 req_clear_remove -= 1; 647 wakeup_one(&proc_waiting); 648 } 649 /* 650 * We do not generally want to stop for buffer space, but if 651 * we are really being a buffer hog, we will stop and wait. 652 */ 653 if (loopcount++ % 128 == 0) 654 bwillwrite(); 655 /* 656 * Never allow processing to run for more than one 657 * second. Otherwise the other syncer tasks may get 658 * excessively backlogged. 659 */ 660 if (starttime != time_second && matchmnt == NULL) { 661 matchcnt = -1; 662 break; 663 } 664 } 665 if (matchmnt == NULL) { 666 softdep_worklist_busy -= 1; 667 if (softdep_worklist_req && softdep_worklist_busy == 0) 668 wakeup(&softdep_worklist_req); 669 } 670 return (matchcnt); 671} 672 673/* 674 * Process one item on the worklist. 675 */ 676static int 677process_worklist_item(matchmnt, flags) 678 struct mount *matchmnt; 679 int flags; 680{ 681 struct worklist *wk, *wkend; 682 struct mount *mp; 683 struct vnode *vp; 684 int matchcnt = 0; 685 686 /* 687 * If we are being called because of a process doing a 688 * copy-on-write, then it is not safe to write as we may 689 * recurse into the copy-on-write routine. 690 */ 691 if (curthread->td_pflags & TDP_COWINPROGRESS) 692 return (-1); 693 ACQUIRE_LOCK(&lk); 694 /* 695 * Normally we just process each item on the worklist in order. 696 * However, if we are in a situation where we cannot lock any 697 * inodes, we have to skip over any dirrem requests whose 698 * vnodes are resident and locked. 699 */ 700 vp = NULL; 701 LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) { 702 if (wk->wk_state & INPROGRESS) 703 continue; 704 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM) 705 break; 706 wk->wk_state |= INPROGRESS; 707 FREE_LOCK(&lk); 708 VFS_VGET(WK_DIRREM(wk)->dm_mnt, WK_DIRREM(wk)->dm_oldinum, 709 LK_NOWAIT | LK_EXCLUSIVE, &vp); 710 ACQUIRE_LOCK(&lk); 711 wk->wk_state &= ~INPROGRESS; 712 if (vp != NULL) 713 break; 714 } 715 if (wk == 0) { 716 FREE_LOCK(&lk); 717 return (-1); 718 } 719 /* 720 * Remove the item to be processed. If we are removing the last 721 * item on the list, we need to recalculate the tail pointer. 722 * As this happens rarely and usually when the list is short, 723 * we just run down the list to find it rather than tracking it 724 * in the above loop. 725 */ 726 WORKLIST_REMOVE(wk); 727 if (wk == worklist_tail) { 728 LIST_FOREACH(wkend, &softdep_workitem_pending, wk_list) 729 if (LIST_NEXT(wkend, wk_list) == NULL) 730 break; 731 worklist_tail = wkend; 732 } 733 num_on_worklist -= 1; 734 FREE_LOCK(&lk); 735 switch (wk->wk_type) { 736 737 case D_DIRREM: 738 /* removal of a directory entry */ 739 mp = WK_DIRREM(wk)->dm_mnt; 740 if (vn_write_suspend_wait(NULL, mp, V_NOWAIT)) 741 panic("%s: dirrem on suspended filesystem", 742 "process_worklist_item"); 743 if (mp == matchmnt) 744 matchcnt += 1; 745 handle_workitem_remove(WK_DIRREM(wk), vp); 746 break; 747 748 case D_FREEBLKS: 749 /* releasing blocks and/or fragments from a file */ 750 mp = WK_FREEBLKS(wk)->fb_mnt; 751 if (vn_write_suspend_wait(NULL, mp, V_NOWAIT)) 752 panic("%s: freeblks on suspended filesystem", 753 "process_worklist_item"); 754 if (mp == matchmnt) 755 matchcnt += 1; 756 handle_workitem_freeblocks(WK_FREEBLKS(wk), flags & LK_NOWAIT); 757 break; 758 759 case D_FREEFRAG: 760 /* releasing a fragment when replaced as a file grows */ 761 mp = WK_FREEFRAG(wk)->ff_mnt; 762 if (vn_write_suspend_wait(NULL, mp, V_NOWAIT)) 763 panic("%s: freefrag on suspended filesystem", 764 "process_worklist_item"); 765 if (mp == matchmnt) 766 matchcnt += 1; 767 handle_workitem_freefrag(WK_FREEFRAG(wk)); 768 break; 769 770 case D_FREEFILE: 771 /* releasing an inode when its link count drops to 0 */ 772 mp = WK_FREEFILE(wk)->fx_mnt; 773 if (vn_write_suspend_wait(NULL, mp, V_NOWAIT)) 774 panic("%s: freefile on suspended filesystem", 775 "process_worklist_item"); 776 if (mp == matchmnt) 777 matchcnt += 1; 778 handle_workitem_freefile(WK_FREEFILE(wk)); 779 break; 780 781 default: 782 panic("%s_process_worklist: Unknown type %s", 783 "softdep", TYPENAME(wk->wk_type)); 784 /* NOTREACHED */ 785 } 786 return (matchcnt); 787} 788 789/* 790 * Move dependencies from one buffer to another. 791 */ 792static void 793softdep_move_dependencies(oldbp, newbp) 794 struct buf *oldbp; 795 struct buf *newbp; 796{ 797 struct worklist *wk, *wktail; 798 799 if (LIST_FIRST(&newbp->b_dep) != NULL) 800 panic("softdep_move_dependencies: need merge code"); 801 wktail = 0; 802 ACQUIRE_LOCK(&lk); 803 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 804 LIST_REMOVE(wk, wk_list); 805 if (wktail == 0) 806 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 807 else 808 LIST_INSERT_AFTER(wktail, wk, wk_list); 809 wktail = wk; 810 } 811 FREE_LOCK(&lk); 812} 813 814/* 815 * Purge the work list of all items associated with a particular mount point. 816 */ 817int 818softdep_flushworklist(oldmnt, countp, td) 819 struct mount *oldmnt; 820 int *countp; 821 struct thread *td; 822{ 823 struct vnode *devvp; 824 int count, error = 0; 825 826 /* 827 * Await our turn to clear out the queue, then serialize access. 828 */ 829 while (softdep_worklist_busy) { 830 softdep_worklist_req += 1; 831 tsleep(&softdep_worklist_req, PRIBIO, "softflush", 0); 832 softdep_worklist_req -= 1; 833 } 834 softdep_worklist_busy = -1; 835 /* 836 * Alternately flush the block device associated with the mount 837 * point and process any dependencies that the flushing 838 * creates. We continue until no more worklist dependencies 839 * are found. 840 */ 841 *countp = 0; 842 devvp = VFSTOUFS(oldmnt)->um_devvp; 843 while ((count = softdep_process_worklist(oldmnt)) > 0) { 844 *countp += count; 845 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, td); 846 error = VOP_FSYNC(devvp, td->td_ucred, MNT_WAIT, td); 847 VOP_UNLOCK(devvp, 0, td); 848 if (error) 849 break; 850 } 851 softdep_worklist_busy = 0; 852 if (softdep_worklist_req) 853 wakeup(&softdep_worklist_req); 854 return (error); 855} 856 857/* 858 * Flush all vnodes and worklist items associated with a specified mount point. 859 */ 860int 861softdep_flushfiles(oldmnt, flags, td) 862 struct mount *oldmnt; 863 int flags; 864 struct thread *td; 865{ 866 int error, count, loopcnt; 867 868 error = 0; 869 870 /* 871 * Alternately flush the vnodes associated with the mount 872 * point and process any dependencies that the flushing 873 * creates. In theory, this loop can happen at most twice, 874 * but we give it a few extra just to be sure. 875 */ 876 for (loopcnt = 10; loopcnt > 0; loopcnt--) { 877 /* 878 * Do another flush in case any vnodes were brought in 879 * as part of the cleanup operations. 880 */ 881 if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0) 882 break; 883 if ((error = softdep_flushworklist(oldmnt, &count, td)) != 0 || 884 count == 0) 885 break; 886 } 887 /* 888 * If we are unmounting then it is an error to fail. If we 889 * are simply trying to downgrade to read-only, then filesystem 890 * activity can keep us busy forever, so we just fail with EBUSY. 891 */ 892 if (loopcnt == 0) { 893 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 894 panic("softdep_flushfiles: looping"); 895 error = EBUSY; 896 } 897 return (error); 898} 899 900/* 901 * Structure hashing. 902 * 903 * There are three types of structures that can be looked up: 904 * 1) pagedep structures identified by mount point, inode number, 905 * and logical block. 906 * 2) inodedep structures identified by mount point and inode number. 907 * 3) newblk structures identified by mount point and 908 * physical block number. 909 * 910 * The "pagedep" and "inodedep" dependency structures are hashed 911 * separately from the file blocks and inodes to which they correspond. 912 * This separation helps when the in-memory copy of an inode or 913 * file block must be replaced. It also obviates the need to access 914 * an inode or file page when simply updating (or de-allocating) 915 * dependency structures. Lookup of newblk structures is needed to 916 * find newly allocated blocks when trying to associate them with 917 * their allocdirect or allocindir structure. 918 * 919 * The lookup routines optionally create and hash a new instance when 920 * an existing entry is not found. 921 */ 922#define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 923#define NODELAY 0x0002 /* cannot do background work */ 924 925/* 926 * Structures and routines associated with pagedep caching. 927 */ 928LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl; 929u_long pagedep_hash; /* size of hash table - 1 */ 930#define PAGEDEP_HASH(mp, inum, lbn) \ 931 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \ 932 pagedep_hash]) 933static struct sema pagedep_in_progress; 934 935/* 936 * Look up a pagedep. Return 1 if found, 0 if not found or found 937 * when asked to allocate but not associated with any buffer. 938 * If not found, allocate if DEPALLOC flag is passed. 939 * Found or allocated entry is returned in pagedeppp. 940 * This routine must be called with splbio interrupts blocked. 941 */ 942static int 943pagedep_lookup(ip, lbn, flags, pagedeppp) 944 struct inode *ip; 945 ufs_lbn_t lbn; 946 int flags; 947 struct pagedep **pagedeppp; 948{ 949 struct pagedep *pagedep; 950 struct pagedep_hashhead *pagedephd; 951 struct mount *mp; 952 int i; 953 954#ifdef DEBUG 955 if (lk.lkt_held == NOHOLDER) 956 panic("pagedep_lookup: lock not held"); 957#endif 958 mp = ITOV(ip)->v_mount; 959 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn); 960top: 961 LIST_FOREACH(pagedep, pagedephd, pd_hash) 962 if (ip->i_number == pagedep->pd_ino && 963 lbn == pagedep->pd_lbn && 964 mp == pagedep->pd_mnt) 965 break; 966 if (pagedep) { 967 *pagedeppp = pagedep; 968 if ((flags & DEPALLOC) != 0 && 969 (pagedep->pd_state & ONWORKLIST) == 0) 970 return (0); 971 return (1); 972 } 973 if ((flags & DEPALLOC) == 0) { 974 *pagedeppp = NULL; 975 return (0); 976 } 977 if (sema_get(&pagedep_in_progress, &lk) == 0) { 978 ACQUIRE_LOCK(&lk); 979 goto top; 980 } 981 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP, 982 M_SOFTDEP_FLAGS|M_ZERO); 983 pagedep->pd_list.wk_type = D_PAGEDEP; 984 pagedep->pd_mnt = mp; 985 pagedep->pd_ino = ip->i_number; 986 pagedep->pd_lbn = lbn; 987 LIST_INIT(&pagedep->pd_dirremhd); 988 LIST_INIT(&pagedep->pd_pendinghd); 989 for (i = 0; i < DAHASHSZ; i++) 990 LIST_INIT(&pagedep->pd_diraddhd[i]); 991 ACQUIRE_LOCK(&lk); 992 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 993 sema_release(&pagedep_in_progress); 994 *pagedeppp = pagedep; 995 return (0); 996} 997 998/* 999 * Structures and routines associated with inodedep caching. 1000 */ 1001LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl; 1002static u_long inodedep_hash; /* size of hash table - 1 */ 1003static long num_inodedep; /* number of inodedep allocated */ 1004#define INODEDEP_HASH(fs, inum) \ 1005 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash]) 1006static struct sema inodedep_in_progress; 1007 1008/* 1009 * Look up an inodedep. Return 1 if found, 0 if not found. 1010 * If not found, allocate if DEPALLOC flag is passed. 1011 * Found or allocated entry is returned in inodedeppp. 1012 * This routine must be called with splbio interrupts blocked. 1013 */ 1014static int 1015inodedep_lookup(fs, inum, flags, inodedeppp) 1016 struct fs *fs; 1017 ino_t inum; 1018 int flags; 1019 struct inodedep **inodedeppp; 1020{ 1021 struct inodedep *inodedep; 1022 struct inodedep_hashhead *inodedephd; 1023 int firsttry; 1024 1025#ifdef DEBUG 1026 if (lk.lkt_held == NOHOLDER) 1027 panic("inodedep_lookup: lock not held"); 1028#endif 1029 firsttry = 1; 1030 inodedephd = INODEDEP_HASH(fs, inum); 1031top: 1032 LIST_FOREACH(inodedep, inodedephd, id_hash) 1033 if (inum == inodedep->id_ino && fs == inodedep->id_fs) 1034 break; 1035 if (inodedep) { 1036 *inodedeppp = inodedep; 1037 return (1); 1038 } 1039 if ((flags & DEPALLOC) == 0) { 1040 *inodedeppp = NULL; 1041 return (0); 1042 } 1043 /* 1044 * If we are over our limit, try to improve the situation. 1045 */ 1046 if (num_inodedep > max_softdeps && firsttry && (flags & NODELAY) == 0 && 1047 request_cleanup(FLUSH_INODES, 1)) { 1048 firsttry = 0; 1049 goto top; 1050 } 1051 if (sema_get(&inodedep_in_progress, &lk) == 0) { 1052 ACQUIRE_LOCK(&lk); 1053 goto top; 1054 } 1055 num_inodedep += 1; 1056 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep), 1057 M_INODEDEP, M_SOFTDEP_FLAGS); 1058 inodedep->id_list.wk_type = D_INODEDEP; 1059 inodedep->id_fs = fs; 1060 inodedep->id_ino = inum; 1061 inodedep->id_state = ALLCOMPLETE; 1062 inodedep->id_nlinkdelta = 0; 1063 inodedep->id_savedino1 = NULL; 1064 inodedep->id_savedsize = -1; 1065 inodedep->id_savedextsize = -1; 1066 inodedep->id_buf = NULL; 1067 LIST_INIT(&inodedep->id_pendinghd); 1068 LIST_INIT(&inodedep->id_inowait); 1069 LIST_INIT(&inodedep->id_bufwait); 1070 TAILQ_INIT(&inodedep->id_inoupdt); 1071 TAILQ_INIT(&inodedep->id_newinoupdt); 1072 TAILQ_INIT(&inodedep->id_extupdt); 1073 TAILQ_INIT(&inodedep->id_newextupdt); 1074 ACQUIRE_LOCK(&lk); 1075 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 1076 sema_release(&inodedep_in_progress); 1077 *inodedeppp = inodedep; 1078 return (0); 1079} 1080 1081/* 1082 * Structures and routines associated with newblk caching. 1083 */ 1084LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl; 1085u_long newblk_hash; /* size of hash table - 1 */ 1086#define NEWBLK_HASH(fs, inum) \ 1087 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash]) 1088static struct sema newblk_in_progress; 1089 1090/* 1091 * Look up a newblk. Return 1 if found, 0 if not found. 1092 * If not found, allocate if DEPALLOC flag is passed. 1093 * Found or allocated entry is returned in newblkpp. 1094 */ 1095static int 1096newblk_lookup(fs, newblkno, flags, newblkpp) 1097 struct fs *fs; 1098 ufs2_daddr_t newblkno; 1099 int flags; 1100 struct newblk **newblkpp; 1101{ 1102 struct newblk *newblk; 1103 struct newblk_hashhead *newblkhd; 1104 1105 newblkhd = NEWBLK_HASH(fs, newblkno); 1106top: 1107 LIST_FOREACH(newblk, newblkhd, nb_hash) 1108 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs) 1109 break; 1110 if (newblk) { 1111 *newblkpp = newblk; 1112 return (1); 1113 } 1114 if ((flags & DEPALLOC) == 0) { 1115 *newblkpp = NULL; 1116 return (0); 1117 } 1118 if (sema_get(&newblk_in_progress, 0) == 0) 1119 goto top; 1120 MALLOC(newblk, struct newblk *, sizeof(struct newblk), 1121 M_NEWBLK, M_SOFTDEP_FLAGS); 1122 newblk->nb_state = 0; 1123 newblk->nb_fs = fs; 1124 newblk->nb_newblkno = newblkno; 1125 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 1126 sema_release(&newblk_in_progress); 1127 *newblkpp = newblk; 1128 return (0); 1129} 1130 1131/* 1132 * Executed during filesystem system initialization before 1133 * mounting any filesystems. 1134 */ 1135void 1136softdep_initialize() 1137{ 1138 1139 LIST_INIT(&mkdirlisthd); 1140 LIST_INIT(&softdep_workitem_pending); 1141 max_softdeps = desiredvnodes * 4; 1142 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 1143 &pagedep_hash); 1144 sema_init(&pagedep_in_progress, "pagedep", PRIBIO, 0); 1145 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash); 1146 sema_init(&inodedep_in_progress, "inodedep", PRIBIO, 0); 1147 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash); 1148 sema_init(&newblk_in_progress, "newblk", PRIBIO, 0); 1149 1150 /* hooks through which the main kernel code calls us */ 1151 softdep_process_worklist_hook = softdep_process_worklist; 1152 softdep_fsync_hook = softdep_fsync; 1153 1154 /* initialise bioops hack */ 1155 bioops.io_start = softdep_disk_io_initiation; 1156 bioops.io_complete = softdep_disk_write_complete; 1157 bioops.io_deallocate = softdep_deallocate_dependencies; 1158 bioops.io_movedeps = softdep_move_dependencies; 1159 bioops.io_countdeps = softdep_count_dependencies; 1160} 1161 1162/* 1163 * Executed after all filesystems have been unmounted during 1164 * filesystem module unload. 1165 */ 1166void 1167softdep_uninitialize() 1168{ 1169 1170 softdep_process_worklist_hook = NULL; 1171 softdep_fsync_hook = NULL; 1172 hashdestroy(pagedep_hashtbl, M_PAGEDEP, pagedep_hash); 1173 hashdestroy(inodedep_hashtbl, M_INODEDEP, inodedep_hash); 1174 hashdestroy(newblk_hashtbl, M_NEWBLK, newblk_hash); 1175} 1176 1177/* 1178 * Called at mount time to notify the dependency code that a 1179 * filesystem wishes to use it. 1180 */ 1181int 1182softdep_mount(devvp, mp, fs, cred) 1183 struct vnode *devvp; 1184 struct mount *mp; 1185 struct fs *fs; 1186 struct ucred *cred; 1187{ 1188 struct csum_total cstotal; 1189 struct cg *cgp; 1190 struct buf *bp; 1191 int error, cyl; 1192 1193 mp->mnt_flag &= ~MNT_ASYNC; 1194 mp->mnt_flag |= MNT_SOFTDEP; 1195 /* 1196 * When doing soft updates, the counters in the 1197 * superblock may have gotten out of sync, so we have 1198 * to scan the cylinder groups and recalculate them. 1199 */ 1200 if (fs->fs_clean != 0) 1201 return (0); 1202 bzero(&cstotal, sizeof cstotal); 1203 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 1204 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 1205 fs->fs_cgsize, cred, &bp)) != 0) { 1206 brelse(bp); 1207 return (error); 1208 } 1209 cgp = (struct cg *)bp->b_data; 1210 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 1211 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 1212 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 1213 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 1214 fs->fs_cs(fs, cyl) = cgp->cg_cs; 1215 brelse(bp); 1216 } 1217#ifdef DEBUG 1218 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 1219 printf("%s: superblock summary recomputed\n", fs->fs_fsmnt); 1220#endif 1221 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 1222 return (0); 1223} 1224 1225/* 1226 * Protecting the freemaps (or bitmaps). 1227 * 1228 * To eliminate the need to execute fsck before mounting a filesystem 1229 * after a power failure, one must (conservatively) guarantee that the 1230 * on-disk copy of the bitmaps never indicate that a live inode or block is 1231 * free. So, when a block or inode is allocated, the bitmap should be 1232 * updated (on disk) before any new pointers. When a block or inode is 1233 * freed, the bitmap should not be updated until all pointers have been 1234 * reset. The latter dependency is handled by the delayed de-allocation 1235 * approach described below for block and inode de-allocation. The former 1236 * dependency is handled by calling the following procedure when a block or 1237 * inode is allocated. When an inode is allocated an "inodedep" is created 1238 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 1239 * Each "inodedep" is also inserted into the hash indexing structure so 1240 * that any additional link additions can be made dependent on the inode 1241 * allocation. 1242 * 1243 * The ufs filesystem maintains a number of free block counts (e.g., per 1244 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 1245 * in addition to the bitmaps. These counts are used to improve efficiency 1246 * during allocation and therefore must be consistent with the bitmaps. 1247 * There is no convenient way to guarantee post-crash consistency of these 1248 * counts with simple update ordering, for two main reasons: (1) The counts 1249 * and bitmaps for a single cylinder group block are not in the same disk 1250 * sector. If a disk write is interrupted (e.g., by power failure), one may 1251 * be written and the other not. (2) Some of the counts are located in the 1252 * superblock rather than the cylinder group block. So, we focus our soft 1253 * updates implementation on protecting the bitmaps. When mounting a 1254 * filesystem, we recompute the auxiliary counts from the bitmaps. 1255 */ 1256 1257/* 1258 * Called just after updating the cylinder group block to allocate an inode. 1259 */ 1260void 1261softdep_setup_inomapdep(bp, ip, newinum) 1262 struct buf *bp; /* buffer for cylgroup block with inode map */ 1263 struct inode *ip; /* inode related to allocation */ 1264 ino_t newinum; /* new inode number being allocated */ 1265{ 1266 struct inodedep *inodedep; 1267 struct bmsafemap *bmsafemap; 1268 1269 /* 1270 * Create a dependency for the newly allocated inode. 1271 * Panic if it already exists as something is seriously wrong. 1272 * Otherwise add it to the dependency list for the buffer holding 1273 * the cylinder group map from which it was allocated. 1274 */ 1275 ACQUIRE_LOCK(&lk); 1276 if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) { 1277 FREE_LOCK(&lk); 1278 panic("softdep_setup_inomapdep: found inode"); 1279 } 1280 inodedep->id_buf = bp; 1281 inodedep->id_state &= ~DEPCOMPLETE; 1282 bmsafemap = bmsafemap_lookup(bp); 1283 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 1284 FREE_LOCK(&lk); 1285} 1286 1287/* 1288 * Called just after updating the cylinder group block to 1289 * allocate block or fragment. 1290 */ 1291void 1292softdep_setup_blkmapdep(bp, fs, newblkno) 1293 struct buf *bp; /* buffer for cylgroup block with block map */ 1294 struct fs *fs; /* filesystem doing allocation */ 1295 ufs2_daddr_t newblkno; /* number of newly allocated block */ 1296{ 1297 struct newblk *newblk; 1298 struct bmsafemap *bmsafemap; 1299 1300 /* 1301 * Create a dependency for the newly allocated block. 1302 * Add it to the dependency list for the buffer holding 1303 * the cylinder group map from which it was allocated. 1304 */ 1305 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0) 1306 panic("softdep_setup_blkmapdep: found block"); 1307 ACQUIRE_LOCK(&lk); 1308 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp); 1309 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 1310 FREE_LOCK(&lk); 1311} 1312 1313/* 1314 * Find the bmsafemap associated with a cylinder group buffer. 1315 * If none exists, create one. The buffer must be locked when 1316 * this routine is called and this routine must be called with 1317 * splbio interrupts blocked. 1318 */ 1319static struct bmsafemap * 1320bmsafemap_lookup(bp) 1321 struct buf *bp; 1322{ 1323 struct bmsafemap *bmsafemap; 1324 struct worklist *wk; 1325 1326#ifdef DEBUG 1327 if (lk.lkt_held == NOHOLDER) 1328 panic("bmsafemap_lookup: lock not held"); 1329#endif 1330 LIST_FOREACH(wk, &bp->b_dep, wk_list) 1331 if (wk->wk_type == D_BMSAFEMAP) 1332 return (WK_BMSAFEMAP(wk)); 1333 FREE_LOCK(&lk); 1334 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap), 1335 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 1336 bmsafemap->sm_list.wk_type = D_BMSAFEMAP; 1337 bmsafemap->sm_list.wk_state = 0; 1338 bmsafemap->sm_buf = bp; 1339 LIST_INIT(&bmsafemap->sm_allocdirecthd); 1340 LIST_INIT(&bmsafemap->sm_allocindirhd); 1341 LIST_INIT(&bmsafemap->sm_inodedephd); 1342 LIST_INIT(&bmsafemap->sm_newblkhd); 1343 ACQUIRE_LOCK(&lk); 1344 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 1345 return (bmsafemap); 1346} 1347 1348/* 1349 * Direct block allocation dependencies. 1350 * 1351 * When a new block is allocated, the corresponding disk locations must be 1352 * initialized (with zeros or new data) before the on-disk inode points to 1353 * them. Also, the freemap from which the block was allocated must be 1354 * updated (on disk) before the inode's pointer. These two dependencies are 1355 * independent of each other and are needed for all file blocks and indirect 1356 * blocks that are pointed to directly by the inode. Just before the 1357 * "in-core" version of the inode is updated with a newly allocated block 1358 * number, a procedure (below) is called to setup allocation dependency 1359 * structures. These structures are removed when the corresponding 1360 * dependencies are satisfied or when the block allocation becomes obsolete 1361 * (i.e., the file is deleted, the block is de-allocated, or the block is a 1362 * fragment that gets upgraded). All of these cases are handled in 1363 * procedures described later. 1364 * 1365 * When a file extension causes a fragment to be upgraded, either to a larger 1366 * fragment or to a full block, the on-disk location may change (if the 1367 * previous fragment could not simply be extended). In this case, the old 1368 * fragment must be de-allocated, but not until after the inode's pointer has 1369 * been updated. In most cases, this is handled by later procedures, which 1370 * will construct a "freefrag" structure to be added to the workitem queue 1371 * when the inode update is complete (or obsolete). The main exception to 1372 * this is when an allocation occurs while a pending allocation dependency 1373 * (for the same block pointer) remains. This case is handled in the main 1374 * allocation dependency setup procedure by immediately freeing the 1375 * unreferenced fragments. 1376 */ 1377void 1378softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 1379 struct inode *ip; /* inode to which block is being added */ 1380 ufs_lbn_t lbn; /* block pointer within inode */ 1381 ufs2_daddr_t newblkno; /* disk block number being added */ 1382 ufs2_daddr_t oldblkno; /* previous block number, 0 unless frag */ 1383 long newsize; /* size of new block */ 1384 long oldsize; /* size of new block */ 1385 struct buf *bp; /* bp for allocated block */ 1386{ 1387 struct allocdirect *adp, *oldadp; 1388 struct allocdirectlst *adphead; 1389 struct bmsafemap *bmsafemap; 1390 struct inodedep *inodedep; 1391 struct pagedep *pagedep; 1392 struct newblk *newblk; 1393 1394 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect), 1395 M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO); 1396 adp->ad_list.wk_type = D_ALLOCDIRECT; 1397 adp->ad_lbn = lbn; 1398 adp->ad_newblkno = newblkno; 1399 adp->ad_oldblkno = oldblkno; 1400 adp->ad_newsize = newsize; 1401 adp->ad_oldsize = oldsize; 1402 adp->ad_state = ATTACHED; 1403 LIST_INIT(&adp->ad_newdirblk); 1404 if (newblkno == oldblkno) 1405 adp->ad_freefrag = NULL; 1406 else 1407 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1408 1409 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1410 panic("softdep_setup_allocdirect: lost block"); 1411 1412 ACQUIRE_LOCK(&lk); 1413 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep); 1414 adp->ad_inodedep = inodedep; 1415 1416 if (newblk->nb_state == DEPCOMPLETE) { 1417 adp->ad_state |= DEPCOMPLETE; 1418 adp->ad_buf = NULL; 1419 } else { 1420 bmsafemap = newblk->nb_bmsafemap; 1421 adp->ad_buf = bmsafemap->sm_buf; 1422 LIST_REMOVE(newblk, nb_deps); 1423 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1424 } 1425 LIST_REMOVE(newblk, nb_hash); 1426 FREE(newblk, M_NEWBLK); 1427 1428 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list); 1429 if (lbn >= NDADDR) { 1430 /* allocating an indirect block */ 1431 if (oldblkno != 0) { 1432 FREE_LOCK(&lk); 1433 panic("softdep_setup_allocdirect: non-zero indir"); 1434 } 1435 } else { 1436 /* 1437 * Allocating a direct block. 1438 * 1439 * If we are allocating a directory block, then we must 1440 * allocate an associated pagedep to track additions and 1441 * deletions. 1442 */ 1443 if ((ip->i_mode & IFMT) == IFDIR && 1444 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1445 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 1446 } 1447 /* 1448 * The list of allocdirects must be kept in sorted and ascending 1449 * order so that the rollback routines can quickly determine the 1450 * first uncommitted block (the size of the file stored on disk 1451 * ends at the end of the lowest committed fragment, or if there 1452 * are no fragments, at the end of the highest committed block). 1453 * Since files generally grow, the typical case is that the new 1454 * block is to be added at the end of the list. We speed this 1455 * special case by checking against the last allocdirect in the 1456 * list before laboriously traversing the list looking for the 1457 * insertion point. 1458 */ 1459 adphead = &inodedep->id_newinoupdt; 1460 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1461 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1462 /* insert at end of list */ 1463 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1464 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1465 allocdirect_merge(adphead, adp, oldadp); 1466 FREE_LOCK(&lk); 1467 return; 1468 } 1469 TAILQ_FOREACH(oldadp, adphead, ad_next) { 1470 if (oldadp->ad_lbn >= lbn) 1471 break; 1472 } 1473 if (oldadp == NULL) { 1474 FREE_LOCK(&lk); 1475 panic("softdep_setup_allocdirect: lost entry"); 1476 } 1477 /* insert in middle of list */ 1478 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1479 if (oldadp->ad_lbn == lbn) 1480 allocdirect_merge(adphead, adp, oldadp); 1481 FREE_LOCK(&lk); 1482} 1483 1484/* 1485 * Replace an old allocdirect dependency with a newer one. 1486 * This routine must be called with splbio interrupts blocked. 1487 */ 1488static void 1489allocdirect_merge(adphead, newadp, oldadp) 1490 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 1491 struct allocdirect *newadp; /* allocdirect being added */ 1492 struct allocdirect *oldadp; /* existing allocdirect being checked */ 1493{ 1494 struct worklist *wk; 1495 struct freefrag *freefrag; 1496 struct newdirblk *newdirblk; 1497 1498#ifdef DEBUG 1499 if (lk.lkt_held == NOHOLDER) 1500 panic("allocdirect_merge: lock not held"); 1501#endif 1502 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 1503 newadp->ad_oldsize != oldadp->ad_newsize || 1504 newadp->ad_lbn >= NDADDR) { 1505 FREE_LOCK(&lk); 1506 panic("%s %jd != new %jd || old size %ld != new %ld", 1507 "allocdirect_merge: old blkno", 1508 (intmax_t)newadp->ad_oldblkno, 1509 (intmax_t)oldadp->ad_newblkno, 1510 newadp->ad_oldsize, oldadp->ad_newsize); 1511 } 1512 newadp->ad_oldblkno = oldadp->ad_oldblkno; 1513 newadp->ad_oldsize = oldadp->ad_oldsize; 1514 /* 1515 * If the old dependency had a fragment to free or had never 1516 * previously had a block allocated, then the new dependency 1517 * can immediately post its freefrag and adopt the old freefrag. 1518 * This action is done by swapping the freefrag dependencies. 1519 * The new dependency gains the old one's freefrag, and the 1520 * old one gets the new one and then immediately puts it on 1521 * the worklist when it is freed by free_allocdirect. It is 1522 * not possible to do this swap when the old dependency had a 1523 * non-zero size but no previous fragment to free. This condition 1524 * arises when the new block is an extension of the old block. 1525 * Here, the first part of the fragment allocated to the new 1526 * dependency is part of the block currently claimed on disk by 1527 * the old dependency, so cannot legitimately be freed until the 1528 * conditions for the new dependency are fulfilled. 1529 */ 1530 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 1531 freefrag = newadp->ad_freefrag; 1532 newadp->ad_freefrag = oldadp->ad_freefrag; 1533 oldadp->ad_freefrag = freefrag; 1534 } 1535 /* 1536 * If we are tracking a new directory-block allocation, 1537 * move it from the old allocdirect to the new allocdirect. 1538 */ 1539 if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) { 1540 newdirblk = WK_NEWDIRBLK(wk); 1541 WORKLIST_REMOVE(&newdirblk->db_list); 1542 if (LIST_FIRST(&oldadp->ad_newdirblk) != NULL) 1543 panic("allocdirect_merge: extra newdirblk"); 1544 WORKLIST_INSERT(&newadp->ad_newdirblk, &newdirblk->db_list); 1545 } 1546 free_allocdirect(adphead, oldadp, 0); 1547} 1548 1549/* 1550 * Allocate a new freefrag structure if needed. 1551 */ 1552static struct freefrag * 1553newfreefrag(ip, blkno, size) 1554 struct inode *ip; 1555 ufs2_daddr_t blkno; 1556 long size; 1557{ 1558 struct freefrag *freefrag; 1559 struct fs *fs; 1560 1561 if (blkno == 0) 1562 return (NULL); 1563 fs = ip->i_fs; 1564 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 1565 panic("newfreefrag: frag size"); 1566 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag), 1567 M_FREEFRAG, M_SOFTDEP_FLAGS); 1568 freefrag->ff_list.wk_type = D_FREEFRAG; 1569 freefrag->ff_state = 0; 1570 freefrag->ff_inum = ip->i_number; 1571 freefrag->ff_mnt = ITOV(ip)->v_mount; 1572 freefrag->ff_blkno = blkno; 1573 freefrag->ff_fragsize = size; 1574 return (freefrag); 1575} 1576 1577/* 1578 * This workitem de-allocates fragments that were replaced during 1579 * file block allocation. 1580 */ 1581static void 1582handle_workitem_freefrag(freefrag) 1583 struct freefrag *freefrag; 1584{ 1585 struct ufsmount *ump = VFSTOUFS(freefrag->ff_mnt); 1586 1587 ffs_blkfree(ump->um_fs, ump->um_devvp, freefrag->ff_blkno, 1588 freefrag->ff_fragsize, freefrag->ff_inum); 1589 FREE(freefrag, M_FREEFRAG); 1590} 1591 1592/* 1593 * Set up a dependency structure for an external attributes data block. 1594 * This routine follows much of the structure of softdep_setup_allocdirect. 1595 * See the description of softdep_setup_allocdirect above for details. 1596 */ 1597void 1598softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 1599 struct inode *ip; 1600 ufs_lbn_t lbn; 1601 ufs2_daddr_t newblkno; 1602 ufs2_daddr_t oldblkno; 1603 long newsize; 1604 long oldsize; 1605 struct buf *bp; 1606{ 1607 struct allocdirect *adp, *oldadp; 1608 struct allocdirectlst *adphead; 1609 struct bmsafemap *bmsafemap; 1610 struct inodedep *inodedep; 1611 struct newblk *newblk; 1612 1613 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect), 1614 M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO); 1615 adp->ad_list.wk_type = D_ALLOCDIRECT; 1616 adp->ad_lbn = lbn; 1617 adp->ad_newblkno = newblkno; 1618 adp->ad_oldblkno = oldblkno; 1619 adp->ad_newsize = newsize; 1620 adp->ad_oldsize = oldsize; 1621 adp->ad_state = ATTACHED | EXTDATA; 1622 LIST_INIT(&adp->ad_newdirblk); 1623 if (newblkno == oldblkno) 1624 adp->ad_freefrag = NULL; 1625 else 1626 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1627 1628 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1629 panic("softdep_setup_allocext: lost block"); 1630 1631 ACQUIRE_LOCK(&lk); 1632 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep); 1633 adp->ad_inodedep = inodedep; 1634 1635 if (newblk->nb_state == DEPCOMPLETE) { 1636 adp->ad_state |= DEPCOMPLETE; 1637 adp->ad_buf = NULL; 1638 } else { 1639 bmsafemap = newblk->nb_bmsafemap; 1640 adp->ad_buf = bmsafemap->sm_buf; 1641 LIST_REMOVE(newblk, nb_deps); 1642 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1643 } 1644 LIST_REMOVE(newblk, nb_hash); 1645 FREE(newblk, M_NEWBLK); 1646 1647 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list); 1648 if (lbn >= NXADDR) { 1649 FREE_LOCK(&lk); 1650 panic("softdep_setup_allocext: lbn %lld > NXADDR", 1651 (long long)lbn); 1652 } 1653 /* 1654 * The list of allocdirects must be kept in sorted and ascending 1655 * order so that the rollback routines can quickly determine the 1656 * first uncommitted block (the size of the file stored on disk 1657 * ends at the end of the lowest committed fragment, or if there 1658 * are no fragments, at the end of the highest committed block). 1659 * Since files generally grow, the typical case is that the new 1660 * block is to be added at the end of the list. We speed this 1661 * special case by checking against the last allocdirect in the 1662 * list before laboriously traversing the list looking for the 1663 * insertion point. 1664 */ 1665 adphead = &inodedep->id_newextupdt; 1666 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1667 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1668 /* insert at end of list */ 1669 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1670 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1671 allocdirect_merge(adphead, adp, oldadp); 1672 FREE_LOCK(&lk); 1673 return; 1674 } 1675 TAILQ_FOREACH(oldadp, adphead, ad_next) { 1676 if (oldadp->ad_lbn >= lbn) 1677 break; 1678 } 1679 if (oldadp == NULL) { 1680 FREE_LOCK(&lk); 1681 panic("softdep_setup_allocext: lost entry"); 1682 } 1683 /* insert in middle of list */ 1684 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1685 if (oldadp->ad_lbn == lbn) 1686 allocdirect_merge(adphead, adp, oldadp); 1687 FREE_LOCK(&lk); 1688} 1689 1690/* 1691 * Indirect block allocation dependencies. 1692 * 1693 * The same dependencies that exist for a direct block also exist when 1694 * a new block is allocated and pointed to by an entry in a block of 1695 * indirect pointers. The undo/redo states described above are also 1696 * used here. Because an indirect block contains many pointers that 1697 * may have dependencies, a second copy of the entire in-memory indirect 1698 * block is kept. The buffer cache copy is always completely up-to-date. 1699 * The second copy, which is used only as a source for disk writes, 1700 * contains only the safe pointers (i.e., those that have no remaining 1701 * update dependencies). The second copy is freed when all pointers 1702 * are safe. The cache is not allowed to replace indirect blocks with 1703 * pending update dependencies. If a buffer containing an indirect 1704 * block with dependencies is written, these routines will mark it 1705 * dirty again. It can only be successfully written once all the 1706 * dependencies are removed. The ffs_fsync routine in conjunction with 1707 * softdep_sync_metadata work together to get all the dependencies 1708 * removed so that a file can be successfully written to disk. Three 1709 * procedures are used when setting up indirect block pointer 1710 * dependencies. The division is necessary because of the organization 1711 * of the "balloc" routine and because of the distinction between file 1712 * pages and file metadata blocks. 1713 */ 1714 1715/* 1716 * Allocate a new allocindir structure. 1717 */ 1718static struct allocindir * 1719newallocindir(ip, ptrno, newblkno, oldblkno) 1720 struct inode *ip; /* inode for file being extended */ 1721 int ptrno; /* offset of pointer in indirect block */ 1722 ufs2_daddr_t newblkno; /* disk block number being added */ 1723 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 1724{ 1725 struct allocindir *aip; 1726 1727 MALLOC(aip, struct allocindir *, sizeof(struct allocindir), 1728 M_ALLOCINDIR, M_SOFTDEP_FLAGS|M_ZERO); 1729 aip->ai_list.wk_type = D_ALLOCINDIR; 1730 aip->ai_state = ATTACHED; 1731 aip->ai_offset = ptrno; 1732 aip->ai_newblkno = newblkno; 1733 aip->ai_oldblkno = oldblkno; 1734 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize); 1735 return (aip); 1736} 1737 1738/* 1739 * Called just before setting an indirect block pointer 1740 * to a newly allocated file page. 1741 */ 1742void 1743softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 1744 struct inode *ip; /* inode for file being extended */ 1745 ufs_lbn_t lbn; /* allocated block number within file */ 1746 struct buf *bp; /* buffer with indirect blk referencing page */ 1747 int ptrno; /* offset of pointer in indirect block */ 1748 ufs2_daddr_t newblkno; /* disk block number being added */ 1749 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 1750 struct buf *nbp; /* buffer holding allocated page */ 1751{ 1752 struct allocindir *aip; 1753 struct pagedep *pagedep; 1754 1755 aip = newallocindir(ip, ptrno, newblkno, oldblkno); 1756 ACQUIRE_LOCK(&lk); 1757 /* 1758 * If we are allocating a directory page, then we must 1759 * allocate an associated pagedep to track additions and 1760 * deletions. 1761 */ 1762 if ((ip->i_mode & IFMT) == IFDIR && 1763 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1764 WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list); 1765 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1766 FREE_LOCK(&lk); 1767 setup_allocindir_phase2(bp, ip, aip); 1768} 1769 1770/* 1771 * Called just before setting an indirect block pointer to a 1772 * newly allocated indirect block. 1773 */ 1774void 1775softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 1776 struct buf *nbp; /* newly allocated indirect block */ 1777 struct inode *ip; /* inode for file being extended */ 1778 struct buf *bp; /* indirect block referencing allocated block */ 1779 int ptrno; /* offset of pointer in indirect block */ 1780 ufs2_daddr_t newblkno; /* disk block number being added */ 1781{ 1782 struct allocindir *aip; 1783 1784 aip = newallocindir(ip, ptrno, newblkno, 0); 1785 ACQUIRE_LOCK(&lk); 1786 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1787 FREE_LOCK(&lk); 1788 setup_allocindir_phase2(bp, ip, aip); 1789} 1790 1791/* 1792 * Called to finish the allocation of the "aip" allocated 1793 * by one of the two routines above. 1794 */ 1795static void 1796setup_allocindir_phase2(bp, ip, aip) 1797 struct buf *bp; /* in-memory copy of the indirect block */ 1798 struct inode *ip; /* inode for file being extended */ 1799 struct allocindir *aip; /* allocindir allocated by the above routines */ 1800{ 1801 struct worklist *wk; 1802 struct indirdep *indirdep, *newindirdep; 1803 struct bmsafemap *bmsafemap; 1804 struct allocindir *oldaip; 1805 struct freefrag *freefrag; 1806 struct newblk *newblk; 1807 ufs2_daddr_t blkno; 1808 1809 if (bp->b_lblkno >= 0) 1810 panic("setup_allocindir_phase2: not indir blk"); 1811 for (indirdep = NULL, newindirdep = NULL; ; ) { 1812 ACQUIRE_LOCK(&lk); 1813 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 1814 if (wk->wk_type != D_INDIRDEP) 1815 continue; 1816 indirdep = WK_INDIRDEP(wk); 1817 break; 1818 } 1819 if (indirdep == NULL && newindirdep) { 1820 indirdep = newindirdep; 1821 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 1822 newindirdep = NULL; 1823 } 1824 FREE_LOCK(&lk); 1825 if (indirdep) { 1826 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0, 1827 &newblk) == 0) 1828 panic("setup_allocindir: lost block"); 1829 ACQUIRE_LOCK(&lk); 1830 if (newblk->nb_state == DEPCOMPLETE) { 1831 aip->ai_state |= DEPCOMPLETE; 1832 aip->ai_buf = NULL; 1833 } else { 1834 bmsafemap = newblk->nb_bmsafemap; 1835 aip->ai_buf = bmsafemap->sm_buf; 1836 LIST_REMOVE(newblk, nb_deps); 1837 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd, 1838 aip, ai_deps); 1839 } 1840 LIST_REMOVE(newblk, nb_hash); 1841 FREE(newblk, M_NEWBLK); 1842 aip->ai_indirdep = indirdep; 1843 /* 1844 * Check to see if there is an existing dependency 1845 * for this block. If there is, merge the old 1846 * dependency into the new one. 1847 */ 1848 if (aip->ai_oldblkno == 0) 1849 oldaip = NULL; 1850 else 1851 1852 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) 1853 if (oldaip->ai_offset == aip->ai_offset) 1854 break; 1855 freefrag = NULL; 1856 if (oldaip != NULL) { 1857 if (oldaip->ai_newblkno != aip->ai_oldblkno) { 1858 FREE_LOCK(&lk); 1859 panic("setup_allocindir_phase2: blkno"); 1860 } 1861 aip->ai_oldblkno = oldaip->ai_oldblkno; 1862 freefrag = aip->ai_freefrag; 1863 aip->ai_freefrag = oldaip->ai_freefrag; 1864 oldaip->ai_freefrag = NULL; 1865 free_allocindir(oldaip, NULL); 1866 } 1867 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 1868 if (ip->i_ump->um_fstype == UFS1) 1869 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data) 1870 [aip->ai_offset] = aip->ai_oldblkno; 1871 else 1872 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data) 1873 [aip->ai_offset] = aip->ai_oldblkno; 1874 FREE_LOCK(&lk); 1875 if (freefrag != NULL) 1876 handle_workitem_freefrag(freefrag); 1877 } 1878 if (newindirdep) { 1879 brelse(newindirdep->ir_savebp); 1880 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP); 1881 } 1882 if (indirdep) 1883 break; 1884 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep), 1885 M_INDIRDEP, M_SOFTDEP_FLAGS); 1886 newindirdep->ir_list.wk_type = D_INDIRDEP; 1887 newindirdep->ir_state = ATTACHED; 1888 if (ip->i_ump->um_fstype == UFS1) 1889 newindirdep->ir_state |= UFS1FMT; 1890 LIST_INIT(&newindirdep->ir_deplisthd); 1891 LIST_INIT(&newindirdep->ir_donehd); 1892 if (bp->b_blkno == bp->b_lblkno) { 1893 ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp, 1894 NULL, NULL); 1895 bp->b_blkno = blkno; 1896 } 1897 newindirdep->ir_savebp = 1898 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0); 1899 BUF_KERNPROC(newindirdep->ir_savebp); 1900 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 1901 } 1902} 1903 1904/* 1905 * Block de-allocation dependencies. 1906 * 1907 * When blocks are de-allocated, the on-disk pointers must be nullified before 1908 * the blocks are made available for use by other files. (The true 1909 * requirement is that old pointers must be nullified before new on-disk 1910 * pointers are set. We chose this slightly more stringent requirement to 1911 * reduce complexity.) Our implementation handles this dependency by updating 1912 * the inode (or indirect block) appropriately but delaying the actual block 1913 * de-allocation (i.e., freemap and free space count manipulation) until 1914 * after the updated versions reach stable storage. After the disk is 1915 * updated, the blocks can be safely de-allocated whenever it is convenient. 1916 * This implementation handles only the common case of reducing a file's 1917 * length to zero. Other cases are handled by the conventional synchronous 1918 * write approach. 1919 * 1920 * The ffs implementation with which we worked double-checks 1921 * the state of the block pointers and file size as it reduces 1922 * a file's length. Some of this code is replicated here in our 1923 * soft updates implementation. The freeblks->fb_chkcnt field is 1924 * used to transfer a part of this information to the procedure 1925 * that eventually de-allocates the blocks. 1926 * 1927 * This routine should be called from the routine that shortens 1928 * a file's length, before the inode's size or block pointers 1929 * are modified. It will save the block pointer information for 1930 * later release and zero the inode so that the calling routine 1931 * can release it. 1932 */ 1933void 1934softdep_setup_freeblocks(ip, length, flags) 1935 struct inode *ip; /* The inode whose length is to be reduced */ 1936 off_t length; /* The new length for the file */ 1937 int flags; /* IO_EXT and/or IO_NORMAL */ 1938{ 1939 struct freeblks *freeblks; 1940 struct inodedep *inodedep; 1941 struct allocdirect *adp; 1942 struct vnode *vp; 1943 struct buf *bp; 1944 struct fs *fs; 1945 ufs2_daddr_t extblocks, datablocks; 1946 int i, delay, error; 1947 1948 fs = ip->i_fs; 1949 if (length != 0) 1950 panic("softdep_setup_freeblocks: non-zero length"); 1951 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks), 1952 M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO); 1953 freeblks->fb_list.wk_type = D_FREEBLKS; 1954 freeblks->fb_uid = ip->i_uid; 1955 freeblks->fb_previousinum = ip->i_number; 1956 freeblks->fb_devvp = ip->i_devvp; 1957 freeblks->fb_mnt = ITOV(ip)->v_mount; 1958 extblocks = 0; 1959 if (fs->fs_magic == FS_UFS2_MAGIC) 1960 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 1961 datablocks = DIP(ip, i_blocks) - extblocks; 1962 if ((flags & IO_NORMAL) == 0) { 1963 freeblks->fb_oldsize = 0; 1964 freeblks->fb_chkcnt = 0; 1965 } else { 1966 freeblks->fb_oldsize = ip->i_size; 1967 ip->i_size = 0; 1968 DIP(ip, i_size) = 0; 1969 freeblks->fb_chkcnt = datablocks; 1970 for (i = 0; i < NDADDR; i++) { 1971 freeblks->fb_dblks[i] = DIP(ip, i_db[i]); 1972 DIP(ip, i_db[i]) = 0; 1973 } 1974 for (i = 0; i < NIADDR; i++) { 1975 freeblks->fb_iblks[i] = DIP(ip, i_ib[i]); 1976 DIP(ip, i_ib[i]) = 0; 1977 } 1978 /* 1979 * If the file was removed, then the space being freed was 1980 * accounted for then (see softdep_filereleased()). If the 1981 * file is merely being truncated, then we account for it now. 1982 */ 1983 if ((ip->i_flag & IN_SPACECOUNTED) == 0) 1984 fs->fs_pendingblocks += datablocks; 1985 } 1986 if ((flags & IO_EXT) == 0) { 1987 freeblks->fb_oldextsize = 0; 1988 } else { 1989 freeblks->fb_oldextsize = ip->i_din2->di_extsize; 1990 ip->i_din2->di_extsize = 0; 1991 freeblks->fb_chkcnt += extblocks; 1992 for (i = 0; i < NXADDR; i++) { 1993 freeblks->fb_eblks[i] = ip->i_din2->di_extb[i]; 1994 ip->i_din2->di_extb[i] = 0; 1995 } 1996 } 1997 DIP(ip, i_blocks) -= freeblks->fb_chkcnt; 1998 /* 1999 * Push the zero'ed inode to to its disk buffer so that we are free 2000 * to delete its dependencies below. Once the dependencies are gone 2001 * the buffer can be safely released. 2002 */ 2003 if ((error = bread(ip->i_devvp, 2004 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 2005 (int)fs->fs_bsize, NOCRED, &bp)) != 0) { 2006 brelse(bp); 2007 softdep_error("softdep_setup_freeblocks", error); 2008 } 2009 if (ip->i_ump->um_fstype == UFS1) 2010 *((struct ufs1_dinode *)bp->b_data + 2011 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1; 2012 else 2013 *((struct ufs2_dinode *)bp->b_data + 2014 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2; 2015 /* 2016 * Find and eliminate any inode dependencies. 2017 */ 2018 ACQUIRE_LOCK(&lk); 2019 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep); 2020 if ((inodedep->id_state & IOSTARTED) != 0) { 2021 FREE_LOCK(&lk); 2022 panic("softdep_setup_freeblocks: inode busy"); 2023 } 2024 /* 2025 * Add the freeblks structure to the list of operations that 2026 * must await the zero'ed inode being written to disk. If we 2027 * still have a bitmap dependency (delay == 0), then the inode 2028 * has never been written to disk, so we can process the 2029 * freeblks below once we have deleted the dependencies. 2030 */ 2031 delay = (inodedep->id_state & DEPCOMPLETE); 2032 if (delay) 2033 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list); 2034 /* 2035 * Because the file length has been truncated to zero, any 2036 * pending block allocation dependency structures associated 2037 * with this inode are obsolete and can simply be de-allocated. 2038 * We must first merge the two dependency lists to get rid of 2039 * any duplicate freefrag structures, then purge the merged list. 2040 * If we still have a bitmap dependency, then the inode has never 2041 * been written to disk, so we can free any fragments without delay. 2042 */ 2043 if (flags & IO_NORMAL) { 2044 merge_inode_lists(&inodedep->id_newinoupdt, 2045 &inodedep->id_inoupdt); 2046 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 2047 free_allocdirect(&inodedep->id_inoupdt, adp, delay); 2048 } 2049 if (flags & IO_EXT) { 2050 merge_inode_lists(&inodedep->id_newextupdt, 2051 &inodedep->id_extupdt); 2052 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0) 2053 free_allocdirect(&inodedep->id_extupdt, adp, delay); 2054 } 2055 FREE_LOCK(&lk); 2056 bdwrite(bp); 2057 /* 2058 * We must wait for any I/O in progress to finish so that 2059 * all potential buffers on the dirty list will be visible. 2060 * Once they are all there, walk the list and get rid of 2061 * any dependencies. 2062 */ 2063 vp = ITOV(ip); 2064 ACQUIRE_LOCK(&lk); 2065 VI_LOCK(vp); 2066 drain_output(vp, 1); 2067restart: 2068 TAILQ_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) { 2069 if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) || 2070 ((flags & IO_NORMAL) == 0 && 2071 (bp->b_xflags & BX_ALTDATA) == 0)) 2072 continue; 2073 if ((bp = getdirtybuf(&bp, VI_MTX(vp), MNT_WAIT)) == NULL) 2074 goto restart; 2075 (void) inodedep_lookup(fs, ip->i_number, 0, &inodedep); 2076 deallocate_dependencies(bp, inodedep); 2077 bp->b_flags |= B_INVAL | B_NOCACHE; 2078 FREE_LOCK(&lk); 2079 brelse(bp); 2080 ACQUIRE_LOCK(&lk); 2081 VI_LOCK(vp); 2082 goto restart; 2083 } 2084 VI_UNLOCK(vp); 2085 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0) 2086 (void) free_inodedep(inodedep); 2087 FREE_LOCK(&lk); 2088 /* 2089 * If the inode has never been written to disk (delay == 0), 2090 * then we can process the freeblks now that we have deleted 2091 * the dependencies. 2092 */ 2093 if (!delay) 2094 handle_workitem_freeblocks(freeblks, 0); 2095} 2096 2097/* 2098 * Reclaim any dependency structures from a buffer that is about to 2099 * be reallocated to a new vnode. The buffer must be locked, thus, 2100 * no I/O completion operations can occur while we are manipulating 2101 * its associated dependencies. The mutex is held so that other I/O's 2102 * associated with related dependencies do not occur. 2103 */ 2104static void 2105deallocate_dependencies(bp, inodedep) 2106 struct buf *bp; 2107 struct inodedep *inodedep; 2108{ 2109 struct worklist *wk; 2110 struct indirdep *indirdep; 2111 struct allocindir *aip; 2112 struct pagedep *pagedep; 2113 struct dirrem *dirrem; 2114 struct diradd *dap; 2115 int i; 2116 2117 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2118 switch (wk->wk_type) { 2119 2120 case D_INDIRDEP: 2121 indirdep = WK_INDIRDEP(wk); 2122 /* 2123 * None of the indirect pointers will ever be visible, 2124 * so they can simply be tossed. GOINGAWAY ensures 2125 * that allocated pointers will be saved in the buffer 2126 * cache until they are freed. Note that they will 2127 * only be able to be found by their physical address 2128 * since the inode mapping the logical address will 2129 * be gone. The save buffer used for the safe copy 2130 * was allocated in setup_allocindir_phase2 using 2131 * the physical address so it could be used for this 2132 * purpose. Hence we swap the safe copy with the real 2133 * copy, allowing the safe copy to be freed and holding 2134 * on to the real copy for later use in indir_trunc. 2135 */ 2136 if (indirdep->ir_state & GOINGAWAY) { 2137 FREE_LOCK(&lk); 2138 panic("deallocate_dependencies: already gone"); 2139 } 2140 indirdep->ir_state |= GOINGAWAY; 2141 VFSTOUFS(bp->b_vp->v_mount)->um_numindirdeps += 1; 2142 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 2143 free_allocindir(aip, inodedep); 2144 if (bp->b_lblkno >= 0 || 2145 bp->b_blkno != indirdep->ir_savebp->b_lblkno) { 2146 FREE_LOCK(&lk); 2147 panic("deallocate_dependencies: not indir"); 2148 } 2149 bcopy(bp->b_data, indirdep->ir_savebp->b_data, 2150 bp->b_bcount); 2151 WORKLIST_REMOVE(wk); 2152 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk); 2153 continue; 2154 2155 case D_PAGEDEP: 2156 pagedep = WK_PAGEDEP(wk); 2157 /* 2158 * None of the directory additions will ever be 2159 * visible, so they can simply be tossed. 2160 */ 2161 for (i = 0; i < DAHASHSZ; i++) 2162 while ((dap = 2163 LIST_FIRST(&pagedep->pd_diraddhd[i]))) 2164 free_diradd(dap); 2165 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0) 2166 free_diradd(dap); 2167 /* 2168 * Copy any directory remove dependencies to the list 2169 * to be processed after the zero'ed inode is written. 2170 * If the inode has already been written, then they 2171 * can be dumped directly onto the work list. 2172 */ 2173 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 2174 LIST_REMOVE(dirrem, dm_next); 2175 dirrem->dm_dirinum = pagedep->pd_ino; 2176 if (inodedep == NULL || 2177 (inodedep->id_state & ALLCOMPLETE) == 2178 ALLCOMPLETE) 2179 add_to_worklist(&dirrem->dm_list); 2180 else 2181 WORKLIST_INSERT(&inodedep->id_bufwait, 2182 &dirrem->dm_list); 2183 } 2184 if ((pagedep->pd_state & NEWBLOCK) != 0) { 2185 LIST_FOREACH(wk, &inodedep->id_bufwait, wk_list) 2186 if (wk->wk_type == D_NEWDIRBLK && 2187 WK_NEWDIRBLK(wk)->db_pagedep == 2188 pagedep) 2189 break; 2190 if (wk != NULL) { 2191 WORKLIST_REMOVE(wk); 2192 free_newdirblk(WK_NEWDIRBLK(wk)); 2193 } else { 2194 FREE_LOCK(&lk); 2195 panic("deallocate_dependencies: " 2196 "lost pagedep"); 2197 } 2198 } 2199 WORKLIST_REMOVE(&pagedep->pd_list); 2200 LIST_REMOVE(pagedep, pd_hash); 2201 WORKITEM_FREE(pagedep, D_PAGEDEP); 2202 continue; 2203 2204 case D_ALLOCINDIR: 2205 free_allocindir(WK_ALLOCINDIR(wk), inodedep); 2206 continue; 2207 2208 case D_ALLOCDIRECT: 2209 case D_INODEDEP: 2210 FREE_LOCK(&lk); 2211 panic("deallocate_dependencies: Unexpected type %s", 2212 TYPENAME(wk->wk_type)); 2213 /* NOTREACHED */ 2214 2215 default: 2216 FREE_LOCK(&lk); 2217 panic("deallocate_dependencies: Unknown type %s", 2218 TYPENAME(wk->wk_type)); 2219 /* NOTREACHED */ 2220 } 2221 } 2222} 2223 2224/* 2225 * Free an allocdirect. Generate a new freefrag work request if appropriate. 2226 * This routine must be called with splbio interrupts blocked. 2227 */ 2228static void 2229free_allocdirect(adphead, adp, delay) 2230 struct allocdirectlst *adphead; 2231 struct allocdirect *adp; 2232 int delay; 2233{ 2234 struct newdirblk *newdirblk; 2235 struct worklist *wk; 2236 2237#ifdef DEBUG 2238 if (lk.lkt_held == NOHOLDER) 2239 panic("free_allocdirect: lock not held"); 2240#endif 2241 if ((adp->ad_state & DEPCOMPLETE) == 0) 2242 LIST_REMOVE(adp, ad_deps); 2243 TAILQ_REMOVE(adphead, adp, ad_next); 2244 if ((adp->ad_state & COMPLETE) == 0) 2245 WORKLIST_REMOVE(&adp->ad_list); 2246 if (adp->ad_freefrag != NULL) { 2247 if (delay) 2248 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 2249 &adp->ad_freefrag->ff_list); 2250 else 2251 add_to_worklist(&adp->ad_freefrag->ff_list); 2252 } 2253 if ((wk = LIST_FIRST(&adp->ad_newdirblk)) != NULL) { 2254 newdirblk = WK_NEWDIRBLK(wk); 2255 WORKLIST_REMOVE(&newdirblk->db_list); 2256 if (LIST_FIRST(&adp->ad_newdirblk) != NULL) 2257 panic("free_allocdirect: extra newdirblk"); 2258 if (delay) 2259 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 2260 &newdirblk->db_list); 2261 else 2262 free_newdirblk(newdirblk); 2263 } 2264 WORKITEM_FREE(adp, D_ALLOCDIRECT); 2265} 2266 2267/* 2268 * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep. 2269 * This routine must be called with splbio interrupts blocked. 2270 */ 2271static void 2272free_newdirblk(newdirblk) 2273 struct newdirblk *newdirblk; 2274{ 2275 struct pagedep *pagedep; 2276 struct diradd *dap; 2277 int i; 2278 2279#ifdef DEBUG 2280 if (lk.lkt_held == NOHOLDER) 2281 panic("free_newdirblk: lock not held"); 2282#endif 2283 /* 2284 * If the pagedep is still linked onto the directory buffer 2285 * dependency chain, then some of the entries on the 2286 * pd_pendinghd list may not be committed to disk yet. In 2287 * this case, we will simply clear the NEWBLOCK flag and 2288 * let the pd_pendinghd list be processed when the pagedep 2289 * is next written. If the pagedep is no longer on the buffer 2290 * dependency chain, then all the entries on the pd_pending 2291 * list are committed to disk and we can free them here. 2292 */ 2293 pagedep = newdirblk->db_pagedep; 2294 pagedep->pd_state &= ~NEWBLOCK; 2295 if ((pagedep->pd_state & ONWORKLIST) == 0) 2296 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 2297 free_diradd(dap); 2298 /* 2299 * If no dependencies remain, the pagedep will be freed. 2300 */ 2301 for (i = 0; i < DAHASHSZ; i++) 2302 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL) 2303 break; 2304 if (i == DAHASHSZ && (pagedep->pd_state & ONWORKLIST) == 0) { 2305 LIST_REMOVE(pagedep, pd_hash); 2306 WORKITEM_FREE(pagedep, D_PAGEDEP); 2307 } 2308 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 2309} 2310 2311/* 2312 * Prepare an inode to be freed. The actual free operation is not 2313 * done until the zero'ed inode has been written to disk. 2314 */ 2315void 2316softdep_freefile(pvp, ino, mode) 2317 struct vnode *pvp; 2318 ino_t ino; 2319 int mode; 2320{ 2321 struct inode *ip = VTOI(pvp); 2322 struct inodedep *inodedep; 2323 struct freefile *freefile; 2324 2325 /* 2326 * This sets up the inode de-allocation dependency. 2327 */ 2328 MALLOC(freefile, struct freefile *, sizeof(struct freefile), 2329 M_FREEFILE, M_SOFTDEP_FLAGS); 2330 freefile->fx_list.wk_type = D_FREEFILE; 2331 freefile->fx_list.wk_state = 0; 2332 freefile->fx_mode = mode; 2333 freefile->fx_oldinum = ino; 2334 freefile->fx_devvp = ip->i_devvp; 2335 freefile->fx_mnt = ITOV(ip)->v_mount; 2336 if ((ip->i_flag & IN_SPACECOUNTED) == 0) 2337 ip->i_fs->fs_pendinginodes += 1; 2338 2339 /* 2340 * If the inodedep does not exist, then the zero'ed inode has 2341 * been written to disk. If the allocated inode has never been 2342 * written to disk, then the on-disk inode is zero'ed. In either 2343 * case we can free the file immediately. 2344 */ 2345 ACQUIRE_LOCK(&lk); 2346 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 || 2347 check_inode_unwritten(inodedep)) { 2348 FREE_LOCK(&lk); 2349 handle_workitem_freefile(freefile); 2350 return; 2351 } 2352 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 2353 FREE_LOCK(&lk); 2354} 2355 2356/* 2357 * Check to see if an inode has never been written to disk. If 2358 * so free the inodedep and return success, otherwise return failure. 2359 * This routine must be called with splbio interrupts blocked. 2360 * 2361 * If we still have a bitmap dependency, then the inode has never 2362 * been written to disk. Drop the dependency as it is no longer 2363 * necessary since the inode is being deallocated. We set the 2364 * ALLCOMPLETE flags since the bitmap now properly shows that the 2365 * inode is not allocated. Even if the inode is actively being 2366 * written, it has been rolled back to its zero'ed state, so we 2367 * are ensured that a zero inode is what is on the disk. For short 2368 * lived files, this change will usually result in removing all the 2369 * dependencies from the inode so that it can be freed immediately. 2370 */ 2371static int 2372check_inode_unwritten(inodedep) 2373 struct inodedep *inodedep; 2374{ 2375 2376 if ((inodedep->id_state & DEPCOMPLETE) != 0 || 2377 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2378 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2379 LIST_FIRST(&inodedep->id_inowait) != NULL || 2380 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2381 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2382 TAILQ_FIRST(&inodedep->id_extupdt) != NULL || 2383 TAILQ_FIRST(&inodedep->id_newextupdt) != NULL || 2384 inodedep->id_nlinkdelta != 0) 2385 return (0); 2386 inodedep->id_state |= ALLCOMPLETE; 2387 LIST_REMOVE(inodedep, id_deps); 2388 inodedep->id_buf = NULL; 2389 if (inodedep->id_state & ONWORKLIST) 2390 WORKLIST_REMOVE(&inodedep->id_list); 2391 if (inodedep->id_savedino1 != NULL) { 2392 FREE(inodedep->id_savedino1, M_INODEDEP); 2393 inodedep->id_savedino1 = NULL; 2394 } 2395 if (free_inodedep(inodedep) == 0) { 2396 FREE_LOCK(&lk); 2397 panic("check_inode_unwritten: busy inode"); 2398 } 2399 return (1); 2400} 2401 2402/* 2403 * Try to free an inodedep structure. Return 1 if it could be freed. 2404 */ 2405static int 2406free_inodedep(inodedep) 2407 struct inodedep *inodedep; 2408{ 2409 2410 if ((inodedep->id_state & ONWORKLIST) != 0 || 2411 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 2412 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2413 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2414 LIST_FIRST(&inodedep->id_inowait) != NULL || 2415 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2416 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2417 TAILQ_FIRST(&inodedep->id_extupdt) != NULL || 2418 TAILQ_FIRST(&inodedep->id_newextupdt) != NULL || 2419 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino1 != NULL) 2420 return (0); 2421 LIST_REMOVE(inodedep, id_hash); 2422 WORKITEM_FREE(inodedep, D_INODEDEP); 2423 num_inodedep -= 1; 2424 return (1); 2425} 2426 2427/* 2428 * This workitem routine performs the block de-allocation. 2429 * The workitem is added to the pending list after the updated 2430 * inode block has been written to disk. As mentioned above, 2431 * checks regarding the number of blocks de-allocated (compared 2432 * to the number of blocks allocated for the file) are also 2433 * performed in this function. 2434 */ 2435static void 2436handle_workitem_freeblocks(freeblks, flags) 2437 struct freeblks *freeblks; 2438 int flags; 2439{ 2440 struct inode *ip; 2441 struct vnode *vp; 2442 struct fs *fs; 2443 int i, nblocks, level, bsize; 2444 ufs2_daddr_t bn, blocksreleased = 0; 2445 int error, allerror = 0; 2446 ufs_lbn_t baselbns[NIADDR], tmpval; 2447 2448 fs = VFSTOUFS(freeblks->fb_mnt)->um_fs; 2449 tmpval = 1; 2450 baselbns[0] = NDADDR; 2451 for (i = 1; i < NIADDR; i++) { 2452 tmpval *= NINDIR(fs); 2453 baselbns[i] = baselbns[i - 1] + tmpval; 2454 } 2455 nblocks = btodb(fs->fs_bsize); 2456 blocksreleased = 0; 2457 /* 2458 * Release all extended attribute blocks or frags. 2459 */ 2460 if (freeblks->fb_oldextsize > 0) { 2461 for (i = (NXADDR - 1); i >= 0; i--) { 2462 if ((bn = freeblks->fb_eblks[i]) == 0) 2463 continue; 2464 bsize = sblksize(fs, freeblks->fb_oldextsize, i); 2465 ffs_blkfree(fs, freeblks->fb_devvp, bn, bsize, 2466 freeblks->fb_previousinum); 2467 blocksreleased += btodb(bsize); 2468 } 2469 } 2470 /* 2471 * Release all data blocks or frags. 2472 */ 2473 if (freeblks->fb_oldsize > 0) { 2474 /* 2475 * Indirect blocks first. 2476 */ 2477 for (level = (NIADDR - 1); level >= 0; level--) { 2478 if ((bn = freeblks->fb_iblks[level]) == 0) 2479 continue; 2480 if ((error = indir_trunc(freeblks, fsbtodb(fs, bn), 2481 level, baselbns[level], &blocksreleased)) == 0) 2482 allerror = error; 2483 ffs_blkfree(fs, freeblks->fb_devvp, bn, fs->fs_bsize, 2484 freeblks->fb_previousinum); 2485 fs->fs_pendingblocks -= nblocks; 2486 blocksreleased += nblocks; 2487 } 2488 /* 2489 * All direct blocks or frags. 2490 */ 2491 for (i = (NDADDR - 1); i >= 0; i--) { 2492 if ((bn = freeblks->fb_dblks[i]) == 0) 2493 continue; 2494 bsize = sblksize(fs, freeblks->fb_oldsize, i); 2495 ffs_blkfree(fs, freeblks->fb_devvp, bn, bsize, 2496 freeblks->fb_previousinum); 2497 fs->fs_pendingblocks -= btodb(bsize); 2498 blocksreleased += btodb(bsize); 2499 } 2500 } 2501 /* 2502 * If we still have not finished background cleanup, then check 2503 * to see if the block count needs to be adjusted. 2504 */ 2505 if (freeblks->fb_chkcnt != blocksreleased && 2506 (fs->fs_flags & FS_UNCLEAN) != 0 && 2507 VFS_VGET(freeblks->fb_mnt, freeblks->fb_previousinum, 2508 (flags & LK_NOWAIT) | LK_EXCLUSIVE, &vp) == 0) { 2509 ip = VTOI(vp); 2510 DIP(ip, i_blocks) += freeblks->fb_chkcnt - blocksreleased; 2511 ip->i_flag |= IN_CHANGE; 2512 vput(vp); 2513 } 2514 2515#ifdef DIAGNOSTIC 2516 if (freeblks->fb_chkcnt != blocksreleased && 2517 ((fs->fs_flags & FS_UNCLEAN) == 0 || (flags & LK_NOWAIT) != 0)) 2518 printf("handle_workitem_freeblocks: block count\n"); 2519 if (allerror) 2520 softdep_error("handle_workitem_freeblks", allerror); 2521#endif /* DIAGNOSTIC */ 2522 2523 WORKITEM_FREE(freeblks, D_FREEBLKS); 2524} 2525 2526/* 2527 * Release blocks associated with the inode ip and stored in the indirect 2528 * block dbn. If level is greater than SINGLE, the block is an indirect block 2529 * and recursive calls to indirtrunc must be used to cleanse other indirect 2530 * blocks. 2531 */ 2532static int 2533indir_trunc(freeblks, dbn, level, lbn, countp) 2534 struct freeblks *freeblks; 2535 ufs2_daddr_t dbn; 2536 int level; 2537 ufs_lbn_t lbn; 2538 ufs2_daddr_t *countp; 2539{ 2540 struct buf *bp; 2541 struct fs *fs; 2542 struct worklist *wk; 2543 struct indirdep *indirdep; 2544 ufs1_daddr_t *bap1 = 0; 2545 ufs2_daddr_t nb, *bap2 = 0; 2546 ufs_lbn_t lbnadd; 2547 int i, nblocks, ufs1fmt; 2548 int error, allerror = 0; 2549 2550 fs = VFSTOUFS(freeblks->fb_mnt)->um_fs; 2551 lbnadd = 1; 2552 for (i = level; i > 0; i--) 2553 lbnadd *= NINDIR(fs); 2554 /* 2555 * Get buffer of block pointers to be freed. This routine is not 2556 * called until the zero'ed inode has been written, so it is safe 2557 * to free blocks as they are encountered. Because the inode has 2558 * been zero'ed, calls to bmap on these blocks will fail. So, we 2559 * have to use the on-disk address and the block device for the 2560 * filesystem to look them up. If the file was deleted before its 2561 * indirect blocks were all written to disk, the routine that set 2562 * us up (deallocate_dependencies) will have arranged to leave 2563 * a complete copy of the indirect block in memory for our use. 2564 * Otherwise we have to read the blocks in from the disk. 2565 */ 2566#ifdef notyet 2567 bp = getblk(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 0, 0, 2568 GB_NOCREAT); 2569#else 2570 bp = incore(freeblks->fb_devvp, dbn); 2571#endif 2572 ACQUIRE_LOCK(&lk); 2573 if (bp != NULL && (wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2574 if (wk->wk_type != D_INDIRDEP || 2575 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp || 2576 (indirdep->ir_state & GOINGAWAY) == 0) { 2577 FREE_LOCK(&lk); 2578 panic("indir_trunc: lost indirdep"); 2579 } 2580 WORKLIST_REMOVE(wk); 2581 WORKITEM_FREE(indirdep, D_INDIRDEP); 2582 if (LIST_FIRST(&bp->b_dep) != NULL) { 2583 FREE_LOCK(&lk); 2584 panic("indir_trunc: dangling dep"); 2585 } 2586 VFSTOUFS(freeblks->fb_mnt)->um_numindirdeps -= 1; 2587 FREE_LOCK(&lk); 2588 } else { 2589#ifdef notyet 2590 if (bp) 2591 brelse(bp); 2592#endif 2593 FREE_LOCK(&lk); 2594 error = bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 2595 NOCRED, &bp); 2596 if (error) { 2597 brelse(bp); 2598 return (error); 2599 } 2600 } 2601 /* 2602 * Recursively free indirect blocks. 2603 */ 2604 if (VFSTOUFS(freeblks->fb_mnt)->um_fstype == UFS1) { 2605 ufs1fmt = 1; 2606 bap1 = (ufs1_daddr_t *)bp->b_data; 2607 } else { 2608 ufs1fmt = 0; 2609 bap2 = (ufs2_daddr_t *)bp->b_data; 2610 } 2611 nblocks = btodb(fs->fs_bsize); 2612 for (i = NINDIR(fs) - 1; i >= 0; i--) { 2613 if (ufs1fmt) 2614 nb = bap1[i]; 2615 else 2616 nb = bap2[i]; 2617 if (nb == 0) 2618 continue; 2619 if (level != 0) { 2620 if ((error = indir_trunc(freeblks, fsbtodb(fs, nb), 2621 level - 1, lbn + (i * lbnadd), countp)) != 0) 2622 allerror = error; 2623 } 2624 ffs_blkfree(fs, freeblks->fb_devvp, nb, fs->fs_bsize, 2625 freeblks->fb_previousinum); 2626 fs->fs_pendingblocks -= nblocks; 2627 *countp += nblocks; 2628 } 2629 bp->b_flags |= B_INVAL | B_NOCACHE; 2630 brelse(bp); 2631 return (allerror); 2632} 2633 2634/* 2635 * Free an allocindir. 2636 * This routine must be called with splbio interrupts blocked. 2637 */ 2638static void 2639free_allocindir(aip, inodedep) 2640 struct allocindir *aip; 2641 struct inodedep *inodedep; 2642{ 2643 struct freefrag *freefrag; 2644 2645#ifdef DEBUG 2646 if (lk.lkt_held == NOHOLDER) 2647 panic("free_allocindir: lock not held"); 2648#endif 2649 if ((aip->ai_state & DEPCOMPLETE) == 0) 2650 LIST_REMOVE(aip, ai_deps); 2651 if (aip->ai_state & ONWORKLIST) 2652 WORKLIST_REMOVE(&aip->ai_list); 2653 LIST_REMOVE(aip, ai_next); 2654 if ((freefrag = aip->ai_freefrag) != NULL) { 2655 if (inodedep == NULL) 2656 add_to_worklist(&freefrag->ff_list); 2657 else 2658 WORKLIST_INSERT(&inodedep->id_bufwait, 2659 &freefrag->ff_list); 2660 } 2661 WORKITEM_FREE(aip, D_ALLOCINDIR); 2662} 2663 2664/* 2665 * Directory entry addition dependencies. 2666 * 2667 * When adding a new directory entry, the inode (with its incremented link 2668 * count) must be written to disk before the directory entry's pointer to it. 2669 * Also, if the inode is newly allocated, the corresponding freemap must be 2670 * updated (on disk) before the directory entry's pointer. These requirements 2671 * are met via undo/redo on the directory entry's pointer, which consists 2672 * simply of the inode number. 2673 * 2674 * As directory entries are added and deleted, the free space within a 2675 * directory block can become fragmented. The ufs filesystem will compact 2676 * a fragmented directory block to make space for a new entry. When this 2677 * occurs, the offsets of previously added entries change. Any "diradd" 2678 * dependency structures corresponding to these entries must be updated with 2679 * the new offsets. 2680 */ 2681 2682/* 2683 * This routine is called after the in-memory inode's link 2684 * count has been incremented, but before the directory entry's 2685 * pointer to the inode has been set. 2686 */ 2687int 2688softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 2689 struct buf *bp; /* buffer containing directory block */ 2690 struct inode *dp; /* inode for directory */ 2691 off_t diroffset; /* offset of new entry in directory */ 2692 ino_t newinum; /* inode referenced by new directory entry */ 2693 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 2694 int isnewblk; /* entry is in a newly allocated block */ 2695{ 2696 int offset; /* offset of new entry within directory block */ 2697 ufs_lbn_t lbn; /* block in directory containing new entry */ 2698 struct fs *fs; 2699 struct diradd *dap; 2700 struct allocdirect *adp; 2701 struct pagedep *pagedep; 2702 struct inodedep *inodedep; 2703 struct newdirblk *newdirblk = 0; 2704 struct mkdir *mkdir1, *mkdir2; 2705 2706 /* 2707 * Whiteouts have no dependencies. 2708 */ 2709 if (newinum == WINO) { 2710 if (newdirbp != NULL) 2711 bdwrite(newdirbp); 2712 return (0); 2713 } 2714 2715 fs = dp->i_fs; 2716 lbn = lblkno(fs, diroffset); 2717 offset = blkoff(fs, diroffset); 2718 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, 2719 M_SOFTDEP_FLAGS|M_ZERO); 2720 dap->da_list.wk_type = D_DIRADD; 2721 dap->da_offset = offset; 2722 dap->da_newinum = newinum; 2723 dap->da_state = ATTACHED; 2724 if (isnewblk && lbn < NDADDR && fragoff(fs, diroffset) == 0) { 2725 MALLOC(newdirblk, struct newdirblk *, sizeof(struct newdirblk), 2726 M_NEWDIRBLK, M_SOFTDEP_FLAGS); 2727 newdirblk->db_list.wk_type = D_NEWDIRBLK; 2728 newdirblk->db_state = 0; 2729 } 2730 if (newdirbp == NULL) { 2731 dap->da_state |= DEPCOMPLETE; 2732 ACQUIRE_LOCK(&lk); 2733 } else { 2734 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 2735 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2736 M_SOFTDEP_FLAGS); 2737 mkdir1->md_list.wk_type = D_MKDIR; 2738 mkdir1->md_state = MKDIR_BODY; 2739 mkdir1->md_diradd = dap; 2740 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2741 M_SOFTDEP_FLAGS); 2742 mkdir2->md_list.wk_type = D_MKDIR; 2743 mkdir2->md_state = MKDIR_PARENT; 2744 mkdir2->md_diradd = dap; 2745 /* 2746 * Dependency on "." and ".." being written to disk. 2747 */ 2748 mkdir1->md_buf = newdirbp; 2749 ACQUIRE_LOCK(&lk); 2750 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 2751 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list); 2752 FREE_LOCK(&lk); 2753 bdwrite(newdirbp); 2754 /* 2755 * Dependency on link count increase for parent directory 2756 */ 2757 ACQUIRE_LOCK(&lk); 2758 if (inodedep_lookup(fs, dp->i_number, 0, &inodedep) == 0 2759 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2760 dap->da_state &= ~MKDIR_PARENT; 2761 WORKITEM_FREE(mkdir2, D_MKDIR); 2762 } else { 2763 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 2764 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 2765 } 2766 } 2767 /* 2768 * Link into parent directory pagedep to await its being written. 2769 */ 2770 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2771 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2772 dap->da_pagedep = pagedep; 2773 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 2774 da_pdlist); 2775 /* 2776 * Link into its inodedep. Put it on the id_bufwait list if the inode 2777 * is not yet written. If it is written, do the post-inode write 2778 * processing to put it on the id_pendinghd list. 2779 */ 2780 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep); 2781 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 2782 diradd_inode_written(dap, inodedep); 2783 else 2784 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2785 if (isnewblk) { 2786 /* 2787 * Directories growing into indirect blocks are rare 2788 * enough and the frequency of new block allocation 2789 * in those cases even more rare, that we choose not 2790 * to bother tracking them. Rather we simply force the 2791 * new directory entry to disk. 2792 */ 2793 if (lbn >= NDADDR) { 2794 FREE_LOCK(&lk); 2795 /* 2796 * We only have a new allocation when at the 2797 * beginning of a new block, not when we are 2798 * expanding into an existing block. 2799 */ 2800 if (blkoff(fs, diroffset) == 0) 2801 return (1); 2802 return (0); 2803 } 2804 /* 2805 * We only have a new allocation when at the beginning 2806 * of a new fragment, not when we are expanding into an 2807 * existing fragment. Also, there is nothing to do if we 2808 * are already tracking this block. 2809 */ 2810 if (fragoff(fs, diroffset) != 0) { 2811 FREE_LOCK(&lk); 2812 return (0); 2813 } 2814 if ((pagedep->pd_state & NEWBLOCK) != 0) { 2815 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 2816 FREE_LOCK(&lk); 2817 return (0); 2818 } 2819 /* 2820 * Find our associated allocdirect and have it track us. 2821 */ 2822 if (inodedep_lookup(fs, dp->i_number, 0, &inodedep) == 0) 2823 panic("softdep_setup_directory_add: lost inodedep"); 2824 adp = TAILQ_LAST(&inodedep->id_newinoupdt, allocdirectlst); 2825 if (adp == NULL || adp->ad_lbn != lbn) { 2826 FREE_LOCK(&lk); 2827 panic("softdep_setup_directory_add: lost entry"); 2828 } 2829 pagedep->pd_state |= NEWBLOCK; 2830 newdirblk->db_pagedep = pagedep; 2831 WORKLIST_INSERT(&adp->ad_newdirblk, &newdirblk->db_list); 2832 } 2833 FREE_LOCK(&lk); 2834 return (0); 2835} 2836 2837/* 2838 * This procedure is called to change the offset of a directory 2839 * entry when compacting a directory block which must be owned 2840 * exclusively by the caller. Note that the actual entry movement 2841 * must be done in this procedure to ensure that no I/O completions 2842 * occur while the move is in progress. 2843 */ 2844void 2845softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize) 2846 struct inode *dp; /* inode for directory */ 2847 caddr_t base; /* address of dp->i_offset */ 2848 caddr_t oldloc; /* address of old directory location */ 2849 caddr_t newloc; /* address of new directory location */ 2850 int entrysize; /* size of directory entry */ 2851{ 2852 int offset, oldoffset, newoffset; 2853 struct pagedep *pagedep; 2854 struct diradd *dap; 2855 ufs_lbn_t lbn; 2856 2857 ACQUIRE_LOCK(&lk); 2858 lbn = lblkno(dp->i_fs, dp->i_offset); 2859 offset = blkoff(dp->i_fs, dp->i_offset); 2860 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) 2861 goto done; 2862 oldoffset = offset + (oldloc - base); 2863 newoffset = offset + (newloc - base); 2864 2865 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) { 2866 if (dap->da_offset != oldoffset) 2867 continue; 2868 dap->da_offset = newoffset; 2869 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) 2870 break; 2871 LIST_REMOVE(dap, da_pdlist); 2872 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], 2873 dap, da_pdlist); 2874 break; 2875 } 2876 if (dap == NULL) { 2877 2878 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) { 2879 if (dap->da_offset == oldoffset) { 2880 dap->da_offset = newoffset; 2881 break; 2882 } 2883 } 2884 } 2885done: 2886 bcopy(oldloc, newloc, entrysize); 2887 FREE_LOCK(&lk); 2888} 2889 2890/* 2891 * Free a diradd dependency structure. This routine must be called 2892 * with splbio interrupts blocked. 2893 */ 2894static void 2895free_diradd(dap) 2896 struct diradd *dap; 2897{ 2898 struct dirrem *dirrem; 2899 struct pagedep *pagedep; 2900 struct inodedep *inodedep; 2901 struct mkdir *mkdir, *nextmd; 2902 2903#ifdef DEBUG 2904 if (lk.lkt_held == NOHOLDER) 2905 panic("free_diradd: lock not held"); 2906#endif 2907 WORKLIST_REMOVE(&dap->da_list); 2908 LIST_REMOVE(dap, da_pdlist); 2909 if ((dap->da_state & DIRCHG) == 0) { 2910 pagedep = dap->da_pagedep; 2911 } else { 2912 dirrem = dap->da_previous; 2913 pagedep = dirrem->dm_pagedep; 2914 dirrem->dm_dirinum = pagedep->pd_ino; 2915 add_to_worklist(&dirrem->dm_list); 2916 } 2917 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum, 2918 0, &inodedep) != 0) 2919 (void) free_inodedep(inodedep); 2920 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2921 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 2922 nextmd = LIST_NEXT(mkdir, md_mkdirs); 2923 if (mkdir->md_diradd != dap) 2924 continue; 2925 dap->da_state &= ~mkdir->md_state; 2926 WORKLIST_REMOVE(&mkdir->md_list); 2927 LIST_REMOVE(mkdir, md_mkdirs); 2928 WORKITEM_FREE(mkdir, D_MKDIR); 2929 } 2930 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2931 FREE_LOCK(&lk); 2932 panic("free_diradd: unfound ref"); 2933 } 2934 } 2935 WORKITEM_FREE(dap, D_DIRADD); 2936} 2937 2938/* 2939 * Directory entry removal dependencies. 2940 * 2941 * When removing a directory entry, the entry's inode pointer must be 2942 * zero'ed on disk before the corresponding inode's link count is decremented 2943 * (possibly freeing the inode for re-use). This dependency is handled by 2944 * updating the directory entry but delaying the inode count reduction until 2945 * after the directory block has been written to disk. After this point, the 2946 * inode count can be decremented whenever it is convenient. 2947 */ 2948 2949/* 2950 * This routine should be called immediately after removing 2951 * a directory entry. The inode's link count should not be 2952 * decremented by the calling procedure -- the soft updates 2953 * code will do this task when it is safe. 2954 */ 2955void 2956softdep_setup_remove(bp, dp, ip, isrmdir) 2957 struct buf *bp; /* buffer containing directory block */ 2958 struct inode *dp; /* inode for the directory being modified */ 2959 struct inode *ip; /* inode for directory entry being removed */ 2960 int isrmdir; /* indicates if doing RMDIR */ 2961{ 2962 struct dirrem *dirrem, *prevdirrem; 2963 2964 /* 2965 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 2966 */ 2967 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2968 2969 /* 2970 * If the COMPLETE flag is clear, then there were no active 2971 * entries and we want to roll back to a zeroed entry until 2972 * the new inode is committed to disk. If the COMPLETE flag is 2973 * set then we have deleted an entry that never made it to 2974 * disk. If the entry we deleted resulted from a name change, 2975 * then the old name still resides on disk. We cannot delete 2976 * its inode (returned to us in prevdirrem) until the zeroed 2977 * directory entry gets to disk. The new inode has never been 2978 * referenced on the disk, so can be deleted immediately. 2979 */ 2980 if ((dirrem->dm_state & COMPLETE) == 0) { 2981 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 2982 dm_next); 2983 FREE_LOCK(&lk); 2984 } else { 2985 if (prevdirrem != NULL) 2986 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 2987 prevdirrem, dm_next); 2988 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 2989 FREE_LOCK(&lk); 2990 handle_workitem_remove(dirrem, NULL); 2991 } 2992} 2993 2994/* 2995 * Allocate a new dirrem if appropriate and return it along with 2996 * its associated pagedep. Called without a lock, returns with lock. 2997 */ 2998static long num_dirrem; /* number of dirrem allocated */ 2999static struct dirrem * 3000newdirrem(bp, dp, ip, isrmdir, prevdirremp) 3001 struct buf *bp; /* buffer containing directory block */ 3002 struct inode *dp; /* inode for the directory being modified */ 3003 struct inode *ip; /* inode for directory entry being removed */ 3004 int isrmdir; /* indicates if doing RMDIR */ 3005 struct dirrem **prevdirremp; /* previously referenced inode, if any */ 3006{ 3007 int offset; 3008 ufs_lbn_t lbn; 3009 struct diradd *dap; 3010 struct dirrem *dirrem; 3011 struct pagedep *pagedep; 3012 3013 /* 3014 * Whiteouts have no deletion dependencies. 3015 */ 3016 if (ip == NULL) 3017 panic("newdirrem: whiteout"); 3018 /* 3019 * If we are over our limit, try to improve the situation. 3020 * Limiting the number of dirrem structures will also limit 3021 * the number of freefile and freeblks structures. 3022 */ 3023 if (num_dirrem > max_softdeps / 2) 3024 (void) request_cleanup(FLUSH_REMOVE, 0); 3025 num_dirrem += 1; 3026 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem), 3027 M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO); 3028 dirrem->dm_list.wk_type = D_DIRREM; 3029 dirrem->dm_state = isrmdir ? RMDIR : 0; 3030 dirrem->dm_mnt = ITOV(ip)->v_mount; 3031 dirrem->dm_oldinum = ip->i_number; 3032 *prevdirremp = NULL; 3033 3034 ACQUIRE_LOCK(&lk); 3035 lbn = lblkno(dp->i_fs, dp->i_offset); 3036 offset = blkoff(dp->i_fs, dp->i_offset); 3037 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 3038 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 3039 dirrem->dm_pagedep = pagedep; 3040 /* 3041 * Check for a diradd dependency for the same directory entry. 3042 * If present, then both dependencies become obsolete and can 3043 * be de-allocated. Check for an entry on both the pd_dirraddhd 3044 * list and the pd_pendinghd list. 3045 */ 3046 3047 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 3048 if (dap->da_offset == offset) 3049 break; 3050 if (dap == NULL) { 3051 3052 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 3053 if (dap->da_offset == offset) 3054 break; 3055 if (dap == NULL) 3056 return (dirrem); 3057 } 3058 /* 3059 * Must be ATTACHED at this point. 3060 */ 3061 if ((dap->da_state & ATTACHED) == 0) { 3062 FREE_LOCK(&lk); 3063 panic("newdirrem: not ATTACHED"); 3064 } 3065 if (dap->da_newinum != ip->i_number) { 3066 FREE_LOCK(&lk); 3067 panic("newdirrem: inum %d should be %d", 3068 ip->i_number, dap->da_newinum); 3069 } 3070 /* 3071 * If we are deleting a changed name that never made it to disk, 3072 * then return the dirrem describing the previous inode (which 3073 * represents the inode currently referenced from this entry on disk). 3074 */ 3075 if ((dap->da_state & DIRCHG) != 0) { 3076 *prevdirremp = dap->da_previous; 3077 dap->da_state &= ~DIRCHG; 3078 dap->da_pagedep = pagedep; 3079 } 3080 /* 3081 * We are deleting an entry that never made it to disk. 3082 * Mark it COMPLETE so we can delete its inode immediately. 3083 */ 3084 dirrem->dm_state |= COMPLETE; 3085 free_diradd(dap); 3086 return (dirrem); 3087} 3088 3089/* 3090 * Directory entry change dependencies. 3091 * 3092 * Changing an existing directory entry requires that an add operation 3093 * be completed first followed by a deletion. The semantics for the addition 3094 * are identical to the description of adding a new entry above except 3095 * that the rollback is to the old inode number rather than zero. Once 3096 * the addition dependency is completed, the removal is done as described 3097 * in the removal routine above. 3098 */ 3099 3100/* 3101 * This routine should be called immediately after changing 3102 * a directory entry. The inode's link count should not be 3103 * decremented by the calling procedure -- the soft updates 3104 * code will perform this task when it is safe. 3105 */ 3106void 3107softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 3108 struct buf *bp; /* buffer containing directory block */ 3109 struct inode *dp; /* inode for the directory being modified */ 3110 struct inode *ip; /* inode for directory entry being removed */ 3111 ino_t newinum; /* new inode number for changed entry */ 3112 int isrmdir; /* indicates if doing RMDIR */ 3113{ 3114 int offset; 3115 struct diradd *dap = NULL; 3116 struct dirrem *dirrem, *prevdirrem; 3117 struct pagedep *pagedep; 3118 struct inodedep *inodedep; 3119 3120 offset = blkoff(dp->i_fs, dp->i_offset); 3121 3122 /* 3123 * Whiteouts do not need diradd dependencies. 3124 */ 3125 if (newinum != WINO) { 3126 MALLOC(dap, struct diradd *, sizeof(struct diradd), 3127 M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO); 3128 dap->da_list.wk_type = D_DIRADD; 3129 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 3130 dap->da_offset = offset; 3131 dap->da_newinum = newinum; 3132 } 3133 3134 /* 3135 * Allocate a new dirrem and ACQUIRE_LOCK. 3136 */ 3137 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 3138 pagedep = dirrem->dm_pagedep; 3139 /* 3140 * The possible values for isrmdir: 3141 * 0 - non-directory file rename 3142 * 1 - directory rename within same directory 3143 * inum - directory rename to new directory of given inode number 3144 * When renaming to a new directory, we are both deleting and 3145 * creating a new directory entry, so the link count on the new 3146 * directory should not change. Thus we do not need the followup 3147 * dirrem which is usually done in handle_workitem_remove. We set 3148 * the DIRCHG flag to tell handle_workitem_remove to skip the 3149 * followup dirrem. 3150 */ 3151 if (isrmdir > 1) 3152 dirrem->dm_state |= DIRCHG; 3153 3154 /* 3155 * Whiteouts have no additional dependencies, 3156 * so just put the dirrem on the correct list. 3157 */ 3158 if (newinum == WINO) { 3159 if ((dirrem->dm_state & COMPLETE) == 0) { 3160 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 3161 dm_next); 3162 } else { 3163 dirrem->dm_dirinum = pagedep->pd_ino; 3164 add_to_worklist(&dirrem->dm_list); 3165 } 3166 FREE_LOCK(&lk); 3167 return; 3168 } 3169 3170 /* 3171 * If the COMPLETE flag is clear, then there were no active 3172 * entries and we want to roll back to the previous inode until 3173 * the new inode is committed to disk. If the COMPLETE flag is 3174 * set, then we have deleted an entry that never made it to disk. 3175 * If the entry we deleted resulted from a name change, then the old 3176 * inode reference still resides on disk. Any rollback that we do 3177 * needs to be to that old inode (returned to us in prevdirrem). If 3178 * the entry we deleted resulted from a create, then there is 3179 * no entry on the disk, so we want to roll back to zero rather 3180 * than the uncommitted inode. In either of the COMPLETE cases we 3181 * want to immediately free the unwritten and unreferenced inode. 3182 */ 3183 if ((dirrem->dm_state & COMPLETE) == 0) { 3184 dap->da_previous = dirrem; 3185 } else { 3186 if (prevdirrem != NULL) { 3187 dap->da_previous = prevdirrem; 3188 } else { 3189 dap->da_state &= ~DIRCHG; 3190 dap->da_pagedep = pagedep; 3191 } 3192 dirrem->dm_dirinum = pagedep->pd_ino; 3193 add_to_worklist(&dirrem->dm_list); 3194 } 3195 /* 3196 * Link into its inodedep. Put it on the id_bufwait list if the inode 3197 * is not yet written. If it is written, do the post-inode write 3198 * processing to put it on the id_pendinghd list. 3199 */ 3200 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 || 3201 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 3202 dap->da_state |= COMPLETE; 3203 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3204 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 3205 } else { 3206 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 3207 dap, da_pdlist); 3208 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 3209 } 3210 FREE_LOCK(&lk); 3211} 3212 3213/* 3214 * Called whenever the link count on an inode is changed. 3215 * It creates an inode dependency so that the new reference(s) 3216 * to the inode cannot be committed to disk until the updated 3217 * inode has been written. 3218 */ 3219void 3220softdep_change_linkcnt(ip) 3221 struct inode *ip; /* the inode with the increased link count */ 3222{ 3223 struct inodedep *inodedep; 3224 3225 ACQUIRE_LOCK(&lk); 3226 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 3227 if (ip->i_nlink < ip->i_effnlink) { 3228 FREE_LOCK(&lk); 3229 panic("softdep_change_linkcnt: bad delta"); 3230 } 3231 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3232 FREE_LOCK(&lk); 3233} 3234 3235/* 3236 * Called when the effective link count and the reference count 3237 * on an inode drops to zero. At this point there are no names 3238 * referencing the file in the filesystem and no active file 3239 * references. The space associated with the file will be freed 3240 * as soon as the necessary soft dependencies are cleared. 3241 */ 3242void 3243softdep_releasefile(ip) 3244 struct inode *ip; /* inode with the zero effective link count */ 3245{ 3246 struct inodedep *inodedep; 3247 struct fs *fs; 3248 int extblocks; 3249 3250 if (ip->i_effnlink > 0) 3251 panic("softdep_filerelease: file still referenced"); 3252 /* 3253 * We may be called several times as the real reference count 3254 * drops to zero. We only want to account for the space once. 3255 */ 3256 if (ip->i_flag & IN_SPACECOUNTED) 3257 return; 3258 /* 3259 * We have to deactivate a snapshot otherwise copyonwrites may 3260 * add blocks and the cleanup may remove blocks after we have 3261 * tried to account for them. 3262 */ 3263 if ((ip->i_flags & SF_SNAPSHOT) != 0) 3264 ffs_snapremove(ITOV(ip)); 3265 /* 3266 * If we are tracking an nlinkdelta, we have to also remember 3267 * whether we accounted for the freed space yet. 3268 */ 3269 ACQUIRE_LOCK(&lk); 3270 if ((inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep))) 3271 inodedep->id_state |= SPACECOUNTED; 3272 FREE_LOCK(&lk); 3273 fs = ip->i_fs; 3274 extblocks = 0; 3275 if (fs->fs_magic == FS_UFS2_MAGIC) 3276 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 3277 ip->i_fs->fs_pendingblocks += DIP(ip, i_blocks) - extblocks; 3278 ip->i_fs->fs_pendinginodes += 1; 3279 ip->i_flag |= IN_SPACECOUNTED; 3280} 3281 3282/* 3283 * This workitem decrements the inode's link count. 3284 * If the link count reaches zero, the file is removed. 3285 */ 3286static void 3287handle_workitem_remove(dirrem, xp) 3288 struct dirrem *dirrem; 3289 struct vnode *xp; 3290{ 3291 struct thread *td = curthread; 3292 struct inodedep *inodedep; 3293 struct vnode *vp; 3294 struct inode *ip; 3295 ino_t oldinum; 3296 int error; 3297 3298 if ((vp = xp) == NULL && 3299 (error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, LK_EXCLUSIVE, 3300 &vp)) != 0) { 3301 softdep_error("handle_workitem_remove: vget", error); 3302 return; 3303 } 3304 ip = VTOI(vp); 3305 ACQUIRE_LOCK(&lk); 3306 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){ 3307 FREE_LOCK(&lk); 3308 panic("handle_workitem_remove: lost inodedep"); 3309 } 3310 /* 3311 * Normal file deletion. 3312 */ 3313 if ((dirrem->dm_state & RMDIR) == 0) { 3314 ip->i_nlink--; 3315 DIP(ip, i_nlink) = ip->i_nlink; 3316 ip->i_flag |= IN_CHANGE; 3317 if (ip->i_nlink < ip->i_effnlink) { 3318 FREE_LOCK(&lk); 3319 panic("handle_workitem_remove: bad file delta"); 3320 } 3321 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3322 FREE_LOCK(&lk); 3323 vput(vp); 3324 num_dirrem -= 1; 3325 WORKITEM_FREE(dirrem, D_DIRREM); 3326 return; 3327 } 3328 /* 3329 * Directory deletion. Decrement reference count for both the 3330 * just deleted parent directory entry and the reference for ".". 3331 * Next truncate the directory to length zero. When the 3332 * truncation completes, arrange to have the reference count on 3333 * the parent decremented to account for the loss of "..". 3334 */ 3335 ip->i_nlink -= 2; 3336 DIP(ip, i_nlink) = ip->i_nlink; 3337 ip->i_flag |= IN_CHANGE; 3338 if (ip->i_nlink < ip->i_effnlink) { 3339 FREE_LOCK(&lk); 3340 panic("handle_workitem_remove: bad dir delta"); 3341 } 3342 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3343 FREE_LOCK(&lk); 3344 if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, td->td_ucred, td)) != 0) 3345 softdep_error("handle_workitem_remove: truncate", error); 3346 /* 3347 * Rename a directory to a new parent. Since, we are both deleting 3348 * and creating a new directory entry, the link count on the new 3349 * directory should not change. Thus we skip the followup dirrem. 3350 */ 3351 if (dirrem->dm_state & DIRCHG) { 3352 vput(vp); 3353 num_dirrem -= 1; 3354 WORKITEM_FREE(dirrem, D_DIRREM); 3355 return; 3356 } 3357 /* 3358 * If the inodedep does not exist, then the zero'ed inode has 3359 * been written to disk. If the allocated inode has never been 3360 * written to disk, then the on-disk inode is zero'ed. In either 3361 * case we can remove the file immediately. 3362 */ 3363 ACQUIRE_LOCK(&lk); 3364 dirrem->dm_state = 0; 3365 oldinum = dirrem->dm_oldinum; 3366 dirrem->dm_oldinum = dirrem->dm_dirinum; 3367 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 || 3368 check_inode_unwritten(inodedep)) { 3369 FREE_LOCK(&lk); 3370 vput(vp); 3371 handle_workitem_remove(dirrem, NULL); 3372 return; 3373 } 3374 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 3375 FREE_LOCK(&lk); 3376 vput(vp); 3377} 3378 3379/* 3380 * Inode de-allocation dependencies. 3381 * 3382 * When an inode's link count is reduced to zero, it can be de-allocated. We 3383 * found it convenient to postpone de-allocation until after the inode is 3384 * written to disk with its new link count (zero). At this point, all of the 3385 * on-disk inode's block pointers are nullified and, with careful dependency 3386 * list ordering, all dependencies related to the inode will be satisfied and 3387 * the corresponding dependency structures de-allocated. So, if/when the 3388 * inode is reused, there will be no mixing of old dependencies with new 3389 * ones. This artificial dependency is set up by the block de-allocation 3390 * procedure above (softdep_setup_freeblocks) and completed by the 3391 * following procedure. 3392 */ 3393static void 3394handle_workitem_freefile(freefile) 3395 struct freefile *freefile; 3396{ 3397 struct fs *fs; 3398 struct inodedep *idp; 3399 int error; 3400 3401 fs = VFSTOUFS(freefile->fx_mnt)->um_fs; 3402#ifdef DEBUG 3403 ACQUIRE_LOCK(&lk); 3404 error = inodedep_lookup(fs, freefile->fx_oldinum, 0, &idp); 3405 FREE_LOCK(&lk); 3406 if (error) 3407 panic("handle_workitem_freefile: inodedep survived"); 3408#endif 3409 fs->fs_pendinginodes -= 1; 3410 if ((error = ffs_freefile(fs, freefile->fx_devvp, freefile->fx_oldinum, 3411 freefile->fx_mode)) != 0) 3412 softdep_error("handle_workitem_freefile", error); 3413 WORKITEM_FREE(freefile, D_FREEFILE); 3414} 3415 3416/* 3417 * Disk writes. 3418 * 3419 * The dependency structures constructed above are most actively used when file 3420 * system blocks are written to disk. No constraints are placed on when a 3421 * block can be written, but unsatisfied update dependencies are made safe by 3422 * modifying (or replacing) the source memory for the duration of the disk 3423 * write. When the disk write completes, the memory block is again brought 3424 * up-to-date. 3425 * 3426 * In-core inode structure reclamation. 3427 * 3428 * Because there are a finite number of "in-core" inode structures, they are 3429 * reused regularly. By transferring all inode-related dependencies to the 3430 * in-memory inode block and indexing them separately (via "inodedep"s), we 3431 * can allow "in-core" inode structures to be reused at any time and avoid 3432 * any increase in contention. 3433 * 3434 * Called just before entering the device driver to initiate a new disk I/O. 3435 * The buffer must be locked, thus, no I/O completion operations can occur 3436 * while we are manipulating its associated dependencies. 3437 */ 3438static void 3439softdep_disk_io_initiation(bp) 3440 struct buf *bp; /* structure describing disk write to occur */ 3441{ 3442 struct worklist *wk, *nextwk; 3443 struct indirdep *indirdep; 3444 struct inodedep *inodedep; 3445 3446 /* 3447 * We only care about write operations. There should never 3448 * be dependencies for reads. 3449 */ 3450 if (bp->b_iocmd == BIO_READ) 3451 panic("softdep_disk_io_initiation: read"); 3452 /* 3453 * Do any necessary pre-I/O processing. 3454 */ 3455 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = nextwk) { 3456 nextwk = LIST_NEXT(wk, wk_list); 3457 switch (wk->wk_type) { 3458 3459 case D_PAGEDEP: 3460 initiate_write_filepage(WK_PAGEDEP(wk), bp); 3461 continue; 3462 3463 case D_INODEDEP: 3464 inodedep = WK_INODEDEP(wk); 3465 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) 3466 initiate_write_inodeblock_ufs1(inodedep, bp); 3467 else 3468 initiate_write_inodeblock_ufs2(inodedep, bp); 3469 continue; 3470 3471 case D_INDIRDEP: 3472 indirdep = WK_INDIRDEP(wk); 3473 if (indirdep->ir_state & GOINGAWAY) 3474 panic("disk_io_initiation: indirdep gone"); 3475 /* 3476 * If there are no remaining dependencies, this 3477 * will be writing the real pointers, so the 3478 * dependency can be freed. 3479 */ 3480 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) { 3481 indirdep->ir_savebp->b_flags |= 3482 B_INVAL | B_NOCACHE; 3483 brelse(indirdep->ir_savebp); 3484 /* inline expand WORKLIST_REMOVE(wk); */ 3485 wk->wk_state &= ~ONWORKLIST; 3486 LIST_REMOVE(wk, wk_list); 3487 WORKITEM_FREE(indirdep, D_INDIRDEP); 3488 continue; 3489 } 3490 /* 3491 * Replace up-to-date version with safe version. 3492 */ 3493 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount, 3494 M_INDIRDEP, M_SOFTDEP_FLAGS); 3495 ACQUIRE_LOCK(&lk); 3496 indirdep->ir_state &= ~ATTACHED; 3497 indirdep->ir_state |= UNDONE; 3498 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 3499 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 3500 bp->b_bcount); 3501 FREE_LOCK(&lk); 3502 continue; 3503 3504 case D_MKDIR: 3505 case D_BMSAFEMAP: 3506 case D_ALLOCDIRECT: 3507 case D_ALLOCINDIR: 3508 continue; 3509 3510 default: 3511 panic("handle_disk_io_initiation: Unexpected type %s", 3512 TYPENAME(wk->wk_type)); 3513 /* NOTREACHED */ 3514 } 3515 } 3516} 3517 3518/* 3519 * Called from within the procedure above to deal with unsatisfied 3520 * allocation dependencies in a directory. The buffer must be locked, 3521 * thus, no I/O completion operations can occur while we are 3522 * manipulating its associated dependencies. 3523 */ 3524static void 3525initiate_write_filepage(pagedep, bp) 3526 struct pagedep *pagedep; 3527 struct buf *bp; 3528{ 3529 struct diradd *dap; 3530 struct direct *ep; 3531 int i; 3532 3533 if (pagedep->pd_state & IOSTARTED) { 3534 /* 3535 * This can only happen if there is a driver that does not 3536 * understand chaining. Here biodone will reissue the call 3537 * to strategy for the incomplete buffers. 3538 */ 3539 printf("initiate_write_filepage: already started\n"); 3540 return; 3541 } 3542 pagedep->pd_state |= IOSTARTED; 3543 ACQUIRE_LOCK(&lk); 3544 for (i = 0; i < DAHASHSZ; i++) { 3545 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 3546 ep = (struct direct *) 3547 ((char *)bp->b_data + dap->da_offset); 3548 if (ep->d_ino != dap->da_newinum) { 3549 FREE_LOCK(&lk); 3550 panic("%s: dir inum %d != new %d", 3551 "initiate_write_filepage", 3552 ep->d_ino, dap->da_newinum); 3553 } 3554 if (dap->da_state & DIRCHG) 3555 ep->d_ino = dap->da_previous->dm_oldinum; 3556 else 3557 ep->d_ino = 0; 3558 dap->da_state &= ~ATTACHED; 3559 dap->da_state |= UNDONE; 3560 } 3561 } 3562 FREE_LOCK(&lk); 3563} 3564 3565/* 3566 * Version of initiate_write_inodeblock that handles UFS1 dinodes. 3567 * Note that any bug fixes made to this routine must be done in the 3568 * version found below. 3569 * 3570 * Called from within the procedure above to deal with unsatisfied 3571 * allocation dependencies in an inodeblock. The buffer must be 3572 * locked, thus, no I/O completion operations can occur while we 3573 * are manipulating its associated dependencies. 3574 */ 3575static void 3576initiate_write_inodeblock_ufs1(inodedep, bp) 3577 struct inodedep *inodedep; 3578 struct buf *bp; /* The inode block */ 3579{ 3580 struct allocdirect *adp, *lastadp; 3581 struct ufs1_dinode *dp; 3582 struct fs *fs; 3583 ufs_lbn_t i, prevlbn = 0; 3584 int deplist; 3585 3586 if (inodedep->id_state & IOSTARTED) 3587 panic("initiate_write_inodeblock_ufs1: already started"); 3588 inodedep->id_state |= IOSTARTED; 3589 fs = inodedep->id_fs; 3590 dp = (struct ufs1_dinode *)bp->b_data + 3591 ino_to_fsbo(fs, inodedep->id_ino); 3592 /* 3593 * If the bitmap is not yet written, then the allocated 3594 * inode cannot be written to disk. 3595 */ 3596 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 3597 if (inodedep->id_savedino1 != NULL) 3598 panic("initiate_write_inodeblock_ufs1: I/O underway"); 3599 MALLOC(inodedep->id_savedino1, struct ufs1_dinode *, 3600 sizeof(struct ufs1_dinode), M_INODEDEP, M_SOFTDEP_FLAGS); 3601 *inodedep->id_savedino1 = *dp; 3602 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 3603 return; 3604 } 3605 /* 3606 * If no dependencies, then there is nothing to roll back. 3607 */ 3608 inodedep->id_savedsize = dp->di_size; 3609 inodedep->id_savedextsize = 0; 3610 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 3611 return; 3612 /* 3613 * Set the dependencies to busy. 3614 */ 3615 ACQUIRE_LOCK(&lk); 3616 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3617 adp = TAILQ_NEXT(adp, ad_next)) { 3618#ifdef DIAGNOSTIC 3619 if (deplist != 0 && prevlbn >= adp->ad_lbn) { 3620 FREE_LOCK(&lk); 3621 panic("softdep_write_inodeblock: lbn order"); 3622 } 3623 prevlbn = adp->ad_lbn; 3624 if (adp->ad_lbn < NDADDR && 3625 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) { 3626 FREE_LOCK(&lk); 3627 panic("%s: direct pointer #%jd mismatch %d != %jd", 3628 "softdep_write_inodeblock", 3629 (intmax_t)adp->ad_lbn, 3630 dp->di_db[adp->ad_lbn], 3631 (intmax_t)adp->ad_newblkno); 3632 } 3633 if (adp->ad_lbn >= NDADDR && 3634 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) { 3635 FREE_LOCK(&lk); 3636 panic("%s: indirect pointer #%jd mismatch %d != %jd", 3637 "softdep_write_inodeblock", 3638 (intmax_t)adp->ad_lbn - NDADDR, 3639 dp->di_ib[adp->ad_lbn - NDADDR], 3640 (intmax_t)adp->ad_newblkno); 3641 } 3642 deplist |= 1 << adp->ad_lbn; 3643 if ((adp->ad_state & ATTACHED) == 0) { 3644 FREE_LOCK(&lk); 3645 panic("softdep_write_inodeblock: Unknown state 0x%x", 3646 adp->ad_state); 3647 } 3648#endif /* DIAGNOSTIC */ 3649 adp->ad_state &= ~ATTACHED; 3650 adp->ad_state |= UNDONE; 3651 } 3652 /* 3653 * The on-disk inode cannot claim to be any larger than the last 3654 * fragment that has been written. Otherwise, the on-disk inode 3655 * might have fragments that were not the last block in the file 3656 * which would corrupt the filesystem. 3657 */ 3658 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3659 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 3660 if (adp->ad_lbn >= NDADDR) 3661 break; 3662 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 3663 /* keep going until hitting a rollback to a frag */ 3664 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 3665 continue; 3666 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 3667 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 3668#ifdef DIAGNOSTIC 3669 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) { 3670 FREE_LOCK(&lk); 3671 panic("softdep_write_inodeblock: lost dep1"); 3672 } 3673#endif /* DIAGNOSTIC */ 3674 dp->di_db[i] = 0; 3675 } 3676 for (i = 0; i < NIADDR; i++) { 3677#ifdef DIAGNOSTIC 3678 if (dp->di_ib[i] != 0 && 3679 (deplist & ((1 << NDADDR) << i)) == 0) { 3680 FREE_LOCK(&lk); 3681 panic("softdep_write_inodeblock: lost dep2"); 3682 } 3683#endif /* DIAGNOSTIC */ 3684 dp->di_ib[i] = 0; 3685 } 3686 FREE_LOCK(&lk); 3687 return; 3688 } 3689 /* 3690 * If we have zero'ed out the last allocated block of the file, 3691 * roll back the size to the last currently allocated block. 3692 * We know that this last allocated block is a full-sized as 3693 * we already checked for fragments in the loop above. 3694 */ 3695 if (lastadp != NULL && 3696 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 3697 for (i = lastadp->ad_lbn; i >= 0; i--) 3698 if (dp->di_db[i] != 0) 3699 break; 3700 dp->di_size = (i + 1) * fs->fs_bsize; 3701 } 3702 /* 3703 * The only dependencies are for indirect blocks. 3704 * 3705 * The file size for indirect block additions is not guaranteed. 3706 * Such a guarantee would be non-trivial to achieve. The conventional 3707 * synchronous write implementation also does not make this guarantee. 3708 * Fsck should catch and fix discrepancies. Arguably, the file size 3709 * can be over-estimated without destroying integrity when the file 3710 * moves into the indirect blocks (i.e., is large). If we want to 3711 * postpone fsck, we are stuck with this argument. 3712 */ 3713 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 3714 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 3715 FREE_LOCK(&lk); 3716} 3717 3718/* 3719 * Version of initiate_write_inodeblock that handles UFS2 dinodes. 3720 * Note that any bug fixes made to this routine must be done in the 3721 * version found above. 3722 * 3723 * Called from within the procedure above to deal with unsatisfied 3724 * allocation dependencies in an inodeblock. The buffer must be 3725 * locked, thus, no I/O completion operations can occur while we 3726 * are manipulating its associated dependencies. 3727 */ 3728static void 3729initiate_write_inodeblock_ufs2(inodedep, bp) 3730 struct inodedep *inodedep; 3731 struct buf *bp; /* The inode block */ 3732{ 3733 struct allocdirect *adp, *lastadp; 3734 struct ufs2_dinode *dp; 3735 struct fs *fs; 3736 ufs_lbn_t i, prevlbn = 0; 3737 int deplist; 3738 3739 if (inodedep->id_state & IOSTARTED) 3740 panic("initiate_write_inodeblock_ufs2: already started"); 3741 inodedep->id_state |= IOSTARTED; 3742 fs = inodedep->id_fs; 3743 dp = (struct ufs2_dinode *)bp->b_data + 3744 ino_to_fsbo(fs, inodedep->id_ino); 3745 /* 3746 * If the bitmap is not yet written, then the allocated 3747 * inode cannot be written to disk. 3748 */ 3749 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 3750 if (inodedep->id_savedino2 != NULL) 3751 panic("initiate_write_inodeblock_ufs2: I/O underway"); 3752 MALLOC(inodedep->id_savedino2, struct ufs2_dinode *, 3753 sizeof(struct ufs2_dinode), M_INODEDEP, M_SOFTDEP_FLAGS); 3754 *inodedep->id_savedino2 = *dp; 3755 bzero((caddr_t)dp, sizeof(struct ufs2_dinode)); 3756 return; 3757 } 3758 /* 3759 * If no dependencies, then there is nothing to roll back. 3760 */ 3761 inodedep->id_savedsize = dp->di_size; 3762 inodedep->id_savedextsize = dp->di_extsize; 3763 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL && 3764 TAILQ_FIRST(&inodedep->id_extupdt) == NULL) 3765 return; 3766 /* 3767 * Set the ext data dependencies to busy. 3768 */ 3769 ACQUIRE_LOCK(&lk); 3770 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 3771 adp = TAILQ_NEXT(adp, ad_next)) { 3772#ifdef DIAGNOSTIC 3773 if (deplist != 0 && prevlbn >= adp->ad_lbn) { 3774 FREE_LOCK(&lk); 3775 panic("softdep_write_inodeblock: lbn order"); 3776 } 3777 prevlbn = adp->ad_lbn; 3778 if (dp->di_extb[adp->ad_lbn] != adp->ad_newblkno) { 3779 FREE_LOCK(&lk); 3780 panic("%s: direct pointer #%jd mismatch %jd != %jd", 3781 "softdep_write_inodeblock", 3782 (intmax_t)adp->ad_lbn, 3783 (intmax_t)dp->di_extb[adp->ad_lbn], 3784 (intmax_t)adp->ad_newblkno); 3785 } 3786 deplist |= 1 << adp->ad_lbn; 3787 if ((adp->ad_state & ATTACHED) == 0) { 3788 FREE_LOCK(&lk); 3789 panic("softdep_write_inodeblock: Unknown state 0x%x", 3790 adp->ad_state); 3791 } 3792#endif /* DIAGNOSTIC */ 3793 adp->ad_state &= ~ATTACHED; 3794 adp->ad_state |= UNDONE; 3795 } 3796 /* 3797 * The on-disk inode cannot claim to be any larger than the last 3798 * fragment that has been written. Otherwise, the on-disk inode 3799 * might have fragments that were not the last block in the ext 3800 * data which would corrupt the filesystem. 3801 */ 3802 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 3803 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 3804 dp->di_extb[adp->ad_lbn] = adp->ad_oldblkno; 3805 /* keep going until hitting a rollback to a frag */ 3806 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 3807 continue; 3808 dp->di_extsize = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 3809 for (i = adp->ad_lbn + 1; i < NXADDR; i++) { 3810#ifdef DIAGNOSTIC 3811 if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0) { 3812 FREE_LOCK(&lk); 3813 panic("softdep_write_inodeblock: lost dep1"); 3814 } 3815#endif /* DIAGNOSTIC */ 3816 dp->di_extb[i] = 0; 3817 } 3818 lastadp = NULL; 3819 break; 3820 } 3821 /* 3822 * If we have zero'ed out the last allocated block of the ext 3823 * data, roll back the size to the last currently allocated block. 3824 * We know that this last allocated block is a full-sized as 3825 * we already checked for fragments in the loop above. 3826 */ 3827 if (lastadp != NULL && 3828 dp->di_extsize <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 3829 for (i = lastadp->ad_lbn; i >= 0; i--) 3830 if (dp->di_extb[i] != 0) 3831 break; 3832 dp->di_extsize = (i + 1) * fs->fs_bsize; 3833 } 3834 /* 3835 * Set the file data dependencies to busy. 3836 */ 3837 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3838 adp = TAILQ_NEXT(adp, ad_next)) { 3839#ifdef DIAGNOSTIC 3840 if (deplist != 0 && prevlbn >= adp->ad_lbn) { 3841 FREE_LOCK(&lk); 3842 panic("softdep_write_inodeblock: lbn order"); 3843 } 3844 prevlbn = adp->ad_lbn; 3845 if (adp->ad_lbn < NDADDR && 3846 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) { 3847 FREE_LOCK(&lk); 3848 panic("%s: direct pointer #%jd mismatch %jd != %jd", 3849 "softdep_write_inodeblock", 3850 (intmax_t)adp->ad_lbn, 3851 (intmax_t)dp->di_db[adp->ad_lbn], 3852 (intmax_t)adp->ad_newblkno); 3853 } 3854 if (adp->ad_lbn >= NDADDR && 3855 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) { 3856 FREE_LOCK(&lk); 3857 panic("%s indirect pointer #%jd mismatch %jd != %jd", 3858 "softdep_write_inodeblock:", 3859 (intmax_t)adp->ad_lbn - NDADDR, 3860 (intmax_t)dp->di_ib[adp->ad_lbn - NDADDR], 3861 (intmax_t)adp->ad_newblkno); 3862 } 3863 deplist |= 1 << adp->ad_lbn; 3864 if ((adp->ad_state & ATTACHED) == 0) { 3865 FREE_LOCK(&lk); 3866 panic("softdep_write_inodeblock: Unknown state 0x%x", 3867 adp->ad_state); 3868 } 3869#endif /* DIAGNOSTIC */ 3870 adp->ad_state &= ~ATTACHED; 3871 adp->ad_state |= UNDONE; 3872 } 3873 /* 3874 * The on-disk inode cannot claim to be any larger than the last 3875 * fragment that has been written. Otherwise, the on-disk inode 3876 * might have fragments that were not the last block in the file 3877 * which would corrupt the filesystem. 3878 */ 3879 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3880 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 3881 if (adp->ad_lbn >= NDADDR) 3882 break; 3883 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 3884 /* keep going until hitting a rollback to a frag */ 3885 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 3886 continue; 3887 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 3888 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 3889#ifdef DIAGNOSTIC 3890 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) { 3891 FREE_LOCK(&lk); 3892 panic("softdep_write_inodeblock: lost dep2"); 3893 } 3894#endif /* DIAGNOSTIC */ 3895 dp->di_db[i] = 0; 3896 } 3897 for (i = 0; i < NIADDR; i++) { 3898#ifdef DIAGNOSTIC 3899 if (dp->di_ib[i] != 0 && 3900 (deplist & ((1 << NDADDR) << i)) == 0) { 3901 FREE_LOCK(&lk); 3902 panic("softdep_write_inodeblock: lost dep3"); 3903 } 3904#endif /* DIAGNOSTIC */ 3905 dp->di_ib[i] = 0; 3906 } 3907 FREE_LOCK(&lk); 3908 return; 3909 } 3910 /* 3911 * If we have zero'ed out the last allocated block of the file, 3912 * roll back the size to the last currently allocated block. 3913 * We know that this last allocated block is a full-sized as 3914 * we already checked for fragments in the loop above. 3915 */ 3916 if (lastadp != NULL && 3917 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 3918 for (i = lastadp->ad_lbn; i >= 0; i--) 3919 if (dp->di_db[i] != 0) 3920 break; 3921 dp->di_size = (i + 1) * fs->fs_bsize; 3922 } 3923 /* 3924 * The only dependencies are for indirect blocks. 3925 * 3926 * The file size for indirect block additions is not guaranteed. 3927 * Such a guarantee would be non-trivial to achieve. The conventional 3928 * synchronous write implementation also does not make this guarantee. 3929 * Fsck should catch and fix discrepancies. Arguably, the file size 3930 * can be over-estimated without destroying integrity when the file 3931 * moves into the indirect blocks (i.e., is large). If we want to 3932 * postpone fsck, we are stuck with this argument. 3933 */ 3934 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 3935 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 3936 FREE_LOCK(&lk); 3937} 3938 3939/* 3940 * This routine is called during the completion interrupt 3941 * service routine for a disk write (from the procedure called 3942 * by the device driver to inform the filesystem caches of 3943 * a request completion). It should be called early in this 3944 * procedure, before the block is made available to other 3945 * processes or other routines are called. 3946 */ 3947static void 3948softdep_disk_write_complete(bp) 3949 struct buf *bp; /* describes the completed disk write */ 3950{ 3951 struct worklist *wk; 3952 struct workhead reattach; 3953 struct newblk *newblk; 3954 struct allocindir *aip; 3955 struct allocdirect *adp; 3956 struct indirdep *indirdep; 3957 struct inodedep *inodedep; 3958 struct bmsafemap *bmsafemap; 3959 3960 /* 3961 * If an error occurred while doing the write, then the data 3962 * has not hit the disk and the dependencies cannot be unrolled. 3963 */ 3964 if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0) 3965 return; 3966#ifdef DEBUG 3967 if (lk.lkt_held != NOHOLDER) 3968 panic("softdep_disk_write_complete: lock is held"); 3969 lk.lkt_held = SPECIAL_FLAG; 3970#endif 3971 LIST_INIT(&reattach); 3972 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 3973 WORKLIST_REMOVE(wk); 3974 switch (wk->wk_type) { 3975 3976 case D_PAGEDEP: 3977 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 3978 WORKLIST_INSERT(&reattach, wk); 3979 continue; 3980 3981 case D_INODEDEP: 3982 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 3983 WORKLIST_INSERT(&reattach, wk); 3984 continue; 3985 3986 case D_BMSAFEMAP: 3987 bmsafemap = WK_BMSAFEMAP(wk); 3988 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) { 3989 newblk->nb_state |= DEPCOMPLETE; 3990 newblk->nb_bmsafemap = NULL; 3991 LIST_REMOVE(newblk, nb_deps); 3992 } 3993 while ((adp = 3994 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) { 3995 adp->ad_state |= DEPCOMPLETE; 3996 adp->ad_buf = NULL; 3997 LIST_REMOVE(adp, ad_deps); 3998 handle_allocdirect_partdone(adp); 3999 } 4000 while ((aip = 4001 LIST_FIRST(&bmsafemap->sm_allocindirhd))) { 4002 aip->ai_state |= DEPCOMPLETE; 4003 aip->ai_buf = NULL; 4004 LIST_REMOVE(aip, ai_deps); 4005 handle_allocindir_partdone(aip); 4006 } 4007 while ((inodedep = 4008 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { 4009 inodedep->id_state |= DEPCOMPLETE; 4010 LIST_REMOVE(inodedep, id_deps); 4011 inodedep->id_buf = NULL; 4012 } 4013 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 4014 continue; 4015 4016 case D_MKDIR: 4017 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 4018 continue; 4019 4020 case D_ALLOCDIRECT: 4021 adp = WK_ALLOCDIRECT(wk); 4022 adp->ad_state |= COMPLETE; 4023 handle_allocdirect_partdone(adp); 4024 continue; 4025 4026 case D_ALLOCINDIR: 4027 aip = WK_ALLOCINDIR(wk); 4028 aip->ai_state |= COMPLETE; 4029 handle_allocindir_partdone(aip); 4030 continue; 4031 4032 case D_INDIRDEP: 4033 indirdep = WK_INDIRDEP(wk); 4034 if (indirdep->ir_state & GOINGAWAY) { 4035 lk.lkt_held = NOHOLDER; 4036 panic("disk_write_complete: indirdep gone"); 4037 } 4038 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 4039 FREE(indirdep->ir_saveddata, M_INDIRDEP); 4040 indirdep->ir_saveddata = 0; 4041 indirdep->ir_state &= ~UNDONE; 4042 indirdep->ir_state |= ATTACHED; 4043 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 4044 handle_allocindir_partdone(aip); 4045 if (aip == LIST_FIRST(&indirdep->ir_donehd)) { 4046 lk.lkt_held = NOHOLDER; 4047 panic("disk_write_complete: not gone"); 4048 } 4049 } 4050 WORKLIST_INSERT(&reattach, wk); 4051 if ((bp->b_flags & B_DELWRI) == 0) 4052 stat_indir_blk_ptrs++; 4053 bdirty(bp); 4054 continue; 4055 4056 default: 4057 lk.lkt_held = NOHOLDER; 4058 panic("handle_disk_write_complete: Unknown type %s", 4059 TYPENAME(wk->wk_type)); 4060 /* NOTREACHED */ 4061 } 4062 } 4063 /* 4064 * Reattach any requests that must be redone. 4065 */ 4066 while ((wk = LIST_FIRST(&reattach)) != NULL) { 4067 WORKLIST_REMOVE(wk); 4068 WORKLIST_INSERT(&bp->b_dep, wk); 4069 } 4070#ifdef DEBUG 4071 if (lk.lkt_held != SPECIAL_FLAG) 4072 panic("softdep_disk_write_complete: lock lost"); 4073 lk.lkt_held = NOHOLDER; 4074#endif 4075} 4076 4077/* 4078 * Called from within softdep_disk_write_complete above. Note that 4079 * this routine is always called from interrupt level with further 4080 * splbio interrupts blocked. 4081 */ 4082static void 4083handle_allocdirect_partdone(adp) 4084 struct allocdirect *adp; /* the completed allocdirect */ 4085{ 4086 struct allocdirectlst *listhead; 4087 struct allocdirect *listadp; 4088 struct inodedep *inodedep; 4089 long bsize, delay; 4090 4091 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 4092 return; 4093 if (adp->ad_buf != NULL) { 4094 lk.lkt_held = NOHOLDER; 4095 panic("handle_allocdirect_partdone: dangling dep"); 4096 } 4097 /* 4098 * The on-disk inode cannot claim to be any larger than the last 4099 * fragment that has been written. Otherwise, the on-disk inode 4100 * might have fragments that were not the last block in the file 4101 * which would corrupt the filesystem. Thus, we cannot free any 4102 * allocdirects after one whose ad_oldblkno claims a fragment as 4103 * these blocks must be rolled back to zero before writing the inode. 4104 * We check the currently active set of allocdirects in id_inoupdt 4105 * or id_extupdt as appropriate. 4106 */ 4107 inodedep = adp->ad_inodedep; 4108 bsize = inodedep->id_fs->fs_bsize; 4109 if (adp->ad_state & EXTDATA) 4110 listhead = &inodedep->id_extupdt; 4111 else 4112 listhead = &inodedep->id_inoupdt; 4113 TAILQ_FOREACH(listadp, listhead, ad_next) { 4114 /* found our block */ 4115 if (listadp == adp) 4116 break; 4117 /* continue if ad_oldlbn is not a fragment */ 4118 if (listadp->ad_oldsize == 0 || 4119 listadp->ad_oldsize == bsize) 4120 continue; 4121 /* hit a fragment */ 4122 return; 4123 } 4124 /* 4125 * If we have reached the end of the current list without 4126 * finding the just finished dependency, then it must be 4127 * on the future dependency list. Future dependencies cannot 4128 * be freed until they are moved to the current list. 4129 */ 4130 if (listadp == NULL) { 4131#ifdef DEBUG 4132 if (adp->ad_state & EXTDATA) 4133 listhead = &inodedep->id_newextupdt; 4134 else 4135 listhead = &inodedep->id_newinoupdt; 4136 TAILQ_FOREACH(listadp, listhead, ad_next) 4137 /* found our block */ 4138 if (listadp == adp) 4139 break; 4140 if (listadp == NULL) { 4141 lk.lkt_held = NOHOLDER; 4142 panic("handle_allocdirect_partdone: lost dep"); 4143 } 4144#endif /* DEBUG */ 4145 return; 4146 } 4147 /* 4148 * If we have found the just finished dependency, then free 4149 * it along with anything that follows it that is complete. 4150 * If the inode still has a bitmap dependency, then it has 4151 * never been written to disk, hence the on-disk inode cannot 4152 * reference the old fragment so we can free it without delay. 4153 */ 4154 delay = (inodedep->id_state & DEPCOMPLETE); 4155 for (; adp; adp = listadp) { 4156 listadp = TAILQ_NEXT(adp, ad_next); 4157 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 4158 return; 4159 free_allocdirect(listhead, adp, delay); 4160 } 4161} 4162 4163/* 4164 * Called from within softdep_disk_write_complete above. Note that 4165 * this routine is always called from interrupt level with further 4166 * splbio interrupts blocked. 4167 */ 4168static void 4169handle_allocindir_partdone(aip) 4170 struct allocindir *aip; /* the completed allocindir */ 4171{ 4172 struct indirdep *indirdep; 4173 4174 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 4175 return; 4176 if (aip->ai_buf != NULL) { 4177 lk.lkt_held = NOHOLDER; 4178 panic("handle_allocindir_partdone: dangling dependency"); 4179 } 4180 indirdep = aip->ai_indirdep; 4181 if (indirdep->ir_state & UNDONE) { 4182 LIST_REMOVE(aip, ai_next); 4183 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 4184 return; 4185 } 4186 if (indirdep->ir_state & UFS1FMT) 4187 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 4188 aip->ai_newblkno; 4189 else 4190 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 4191 aip->ai_newblkno; 4192 LIST_REMOVE(aip, ai_next); 4193 if (aip->ai_freefrag != NULL) 4194 add_to_worklist(&aip->ai_freefrag->ff_list); 4195 WORKITEM_FREE(aip, D_ALLOCINDIR); 4196} 4197 4198/* 4199 * Called from within softdep_disk_write_complete above to restore 4200 * in-memory inode block contents to their most up-to-date state. Note 4201 * that this routine is always called from interrupt level with further 4202 * splbio interrupts blocked. 4203 */ 4204static int 4205handle_written_inodeblock(inodedep, bp) 4206 struct inodedep *inodedep; 4207 struct buf *bp; /* buffer containing the inode block */ 4208{ 4209 struct worklist *wk, *filefree; 4210 struct allocdirect *adp, *nextadp; 4211 struct ufs1_dinode *dp1 = NULL; 4212 struct ufs2_dinode *dp2 = NULL; 4213 int hadchanges, fstype; 4214 4215 if ((inodedep->id_state & IOSTARTED) == 0) { 4216 lk.lkt_held = NOHOLDER; 4217 panic("handle_written_inodeblock: not started"); 4218 } 4219 inodedep->id_state &= ~IOSTARTED; 4220 inodedep->id_state |= COMPLETE; 4221 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) { 4222 fstype = UFS1; 4223 dp1 = (struct ufs1_dinode *)bp->b_data + 4224 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 4225 } else { 4226 fstype = UFS2; 4227 dp2 = (struct ufs2_dinode *)bp->b_data + 4228 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 4229 } 4230 /* 4231 * If we had to rollback the inode allocation because of 4232 * bitmaps being incomplete, then simply restore it. 4233 * Keep the block dirty so that it will not be reclaimed until 4234 * all associated dependencies have been cleared and the 4235 * corresponding updates written to disk. 4236 */ 4237 if (inodedep->id_savedino1 != NULL) { 4238 if (fstype == UFS1) 4239 *dp1 = *inodedep->id_savedino1; 4240 else 4241 *dp2 = *inodedep->id_savedino2; 4242 FREE(inodedep->id_savedino1, M_INODEDEP); 4243 inodedep->id_savedino1 = NULL; 4244 if ((bp->b_flags & B_DELWRI) == 0) 4245 stat_inode_bitmap++; 4246 bdirty(bp); 4247 return (1); 4248 } 4249 /* 4250 * Roll forward anything that had to be rolled back before 4251 * the inode could be updated. 4252 */ 4253 hadchanges = 0; 4254 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 4255 nextadp = TAILQ_NEXT(adp, ad_next); 4256 if (adp->ad_state & ATTACHED) { 4257 lk.lkt_held = NOHOLDER; 4258 panic("handle_written_inodeblock: new entry"); 4259 } 4260 if (fstype == UFS1) { 4261 if (adp->ad_lbn < NDADDR) { 4262 if (dp1->di_db[adp->ad_lbn]!=adp->ad_oldblkno) { 4263 lk.lkt_held = NOHOLDER; 4264 panic("%s %s #%jd mismatch %d != %jd", 4265 "handle_written_inodeblock:", 4266 "direct pointer", 4267 (intmax_t)adp->ad_lbn, 4268 dp1->di_db[adp->ad_lbn], 4269 (intmax_t)adp->ad_oldblkno); 4270 } 4271 dp1->di_db[adp->ad_lbn] = adp->ad_newblkno; 4272 } else { 4273 if (dp1->di_ib[adp->ad_lbn - NDADDR] != 0) { 4274 lk.lkt_held = NOHOLDER; 4275 panic("%s: %s #%jd allocated as %d", 4276 "handle_written_inodeblock", 4277 "indirect pointer", 4278 (intmax_t)adp->ad_lbn - NDADDR, 4279 dp1->di_ib[adp->ad_lbn - NDADDR]); 4280 } 4281 dp1->di_ib[adp->ad_lbn - NDADDR] = 4282 adp->ad_newblkno; 4283 } 4284 } else { 4285 if (adp->ad_lbn < NDADDR) { 4286 if (dp2->di_db[adp->ad_lbn]!=adp->ad_oldblkno) { 4287 lk.lkt_held = NOHOLDER; 4288 panic("%s: %s #%jd %s %jd != %jd", 4289 "handle_written_inodeblock", 4290 "direct pointer", 4291 (intmax_t)adp->ad_lbn, "mismatch", 4292 (intmax_t)dp2->di_db[adp->ad_lbn], 4293 (intmax_t)adp->ad_oldblkno); 4294 } 4295 dp2->di_db[adp->ad_lbn] = adp->ad_newblkno; 4296 } else { 4297 if (dp2->di_ib[adp->ad_lbn - NDADDR] != 0) { 4298 lk.lkt_held = NOHOLDER; 4299 panic("%s: %s #%jd allocated as %jd", 4300 "handle_written_inodeblock", 4301 "indirect pointer", 4302 (intmax_t)adp->ad_lbn - NDADDR, 4303 (intmax_t) 4304 dp2->di_ib[adp->ad_lbn - NDADDR]); 4305 } 4306 dp2->di_ib[adp->ad_lbn - NDADDR] = 4307 adp->ad_newblkno; 4308 } 4309 } 4310 adp->ad_state &= ~UNDONE; 4311 adp->ad_state |= ATTACHED; 4312 hadchanges = 1; 4313 } 4314 for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) { 4315 nextadp = TAILQ_NEXT(adp, ad_next); 4316 if (adp->ad_state & ATTACHED) { 4317 lk.lkt_held = NOHOLDER; 4318 panic("handle_written_inodeblock: new entry"); 4319 } 4320 if (dp2->di_extb[adp->ad_lbn] != adp->ad_oldblkno) { 4321 lk.lkt_held = NOHOLDER; 4322 panic("%s: direct pointers #%jd %s %jd != %jd", 4323 "handle_written_inodeblock", 4324 (intmax_t)adp->ad_lbn, "mismatch", 4325 (intmax_t)dp2->di_extb[adp->ad_lbn], 4326 (intmax_t)adp->ad_oldblkno); 4327 } 4328 dp2->di_extb[adp->ad_lbn] = adp->ad_newblkno; 4329 adp->ad_state &= ~UNDONE; 4330 adp->ad_state |= ATTACHED; 4331 hadchanges = 1; 4332 } 4333 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 4334 stat_direct_blk_ptrs++; 4335 /* 4336 * Reset the file size to its most up-to-date value. 4337 */ 4338 if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1) { 4339 lk.lkt_held = NOHOLDER; 4340 panic("handle_written_inodeblock: bad size"); 4341 } 4342 if (fstype == UFS1) { 4343 if (dp1->di_size != inodedep->id_savedsize) { 4344 dp1->di_size = inodedep->id_savedsize; 4345 hadchanges = 1; 4346 } 4347 } else { 4348 if (dp2->di_size != inodedep->id_savedsize) { 4349 dp2->di_size = inodedep->id_savedsize; 4350 hadchanges = 1; 4351 } 4352 if (dp2->di_extsize != inodedep->id_savedextsize) { 4353 dp2->di_extsize = inodedep->id_savedextsize; 4354 hadchanges = 1; 4355 } 4356 } 4357 inodedep->id_savedsize = -1; 4358 inodedep->id_savedextsize = -1; 4359 /* 4360 * If there were any rollbacks in the inode block, then it must be 4361 * marked dirty so that its will eventually get written back in 4362 * its correct form. 4363 */ 4364 if (hadchanges) 4365 bdirty(bp); 4366 /* 4367 * Process any allocdirects that completed during the update. 4368 */ 4369 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 4370 handle_allocdirect_partdone(adp); 4371 if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 4372 handle_allocdirect_partdone(adp); 4373 /* 4374 * Process deallocations that were held pending until the 4375 * inode had been written to disk. Freeing of the inode 4376 * is delayed until after all blocks have been freed to 4377 * avoid creation of new <vfsid, inum, lbn> triples 4378 * before the old ones have been deleted. 4379 */ 4380 filefree = NULL; 4381 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 4382 WORKLIST_REMOVE(wk); 4383 switch (wk->wk_type) { 4384 4385 case D_FREEFILE: 4386 /* 4387 * We defer adding filefree to the worklist until 4388 * all other additions have been made to ensure 4389 * that it will be done after all the old blocks 4390 * have been freed. 4391 */ 4392 if (filefree != NULL) { 4393 lk.lkt_held = NOHOLDER; 4394 panic("handle_written_inodeblock: filefree"); 4395 } 4396 filefree = wk; 4397 continue; 4398 4399 case D_MKDIR: 4400 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 4401 continue; 4402 4403 case D_DIRADD: 4404 diradd_inode_written(WK_DIRADD(wk), inodedep); 4405 continue; 4406 4407 case D_FREEBLKS: 4408 case D_FREEFRAG: 4409 case D_DIRREM: 4410 add_to_worklist(wk); 4411 continue; 4412 4413 case D_NEWDIRBLK: 4414 free_newdirblk(WK_NEWDIRBLK(wk)); 4415 continue; 4416 4417 default: 4418 lk.lkt_held = NOHOLDER; 4419 panic("handle_written_inodeblock: Unknown type %s", 4420 TYPENAME(wk->wk_type)); 4421 /* NOTREACHED */ 4422 } 4423 } 4424 if (filefree != NULL) { 4425 if (free_inodedep(inodedep) == 0) { 4426 lk.lkt_held = NOHOLDER; 4427 panic("handle_written_inodeblock: live inodedep"); 4428 } 4429 add_to_worklist(filefree); 4430 return (0); 4431 } 4432 4433 /* 4434 * If no outstanding dependencies, free it. 4435 */ 4436 if (free_inodedep(inodedep) || 4437 (TAILQ_FIRST(&inodedep->id_inoupdt) == 0 && 4438 TAILQ_FIRST(&inodedep->id_extupdt) == 0)) 4439 return (0); 4440 return (hadchanges); 4441} 4442 4443/* 4444 * Process a diradd entry after its dependent inode has been written. 4445 * This routine must be called with splbio interrupts blocked. 4446 */ 4447static void 4448diradd_inode_written(dap, inodedep) 4449 struct diradd *dap; 4450 struct inodedep *inodedep; 4451{ 4452 struct pagedep *pagedep; 4453 4454 dap->da_state |= COMPLETE; 4455 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 4456 if (dap->da_state & DIRCHG) 4457 pagedep = dap->da_previous->dm_pagedep; 4458 else 4459 pagedep = dap->da_pagedep; 4460 LIST_REMOVE(dap, da_pdlist); 4461 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 4462 } 4463 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 4464} 4465 4466/* 4467 * Handle the completion of a mkdir dependency. 4468 */ 4469static void 4470handle_written_mkdir(mkdir, type) 4471 struct mkdir *mkdir; 4472 int type; 4473{ 4474 struct diradd *dap; 4475 struct pagedep *pagedep; 4476 4477 if (mkdir->md_state != type) { 4478 lk.lkt_held = NOHOLDER; 4479 panic("handle_written_mkdir: bad type"); 4480 } 4481 dap = mkdir->md_diradd; 4482 dap->da_state &= ~type; 4483 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 4484 dap->da_state |= DEPCOMPLETE; 4485 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 4486 if (dap->da_state & DIRCHG) 4487 pagedep = dap->da_previous->dm_pagedep; 4488 else 4489 pagedep = dap->da_pagedep; 4490 LIST_REMOVE(dap, da_pdlist); 4491 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 4492 } 4493 LIST_REMOVE(mkdir, md_mkdirs); 4494 WORKITEM_FREE(mkdir, D_MKDIR); 4495} 4496 4497/* 4498 * Called from within softdep_disk_write_complete above. 4499 * A write operation was just completed. Removed inodes can 4500 * now be freed and associated block pointers may be committed. 4501 * Note that this routine is always called from interrupt level 4502 * with further splbio interrupts blocked. 4503 */ 4504static int 4505handle_written_filepage(pagedep, bp) 4506 struct pagedep *pagedep; 4507 struct buf *bp; /* buffer containing the written page */ 4508{ 4509 struct dirrem *dirrem; 4510 struct diradd *dap, *nextdap; 4511 struct direct *ep; 4512 int i, chgs; 4513 4514 if ((pagedep->pd_state & IOSTARTED) == 0) { 4515 lk.lkt_held = NOHOLDER; 4516 panic("handle_written_filepage: not started"); 4517 } 4518 pagedep->pd_state &= ~IOSTARTED; 4519 /* 4520 * Process any directory removals that have been committed. 4521 */ 4522 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 4523 LIST_REMOVE(dirrem, dm_next); 4524 dirrem->dm_dirinum = pagedep->pd_ino; 4525 add_to_worklist(&dirrem->dm_list); 4526 } 4527 /* 4528 * Free any directory additions that have been committed. 4529 * If it is a newly allocated block, we have to wait until 4530 * the on-disk directory inode claims the new block. 4531 */ 4532 if ((pagedep->pd_state & NEWBLOCK) == 0) 4533 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 4534 free_diradd(dap); 4535 /* 4536 * Uncommitted directory entries must be restored. 4537 */ 4538 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 4539 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 4540 dap = nextdap) { 4541 nextdap = LIST_NEXT(dap, da_pdlist); 4542 if (dap->da_state & ATTACHED) { 4543 lk.lkt_held = NOHOLDER; 4544 panic("handle_written_filepage: attached"); 4545 } 4546 ep = (struct direct *) 4547 ((char *)bp->b_data + dap->da_offset); 4548 ep->d_ino = dap->da_newinum; 4549 dap->da_state &= ~UNDONE; 4550 dap->da_state |= ATTACHED; 4551 chgs = 1; 4552 /* 4553 * If the inode referenced by the directory has 4554 * been written out, then the dependency can be 4555 * moved to the pending list. 4556 */ 4557 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 4558 LIST_REMOVE(dap, da_pdlist); 4559 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 4560 da_pdlist); 4561 } 4562 } 4563 } 4564 /* 4565 * If there were any rollbacks in the directory, then it must be 4566 * marked dirty so that its will eventually get written back in 4567 * its correct form. 4568 */ 4569 if (chgs) { 4570 if ((bp->b_flags & B_DELWRI) == 0) 4571 stat_dir_entry++; 4572 bdirty(bp); 4573 return (1); 4574 } 4575 /* 4576 * If we are not waiting for a new directory block to be 4577 * claimed by its inode, then the pagedep will be freed. 4578 * Otherwise it will remain to track any new entries on 4579 * the page in case they are fsync'ed. 4580 */ 4581 if ((pagedep->pd_state & NEWBLOCK) == 0) { 4582 LIST_REMOVE(pagedep, pd_hash); 4583 WORKITEM_FREE(pagedep, D_PAGEDEP); 4584 } 4585 return (0); 4586} 4587 4588/* 4589 * Writing back in-core inode structures. 4590 * 4591 * The filesystem only accesses an inode's contents when it occupies an 4592 * "in-core" inode structure. These "in-core" structures are separate from 4593 * the page frames used to cache inode blocks. Only the latter are 4594 * transferred to/from the disk. So, when the updated contents of the 4595 * "in-core" inode structure are copied to the corresponding in-memory inode 4596 * block, the dependencies are also transferred. The following procedure is 4597 * called when copying a dirty "in-core" inode to a cached inode block. 4598 */ 4599 4600/* 4601 * Called when an inode is loaded from disk. If the effective link count 4602 * differed from the actual link count when it was last flushed, then we 4603 * need to ensure that the correct effective link count is put back. 4604 */ 4605void 4606softdep_load_inodeblock(ip) 4607 struct inode *ip; /* the "in_core" copy of the inode */ 4608{ 4609 struct inodedep *inodedep; 4610 4611 /* 4612 * Check for alternate nlink count. 4613 */ 4614 ip->i_effnlink = ip->i_nlink; 4615 ACQUIRE_LOCK(&lk); 4616 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 4617 FREE_LOCK(&lk); 4618 return; 4619 } 4620 ip->i_effnlink -= inodedep->id_nlinkdelta; 4621 if (inodedep->id_state & SPACECOUNTED) 4622 ip->i_flag |= IN_SPACECOUNTED; 4623 FREE_LOCK(&lk); 4624} 4625 4626/* 4627 * This routine is called just before the "in-core" inode 4628 * information is to be copied to the in-memory inode block. 4629 * Recall that an inode block contains several inodes. If 4630 * the force flag is set, then the dependencies will be 4631 * cleared so that the update can always be made. Note that 4632 * the buffer is locked when this routine is called, so we 4633 * will never be in the middle of writing the inode block 4634 * to disk. 4635 */ 4636void 4637softdep_update_inodeblock(ip, bp, waitfor) 4638 struct inode *ip; /* the "in_core" copy of the inode */ 4639 struct buf *bp; /* the buffer containing the inode block */ 4640 int waitfor; /* nonzero => update must be allowed */ 4641{ 4642 struct inodedep *inodedep; 4643 struct worklist *wk; 4644 struct buf *ibp; 4645 int error; 4646 4647 /* 4648 * If the effective link count is not equal to the actual link 4649 * count, then we must track the difference in an inodedep while 4650 * the inode is (potentially) tossed out of the cache. Otherwise, 4651 * if there is no existing inodedep, then there are no dependencies 4652 * to track. 4653 */ 4654 ACQUIRE_LOCK(&lk); 4655 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 4656 FREE_LOCK(&lk); 4657 if (ip->i_effnlink != ip->i_nlink) 4658 panic("softdep_update_inodeblock: bad link count"); 4659 return; 4660 } 4661 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) { 4662 FREE_LOCK(&lk); 4663 panic("softdep_update_inodeblock: bad delta"); 4664 } 4665 /* 4666 * Changes have been initiated. Anything depending on these 4667 * changes cannot occur until this inode has been written. 4668 */ 4669 inodedep->id_state &= ~COMPLETE; 4670 if ((inodedep->id_state & ONWORKLIST) == 0) 4671 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 4672 /* 4673 * Any new dependencies associated with the incore inode must 4674 * now be moved to the list associated with the buffer holding 4675 * the in-memory copy of the inode. Once merged process any 4676 * allocdirects that are completed by the merger. 4677 */ 4678 merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt); 4679 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL) 4680 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); 4681 merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt); 4682 if (TAILQ_FIRST(&inodedep->id_extupdt) != NULL) 4683 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt)); 4684 /* 4685 * Now that the inode has been pushed into the buffer, the 4686 * operations dependent on the inode being written to disk 4687 * can be moved to the id_bufwait so that they will be 4688 * processed when the buffer I/O completes. 4689 */ 4690 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 4691 WORKLIST_REMOVE(wk); 4692 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 4693 } 4694 /* 4695 * Newly allocated inodes cannot be written until the bitmap 4696 * that allocates them have been written (indicated by 4697 * DEPCOMPLETE being set in id_state). If we are doing a 4698 * forced sync (e.g., an fsync on a file), we force the bitmap 4699 * to be written so that the update can be done. 4700 */ 4701 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) { 4702 FREE_LOCK(&lk); 4703 return; 4704 } 4705 ibp = inodedep->id_buf; 4706 ibp = getdirtybuf(&ibp, NULL, MNT_WAIT); 4707 FREE_LOCK(&lk); 4708 if (ibp && (error = bwrite(ibp)) != 0) 4709 softdep_error("softdep_update_inodeblock: bwrite", error); 4710 if ((inodedep->id_state & DEPCOMPLETE) == 0) 4711 panic("softdep_update_inodeblock: update failed"); 4712} 4713 4714/* 4715 * Merge the a new inode dependency list (such as id_newinoupdt) into an 4716 * old inode dependency list (such as id_inoupdt). This routine must be 4717 * called with splbio interrupts blocked. 4718 */ 4719static void 4720merge_inode_lists(newlisthead, oldlisthead) 4721 struct allocdirectlst *newlisthead; 4722 struct allocdirectlst *oldlisthead; 4723{ 4724 struct allocdirect *listadp, *newadp; 4725 4726 newadp = TAILQ_FIRST(newlisthead); 4727 for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) { 4728 if (listadp->ad_lbn < newadp->ad_lbn) { 4729 listadp = TAILQ_NEXT(listadp, ad_next); 4730 continue; 4731 } 4732 TAILQ_REMOVE(newlisthead, newadp, ad_next); 4733 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 4734 if (listadp->ad_lbn == newadp->ad_lbn) { 4735 allocdirect_merge(oldlisthead, newadp, 4736 listadp); 4737 listadp = newadp; 4738 } 4739 newadp = TAILQ_FIRST(newlisthead); 4740 } 4741 while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) { 4742 TAILQ_REMOVE(newlisthead, newadp, ad_next); 4743 TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next); 4744 } 4745} 4746 4747/* 4748 * If we are doing an fsync, then we must ensure that any directory 4749 * entries for the inode have been written after the inode gets to disk. 4750 */ 4751int 4752softdep_fsync(vp) 4753 struct vnode *vp; /* the "in_core" copy of the inode */ 4754{ 4755 struct inodedep *inodedep; 4756 struct pagedep *pagedep; 4757 struct worklist *wk; 4758 struct diradd *dap; 4759 struct mount *mnt; 4760 struct vnode *pvp; 4761 struct inode *ip; 4762 struct buf *bp; 4763 struct fs *fs; 4764 struct thread *td = curthread; 4765 int error, flushparent; 4766 ino_t parentino; 4767 ufs_lbn_t lbn; 4768 4769 ip = VTOI(vp); 4770 fs = ip->i_fs; 4771 ACQUIRE_LOCK(&lk); 4772 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) { 4773 FREE_LOCK(&lk); 4774 return (0); 4775 } 4776 if (LIST_FIRST(&inodedep->id_inowait) != NULL || 4777 LIST_FIRST(&inodedep->id_bufwait) != NULL || 4778 TAILQ_FIRST(&inodedep->id_extupdt) != NULL || 4779 TAILQ_FIRST(&inodedep->id_newextupdt) != NULL || 4780 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 4781 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) { 4782 FREE_LOCK(&lk); 4783 panic("softdep_fsync: pending ops"); 4784 } 4785 for (error = 0, flushparent = 0; ; ) { 4786 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 4787 break; 4788 if (wk->wk_type != D_DIRADD) { 4789 FREE_LOCK(&lk); 4790 panic("softdep_fsync: Unexpected type %s", 4791 TYPENAME(wk->wk_type)); 4792 } 4793 dap = WK_DIRADD(wk); 4794 /* 4795 * Flush our parent if this directory entry has a MKDIR_PARENT 4796 * dependency or is contained in a newly allocated block. 4797 */ 4798 if (dap->da_state & DIRCHG) 4799 pagedep = dap->da_previous->dm_pagedep; 4800 else 4801 pagedep = dap->da_pagedep; 4802 mnt = pagedep->pd_mnt; 4803 parentino = pagedep->pd_ino; 4804 lbn = pagedep->pd_lbn; 4805 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) { 4806 FREE_LOCK(&lk); 4807 panic("softdep_fsync: dirty"); 4808 } 4809 if ((dap->da_state & MKDIR_PARENT) || 4810 (pagedep->pd_state & NEWBLOCK)) 4811 flushparent = 1; 4812 else 4813 flushparent = 0; 4814 /* 4815 * If we are being fsync'ed as part of vgone'ing this vnode, 4816 * then we will not be able to release and recover the 4817 * vnode below, so we just have to give up on writing its 4818 * directory entry out. It will eventually be written, just 4819 * not now, but then the user was not asking to have it 4820 * written, so we are not breaking any promises. 4821 */ 4822 if (vp->v_iflag & VI_XLOCK) 4823 break; 4824 /* 4825 * We prevent deadlock by always fetching inodes from the 4826 * root, moving down the directory tree. Thus, when fetching 4827 * our parent directory, we first try to get the lock. If 4828 * that fails, we must unlock ourselves before requesting 4829 * the lock on our parent. See the comment in ufs_lookup 4830 * for details on possible races. 4831 */ 4832 FREE_LOCK(&lk); 4833 if (VFS_VGET(mnt, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp)) { 4834 VOP_UNLOCK(vp, 0, td); 4835 error = VFS_VGET(mnt, parentino, LK_EXCLUSIVE, &pvp); 4836 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 4837 if (error != 0) 4838 return (error); 4839 } 4840 /* 4841 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps 4842 * that are contained in direct blocks will be resolved by 4843 * doing a UFS_UPDATE. Pagedeps contained in indirect blocks 4844 * may require a complete sync'ing of the directory. So, we 4845 * try the cheap and fast UFS_UPDATE first, and if that fails, 4846 * then we do the slower VOP_FSYNC of the directory. 4847 */ 4848 if (flushparent) { 4849 if ((error = UFS_UPDATE(pvp, 1)) != 0) { 4850 vput(pvp); 4851 return (error); 4852 } 4853 if ((pagedep->pd_state & NEWBLOCK) && 4854 (error = VOP_FSYNC(pvp, td->td_ucred, MNT_WAIT, td))) { 4855 vput(pvp); 4856 return (error); 4857 } 4858 } 4859 /* 4860 * Flush directory page containing the inode's name. 4861 */ 4862 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred, 4863 &bp); 4864 if (error == 0) 4865 error = bwrite(bp); 4866 else 4867 brelse(bp); 4868 vput(pvp); 4869 if (error != 0) 4870 return (error); 4871 ACQUIRE_LOCK(&lk); 4872 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) 4873 break; 4874 } 4875 FREE_LOCK(&lk); 4876 return (0); 4877} 4878 4879/* 4880 * Flush all the dirty bitmaps associated with the block device 4881 * before flushing the rest of the dirty blocks so as to reduce 4882 * the number of dependencies that will have to be rolled back. 4883 */ 4884void 4885softdep_fsync_mountdev(vp) 4886 struct vnode *vp; 4887{ 4888 struct buf *bp, *nbp; 4889 struct worklist *wk; 4890 4891 if (!vn_isdisk(vp, NULL)) 4892 panic("softdep_fsync_mountdev: vnode not a disk"); 4893 ACQUIRE_LOCK(&lk); 4894 VI_LOCK(vp); 4895 for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 4896 nbp = TAILQ_NEXT(bp, b_vnbufs); 4897 /* 4898 * If it is already scheduled, skip to the next buffer. 4899 */ 4900 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 4901 continue; 4902 4903 if ((bp->b_flags & B_DELWRI) == 0) { 4904 FREE_LOCK(&lk); 4905 panic("softdep_fsync_mountdev: not dirty"); 4906 } 4907 /* 4908 * We are only interested in bitmaps with outstanding 4909 * dependencies. 4910 */ 4911 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 4912 wk->wk_type != D_BMSAFEMAP || 4913 (bp->b_vflags & BV_BKGRDINPROG)) { 4914 BUF_UNLOCK(bp); 4915 continue; 4916 } 4917 VI_UNLOCK(vp); 4918 bremfree(bp); 4919 FREE_LOCK(&lk); 4920 (void) bawrite(bp); 4921 ACQUIRE_LOCK(&lk); 4922 /* 4923 * Since we may have slept during the I/O, we need 4924 * to start from a known point. 4925 */ 4926 VI_LOCK(vp); 4927 nbp = TAILQ_FIRST(&vp->v_dirtyblkhd); 4928 } 4929 drain_output(vp, 1); 4930 VI_UNLOCK(vp); 4931 FREE_LOCK(&lk); 4932} 4933 4934/* 4935 * This routine is called when we are trying to synchronously flush a 4936 * file. This routine must eliminate any filesystem metadata dependencies 4937 * so that the syncing routine can succeed by pushing the dirty blocks 4938 * associated with the file. If any I/O errors occur, they are returned. 4939 */ 4940int 4941softdep_sync_metadata(ap) 4942 struct vop_fsync_args /* { 4943 struct vnode *a_vp; 4944 struct ucred *a_cred; 4945 int a_waitfor; 4946 struct thread *a_td; 4947 } */ *ap; 4948{ 4949 struct vnode *vp = ap->a_vp; 4950 struct pagedep *pagedep; 4951 struct allocdirect *adp; 4952 struct allocindir *aip; 4953 struct buf *bp, *nbp; 4954 struct worklist *wk; 4955 int i, error, waitfor; 4956 4957 /* 4958 * Check whether this vnode is involved in a filesystem 4959 * that is doing soft dependency processing. 4960 */ 4961 if (!vn_isdisk(vp, NULL)) { 4962 if (!DOINGSOFTDEP(vp)) 4963 return (0); 4964 } else 4965 if (vp->v_rdev->si_mountpoint == NULL || 4966 (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0) 4967 return (0); 4968 /* 4969 * Ensure that any direct block dependencies have been cleared. 4970 */ 4971 ACQUIRE_LOCK(&lk); 4972 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) { 4973 FREE_LOCK(&lk); 4974 return (error); 4975 } 4976 /* 4977 * For most files, the only metadata dependencies are the 4978 * cylinder group maps that allocate their inode or blocks. 4979 * The block allocation dependencies can be found by traversing 4980 * the dependency lists for any buffers that remain on their 4981 * dirty buffer list. The inode allocation dependency will 4982 * be resolved when the inode is updated with MNT_WAIT. 4983 * This work is done in two passes. The first pass grabs most 4984 * of the buffers and begins asynchronously writing them. The 4985 * only way to wait for these asynchronous writes is to sleep 4986 * on the filesystem vnode which may stay busy for a long time 4987 * if the filesystem is active. So, instead, we make a second 4988 * pass over the dependencies blocking on each write. In the 4989 * usual case we will be blocking against a write that we 4990 * initiated, so when it is done the dependency will have been 4991 * resolved. Thus the second pass is expected to end quickly. 4992 */ 4993 waitfor = MNT_NOWAIT; 4994top: 4995 /* 4996 * We must wait for any I/O in progress to finish so that 4997 * all potential buffers on the dirty list will be visible. 4998 */ 4999 VI_LOCK(vp); 5000 drain_output(vp, 1); 5001 bp = getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), 5002 VI_MTX(vp), MNT_WAIT); 5003 if (bp == NULL) { 5004 VI_UNLOCK(vp); 5005 FREE_LOCK(&lk); 5006 return (0); 5007 } 5008 /* While syncing snapshots, we must allow recursive lookups */ 5009 bp->b_lock.lk_flags |= LK_CANRECURSE; 5010loop: 5011 /* 5012 * As we hold the buffer locked, none of its dependencies 5013 * will disappear. 5014 */ 5015 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5016 switch (wk->wk_type) { 5017 5018 case D_ALLOCDIRECT: 5019 adp = WK_ALLOCDIRECT(wk); 5020 if (adp->ad_state & DEPCOMPLETE) 5021 continue; 5022 nbp = adp->ad_buf; 5023 nbp = getdirtybuf(&nbp, NULL, waitfor); 5024 if (nbp == NULL) 5025 continue; 5026 FREE_LOCK(&lk); 5027 if (waitfor == MNT_NOWAIT) { 5028 bawrite(nbp); 5029 } else if ((error = bwrite(nbp)) != 0) { 5030 break; 5031 } 5032 ACQUIRE_LOCK(&lk); 5033 continue; 5034 5035 case D_ALLOCINDIR: 5036 aip = WK_ALLOCINDIR(wk); 5037 if (aip->ai_state & DEPCOMPLETE) 5038 continue; 5039 nbp = aip->ai_buf; 5040 nbp = getdirtybuf(&nbp, NULL, waitfor); 5041 if (nbp == NULL) 5042 continue; 5043 FREE_LOCK(&lk); 5044 if (waitfor == MNT_NOWAIT) { 5045 bawrite(nbp); 5046 } else if ((error = bwrite(nbp)) != 0) { 5047 break; 5048 } 5049 ACQUIRE_LOCK(&lk); 5050 continue; 5051 5052 case D_INDIRDEP: 5053 restart: 5054 5055 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) { 5056 if (aip->ai_state & DEPCOMPLETE) 5057 continue; 5058 nbp = aip->ai_buf; 5059 nbp = getdirtybuf(&nbp, NULL, MNT_WAIT); 5060 if (nbp == NULL) 5061 goto restart; 5062 FREE_LOCK(&lk); 5063 if ((error = bwrite(nbp)) != 0) { 5064 break; 5065 } 5066 ACQUIRE_LOCK(&lk); 5067 goto restart; 5068 } 5069 continue; 5070 5071 case D_INODEDEP: 5072 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs, 5073 WK_INODEDEP(wk)->id_ino)) != 0) { 5074 FREE_LOCK(&lk); 5075 break; 5076 } 5077 continue; 5078 5079 case D_PAGEDEP: 5080 /* 5081 * We are trying to sync a directory that may 5082 * have dependencies on both its own metadata 5083 * and/or dependencies on the inodes of any 5084 * recently allocated files. We walk its diradd 5085 * lists pushing out the associated inode. 5086 */ 5087 pagedep = WK_PAGEDEP(wk); 5088 for (i = 0; i < DAHASHSZ; i++) { 5089 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 5090 continue; 5091 if ((error = 5092 flush_pagedep_deps(vp, pagedep->pd_mnt, 5093 &pagedep->pd_diraddhd[i]))) { 5094 FREE_LOCK(&lk); 5095 break; 5096 } 5097 } 5098 continue; 5099 5100 case D_MKDIR: 5101 /* 5102 * This case should never happen if the vnode has 5103 * been properly sync'ed. However, if this function 5104 * is used at a place where the vnode has not yet 5105 * been sync'ed, this dependency can show up. So, 5106 * rather than panic, just flush it. 5107 */ 5108 nbp = WK_MKDIR(wk)->md_buf; 5109 nbp = getdirtybuf(&nbp, NULL, waitfor); 5110 if (nbp == NULL) 5111 continue; 5112 FREE_LOCK(&lk); 5113 if (waitfor == MNT_NOWAIT) { 5114 bawrite(nbp); 5115 } else if ((error = bwrite(nbp)) != 0) { 5116 break; 5117 } 5118 ACQUIRE_LOCK(&lk); 5119 continue; 5120 5121 case D_BMSAFEMAP: 5122 /* 5123 * This case should never happen if the vnode has 5124 * been properly sync'ed. However, if this function 5125 * is used at a place where the vnode has not yet 5126 * been sync'ed, this dependency can show up. So, 5127 * rather than panic, just flush it. 5128 */ 5129 nbp = WK_BMSAFEMAP(wk)->sm_buf; 5130 nbp = getdirtybuf(&nbp, NULL, waitfor); 5131 if (nbp == NULL) 5132 continue; 5133 FREE_LOCK(&lk); 5134 if (waitfor == MNT_NOWAIT) { 5135 bawrite(nbp); 5136 } else if ((error = bwrite(nbp)) != 0) { 5137 break; 5138 } 5139 ACQUIRE_LOCK(&lk); 5140 continue; 5141 5142 default: 5143 FREE_LOCK(&lk); 5144 panic("softdep_sync_metadata: Unknown type %s", 5145 TYPENAME(wk->wk_type)); 5146 /* NOTREACHED */ 5147 } 5148 /* We reach here only in error and unlocked */ 5149 if (error == 0) 5150 panic("softdep_sync_metadata: zero error"); 5151 bp->b_lock.lk_flags &= ~LK_CANRECURSE; 5152 bawrite(bp); 5153 return (error); 5154 } 5155 VI_LOCK(vp); 5156 nbp = getdirtybuf(&TAILQ_NEXT(bp, b_vnbufs), VI_MTX(vp), MNT_WAIT); 5157 if (nbp == NULL) 5158 VI_UNLOCK(vp); 5159 FREE_LOCK(&lk); 5160 bp->b_lock.lk_flags &= ~LK_CANRECURSE; 5161 bawrite(bp); 5162 ACQUIRE_LOCK(&lk); 5163 if (nbp != NULL) { 5164 bp = nbp; 5165 goto loop; 5166 } 5167 /* 5168 * The brief unlock is to allow any pent up dependency 5169 * processing to be done. Then proceed with the second pass. 5170 */ 5171 if (waitfor == MNT_NOWAIT) { 5172 waitfor = MNT_WAIT; 5173 FREE_LOCK(&lk); 5174 ACQUIRE_LOCK(&lk); 5175 goto top; 5176 } 5177 5178 /* 5179 * If we have managed to get rid of all the dirty buffers, 5180 * then we are done. For certain directories and block 5181 * devices, we may need to do further work. 5182 * 5183 * We must wait for any I/O in progress to finish so that 5184 * all potential buffers on the dirty list will be visible. 5185 */ 5186 VI_LOCK(vp); 5187 drain_output(vp, 1); 5188 if (TAILQ_FIRST(&vp->v_dirtyblkhd) == NULL) { 5189 VI_UNLOCK(vp); 5190 FREE_LOCK(&lk); 5191 return (0); 5192 } 5193 VI_UNLOCK(vp); 5194 5195 FREE_LOCK(&lk); 5196 /* 5197 * If we are trying to sync a block device, some of its buffers may 5198 * contain metadata that cannot be written until the contents of some 5199 * partially written files have been written to disk. The only easy 5200 * way to accomplish this is to sync the entire filesystem (luckily 5201 * this happens rarely). 5202 */ 5203 if (vn_isdisk(vp, NULL) && 5204 vp->v_rdev->si_mountpoint && !VOP_ISLOCKED(vp, NULL) && 5205 (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT, ap->a_cred, 5206 ap->a_td)) != 0) 5207 return (error); 5208 return (0); 5209} 5210 5211/* 5212 * Flush the dependencies associated with an inodedep. 5213 * Called with splbio blocked. 5214 */ 5215static int 5216flush_inodedep_deps(fs, ino) 5217 struct fs *fs; 5218 ino_t ino; 5219{ 5220 struct inodedep *inodedep; 5221 int error, waitfor; 5222 5223 /* 5224 * This work is done in two passes. The first pass grabs most 5225 * of the buffers and begins asynchronously writing them. The 5226 * only way to wait for these asynchronous writes is to sleep 5227 * on the filesystem vnode which may stay busy for a long time 5228 * if the filesystem is active. So, instead, we make a second 5229 * pass over the dependencies blocking on each write. In the 5230 * usual case we will be blocking against a write that we 5231 * initiated, so when it is done the dependency will have been 5232 * resolved. Thus the second pass is expected to end quickly. 5233 * We give a brief window at the top of the loop to allow 5234 * any pending I/O to complete. 5235 */ 5236 for (error = 0, waitfor = MNT_NOWAIT; ; ) { 5237 if (error) 5238 return (error); 5239 FREE_LOCK(&lk); 5240 ACQUIRE_LOCK(&lk); 5241 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 5242 return (0); 5243 if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) || 5244 flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) || 5245 flush_deplist(&inodedep->id_extupdt, waitfor, &error) || 5246 flush_deplist(&inodedep->id_newextupdt, waitfor, &error)) 5247 continue; 5248 /* 5249 * If pass2, we are done, otherwise do pass 2. 5250 */ 5251 if (waitfor == MNT_WAIT) 5252 break; 5253 waitfor = MNT_WAIT; 5254 } 5255 /* 5256 * Try freeing inodedep in case all dependencies have been removed. 5257 */ 5258 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0) 5259 (void) free_inodedep(inodedep); 5260 return (0); 5261} 5262 5263/* 5264 * Flush an inode dependency list. 5265 * Called with splbio blocked. 5266 */ 5267static int 5268flush_deplist(listhead, waitfor, errorp) 5269 struct allocdirectlst *listhead; 5270 int waitfor; 5271 int *errorp; 5272{ 5273 struct allocdirect *adp; 5274 struct buf *bp; 5275 5276 TAILQ_FOREACH(adp, listhead, ad_next) { 5277 if (adp->ad_state & DEPCOMPLETE) 5278 continue; 5279 bp = adp->ad_buf; 5280 bp = getdirtybuf(&bp, NULL, waitfor); 5281 if (bp == NULL) { 5282 if (waitfor == MNT_NOWAIT) 5283 continue; 5284 return (1); 5285 } 5286 FREE_LOCK(&lk); 5287 if (waitfor == MNT_NOWAIT) { 5288 bawrite(bp); 5289 } else if ((*errorp = bwrite(bp)) != 0) { 5290 ACQUIRE_LOCK(&lk); 5291 return (1); 5292 } 5293 ACQUIRE_LOCK(&lk); 5294 return (1); 5295 } 5296 return (0); 5297} 5298 5299/* 5300 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 5301 * Called with splbio blocked. 5302 */ 5303static int 5304flush_pagedep_deps(pvp, mp, diraddhdp) 5305 struct vnode *pvp; 5306 struct mount *mp; 5307 struct diraddhd *diraddhdp; 5308{ 5309 struct thread *td = curthread; 5310 struct inodedep *inodedep; 5311 struct ufsmount *ump; 5312 struct diradd *dap; 5313 struct vnode *vp; 5314 int error = 0; 5315 struct buf *bp; 5316 ino_t inum; 5317 5318 ump = VFSTOUFS(mp); 5319 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 5320 /* 5321 * Flush ourselves if this directory entry 5322 * has a MKDIR_PARENT dependency. 5323 */ 5324 if (dap->da_state & MKDIR_PARENT) { 5325 FREE_LOCK(&lk); 5326 if ((error = UFS_UPDATE(pvp, 1)) != 0) 5327 break; 5328 ACQUIRE_LOCK(&lk); 5329 /* 5330 * If that cleared dependencies, go on to next. 5331 */ 5332 if (dap != LIST_FIRST(diraddhdp)) 5333 continue; 5334 if (dap->da_state & MKDIR_PARENT) { 5335 FREE_LOCK(&lk); 5336 panic("flush_pagedep_deps: MKDIR_PARENT"); 5337 } 5338 } 5339 /* 5340 * A newly allocated directory must have its "." and 5341 * ".." entries written out before its name can be 5342 * committed in its parent. We do not want or need 5343 * the full semantics of a synchronous VOP_FSYNC as 5344 * that may end up here again, once for each directory 5345 * level in the filesystem. Instead, we push the blocks 5346 * and wait for them to clear. We have to fsync twice 5347 * because the first call may choose to defer blocks 5348 * that still have dependencies, but deferral will 5349 * happen at most once. 5350 */ 5351 inum = dap->da_newinum; 5352 if (dap->da_state & MKDIR_BODY) { 5353 FREE_LOCK(&lk); 5354 if ((error = VFS_VGET(mp, inum, LK_EXCLUSIVE, &vp))) 5355 break; 5356 if ((error=VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td)) || 5357 (error=VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td))) { 5358 vput(vp); 5359 break; 5360 } 5361 VI_LOCK(vp); 5362 drain_output(vp, 0); 5363 VI_UNLOCK(vp); 5364 vput(vp); 5365 ACQUIRE_LOCK(&lk); 5366 /* 5367 * If that cleared dependencies, go on to next. 5368 */ 5369 if (dap != LIST_FIRST(diraddhdp)) 5370 continue; 5371 if (dap->da_state & MKDIR_BODY) { 5372 FREE_LOCK(&lk); 5373 panic("flush_pagedep_deps: MKDIR_BODY"); 5374 } 5375 } 5376 /* 5377 * Flush the inode on which the directory entry depends. 5378 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 5379 * the only remaining dependency is that the updated inode 5380 * count must get pushed to disk. The inode has already 5381 * been pushed into its inode buffer (via VOP_UPDATE) at 5382 * the time of the reference count change. So we need only 5383 * locate that buffer, ensure that there will be no rollback 5384 * caused by a bitmap dependency, then write the inode buffer. 5385 */ 5386 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) { 5387 FREE_LOCK(&lk); 5388 panic("flush_pagedep_deps: lost inode"); 5389 } 5390 /* 5391 * If the inode still has bitmap dependencies, 5392 * push them to disk. 5393 */ 5394 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 5395 bp = inodedep->id_buf; 5396 bp = getdirtybuf(&bp, NULL, MNT_WAIT); 5397 FREE_LOCK(&lk); 5398 if (bp && (error = bwrite(bp)) != 0) 5399 break; 5400 ACQUIRE_LOCK(&lk); 5401 if (dap != LIST_FIRST(diraddhdp)) 5402 continue; 5403 } 5404 /* 5405 * If the inode is still sitting in a buffer waiting 5406 * to be written, push it to disk. 5407 */ 5408 FREE_LOCK(&lk); 5409 if ((error = bread(ump->um_devvp, 5410 fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), 5411 (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0) { 5412 brelse(bp); 5413 break; 5414 } 5415 if ((error = bwrite(bp)) != 0) 5416 break; 5417 ACQUIRE_LOCK(&lk); 5418 /* 5419 * If we have failed to get rid of all the dependencies 5420 * then something is seriously wrong. 5421 */ 5422 if (dap == LIST_FIRST(diraddhdp)) { 5423 FREE_LOCK(&lk); 5424 panic("flush_pagedep_deps: flush failed"); 5425 } 5426 } 5427 if (error) 5428 ACQUIRE_LOCK(&lk); 5429 return (error); 5430} 5431 5432/* 5433 * A large burst of file addition or deletion activity can drive the 5434 * memory load excessively high. First attempt to slow things down 5435 * using the techniques below. If that fails, this routine requests 5436 * the offending operations to fall back to running synchronously 5437 * until the memory load returns to a reasonable level. 5438 */ 5439int 5440softdep_slowdown(vp) 5441 struct vnode *vp; 5442{ 5443 int max_softdeps_hard; 5444 5445 max_softdeps_hard = max_softdeps * 11 / 10; 5446 if (num_dirrem < max_softdeps_hard / 2 && 5447 num_inodedep < max_softdeps_hard && 5448 VFSTOUFS(vp->v_mount)->um_numindirdeps < maxindirdeps) 5449 return (0); 5450 if (VFSTOUFS(vp->v_mount)->um_numindirdeps >= maxindirdeps) 5451 speedup_syncer(); 5452 stat_sync_limit_hit += 1; 5453 return (1); 5454} 5455 5456/* 5457 * Called by the allocation routines when they are about to fail 5458 * in the hope that we can free up some disk space. 5459 * 5460 * First check to see if the work list has anything on it. If it has, 5461 * clean up entries until we successfully free some space. Because this 5462 * process holds inodes locked, we cannot handle any remove requests 5463 * that might block on a locked inode as that could lead to deadlock. 5464 * If the worklist yields no free space, encourage the syncer daemon 5465 * to help us. In no event will we try for longer than tickdelay seconds. 5466 */ 5467int 5468softdep_request_cleanup(fs, vp) 5469 struct fs *fs; 5470 struct vnode *vp; 5471{ 5472 long starttime; 5473 ufs2_daddr_t needed; 5474 5475 needed = fs->fs_cstotal.cs_nbfree + fs->fs_contigsumsize; 5476 starttime = time_second + tickdelay; 5477 /* 5478 * If we are being called because of a process doing a 5479 * copy-on-write, then it is not safe to update the vnode 5480 * as we may recurse into the copy-on-write routine. 5481 */ 5482 if (!(curthread->td_pflags & TDP_COWINPROGRESS) && 5483 UFS_UPDATE(vp, 1) != 0) 5484 return (0); 5485 while (fs->fs_pendingblocks > 0 && fs->fs_cstotal.cs_nbfree <= needed) { 5486 if (time_second > starttime) 5487 return (0); 5488 if (num_on_worklist > 0 && 5489 process_worklist_item(NULL, LK_NOWAIT) != -1) { 5490 stat_worklist_push += 1; 5491 continue; 5492 } 5493 request_cleanup(FLUSH_REMOVE_WAIT, 0); 5494 } 5495 return (1); 5496} 5497 5498/* 5499 * If memory utilization has gotten too high, deliberately slow things 5500 * down and speed up the I/O processing. 5501 */ 5502static int 5503request_cleanup(resource, islocked) 5504 int resource; 5505 int islocked; 5506{ 5507 struct thread *td = curthread; 5508 5509 /* 5510 * We never hold up the filesystem syncer process. 5511 */ 5512 if (td == filesys_syncer) 5513 return (0); 5514 /* 5515 * First check to see if the work list has gotten backlogged. 5516 * If it has, co-opt this process to help clean up two entries. 5517 * Because this process may hold inodes locked, we cannot 5518 * handle any remove requests that might block on a locked 5519 * inode as that could lead to deadlock. 5520 */ 5521 if (num_on_worklist > max_softdeps / 10) { 5522 if (islocked) 5523 FREE_LOCK(&lk); 5524 process_worklist_item(NULL, LK_NOWAIT); 5525 process_worklist_item(NULL, LK_NOWAIT); 5526 stat_worklist_push += 2; 5527 if (islocked) 5528 ACQUIRE_LOCK(&lk); 5529 return(1); 5530 } 5531 /* 5532 * Next, we attempt to speed up the syncer process. If that 5533 * is successful, then we allow the process to continue. 5534 */ 5535 if (speedup_syncer() && resource != FLUSH_REMOVE_WAIT) 5536 return(0); 5537 /* 5538 * If we are resource constrained on inode dependencies, try 5539 * flushing some dirty inodes. Otherwise, we are constrained 5540 * by file deletions, so try accelerating flushes of directories 5541 * with removal dependencies. We would like to do the cleanup 5542 * here, but we probably hold an inode locked at this point and 5543 * that might deadlock against one that we try to clean. So, 5544 * the best that we can do is request the syncer daemon to do 5545 * the cleanup for us. 5546 */ 5547 switch (resource) { 5548 5549 case FLUSH_INODES: 5550 stat_ino_limit_push += 1; 5551 req_clear_inodedeps += 1; 5552 stat_countp = &stat_ino_limit_hit; 5553 break; 5554 5555 case FLUSH_REMOVE: 5556 case FLUSH_REMOVE_WAIT: 5557 stat_blk_limit_push += 1; 5558 req_clear_remove += 1; 5559 stat_countp = &stat_blk_limit_hit; 5560 break; 5561 5562 default: 5563 if (islocked) 5564 FREE_LOCK(&lk); 5565 panic("request_cleanup: unknown type"); 5566 } 5567 /* 5568 * Hopefully the syncer daemon will catch up and awaken us. 5569 * We wait at most tickdelay before proceeding in any case. 5570 */ 5571 if (islocked == 0) 5572 ACQUIRE_LOCK(&lk); 5573 proc_waiting += 1; 5574 if (handle.callout == NULL) 5575 handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2); 5576 interlocked_sleep(&lk, SLEEP, (caddr_t)&proc_waiting, NULL, PPAUSE, 5577 "softupdate", 0); 5578 proc_waiting -= 1; 5579 if (islocked == 0) 5580 FREE_LOCK(&lk); 5581 return (1); 5582} 5583 5584/* 5585 * Awaken processes pausing in request_cleanup and clear proc_waiting 5586 * to indicate that there is no longer a timer running. 5587 */ 5588static void 5589pause_timer(arg) 5590 void *arg; 5591{ 5592 5593 *stat_countp += 1; 5594 wakeup_one(&proc_waiting); 5595 if (proc_waiting > 0) 5596 handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2); 5597 else 5598 handle.callout = NULL; 5599} 5600 5601/* 5602 * Flush out a directory with at least one removal dependency in an effort to 5603 * reduce the number of dirrem, freefile, and freeblks dependency structures. 5604 */ 5605static void 5606clear_remove(td) 5607 struct thread *td; 5608{ 5609 struct pagedep_hashhead *pagedephd; 5610 struct pagedep *pagedep; 5611 static int next = 0; 5612 struct mount *mp; 5613 struct vnode *vp; 5614 int error, cnt; 5615 ino_t ino; 5616 5617 ACQUIRE_LOCK(&lk); 5618 for (cnt = 0; cnt < pagedep_hash; cnt++) { 5619 pagedephd = &pagedep_hashtbl[next++]; 5620 if (next >= pagedep_hash) 5621 next = 0; 5622 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 5623 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL) 5624 continue; 5625 mp = pagedep->pd_mnt; 5626 ino = pagedep->pd_ino; 5627 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 5628 continue; 5629 FREE_LOCK(&lk); 5630 if ((error = VFS_VGET(mp, ino, LK_EXCLUSIVE, &vp))) { 5631 softdep_error("clear_remove: vget", error); 5632 vn_finished_write(mp); 5633 return; 5634 } 5635 if ((error = VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td))) 5636 softdep_error("clear_remove: fsync", error); 5637 VI_LOCK(vp); 5638 drain_output(vp, 0); 5639 VI_UNLOCK(vp); 5640 vput(vp); 5641 vn_finished_write(mp); 5642 return; 5643 } 5644 } 5645 FREE_LOCK(&lk); 5646} 5647 5648/* 5649 * Clear out a block of dirty inodes in an effort to reduce 5650 * the number of inodedep dependency structures. 5651 */ 5652static void 5653clear_inodedeps(td) 5654 struct thread *td; 5655{ 5656 struct inodedep_hashhead *inodedephd; 5657 struct inodedep *inodedep; 5658 static int next = 0; 5659 struct mount *mp; 5660 struct vnode *vp; 5661 struct fs *fs; 5662 int error, cnt; 5663 ino_t firstino, lastino, ino; 5664 5665 ACQUIRE_LOCK(&lk); 5666 /* 5667 * Pick a random inode dependency to be cleared. 5668 * We will then gather up all the inodes in its block 5669 * that have dependencies and flush them out. 5670 */ 5671 for (cnt = 0; cnt < inodedep_hash; cnt++) { 5672 inodedephd = &inodedep_hashtbl[next++]; 5673 if (next >= inodedep_hash) 5674 next = 0; 5675 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 5676 break; 5677 } 5678 if (inodedep == NULL) 5679 return; 5680 /* 5681 * Ugly code to find mount point given pointer to superblock. 5682 */ 5683 fs = inodedep->id_fs; 5684 TAILQ_FOREACH(mp, &mountlist, mnt_list) 5685 if ((mp->mnt_flag & MNT_SOFTDEP) && fs == VFSTOUFS(mp)->um_fs) 5686 break; 5687 /* 5688 * Find the last inode in the block with dependencies. 5689 */ 5690 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 5691 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 5692 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0) 5693 break; 5694 /* 5695 * Asynchronously push all but the last inode with dependencies. 5696 * Synchronously push the last inode with dependencies to ensure 5697 * that the inode block gets written to free up the inodedeps. 5698 */ 5699 for (ino = firstino; ino <= lastino; ino++) { 5700 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 5701 continue; 5702 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 5703 continue; 5704 FREE_LOCK(&lk); 5705 if ((error = VFS_VGET(mp, ino, LK_EXCLUSIVE, &vp)) != 0) { 5706 softdep_error("clear_inodedeps: vget", error); 5707 vn_finished_write(mp); 5708 return; 5709 } 5710 if (ino == lastino) { 5711 if ((error = VOP_FSYNC(vp, td->td_ucred, MNT_WAIT, td))) 5712 softdep_error("clear_inodedeps: fsync1", error); 5713 } else { 5714 if ((error = VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td))) 5715 softdep_error("clear_inodedeps: fsync2", error); 5716 VI_LOCK(vp); 5717 drain_output(vp, 0); 5718 VI_UNLOCK(vp); 5719 } 5720 vput(vp); 5721 vn_finished_write(mp); 5722 ACQUIRE_LOCK(&lk); 5723 } 5724 FREE_LOCK(&lk); 5725} 5726 5727/* 5728 * Function to determine if the buffer has outstanding dependencies 5729 * that will cause a roll-back if the buffer is written. If wantcount 5730 * is set, return number of dependencies, otherwise just yes or no. 5731 */ 5732static int 5733softdep_count_dependencies(bp, wantcount) 5734 struct buf *bp; 5735 int wantcount; 5736{ 5737 struct worklist *wk; 5738 struct inodedep *inodedep; 5739 struct indirdep *indirdep; 5740 struct allocindir *aip; 5741 struct pagedep *pagedep; 5742 struct diradd *dap; 5743 int i, retval; 5744 5745 retval = 0; 5746 ACQUIRE_LOCK(&lk); 5747 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5748 switch (wk->wk_type) { 5749 5750 case D_INODEDEP: 5751 inodedep = WK_INODEDEP(wk); 5752 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 5753 /* bitmap allocation dependency */ 5754 retval += 1; 5755 if (!wantcount) 5756 goto out; 5757 } 5758 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 5759 /* direct block pointer dependency */ 5760 retval += 1; 5761 if (!wantcount) 5762 goto out; 5763 } 5764 if (TAILQ_FIRST(&inodedep->id_extupdt)) { 5765 /* direct block pointer dependency */ 5766 retval += 1; 5767 if (!wantcount) 5768 goto out; 5769 } 5770 continue; 5771 5772 case D_INDIRDEP: 5773 indirdep = WK_INDIRDEP(wk); 5774 5775 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 5776 /* indirect block pointer dependency */ 5777 retval += 1; 5778 if (!wantcount) 5779 goto out; 5780 } 5781 continue; 5782 5783 case D_PAGEDEP: 5784 pagedep = WK_PAGEDEP(wk); 5785 for (i = 0; i < DAHASHSZ; i++) { 5786 5787 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 5788 /* directory entry dependency */ 5789 retval += 1; 5790 if (!wantcount) 5791 goto out; 5792 } 5793 } 5794 continue; 5795 5796 case D_BMSAFEMAP: 5797 case D_ALLOCDIRECT: 5798 case D_ALLOCINDIR: 5799 case D_MKDIR: 5800 /* never a dependency on these blocks */ 5801 continue; 5802 5803 default: 5804 FREE_LOCK(&lk); 5805 panic("softdep_check_for_rollback: Unexpected type %s", 5806 TYPENAME(wk->wk_type)); 5807 /* NOTREACHED */ 5808 } 5809 } 5810out: 5811 FREE_LOCK(&lk); 5812 return retval; 5813} 5814 5815/* 5816 * Acquire exclusive access to a buffer. 5817 * Must be called with splbio blocked. 5818 * Return acquired buffer or NULL on failure. mtx, if provided, will be 5819 * released on success but held on failure. 5820 */ 5821static struct buf * 5822getdirtybuf(bpp, mtx, waitfor) 5823 struct buf **bpp; 5824 struct mtx *mtx; 5825 int waitfor; 5826{ 5827 struct buf *bp; 5828 int error; 5829 5830 /* 5831 * XXX This code and the code that calls it need to be reviewed to 5832 * verify its use of the vnode interlock. 5833 */ 5834 5835 for (;;) { 5836 if ((bp = *bpp) == NULL) 5837 return (0); 5838 if (bp->b_vp == NULL) 5839 backtrace(); 5840 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0) { 5841 if ((bp->b_vflags & BV_BKGRDINPROG) == 0) 5842 break; 5843 BUF_UNLOCK(bp); 5844 if (waitfor != MNT_WAIT) 5845 return (NULL); 5846 /* 5847 * The mtx argument must be bp->b_vp's mutex in 5848 * this case. 5849 */ 5850#ifdef DEBUG_VFS_LOCKS 5851 if (bp->b_vp->v_type != VCHR) 5852 ASSERT_VI_LOCKED(bp->b_vp, "getdirtybuf"); 5853#endif 5854 bp->b_vflags |= BV_BKGRDWAIT; 5855 interlocked_sleep(&lk, SLEEP, &bp->b_xflags, mtx, 5856 PRIBIO, "getbuf", 0); 5857 continue; 5858 } 5859 if (waitfor != MNT_WAIT) 5860 return (NULL); 5861 if (mtx) { 5862 error = interlocked_sleep(&lk, LOCKBUF, bp, mtx, 5863 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 0, 0); 5864 mtx_lock(mtx); 5865 } else 5866 error = interlocked_sleep(&lk, LOCKBUF, bp, NULL, 5867 LK_EXCLUSIVE | LK_SLEEPFAIL, 0, 0); 5868 if (error != ENOLCK) { 5869 FREE_LOCK(&lk); 5870 panic("getdirtybuf: inconsistent lock"); 5871 } 5872 } 5873 if ((bp->b_flags & B_DELWRI) == 0) { 5874 BUF_UNLOCK(bp); 5875 return (NULL); 5876 } 5877 if (mtx) 5878 mtx_unlock(mtx); 5879 bremfree(bp); 5880 return (bp); 5881} 5882 5883/* 5884 * Wait for pending output on a vnode to complete. 5885 * Must be called with vnode lock and interlock locked. 5886 */ 5887static void 5888drain_output(vp, islocked) 5889 struct vnode *vp; 5890 int islocked; 5891{ 5892 ASSERT_VOP_LOCKED(vp, "drain_output"); 5893 ASSERT_VI_LOCKED(vp, "drain_output"); 5894 5895 if (!islocked) 5896 ACQUIRE_LOCK(&lk); 5897 while (vp->v_numoutput) { 5898 vp->v_iflag |= VI_BWAIT; 5899 interlocked_sleep(&lk, SLEEP, (caddr_t)&vp->v_numoutput, 5900 VI_MTX(vp), PRIBIO + 1, "drainvp", 0); 5901 } 5902 if (!islocked) 5903 FREE_LOCK(&lk); 5904} 5905 5906/* 5907 * Called whenever a buffer that is being invalidated or reallocated 5908 * contains dependencies. This should only happen if an I/O error has 5909 * occurred. The routine is called with the buffer locked. 5910 */ 5911static void 5912softdep_deallocate_dependencies(bp) 5913 struct buf *bp; 5914{ 5915 5916 if ((bp->b_ioflags & BIO_ERROR) == 0) 5917 panic("softdep_deallocate_dependencies: dangling deps"); 5918 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); 5919 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 5920} 5921 5922/* 5923 * Function to handle asynchronous write errors in the filesystem. 5924 */ 5925static void 5926softdep_error(func, error) 5927 char *func; 5928 int error; 5929{ 5930 5931 /* XXX should do something better! */ 5932 printf("%s: got error %d while accessing filesystem\n", func, error); 5933} 5934