ffs_softdep.c revision 36900
1 2/* 3 * Copyright 1998 Marshall Kirk McKusick. All Rights Reserved. 4 * 5 * The soft updates code is derived from the appendix of a University 6 * of Michigan technical report (Gregory R. Ganger and Yale N. Patt, 7 * "Soft Updates: A Solution to the Metadata Update Problem in File 8 * Systems", CSE-TR-254-95, August 1995). 9 * 10 * The following are the copyrights and redistribution conditions that 11 * apply to this copy of the soft update software. For a license 12 * to use, redistribute or sell the soft update software under 13 * conditions other than those described here, please contact the 14 * author at one of the following addresses: 15 * 16 * Marshall Kirk McKusick mckusick@mckusick.com 17 * 1614 Oxford Street +1-510-843-9542 18 * Berkeley, CA 94709-1608 19 * USA 20 * 21 * Redistribution and use in source and binary forms, with or without 22 * modification, are permitted provided that the following conditions 23 * are met: 24 * 25 * 1. Redistributions of source code must retain the above copyright 26 * notice, this list of conditions and the following disclaimer. 27 * 2. Redistributions in binary form must reproduce the above copyright 28 * notice, this list of conditions and the following disclaimer in the 29 * documentation and/or other materials provided with the distribution. 30 * 3. None of the names of McKusick, Ganger, Patt, or the University of 31 * Michigan may be used to endorse or promote products derived from 32 * this software without specific prior written permission. 33 * 4. Redistributions in any form must be accompanied by information on 34 * how to obtain complete source code for any accompanying software 35 * that uses this software. This source code must either be included 36 * in the distribution or be available for no more than the cost of 37 * distribution plus a nominal fee, and must be freely redistributable 38 * under reasonable conditions. For an executable file, complete 39 * source code means the source code for all modules it contains. 40 * It does not mean source code for modules or files that typically 41 * accompany the operating system on which the executable file runs, 42 * e.g., standard library modules or system header files. 43 * 44 * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY 45 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 46 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 47 * DISCLAIMED. IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR 48 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 49 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 50 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 51 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 52 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 53 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 54 * SUCH DAMAGE. 55 * 56 * @(#)ffs_softdep.c 9.23 (McKusick) 2/20/98 57 * $Id: ffs_softdep.c,v 1.9 1998/06/10 20:45:46 julian Exp $ 58 */ 59 60/* 61 * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide. 62 */ 63#ifndef DIAGNOSTIC 64#define DIAGNOSTIC 65#endif 66#ifndef DEBUG 67#define DEBUG 68#endif 69 70#include <sys/param.h> 71#include <sys/buf.h> 72#include <sys/kernel.h> 73#include <sys/malloc.h> 74#include <sys/mount.h> 75#include <sys/proc.h> 76#include <sys/syslog.h> 77#include <sys/systm.h> 78#include <sys/vnode.h> 79#include <miscfs/specfs/specdev.h> 80#include <ufs/ufs/dir.h> 81#include <ufs/ufs/quota.h> 82#include <ufs/ufs/inode.h> 83#include <ufs/ufs/ufsmount.h> 84#include <ufs/ffs/fs.h> 85#include <ufs/ffs/softdep.h> 86#include <ufs/ffs/ffs_extern.h> 87#include <ufs/ufs/ufs_extern.h> 88 89/* 90 * These definitions need to be adapted to the system to which 91 * this file is being ported. 92 */ 93/* 94 * malloc types defined for the softdep system. 95 */ 96MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies"); 97MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies"); 98MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation"); 99MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map"); 100MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode"); 101MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies"); 102MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block"); 103MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode"); 104MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode"); 105MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated"); 106MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry"); 107MALLOC_DEFINE(M_MKDIR, "mkdir","New directory"); 108MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted"); 109 110#define D_PAGEDEP 0 111#define D_INODEDEP 1 112#define D_NEWBLK 2 113#define D_BMSAFEMAP 3 114#define D_ALLOCDIRECT 4 115#define D_INDIRDEP 5 116#define D_ALLOCINDIR 6 117#define D_FREEFRAG 7 118#define D_FREEBLKS 8 119#define D_FREEFILE 9 120#define D_DIRADD 10 121#define D_MKDIR 11 122#define D_DIRREM 12 123#define D_LAST D_DIRREM 124 125/* 126 * translate from workitem type to memory type 127 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 128 */ 129static struct malloc_type *memtype[] = { 130 M_PAGEDEP, 131 M_INODEDEP, 132 M_NEWBLK, 133 M_BMSAFEMAP, 134 M_ALLOCDIRECT, 135 M_INDIRDEP, 136 M_ALLOCINDIR, 137 M_FREEFRAG, 138 M_FREEBLKS, 139 M_FREEFILE, 140 M_DIRADD, 141 M_MKDIR, 142 M_DIRREM 143}; 144 145#define DtoM(type) (memtype[type]) 146 147/* 148 * Names of malloc types. 149 */ 150#define TYPENAME(type) \ 151 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???") 152#define CURPROC curproc 153/* 154 * End system adaptaion definitions. 155 */ 156 157/* 158 * Internal function prototypes. 159 */ 160static void softdep_error __P((char *, int)); 161static int getdirtybuf __P((struct buf **, int)); 162static int flush_pagedep_deps __P((struct vnode *, struct mount *, 163 struct diraddhd *)); 164static int flush_inodedep_deps __P((struct fs *, ino_t)); 165static int handle_written_filepage __P((struct pagedep *, struct buf *)); 166static void diradd_inode_written __P((struct diradd *, struct inodedep *)); 167static int handle_written_inodeblock __P((struct inodedep *, struct buf *)); 168static void handle_allocdirect_partdone __P((struct allocdirect *)); 169static void handle_allocindir_partdone __P((struct allocindir *)); 170static void initiate_write_filepage __P((struct pagedep *, struct buf *)); 171static void handle_written_mkdir __P((struct mkdir *, int)); 172static void initiate_write_inodeblock __P((struct inodedep *, struct buf *)); 173static void handle_workitem_freefile __P((struct freefile *)); 174static void handle_workitem_remove __P((struct dirrem *)); 175static struct dirrem *newdirrem __P((struct buf *, struct inode *, 176 struct inode *, int)); 177static void free_diradd __P((struct diradd *)); 178static void free_allocindir __P((struct allocindir *, struct inodedep *)); 179static int indir_trunc __P((struct inode *, ufs_daddr_t, int, ufs_lbn_t, 180 long *)); 181static void deallocate_dependencies __P((struct buf *, struct inodedep *)); 182static void free_allocdirect __P((struct allocdirectlst *, 183 struct allocdirect *, int)); 184static int free_inodedep __P((struct inodedep *)); 185static void handle_workitem_freeblocks __P((struct freeblks *)); 186static void merge_inode_lists __P((struct inodedep *)); 187static void setup_allocindir_phase2 __P((struct buf *, struct inode *, 188 struct allocindir *)); 189static struct allocindir *newallocindir __P((struct inode *, int, ufs_daddr_t, 190 ufs_daddr_t)); 191static void handle_workitem_freefrag __P((struct freefrag *)); 192static struct freefrag *newfreefrag __P((struct inode *, ufs_daddr_t, long)); 193static void allocdirect_merge __P((struct allocdirectlst *, 194 struct allocdirect *, struct allocdirect *)); 195static struct bmsafemap *bmsafemap_lookup __P((struct buf *)); 196static int newblk_lookup __P((struct fs *, ufs_daddr_t, int, 197 struct newblk **)); 198static int inodedep_lookup __P((struct fs *, ino_t, int, struct inodedep **)); 199static int pagedep_lookup __P((struct inode *, ufs_lbn_t, int, 200 struct pagedep **)); 201static void pause_timer __P((void *)); 202static int checklimit __P((long *, int)); 203static void add_to_worklist __P((struct worklist *)); 204 205/* 206 * Exported softdep operations. 207 */ 208struct bio_ops bioops = { 209 softdep_disk_io_initiation, /* io_start */ 210 softdep_disk_write_complete, /* io_complete */ 211 softdep_deallocate_dependencies, /* io_deallocate */ 212 softdep_process_worklist, /* io_sync */ 213}; 214 215/* 216 * Locking primitives. 217 * 218 * For a uniprocessor, all we need to do is protect against disk 219 * interrupts. For a multiprocessor, this lock would have to be 220 * a mutex. A single mutex is used throughout this file, though 221 * finer grain locking could be used if contention warranted it. 222 * 223 * For a multiprocessor, the sleep call would accept a lock and 224 * release it after the sleep processing was complete. In a uniprocessor 225 * implementation there is no such interlock, so we simple mark 226 * the places where it needs to be done with the `interlocked' form 227 * of the lock calls. Since the uniprocessor sleep already interlocks 228 * the spl, there is nothing that really needs to be done. 229 */ 230#ifndef /* NOT */ DEBUG 231static struct lockit { 232 int lkt_spl; 233} lk = { 0 }; 234#define ACQUIRE_LOCK(lk) (lk)->lkt_spl = splbio() 235#define FREE_LOCK(lk) splx((lk)->lkt_spl) 236#define ACQUIRE_LOCK_INTERLOCKED(lk) 237#define FREE_LOCK_INTERLOCKED(lk) 238 239#else /* DEBUG */ 240static struct lockit { 241 int lkt_spl; 242 pid_t lkt_held; 243} lk = { 0, -1 }; 244static int lockcnt; 245 246static void acquire_lock __P((struct lockit *)); 247static void free_lock __P((struct lockit *)); 248static void acquire_lock_interlocked __P((struct lockit *)); 249static void free_lock_interlocked __P((struct lockit *)); 250 251#define ACQUIRE_LOCK(lk) acquire_lock(lk) 252#define FREE_LOCK(lk) free_lock(lk) 253#define ACQUIRE_LOCK_INTERLOCKED(lk) acquire_lock_interlocked(lk) 254#define FREE_LOCK_INTERLOCKED(lk) free_lock_interlocked(lk) 255 256static void 257acquire_lock(lk) 258 struct lockit *lk; 259{ 260 261 if (lk->lkt_held != -1) 262 if (lk->lkt_held == CURPROC->p_pid) 263 panic("softdep_lock: locking against myself"); 264 else 265 panic("softdep_lock: lock held by %d", lk->lkt_held); 266 lk->lkt_spl = splbio(); 267 lk->lkt_held = CURPROC->p_pid; 268 lockcnt++; 269} 270 271static void 272free_lock(lk) 273 struct lockit *lk; 274{ 275 276 if (lk->lkt_held == -1) 277 panic("softdep_unlock: lock not held"); 278 lk->lkt_held = -1; 279 splx(lk->lkt_spl); 280} 281 282static void 283acquire_lock_interlocked(lk) 284 struct lockit *lk; 285{ 286 287 if (lk->lkt_held != -1) 288 if (lk->lkt_held == CURPROC->p_pid) 289 panic("softdep_lock_interlocked: locking against self"); 290 else 291 panic("softdep_lock_interlocked: lock held by %d", 292 lk->lkt_held); 293 lk->lkt_held = CURPROC->p_pid; 294 lockcnt++; 295} 296 297static void 298free_lock_interlocked(lk) 299 struct lockit *lk; 300{ 301 302 if (lk->lkt_held == -1) 303 panic("softdep_unlock_interlocked: lock not held"); 304 lk->lkt_held = -1; 305} 306#endif /* DEBUG */ 307 308/* 309 * Place holder for real semaphores. 310 */ 311struct sema { 312 int value; 313 pid_t holder; 314 char *name; 315 int prio; 316 int timo; 317}; 318static void sema_init __P((struct sema *, char *, int, int)); 319static int sema_get __P((struct sema *, struct lockit *)); 320static void sema_release __P((struct sema *)); 321 322static void 323sema_init(semap, name, prio, timo) 324 struct sema *semap; 325 char *name; 326 int prio, timo; 327{ 328 329 semap->holder = -1; 330 semap->value = 0; 331 semap->name = name; 332 semap->prio = prio; 333 semap->timo = timo; 334} 335 336static int 337sema_get(semap, interlock) 338 struct sema *semap; 339 struct lockit *interlock; 340{ 341 342 if (semap->value++ > 0) { 343 if (interlock != NULL) 344 FREE_LOCK_INTERLOCKED(interlock); 345 tsleep((caddr_t)semap, semap->prio, semap->name, semap->timo); 346 if (interlock != NULL) { 347 ACQUIRE_LOCK_INTERLOCKED(interlock); 348 FREE_LOCK(interlock); 349 } 350 return (0); 351 } 352 semap->holder = CURPROC->p_pid; 353 if (interlock != NULL) 354 FREE_LOCK(interlock); 355 return (1); 356} 357 358static void 359sema_release(semap) 360 struct sema *semap; 361{ 362 363 if (semap->value <= 0 || semap->holder != CURPROC->p_pid) 364 panic("sema_release: not held"); 365 if (--semap->value > 0) { 366 semap->value = 0; 367 wakeup(semap); 368 } 369 semap->holder = -1; 370} 371 372/* 373 * Worklist queue management. 374 * These routines require that the lock be held. 375 */ 376#ifndef /* NOT */ DEBUG 377#define WORKLIST_INSERT(head, item) do { \ 378 (item)->wk_state |= ONWORKLIST; \ 379 LIST_INSERT_HEAD(head, item, wk_list); \ 380} while (0) 381#define WORKLIST_REMOVE(item) do { \ 382 (item)->wk_state &= ~ONWORKLIST; \ 383 LIST_REMOVE(item, wk_list); \ 384} while (0) 385#define WORKITEM_FREE(item, type) FREE(item, DtoM(type)) 386 387#else /* DEBUG */ 388static void worklist_insert __P((struct workhead *, struct worklist *)); 389static void worklist_remove __P((struct worklist *)); 390static void workitem_free __P((struct worklist *, int)); 391 392#define WORKLIST_INSERT(head, item) worklist_insert(head, item) 393#define WORKLIST_REMOVE(item) worklist_remove(item) 394#define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type) 395 396static void 397worklist_insert(head, item) 398 struct workhead *head; 399 struct worklist *item; 400{ 401 402 if (lk.lkt_held == -1) 403 panic("worklist_insert: lock not held"); 404 if (item->wk_state & ONWORKLIST) 405 panic("worklist_insert: already on list"); 406 item->wk_state |= ONWORKLIST; 407 LIST_INSERT_HEAD(head, item, wk_list); 408} 409 410static void 411worklist_remove(item) 412 struct worklist *item; 413{ 414 415 if (lk.lkt_held == -1) 416 panic("worklist_remove: lock not held"); 417 if ((item->wk_state & ONWORKLIST) == 0) 418 panic("worklist_remove: not on list"); 419 item->wk_state &= ~ONWORKLIST; 420 LIST_REMOVE(item, wk_list); 421} 422 423static void 424workitem_free(item, type) 425 struct worklist *item; 426 int type; 427{ 428 429 if (item->wk_state & ONWORKLIST) 430 panic("workitem_free: still on list"); 431 if (item->wk_type != type) 432 panic("workitem_free: type mismatch"); 433 FREE(item, DtoM(type)); 434} 435#endif /* DEBUG */ 436 437/* 438 * Workitem queue management 439 */ 440static struct workhead softdep_workitem_pending; 441static int softdep_worklist_busy; 442static int max_softdeps; /* maximum number of structs before slowdown */ 443static int tickdelay = 2; /* number of ticks to pause during slowdown */ 444static int max_limit_hit; /* number of times slowdown imposed */ 445static int rush_requests; /* number of times I/O speeded up */ 446static int proc_waiting; /* tracks whether we have a timeout posted */ 447static pid_t filesys_syncer_pid;/* records pid of filesystem syncer process */ 448#ifdef DEBUG 449#include <vm/vm.h> 450#include <sys/sysctl.h> 451#if defined(__FreeBSD__) 452SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, ""); 453SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, ""); 454SYSCTL_INT(_debug, OID_AUTO, max_limit_hit, CTLFLAG_RW, &max_limit_hit, 0, ""); 455SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &rush_requests, 0, ""); 456#else /* !__FreeBSD__ */ 457struct ctldebug debug8 = { "max_softdeps", &max_softdeps }; 458struct ctldebug debug9 = { "tickdelay", &tickdelay }; 459struct ctldebug debug10 = { "max_limit_hit", &max_limit_hit }; 460struct ctldebug debug11 = { "rush_requests", &rush_requests }; 461#endif /* !__FreeBSD__ */ 462 463#endif /* DEBUG */ 464 465/* 466 * Add an item to the end of the work queue. 467 * This routine requires that the lock be held. 468 * This is the only routine that adds items to the list. 469 * The following routine is the only one that removes items 470 * and does so in order from first to last. 471 */ 472static void 473add_to_worklist(wk) 474 struct worklist *wk; 475{ 476 static struct worklist *worklist_tail; 477 478 if (wk->wk_state & ONWORKLIST) 479 panic("add_to_worklist: already on list"); 480 wk->wk_state |= ONWORKLIST; 481 if (LIST_FIRST(&softdep_workitem_pending) == NULL) { 482 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list); 483 } else { 484 LIST_INSERT_AFTER(worklist_tail, wk, wk_list); 485 } 486 worklist_tail = wk; 487} 488 489/* 490 * Process that runs once per second to handle items in the background queue. 491 * 492 * Note that we ensure that everything is done in the order in which they 493 * appear in the queue. The code below depends on this property to ensure 494 * that blocks of a file are freed before the inode itself is freed. This 495 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 496 * until all the old ones have been purged from the dependency lists. 497 */ 498int 499softdep_process_worklist(matchmnt) 500 struct mount *matchmnt; 501{ 502 struct proc *p = CURPROC; 503 struct worklist *wk; 504 struct fs *matchfs; 505 int matchcnt; 506 507 /* 508 * Record the process identifier of our caller so that we can 509 * give this process preferential treatment in checklimit below. 510 */ 511 filesys_syncer_pid = p->p_pid; 512 matchcnt = 0; 513 matchfs = NULL; 514 if (matchmnt != NULL) 515 matchfs = VFSTOUFS(matchmnt)->um_fs; 516 /* 517 * There is no danger of having multiple processes run this 518 * code. It is single threaded solely so that softdep_flushfiles 519 * (below) can get an accurate count of the number of items 520 * related to its mount point that are in the list. 521 */ 522 if (softdep_worklist_busy && matchmnt == NULL) 523 return (-1); 524 ACQUIRE_LOCK(&lk); 525 while ((wk = LIST_FIRST(&softdep_workitem_pending)) != 0) { 526 WORKLIST_REMOVE(wk); 527 FREE_LOCK(&lk); 528 switch (wk->wk_type) { 529 530 case D_DIRREM: 531 /* removal of a directory entry */ 532 if (WK_DIRREM(wk)->dm_mnt == matchmnt) 533 matchcnt += 1; 534 handle_workitem_remove(WK_DIRREM(wk)); 535 break; 536 537 case D_FREEBLKS: 538 /* releasing blocks and/or fragments from a file */ 539 if (WK_FREEBLKS(wk)->fb_fs == matchfs) 540 matchcnt += 1; 541 handle_workitem_freeblocks(WK_FREEBLKS(wk)); 542 break; 543 544 case D_FREEFRAG: 545 /* releasing a fragment when replaced as a file grows */ 546 if (WK_FREEFRAG(wk)->ff_fs == matchfs) 547 matchcnt += 1; 548 handle_workitem_freefrag(WK_FREEFRAG(wk)); 549 break; 550 551 case D_FREEFILE: 552 /* releasing an inode when its link count drops to 0 */ 553 if (WK_FREEFILE(wk)->fx_fs == matchfs) 554 matchcnt += 1; 555 handle_workitem_freefile(WK_FREEFILE(wk)); 556 break; 557 558 default: 559 panic("%s_process_worklist: Unknown type %s", 560 "softdep", TYPENAME(wk->wk_type)); 561 /* NOTREACHED */ 562 } 563 if (softdep_worklist_busy && matchmnt == NULL) 564 return (-1); 565 ACQUIRE_LOCK(&lk); 566 } 567 FREE_LOCK(&lk); 568 return (matchcnt); 569} 570 571/* 572 * Purge the work list of all items associated with a particular mount point. 573 */ 574int 575softdep_flushfiles(oldmnt, flags, p) 576 struct mount *oldmnt; 577 int flags; 578 struct proc *p; 579{ 580 struct vnode *devvp; 581 int error, loopcnt; 582 583 /* 584 * Await our turn to clear out the queue. 585 */ 586 while (softdep_worklist_busy) 587 tsleep(&lbolt, PRIBIO, "softflush", 0); 588 softdep_worklist_busy = 1; 589 if ((error = ffs_flushfiles(oldmnt, flags, p)) != 0) { 590 softdep_worklist_busy = 0; 591 return (error); 592 } 593 /* 594 * Alternately flush the block device associated with the mount 595 * point and process any dependencies that the flushing 596 * creates. In theory, this loop can happen at most twice, 597 * but we give it a few extra just to be sure. 598 */ 599 devvp = VFSTOUFS(oldmnt)->um_devvp; 600 for (loopcnt = 10; loopcnt > 0; loopcnt--) { 601 if (softdep_process_worklist(oldmnt) == 0) { 602 /* 603 * Do another flush in case any vnodes were brought in 604 * as part of the cleanup operations. 605 */ 606 if ((error = ffs_flushfiles(oldmnt, flags, p)) != 0) 607 break; 608 /* 609 * If we still found nothing to do, we are really done. 610 */ 611 if (softdep_process_worklist(oldmnt) == 0) 612 break; 613 } 614 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, p); 615 error = VOP_FSYNC(devvp, p->p_ucred, MNT_WAIT, p); 616 VOP_UNLOCK(devvp, 0, p); 617 if (error) 618 break; 619 } 620 softdep_worklist_busy = 0; 621 /* 622 * If we are unmounting then it is an error to fail. If we 623 * are simply trying to downgrade to read-only, then filesystem 624 * activity can keep us busy forever, so we just fail with EBUSY. 625 */ 626 if (loopcnt == 0) { 627 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 628 panic("softdep_flushfiles: looping"); 629 error = EBUSY; 630 } 631 return (error); 632} 633 634/* 635 * A large burst of file addition or deletion activity can drive the 636 * memory load excessively high. Therefore we deliberately slow things 637 * down and speed up the I/O processing if we find ourselves with too 638 * many dependencies in progress. 639 */ 640static int 641checklimit(resource, islocked) 642 long *resource; 643 int islocked; 644{ 645 struct proc *p = CURPROC; 646 647 /* 648 * If we are under our limit, just proceed. 649 */ 650 if (*resource < max_softdeps) 651 return (0); 652 /* 653 * We never hold up the filesystem syncer process. 654 */ 655 if (p->p_pid == filesys_syncer_pid) 656 return (0); 657 /* 658 * Our first approach is to speed up the syncer process. 659 * We never push it to speed up more than half of its 660 * normal turn time, otherwise it could take over the cpu. 661 */ 662 if (rushjob < syncdelay / 2) { 663 rushjob += 1; 664 rush_requests += 1; 665 return (0); 666 } 667 /* 668 * Every trick has failed, so we pause momentarily to let 669 * the filesystem syncer process catch up. 670 */ 671 if (islocked == 0) 672 ACQUIRE_LOCK(&lk); 673 if (proc_waiting == 0) { 674 proc_waiting = 1; 675 timeout(pause_timer, NULL, tickdelay > 2 ? tickdelay : 2); 676 } 677 FREE_LOCK_INTERLOCKED(&lk); 678 (void) tsleep((caddr_t)&proc_waiting, PPAUSE | PCATCH, "softupdate", 0); 679 ACQUIRE_LOCK_INTERLOCKED(&lk); 680 if (islocked == 0) 681 FREE_LOCK(&lk); 682 max_limit_hit += 1; 683 return (1); 684} 685 686/* 687 * Awaken processes pausing in checklimit and clear proc_waiting 688 * to indicate that there is no longer a timer running. 689 */ 690void 691pause_timer(arg) 692 void *arg; 693{ 694 695 proc_waiting = 0; 696 wakeup(&proc_waiting); 697} 698 699/* 700 * Structure hashing. 701 * 702 * There are three types of structures that can be looked up: 703 * 1) pagedep structures identified by mount point, inode number, 704 * and logical block. 705 * 2) inodedep structures identified by mount point and inode number. 706 * 3) newblk structures identified by mount point and 707 * physical block number. 708 * 709 * The "pagedep" and "inodedep" dependency structures are hashed 710 * separately from the file blocks and inodes to which they correspond. 711 * This separation helps when the in-memory copy of an inode or 712 * file block must be replaced. It also obviates the need to access 713 * an inode or file page when simply updating (or de-allocating) 714 * dependency structures. Lookup of newblk structures is needed to 715 * find newly allocated blocks when trying to associate them with 716 * their allocdirect or allocindir structure. 717 * 718 * The lookup routines optionally create and hash a new instance when 719 * an existing entry is not found. 720 */ 721#define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 722 723/* 724 * Structures and routines associated with pagedep caching. 725 */ 726LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl; 727u_long pagedep_hash; /* size of hash table - 1 */ 728#define PAGEDEP_HASH(mp, inum, lbn) \ 729 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \ 730 pagedep_hash]) 731static struct sema pagedep_in_progress; 732 733/* 734 * Look up a pagedep. Return 1 if found, 0 if not found. 735 * If not found, allocate if DEPALLOC flag is passed. 736 * Found or allocated entry is returned in pagedeppp. 737 * This routine must be called with splbio interrupts blocked. 738 */ 739static int 740pagedep_lookup(ip, lbn, flags, pagedeppp) 741 struct inode *ip; 742 ufs_lbn_t lbn; 743 int flags; 744 struct pagedep **pagedeppp; 745{ 746 struct pagedep *pagedep; 747 struct pagedep_hashhead *pagedephd; 748 struct mount *mp; 749 int i; 750 751#ifdef DEBUG 752 if (lk.lkt_held == -1) 753 panic("pagedep_lookup: lock not held"); 754#endif 755 mp = ITOV(ip)->v_mount; 756 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn); 757top: 758 for (pagedep = LIST_FIRST(pagedephd); pagedep; 759 pagedep = LIST_NEXT(pagedep, pd_hash)) 760 if (ip->i_number == pagedep->pd_ino && 761 lbn == pagedep->pd_lbn && 762 mp == pagedep->pd_mnt) 763 break; 764 if (pagedep) { 765 *pagedeppp = pagedep; 766 return (1); 767 } 768 if ((flags & DEPALLOC) == 0) { 769 *pagedeppp = NULL; 770 return (0); 771 } 772 if (sema_get(&pagedep_in_progress, &lk) == 0) { 773 ACQUIRE_LOCK(&lk); 774 goto top; 775 } 776 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP, 777 M_WAITOK); 778 bzero(pagedep, sizeof(struct pagedep)); 779 pagedep->pd_list.wk_type = D_PAGEDEP; 780 pagedep->pd_mnt = mp; 781 pagedep->pd_ino = ip->i_number; 782 pagedep->pd_lbn = lbn; 783 LIST_INIT(&pagedep->pd_dirremhd); 784 LIST_INIT(&pagedep->pd_pendinghd); 785 for (i = 0; i < DAHASHSZ; i++) 786 LIST_INIT(&pagedep->pd_diraddhd[i]); 787 ACQUIRE_LOCK(&lk); 788 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 789 sema_release(&pagedep_in_progress); 790 *pagedeppp = pagedep; 791 return (0); 792} 793 794/* 795 * Structures and routines associated with inodedep caching. 796 */ 797LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl; 798static u_long inodedep_hash; /* size of hash table - 1 */ 799static long num_inodedep; /* number of inodedep allocated */ 800#define INODEDEP_HASH(fs, inum) \ 801 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash]) 802static struct sema inodedep_in_progress; 803 804/* 805 * Look up a inodedep. Return 1 if found, 0 if not found. 806 * If not found, allocate if DEPALLOC flag is passed. 807 * Found or allocated entry is returned in inodedeppp. 808 * This routine must be called with splbio interrupts blocked. 809 */ 810static int 811inodedep_lookup(fs, inum, flags, inodedeppp) 812 struct fs *fs; 813 ino_t inum; 814 int flags; 815 struct inodedep **inodedeppp; 816{ 817 struct inodedep *inodedep; 818 struct inodedep_hashhead *inodedephd; 819 int firsttry; 820 821#ifdef DEBUG 822 if (lk.lkt_held == -1) 823 panic("inodedep_lookup: lock not held"); 824#endif 825 firsttry = 1; 826 inodedephd = INODEDEP_HASH(fs, inum); 827top: 828 for (inodedep = LIST_FIRST(inodedephd); inodedep; 829 inodedep = LIST_NEXT(inodedep, id_hash)) 830 if (inum == inodedep->id_ino && fs == inodedep->id_fs) 831 break; 832 if (inodedep) { 833 *inodedeppp = inodedep; 834 return (1); 835 } 836 if ((flags & DEPALLOC) == 0) { 837 *inodedeppp = NULL; 838 return (0); 839 } 840 if (firsttry && checklimit(&num_inodedep, 1) == 1) { 841 firsttry = 0; 842 goto top; 843 } 844 if (sema_get(&inodedep_in_progress, &lk) == 0) { 845 ACQUIRE_LOCK(&lk); 846 goto top; 847 } 848 num_inodedep += 1; 849 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep), 850 M_INODEDEP, M_WAITOK); 851 inodedep->id_list.wk_type = D_INODEDEP; 852 inodedep->id_fs = fs; 853 inodedep->id_ino = inum; 854 inodedep->id_state = ALLCOMPLETE; 855 inodedep->id_nlinkdelta = 0; 856 inodedep->id_savedino = NULL; 857 inodedep->id_savedsize = -1; 858 inodedep->id_buf = NULL; 859 LIST_INIT(&inodedep->id_pendinghd); 860 LIST_INIT(&inodedep->id_inowait); 861 LIST_INIT(&inodedep->id_bufwait); 862 TAILQ_INIT(&inodedep->id_inoupdt); 863 TAILQ_INIT(&inodedep->id_newinoupdt); 864 ACQUIRE_LOCK(&lk); 865 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 866 sema_release(&inodedep_in_progress); 867 *inodedeppp = inodedep; 868 return (0); 869} 870 871/* 872 * Structures and routines associated with newblk caching. 873 */ 874LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl; 875u_long newblk_hash; /* size of hash table - 1 */ 876#define NEWBLK_HASH(fs, inum) \ 877 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash]) 878static struct sema newblk_in_progress; 879 880/* 881 * Look up a newblk. Return 1 if found, 0 if not found. 882 * If not found, allocate if DEPALLOC flag is passed. 883 * Found or allocated entry is returned in newblkpp. 884 */ 885static int 886newblk_lookup(fs, newblkno, flags, newblkpp) 887 struct fs *fs; 888 ufs_daddr_t newblkno; 889 int flags; 890 struct newblk **newblkpp; 891{ 892 struct newblk *newblk; 893 struct newblk_hashhead *newblkhd; 894 895 newblkhd = NEWBLK_HASH(fs, newblkno); 896top: 897 for (newblk = LIST_FIRST(newblkhd); newblk; 898 newblk = LIST_NEXT(newblk, nb_hash)) 899 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs) 900 break; 901 if (newblk) { 902 *newblkpp = newblk; 903 return (1); 904 } 905 if ((flags & DEPALLOC) == 0) { 906 *newblkpp = NULL; 907 return (0); 908 } 909 if (sema_get(&newblk_in_progress, 0) == 0) 910 goto top; 911 MALLOC(newblk, struct newblk *, sizeof(struct newblk), 912 M_NEWBLK, M_WAITOK); 913 newblk->nb_state = 0; 914 newblk->nb_fs = fs; 915 newblk->nb_newblkno = newblkno; 916 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 917 sema_release(&newblk_in_progress); 918 *newblkpp = newblk; 919 return (0); 920} 921 922/* 923 * Executed during filesystem system initialization before 924 * mounting any file systems. 925 */ 926void 927softdep_initialize() 928{ 929 930 LIST_INIT(&mkdirlisthd); 931 LIST_INIT(&softdep_workitem_pending); 932 max_softdeps = desiredvnodes * 8; 933 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 934 &pagedep_hash); 935 sema_init(&pagedep_in_progress, "pagedep", PRIBIO, 0); 936 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash); 937 sema_init(&inodedep_in_progress, "inodedep", PRIBIO, 0); 938 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash); 939 sema_init(&newblk_in_progress, "newblk", PRIBIO, 0); 940} 941 942/* 943 * Called at mount time to notify the dependency code that a 944 * filesystem wishes to use it. 945 */ 946int 947softdep_mount(devvp, mp, fs, cred) 948 struct vnode *devvp; 949 struct mount *mp; 950 struct fs *fs; 951 struct ucred *cred; 952{ 953 struct csum cstotal; 954 struct cg *cgp; 955 struct buf *bp; 956 int error, cyl; 957 958 mp->mnt_flag &= ~MNT_ASYNC; 959 mp->mnt_flag |= MNT_SOFTDEP; 960 /* 961 * When doing soft updates, the counters in the 962 * superblock may have gotten out of sync, so we have 963 * to scan the cylinder groups and recalculate them. 964 */ 965 if (fs->fs_clean != 0) 966 return (0); 967 bzero(&cstotal, sizeof cstotal); 968 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 969 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 970 fs->fs_cgsize, cred, &bp)) != 0) { 971 brelse(bp); 972 return (error); 973 } 974 cgp = (struct cg *)bp->b_data; 975 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 976 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 977 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 978 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 979 fs->fs_cs(fs, cyl) = cgp->cg_cs; 980 brelse(bp); 981 } 982#ifdef DEBUG 983 if (!bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 984 printf("ffs_mountfs: superblock updated\n"); 985#endif 986 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 987 return (0); 988} 989 990/* 991 * Protecting the freemaps (or bitmaps). 992 * 993 * To eliminate the need to execute fsck before mounting a file system 994 * after a power failure, one must (conservatively) guarantee that the 995 * on-disk copy of the bitmaps never indicate that a live inode or block is 996 * free. So, when a block or inode is allocated, the bitmap should be 997 * updated (on disk) before any new pointers. When a block or inode is 998 * freed, the bitmap should not be updated until all pointers have been 999 * reset. The latter dependency is handled by the delayed de-allocation 1000 * approach described below for block and inode de-allocation. The former 1001 * dependency is handled by calling the following procedure when a block or 1002 * inode is allocated. When an inode is allocated an "inodedep" is created 1003 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 1004 * Each "inodedep" is also inserted into the hash indexing structure so 1005 * that any additional link additions can be made dependent on the inode 1006 * allocation. 1007 * 1008 * The ufs file system maintains a number of free block counts (e.g., per 1009 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 1010 * in addition to the bitmaps. These counts are used to improve efficiency 1011 * during allocation and therefore must be consistent with the bitmaps. 1012 * There is no convenient way to guarantee post-crash consistency of these 1013 * counts with simple update ordering, for two main reasons: (1) The counts 1014 * and bitmaps for a single cylinder group block are not in the same disk 1015 * sector. If a disk write is interrupted (e.g., by power failure), one may 1016 * be written and the other not. (2) Some of the counts are located in the 1017 * superblock rather than the cylinder group block. So, we focus our soft 1018 * updates implementation on protecting the bitmaps. When mounting a 1019 * filesystem, we recompute the auxiliary counts from the bitmaps. 1020 */ 1021 1022/* 1023 * Called just after updating the cylinder group block to allocate an inode. 1024 */ 1025void 1026softdep_setup_inomapdep(bp, ip, newinum) 1027 struct buf *bp; /* buffer for cylgroup block with inode map */ 1028 struct inode *ip; /* inode related to allocation */ 1029 ino_t newinum; /* new inode number being allocated */ 1030{ 1031 struct inodedep *inodedep; 1032 struct bmsafemap *bmsafemap; 1033 1034 /* 1035 * Create a dependency for the newly allocated inode. 1036 * Panic if it already exists as something is seriously wrong. 1037 * Otherwise add it to the dependency list for the buffer holding 1038 * the cylinder group map from which it was allocated. 1039 */ 1040 ACQUIRE_LOCK(&lk); 1041 if (inodedep_lookup(ip->i_fs, newinum, DEPALLOC, &inodedep) != 0) 1042 panic("softdep_setup_inomapdep: found inode"); 1043 inodedep->id_buf = bp; 1044 inodedep->id_state &= ~DEPCOMPLETE; 1045 bmsafemap = bmsafemap_lookup(bp); 1046 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 1047 FREE_LOCK(&lk); 1048} 1049 1050/* 1051 * Called just after updating the cylinder group block to 1052 * allocate block or fragment. 1053 */ 1054void 1055softdep_setup_blkmapdep(bp, fs, newblkno) 1056 struct buf *bp; /* buffer for cylgroup block with block map */ 1057 struct fs *fs; /* filesystem doing allocation */ 1058 ufs_daddr_t newblkno; /* number of newly allocated block */ 1059{ 1060 struct newblk *newblk; 1061 struct bmsafemap *bmsafemap; 1062 1063 /* 1064 * Create a dependency for the newly allocated block. 1065 * Add it to the dependency list for the buffer holding 1066 * the cylinder group map from which it was allocated. 1067 */ 1068 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0) 1069 panic("softdep_setup_blkmapdep: found block"); 1070 ACQUIRE_LOCK(&lk); 1071 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp); 1072 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 1073 FREE_LOCK(&lk); 1074} 1075 1076/* 1077 * Find the bmsafemap associated with a cylinder group buffer. 1078 * If none exists, create one. The buffer must be locked when 1079 * this routine is called and this routine must be called with 1080 * splbio interrupts blocked. 1081 */ 1082static struct bmsafemap * 1083bmsafemap_lookup(bp) 1084 struct buf *bp; 1085{ 1086 struct bmsafemap *bmsafemap; 1087 struct worklist *wk; 1088 1089#ifdef DEBUG 1090 if (lk.lkt_held == -1) 1091 panic("bmsafemap_lookup: lock not held"); 1092#endif 1093 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = LIST_NEXT(wk, wk_list)) 1094 if (wk->wk_type == D_BMSAFEMAP) 1095 return (WK_BMSAFEMAP(wk)); 1096 FREE_LOCK(&lk); 1097 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap), 1098 M_BMSAFEMAP, M_WAITOK); 1099 bmsafemap->sm_list.wk_type = D_BMSAFEMAP; 1100 bmsafemap->sm_list.wk_state = 0; 1101 bmsafemap->sm_buf = bp; 1102 LIST_INIT(&bmsafemap->sm_allocdirecthd); 1103 LIST_INIT(&bmsafemap->sm_allocindirhd); 1104 LIST_INIT(&bmsafemap->sm_inodedephd); 1105 LIST_INIT(&bmsafemap->sm_newblkhd); 1106 ACQUIRE_LOCK(&lk); 1107 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 1108 return (bmsafemap); 1109} 1110 1111/* 1112 * Direct block allocation dependencies. 1113 * 1114 * When a new block is allocated, the corresponding disk locations must be 1115 * initialized (with zeros or new data) before the on-disk inode points to 1116 * them. Also, the freemap from which the block was allocated must be 1117 * updated (on disk) before the inode's pointer. These two dependencies are 1118 * independent of each other and are needed for all file blocks and indirect 1119 * blocks that are pointed to directly by the inode. Just before the 1120 * "in-core" version of the inode is updated with a newly allocated block 1121 * number, a procedure (below) is called to setup allocation dependency 1122 * structures. These structures are removed when the corresponding 1123 * dependencies are satisfied or when the block allocation becomes obsolete 1124 * (i.e., the file is deleted, the block is de-allocated, or the block is a 1125 * fragment that gets upgraded). All of these cases are handled in 1126 * procedures described later. 1127 * 1128 * When a file extension causes a fragment to be upgraded, either to a larger 1129 * fragment or to a full block, the on-disk location may change (if the 1130 * previous fragment could not simply be extended). In this case, the old 1131 * fragment must be de-allocated, but not until after the inode's pointer has 1132 * been updated. In most cases, this is handled by later procedures, which 1133 * will construct a "freefrag" structure to be added to the workitem queue 1134 * when the inode update is complete (or obsolete). The main exception to 1135 * this is when an allocation occurs while a pending allocation dependency 1136 * (for the same block pointer) remains. This case is handled in the main 1137 * allocation dependency setup procedure by immediately freeing the 1138 * unreferenced fragments. 1139 */ 1140void 1141softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 1142 struct inode *ip; /* inode to which block is being added */ 1143 ufs_lbn_t lbn; /* block pointer within inode */ 1144 ufs_daddr_t newblkno; /* disk block number being added */ 1145 ufs_daddr_t oldblkno; /* previous block number, 0 unless frag */ 1146 long newsize; /* size of new block */ 1147 long oldsize; /* size of new block */ 1148 struct buf *bp; /* bp for allocated block */ 1149{ 1150 struct allocdirect *adp, *oldadp; 1151 struct allocdirectlst *adphead; 1152 struct bmsafemap *bmsafemap; 1153 struct inodedep *inodedep; 1154 struct pagedep *pagedep; 1155 struct newblk *newblk; 1156 1157 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect), 1158 M_ALLOCDIRECT, M_WAITOK); 1159 bzero(adp, sizeof(struct allocdirect)); 1160 adp->ad_list.wk_type = D_ALLOCDIRECT; 1161 adp->ad_lbn = lbn; 1162 adp->ad_newblkno = newblkno; 1163 adp->ad_oldblkno = oldblkno; 1164 adp->ad_newsize = newsize; 1165 adp->ad_oldsize = oldsize; 1166 adp->ad_state = ATTACHED; 1167 if (newblkno == oldblkno) 1168 adp->ad_freefrag = NULL; 1169 else 1170 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1171 1172 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1173 panic("softdep_setup_allocdirect: lost block"); 1174 1175 ACQUIRE_LOCK(&lk); 1176 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 1177 adp->ad_inodedep = inodedep; 1178 1179 if (newblk->nb_state == DEPCOMPLETE) { 1180 adp->ad_state |= DEPCOMPLETE; 1181 adp->ad_buf = NULL; 1182 } else { 1183 bmsafemap = newblk->nb_bmsafemap; 1184 adp->ad_buf = bmsafemap->sm_buf; 1185 LIST_REMOVE(newblk, nb_deps); 1186 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1187 } 1188 LIST_REMOVE(newblk, nb_hash); 1189 FREE(newblk, M_NEWBLK); 1190 1191 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list); 1192 if (lbn >= NDADDR) { 1193 /* allocating an indirect block */ 1194 if (oldblkno != 0) 1195 panic("softdep_setup_allocdirect: non-zero indir"); 1196 } else { 1197 /* 1198 * Allocating a direct block. 1199 * 1200 * If we are allocating a directory block, then we must 1201 * allocate an associated pagedep to track additions and 1202 * deletions. 1203 */ 1204 if ((ip->i_mode & IFMT) == IFDIR && 1205 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1206 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 1207 } 1208 /* 1209 * The list of allocdirects must be kept in sorted and ascending 1210 * order so that the rollback routines can quickly determine the 1211 * first uncommitted block (the size of the file stored on disk 1212 * ends at the end of the lowest committed fragment, or if there 1213 * are no fragments, at the end of the highest committed block). 1214 * Since files generally grow, the typical case is that the new 1215 * block is to be added at the end of the list. We speed this 1216 * special case by checking against the last allocdirect in the 1217 * list before laboriously traversing the list looking for the 1218 * insertion point. 1219 */ 1220 adphead = &inodedep->id_newinoupdt; 1221 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1222 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1223 /* insert at end of list */ 1224 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1225 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1226 allocdirect_merge(adphead, adp, oldadp); 1227 FREE_LOCK(&lk); 1228 return; 1229 } 1230 for (oldadp = TAILQ_FIRST(adphead); oldadp; 1231 oldadp = TAILQ_NEXT(oldadp, ad_next)) { 1232 if (oldadp->ad_lbn >= lbn) 1233 break; 1234 } 1235 if (oldadp == NULL) 1236 panic("softdep_setup_allocdirect: lost entry"); 1237 /* insert in middle of list */ 1238 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1239 if (oldadp->ad_lbn == lbn) 1240 allocdirect_merge(adphead, adp, oldadp); 1241 FREE_LOCK(&lk); 1242} 1243 1244/* 1245 * Replace an old allocdirect dependency with a newer one. 1246 * This routine must be called with splbio interrupts blocked. 1247 */ 1248static void 1249allocdirect_merge(adphead, newadp, oldadp) 1250 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 1251 struct allocdirect *newadp; /* allocdirect being added */ 1252 struct allocdirect *oldadp; /* existing allocdirect being checked */ 1253{ 1254 struct freefrag *freefrag; 1255 1256#ifdef DEBUG 1257 if (lk.lkt_held == -1) 1258 panic("allocdirect_merge: lock not held"); 1259#endif 1260 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 1261 newadp->ad_oldsize != oldadp->ad_newsize || 1262 newadp->ad_lbn >= NDADDR) 1263 panic("allocdirect_check: old %d != new %d || lbn %d >= %d", 1264 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn, 1265 NDADDR); 1266 newadp->ad_oldblkno = oldadp->ad_oldblkno; 1267 newadp->ad_oldsize = oldadp->ad_oldsize; 1268 /* 1269 * If the old dependency had a fragment to free or had never 1270 * previously had a block allocated, then the new dependency 1271 * can immediately post its freefrag and adopt the old freefrag. 1272 * This action is done by swapping the freefrag dependencies. 1273 * The new dependency gains the old one's freefrag, and the 1274 * old one gets the new one and then immediately puts it on 1275 * the worklist when it is freed by free_allocdirect. It is 1276 * not possible to do this swap when the old dependency had a 1277 * non-zero size but no previous fragment to free. This condition 1278 * arises when the new block is an extension of the old block. 1279 * Here, the first part of the fragment allocated to the new 1280 * dependency is part of the block currently claimed on disk by 1281 * the old dependency, so cannot legitimately be freed until the 1282 * conditions for the new dependency are fulfilled. 1283 */ 1284 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 1285 freefrag = newadp->ad_freefrag; 1286 newadp->ad_freefrag = oldadp->ad_freefrag; 1287 oldadp->ad_freefrag = freefrag; 1288 } 1289 free_allocdirect(adphead, oldadp, 0); 1290} 1291 1292/* 1293 * Allocate a new freefrag structure if needed. 1294 */ 1295static struct freefrag * 1296newfreefrag(ip, blkno, size) 1297 struct inode *ip; 1298 ufs_daddr_t blkno; 1299 long size; 1300{ 1301 struct freefrag *freefrag; 1302 struct fs *fs; 1303 1304 if (blkno == 0) 1305 return (NULL); 1306 fs = ip->i_fs; 1307 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 1308 panic("newfreefrag: frag size"); 1309 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag), 1310 M_FREEFRAG, M_WAITOK); 1311 freefrag->ff_list.wk_type = D_FREEFRAG; 1312 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */ 1313 freefrag->ff_inum = ip->i_number; 1314 freefrag->ff_fs = fs; 1315 freefrag->ff_devvp = ip->i_devvp; 1316 freefrag->ff_blkno = blkno; 1317 freefrag->ff_fragsize = size; 1318 return (freefrag); 1319} 1320 1321/* 1322 * This workitem de-allocates fragments that were replaced during 1323 * file block allocation. 1324 */ 1325static void 1326handle_workitem_freefrag(freefrag) 1327 struct freefrag *freefrag; 1328{ 1329 struct inode tip; 1330 1331 tip.i_fs = freefrag->ff_fs; 1332 tip.i_devvp = freefrag->ff_devvp; 1333 tip.i_dev = freefrag->ff_devvp->v_rdev; 1334 tip.i_number = freefrag->ff_inum; 1335 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */ 1336 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize); 1337 FREE(freefrag, M_FREEFRAG); 1338} 1339 1340/* 1341 * Indirect block allocation dependencies. 1342 * 1343 * The same dependencies that exist for a direct block also exist when 1344 * a new block is allocated and pointed to by an entry in a block of 1345 * indirect pointers. The undo/redo states described above are also 1346 * used here. Because an indirect block contains many pointers that 1347 * may have dependencies, a second copy of the entire in-memory indirect 1348 * block is kept. The buffer cache copy is always completely up-to-date. 1349 * The second copy, which is used only as a source for disk writes, 1350 * contains only the safe pointers (i.e., those that have no remaining 1351 * update dependencies). The second copy is freed when all pointers 1352 * are safe. The cache is not allowed to replace indirect blocks with 1353 * pending update dependencies. If a buffer containing an indirect 1354 * block with dependencies is written, these routines will mark it 1355 * dirty again. It can only be successfully written once all the 1356 * dependencies are removed. The ffs_fsync routine in conjunction with 1357 * softdep_sync_metadata work together to get all the dependencies 1358 * removed so that a file can be successfully written to disk. Three 1359 * procedures are used when setting up indirect block pointer 1360 * dependencies. The division is necessary because of the organization 1361 * of the "balloc" routine and because of the distinction between file 1362 * pages and file metadata blocks. 1363 */ 1364 1365/* 1366 * Allocate a new allocindir structure. 1367 */ 1368static struct allocindir * 1369newallocindir(ip, ptrno, newblkno, oldblkno) 1370 struct inode *ip; /* inode for file being extended */ 1371 int ptrno; /* offset of pointer in indirect block */ 1372 ufs_daddr_t newblkno; /* disk block number being added */ 1373 ufs_daddr_t oldblkno; /* previous block number, 0 if none */ 1374{ 1375 struct allocindir *aip; 1376 1377 MALLOC(aip, struct allocindir *, sizeof(struct allocindir), 1378 M_ALLOCINDIR, M_WAITOK); 1379 bzero(aip, sizeof(struct allocindir)); 1380 aip->ai_list.wk_type = D_ALLOCINDIR; 1381 aip->ai_state = ATTACHED; 1382 aip->ai_offset = ptrno; 1383 aip->ai_newblkno = newblkno; 1384 aip->ai_oldblkno = oldblkno; 1385 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize); 1386 return (aip); 1387} 1388 1389/* 1390 * Called just before setting an indirect block pointer 1391 * to a newly allocated file page. 1392 */ 1393void 1394softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 1395 struct inode *ip; /* inode for file being extended */ 1396 ufs_lbn_t lbn; /* allocated block number within file */ 1397 struct buf *bp; /* buffer with indirect blk referencing page */ 1398 int ptrno; /* offset of pointer in indirect block */ 1399 ufs_daddr_t newblkno; /* disk block number being added */ 1400 ufs_daddr_t oldblkno; /* previous block number, 0 if none */ 1401 struct buf *nbp; /* buffer holding allocated page */ 1402{ 1403 struct allocindir *aip; 1404 struct pagedep *pagedep; 1405 1406 aip = newallocindir(ip, ptrno, newblkno, oldblkno); 1407 ACQUIRE_LOCK(&lk); 1408 /* 1409 * If we are allocating a directory page, then we must 1410 * allocate an associated pagedep to track additions and 1411 * deletions. 1412 */ 1413 if ((ip->i_mode & IFMT) == IFDIR && 1414 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1415 WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list); 1416 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1417 FREE_LOCK(&lk); 1418 setup_allocindir_phase2(bp, ip, aip); 1419} 1420 1421/* 1422 * Called just before setting an indirect block pointer to a 1423 * newly allocated indirect block. 1424 */ 1425void 1426softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 1427 struct buf *nbp; /* newly allocated indirect block */ 1428 struct inode *ip; /* inode for file being extended */ 1429 struct buf *bp; /* indirect block referencing allocated block */ 1430 int ptrno; /* offset of pointer in indirect block */ 1431 ufs_daddr_t newblkno; /* disk block number being added */ 1432{ 1433 struct allocindir *aip; 1434 1435 aip = newallocindir(ip, ptrno, newblkno, 0); 1436 ACQUIRE_LOCK(&lk); 1437 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1438 FREE_LOCK(&lk); 1439 setup_allocindir_phase2(bp, ip, aip); 1440} 1441 1442/* 1443 * Called to finish the allocation of the "aip" allocated 1444 * by one of the two routines above. 1445 */ 1446static void 1447setup_allocindir_phase2(bp, ip, aip) 1448 struct buf *bp; /* in-memory copy of the indirect block */ 1449 struct inode *ip; /* inode for file being extended */ 1450 struct allocindir *aip; /* allocindir allocated by the above routines */ 1451{ 1452 struct worklist *wk; 1453 struct indirdep *indirdep, *newindirdep; 1454 struct bmsafemap *bmsafemap; 1455 struct allocindir *oldaip; 1456 struct freefrag *freefrag; 1457 struct newblk *newblk; 1458 1459 if (bp->b_lblkno >= 0) 1460 panic("setup_allocindir_phase2: not indir blk"); 1461 for (indirdep = NULL, newindirdep = NULL; ; ) { 1462 ACQUIRE_LOCK(&lk); 1463 for (wk = LIST_FIRST(&bp->b_dep); wk; 1464 wk = LIST_NEXT(wk, wk_list)) { 1465 if (wk->wk_type != D_INDIRDEP) 1466 continue; 1467 indirdep = WK_INDIRDEP(wk); 1468 break; 1469 } 1470 if (indirdep == NULL && newindirdep) { 1471 indirdep = newindirdep; 1472 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 1473 newindirdep = NULL; 1474 } 1475 FREE_LOCK(&lk); 1476 if (indirdep) { 1477 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0, 1478 &newblk) == 0) 1479 panic("setup_allocindir: lost block"); 1480 ACQUIRE_LOCK(&lk); 1481 if (newblk->nb_state == DEPCOMPLETE) { 1482 aip->ai_state |= DEPCOMPLETE; 1483 aip->ai_buf = NULL; 1484 } else { 1485 bmsafemap = newblk->nb_bmsafemap; 1486 aip->ai_buf = bmsafemap->sm_buf; 1487 LIST_REMOVE(newblk, nb_deps); 1488 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd, 1489 aip, ai_deps); 1490 } 1491 LIST_REMOVE(newblk, nb_hash); 1492 FREE(newblk, M_NEWBLK); 1493 aip->ai_indirdep = indirdep; 1494 /* 1495 * Check to see if there is an existing dependency 1496 * for this block. If there is, merge the old 1497 * dependency into the new one. 1498 */ 1499 if (aip->ai_oldblkno == 0) 1500 oldaip = NULL; 1501 else 1502 for (oldaip=LIST_FIRST(&indirdep->ir_deplisthd); 1503 oldaip; oldaip = LIST_NEXT(oldaip, ai_next)) 1504 if (oldaip->ai_offset == aip->ai_offset) 1505 break; 1506 if (oldaip != NULL) { 1507 if (oldaip->ai_newblkno != aip->ai_oldblkno) 1508 panic("setup_allocindir_phase2: blkno"); 1509 aip->ai_oldblkno = oldaip->ai_oldblkno; 1510 freefrag = oldaip->ai_freefrag; 1511 oldaip->ai_freefrag = aip->ai_freefrag; 1512 aip->ai_freefrag = freefrag; 1513 free_allocindir(oldaip, NULL); 1514 } 1515 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 1516 ((ufs_daddr_t *)indirdep->ir_savebp->b_data) 1517 [aip->ai_offset] = aip->ai_oldblkno; 1518 FREE_LOCK(&lk); 1519 } 1520 if (newindirdep) { 1521 if (indirdep->ir_savebp != NULL) 1522 brelse(newindirdep->ir_savebp); 1523 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP); 1524 } 1525 if (indirdep) 1526 break; 1527 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep), 1528 M_INDIRDEP, M_WAITOK); 1529 newindirdep->ir_list.wk_type = D_INDIRDEP; 1530 newindirdep->ir_state = ATTACHED; 1531 LIST_INIT(&newindirdep->ir_deplisthd); 1532 LIST_INIT(&newindirdep->ir_donehd); 1533#ifdef __FreeBSD__ 1534 if (bp->b_blkno == bp->b_lblkno) { 1535#if 0 /* we know this happens.. research suggested.. */ 1536 printf("setup_allocindir_phase2: need bmap, blk %d\n", 1537 bp->b_lblkno); 1538#endif 1539 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, 1540 NULL, NULL); 1541 } 1542#endif /* __FreeBSD__ */ 1543 newindirdep->ir_savebp = 1544 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0); 1545 bp->b_flags |= B_XXX; 1546 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 1547 } 1548} 1549 1550/* 1551 * Block de-allocation dependencies. 1552 * 1553 * When blocks are de-allocated, the on-disk pointers must be nullified before 1554 * the blocks are made available for use by other files. (The true 1555 * requirement is that old pointers must be nullified before new on-disk 1556 * pointers are set. We chose this slightly more stringent requirement to 1557 * reduce complexity.) Our implementation handles this dependency by updating 1558 * the inode (or indirect block) appropriately but delaying the actual block 1559 * de-allocation (i.e., freemap and free space count manipulation) until 1560 * after the updated versions reach stable storage. After the disk is 1561 * updated, the blocks can be safely de-allocated whenever it is convenient. 1562 * This implementation handles only the common case of reducing a file's 1563 * length to zero. Other cases are handled by the conventional synchronous 1564 * write approach. 1565 * 1566 * The ffs implementation with which we worked double-checks 1567 * the state of the block pointers and file size as it reduces 1568 * a file's length. Some of this code is replicated here in our 1569 * soft updates implementation. The freeblks->fb_chkcnt field is 1570 * used to transfer a part of this information to the procedure 1571 * that eventually de-allocates the blocks. 1572 * 1573 * This routine should be called from the routine that shortens 1574 * a file's length, before the inode's size or block pointers 1575 * are modified. It will save the block pointer information for 1576 * later release and zero the inode so that the calling routine 1577 * can release it. 1578 */ 1579static long num_freeblks; /* number of freeblks allocated */ 1580void 1581softdep_setup_freeblocks(ip, length) 1582 struct inode *ip; /* The inode whose length is to be reduced */ 1583 off_t length; /* The new length for the file */ 1584{ 1585 struct freeblks *freeblks; 1586 struct inodedep *inodedep; 1587 struct allocdirect *adp; 1588 struct vnode *vp; 1589 struct buf *bp; 1590 struct fs *fs; 1591 int i, error; 1592 1593 fs = ip->i_fs; 1594 if (length != 0) 1595 panic("softde_setup_freeblocks: non-zero length"); 1596 (void) checklimit(&num_freeblks, 0); 1597 num_freeblks += 1; 1598 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks), 1599 M_FREEBLKS, M_WAITOK); 1600 bzero(freeblks, sizeof(struct freeblks)); 1601 freeblks->fb_list.wk_type = D_FREEBLKS; 1602 freeblks->fb_uid = ip->i_uid; 1603 freeblks->fb_previousinum = ip->i_number; 1604 freeblks->fb_devvp = ip->i_devvp; 1605 freeblks->fb_fs = fs; 1606 freeblks->fb_oldsize = ip->i_size; 1607 freeblks->fb_newsize = length; 1608 freeblks->fb_chkcnt = ip->i_blocks; 1609 for (i = 0; i < NDADDR; i++) { 1610 freeblks->fb_dblks[i] = ip->i_db[i]; 1611 ip->i_db[i] = 0; 1612 } 1613 for (i = 0; i < NIADDR; i++) { 1614 freeblks->fb_iblks[i] = ip->i_ib[i]; 1615 ip->i_ib[i] = 0; 1616 } 1617 ip->i_blocks = 0; 1618 ip->i_size = 0; 1619 /* 1620 * Push the zero'ed inode to to its disk buffer so that we are free 1621 * to delete its dependencies below. Once the dependencies are gone 1622 * the buffer can be safely released. 1623 */ 1624 if ((error = bread(ip->i_devvp, 1625 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 1626 (int)fs->fs_bsize, NOCRED, &bp)) != 0) 1627 softdep_error("softdep_setup_freeblocks", error); 1628 *((struct dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) = 1629 ip->i_din; 1630 /* 1631 * Find and eliminate any inode dependencies. 1632 */ 1633 ACQUIRE_LOCK(&lk); 1634 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep); 1635 if ((inodedep->id_state & IOSTARTED) != 0) 1636 panic("softdep_setup_freeblocks: inode busy"); 1637 /* 1638 * Add the freeblks structure to the list of operations that 1639 * must await the zero'ed inode being written to disk. 1640 */ 1641 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list); 1642 /* 1643 * Because the file length has been truncated to zero, any 1644 * pending block allocation dependency structures associated 1645 * with this inode are obsolete and can simply be de-allocated. 1646 * We must first merge the two dependency lists to get rid of 1647 * any duplicate freefrag structures, then purge the merged list. 1648 */ 1649 merge_inode_lists(inodedep); 1650 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 1651 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 1652 bdwrite(bp); 1653 /* 1654 * We must wait for any I/O in progress to finish so that 1655 * all potential buffers on the dirty list will be visible. 1656 * Once they are all there, walk the list and get rid of 1657 * any dependencies. 1658 */ 1659 vp = ITOV(ip); 1660 while (vp->v_numoutput) { 1661 vp->v_flag |= VBWAIT; 1662 FREE_LOCK_INTERLOCKED(&lk); 1663 tsleep((caddr_t)&vp->v_numoutput, PRIBIO + 1, "sdsetf", 0); 1664 ACQUIRE_LOCK_INTERLOCKED(&lk); 1665 } 1666 while (getdirtybuf(&LIST_FIRST(&vp->v_dirtyblkhd), MNT_WAIT)) { 1667 bp = LIST_FIRST(&vp->v_dirtyblkhd); 1668 (void) inodedep_lookup(fs, ip->i_number, 0, &inodedep); 1669 deallocate_dependencies(bp, inodedep); 1670 bp->b_flags |= B_INVAL | B_NOCACHE; 1671 brelse(bp); 1672 } 1673 /* 1674 * Try freeing the inodedep in case that was the last dependency. 1675 */ 1676 if ((inodedep_lookup(fs, ip->i_number, 0, &inodedep)) != 0) 1677 (void) free_inodedep(inodedep); 1678 FREE_LOCK(&lk); 1679} 1680 1681/* 1682 * Reclaim any dependency structures from a buffer that is about to 1683 * be reallocated to a new vnode. The buffer must be locked, thus, 1684 * no I/O completion operations can occur while we are manipulating 1685 * its associated dependencies. The mutex is held so that other I/O's 1686 * associated with related dependencies do not occur. 1687 */ 1688static void 1689deallocate_dependencies(bp, inodedep) 1690 struct buf *bp; 1691 struct inodedep *inodedep; 1692{ 1693 struct worklist *wk; 1694 struct indirdep *indirdep; 1695 struct allocindir *aip; 1696 struct pagedep *pagedep; 1697 struct dirrem *dirrem; 1698 struct diradd *dap; 1699 int i; 1700 1701 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 1702 switch (wk->wk_type) { 1703 1704 case D_INDIRDEP: 1705 indirdep = WK_INDIRDEP(wk); 1706 /* 1707 * None of the indirect pointers will ever be visible, 1708 * so they can simply be tossed. GOINGAWAY ensures 1709 * that allocated pointers will be saved in the buffer 1710 * cache until they are freed. Note that they will 1711 * only be able to be found by their physical address 1712 * since the inode mapping the logical address will 1713 * be gone. The save buffer used for the safe copy 1714 * was allocated in setup_allocindir_phase2 using 1715 * the physical address so it could be used for this 1716 * purpose. Hence we swap the safe copy with the real 1717 * copy, allowing the safe copy to be freed and holding 1718 * on to the real copy for later use in indir_trunc. 1719 */ 1720 if (indirdep->ir_state & GOINGAWAY) 1721 panic("deallocate_dependencies: already gone"); 1722 indirdep->ir_state |= GOINGAWAY; 1723 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 1724 free_allocindir(aip, inodedep); 1725 if (bp->b_lblkno >= 0 || 1726 bp->b_blkno != indirdep->ir_savebp->b_lblkno) 1727 panic("deallocate_dependencies: not indir"); 1728 bcopy(bp->b_data, indirdep->ir_savebp->b_data, 1729 bp->b_bcount); 1730 WORKLIST_REMOVE(wk); 1731 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk); 1732 continue; 1733 1734 case D_PAGEDEP: 1735 pagedep = WK_PAGEDEP(wk); 1736 /* 1737 * None of the directory additions will ever be 1738 * visible, so they can simply be tossed. 1739 */ 1740 for (i = 0; i < DAHASHSZ; i++) 1741 while (dap=LIST_FIRST(&pagedep->pd_diraddhd[i])) 1742 free_diradd(dap); 1743 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0) 1744 free_diradd(dap); 1745 /* 1746 * Copy any directory remove dependencies to the list 1747 * to be processed after the zero'ed inode is written. 1748 * If the inode has already been written, then they 1749 * can be dumped directly onto the work list. 1750 */ 1751 for (dirrem = LIST_FIRST(&pagedep->pd_dirremhd); dirrem; 1752 dirrem = LIST_NEXT(dirrem, dm_next)) { 1753 LIST_REMOVE(dirrem, dm_next); 1754 dirrem->dm_dirinum = pagedep->pd_ino; 1755 if (inodedep == NULL) 1756 add_to_worklist(&dirrem->dm_list); 1757 else 1758 WORKLIST_INSERT(&inodedep->id_bufwait, 1759 &dirrem->dm_list); 1760 } 1761 WORKLIST_REMOVE(&pagedep->pd_list); 1762 LIST_REMOVE(pagedep, pd_hash); 1763 WORKITEM_FREE(pagedep, D_PAGEDEP); 1764 continue; 1765 1766 case D_ALLOCINDIR: 1767 free_allocindir(WK_ALLOCINDIR(wk), inodedep); 1768 continue; 1769 1770 case D_ALLOCDIRECT: 1771 case D_INODEDEP: 1772 panic("deallocate_dependencies: Unexpected type %s", 1773 TYPENAME(wk->wk_type)); 1774 /* NOTREACHED */ 1775 1776 default: 1777 panic("deallocate_dependencies: Unknown type %s", 1778 TYPENAME(wk->wk_type)); 1779 /* NOTREACHED */ 1780 } 1781 } 1782} 1783 1784/* 1785 * Free an allocdirect. Generate a new freefrag work request if appropriate. 1786 * This routine must be called with splbio interrupts blocked. 1787 */ 1788static void 1789free_allocdirect(adphead, adp, delay) 1790 struct allocdirectlst *adphead; 1791 struct allocdirect *adp; 1792 int delay; 1793{ 1794 1795#ifdef DEBUG 1796 if (lk.lkt_held == -1) 1797 panic("free_allocdirect: lock not held"); 1798#endif 1799 if ((adp->ad_state & DEPCOMPLETE) == 0) 1800 LIST_REMOVE(adp, ad_deps); 1801 TAILQ_REMOVE(adphead, adp, ad_next); 1802 if ((adp->ad_state & COMPLETE) == 0) 1803 WORKLIST_REMOVE(&adp->ad_list); 1804 if (adp->ad_freefrag != NULL) { 1805 if (delay) 1806 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 1807 &adp->ad_freefrag->ff_list); 1808 else 1809 add_to_worklist(&adp->ad_freefrag->ff_list); 1810 } 1811 WORKITEM_FREE(adp, D_ALLOCDIRECT); 1812} 1813 1814/* 1815 * Prepare an inode to be freed. The actual free operation is not 1816 * done until the zero'ed inode has been written to disk. 1817 */ 1818static long num_freefile; /* number of freefile allocated */ 1819void 1820softdep_freefile(pvp, ino, mode) 1821 struct vnode *pvp; 1822 ino_t ino; 1823 int mode; 1824{ 1825 struct inode *ip = VTOI(pvp); 1826 struct inodedep *inodedep; 1827 struct freefile *freefile; 1828 1829 /* 1830 * This sets up the inode de-allocation dependency. 1831 */ 1832 (void) checklimit(&num_freefile, 0); 1833 num_freefile += 1; 1834 MALLOC(freefile, struct freefile *, sizeof(struct freefile), 1835 M_FREEFILE, M_WAITOK); 1836 freefile->fx_list.wk_type = D_FREEFILE; 1837 freefile->fx_list.wk_state = 0; 1838 freefile->fx_mode = mode; 1839 freefile->fx_oldinum = ino; 1840 freefile->fx_devvp = ip->i_devvp; 1841 freefile->fx_fs = ip->i_fs; 1842 1843 /* 1844 * If the inodedep does not exist, then the zero'ed inode has 1845 * been written to disk and we can free the file immediately. 1846 */ 1847 ACQUIRE_LOCK(&lk); 1848 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0) { 1849 add_to_worklist(&freefile->fx_list); 1850 FREE_LOCK(&lk); 1851 return; 1852 } 1853 1854 /* 1855 * If we still have a bitmap dependency, then the inode has never 1856 * been written to disk. Drop the dependency as it is no longer 1857 * necessary since the inode is being deallocated. We could process 1858 * the freefile immediately, but then we would have to clear the 1859 * id_inowait dependencies here and it is easier just to let the 1860 * zero'ed inode be written and let them be cleaned up in the 1861 * normal followup actions that follow the inode write. 1862 */ 1863 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 1864 inodedep->id_state |= DEPCOMPLETE; 1865 LIST_REMOVE(inodedep, id_deps); 1866 inodedep->id_buf = NULL; 1867 } 1868 /* 1869 * If the inodedep has no dependencies associated with it, 1870 * then we must free it here and free the file immediately. 1871 * This case arises when an early allocation fails (for 1872 * example, the user is over their file quota). 1873 */ 1874 if (free_inodedep(inodedep) == 0) 1875 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 1876 else 1877 add_to_worklist(&freefile->fx_list); 1878 FREE_LOCK(&lk); 1879} 1880 1881/* 1882 * Try to free an inodedep structure. Return 1 if it could be freed. 1883 */ 1884static int 1885free_inodedep(inodedep) 1886 struct inodedep *inodedep; 1887{ 1888 1889 if ((inodedep->id_state & ONWORKLIST) != 0 || 1890 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 1891 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 1892 LIST_FIRST(&inodedep->id_bufwait) != NULL || 1893 LIST_FIRST(&inodedep->id_inowait) != NULL || 1894 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 1895 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 1896 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL) 1897 return (0); 1898 LIST_REMOVE(inodedep, id_hash); 1899 WORKITEM_FREE(inodedep, D_INODEDEP); 1900 num_inodedep -= 1; 1901 return (1); 1902} 1903 1904/* 1905 * This workitem routine performs the block de-allocation. 1906 * The workitem is added to the pending list after the updated 1907 * inode block has been written to disk. As mentioned above, 1908 * checks regarding the number of blocks de-allocated (compared 1909 * to the number of blocks allocated for the file) are also 1910 * performed in this function. 1911 */ 1912static void 1913handle_workitem_freeblocks(freeblks) 1914 struct freeblks *freeblks; 1915{ 1916 struct inode tip; 1917 ufs_daddr_t bn; 1918 struct fs *fs; 1919 int i, level, bsize; 1920 long nblocks, blocksreleased = 0; 1921 int error, allerror = 0; 1922 ufs_lbn_t baselbns[NIADDR], tmpval; 1923 1924 tip.i_number = freeblks->fb_previousinum; 1925 tip.i_devvp = freeblks->fb_devvp; 1926 tip.i_dev = freeblks->fb_devvp->v_rdev; 1927 tip.i_fs = freeblks->fb_fs; 1928 tip.i_size = freeblks->fb_oldsize; 1929 tip.i_uid = freeblks->fb_uid; 1930 fs = freeblks->fb_fs; 1931 tmpval = 1; 1932 baselbns[0] = NDADDR; 1933 for (i = 1; i < NIADDR; i++) { 1934 tmpval *= NINDIR(fs); 1935 baselbns[i] = baselbns[i - 1] + tmpval; 1936 } 1937 nblocks = btodb(fs->fs_bsize); 1938 blocksreleased = 0; 1939 /* 1940 * Indirect blocks first. 1941 */ 1942 for (level = (NIADDR - 1); level >= 0; level--) { 1943 if ((bn = freeblks->fb_iblks[level]) == 0) 1944 continue; 1945 if ((error = indir_trunc(&tip, fsbtodb(fs, bn), level, 1946 baselbns[level], &blocksreleased)) == 0) 1947 allerror = error; 1948 ffs_blkfree(&tip, bn, fs->fs_bsize); 1949 blocksreleased += nblocks; 1950 } 1951 /* 1952 * All direct blocks or frags. 1953 */ 1954 for (i = (NDADDR - 1); i >= 0; i--) { 1955 if ((bn = freeblks->fb_dblks[i]) == 0) 1956 continue; 1957 bsize = blksize(fs, &tip, i); 1958 ffs_blkfree(&tip, bn, bsize); 1959 blocksreleased += btodb(bsize); 1960 } 1961 1962#ifdef DIAGNOSTIC 1963 if (freeblks->fb_chkcnt != blocksreleased) 1964 panic("handle_workitem_freeblocks: block count"); 1965 if (allerror) 1966 softdep_error("handle_workitem_freeblks", allerror); 1967#endif /* DIAGNOSTIC */ 1968 WORKITEM_FREE(freeblks, D_FREEBLKS); 1969 num_freeblks -= 1; 1970} 1971 1972/* 1973 * Release blocks associated with the inode ip and stored in the indirect 1974 * block dbn. If level is greater than SINGLE, the block is an indirect block 1975 * and recursive calls to indirtrunc must be used to cleanse other indirect 1976 * blocks. 1977 */ 1978static int 1979indir_trunc(ip, dbn, level, lbn, countp) 1980 struct inode *ip; 1981 ufs_daddr_t dbn; 1982 int level; 1983 ufs_lbn_t lbn; 1984 long *countp; 1985{ 1986 struct buf *bp; 1987 ufs_daddr_t *bap; 1988 ufs_daddr_t nb; 1989 struct fs *fs; 1990 struct worklist *wk; 1991 struct indirdep *indirdep; 1992 int i, lbnadd, nblocks; 1993 int error, allerror = 0; 1994 1995 fs = ip->i_fs; 1996 lbnadd = 1; 1997 for (i = level; i > 0; i--) 1998 lbnadd *= NINDIR(fs); 1999 /* 2000 * Get buffer of block pointers to be freed. This routine is not 2001 * called until the zero'ed inode has been written, so it is safe 2002 * to free blocks as they are encountered. Because the inode has 2003 * been zero'ed, calls to bmap on these blocks will fail. So, we 2004 * have to use the on-disk address and the block device for the 2005 * filesystem to look them up. If the file was deleted before its 2006 * indirect blocks were all written to disk, the routine that set 2007 * us up (deallocate_dependencies) will have arranged to leave 2008 * a complete copy of the indirect block in memory for our use. 2009 * Otherwise we have to read the blocks in from the disk. 2010 */ 2011 ACQUIRE_LOCK(&lk); 2012 if ((bp = incore(ip->i_devvp, dbn)) != NULL && 2013 (wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2014 if (wk->wk_type != D_INDIRDEP || 2015 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp || 2016 (indirdep->ir_state & GOINGAWAY) == 0) 2017 panic("indir_trunc: lost indirdep"); 2018 WORKLIST_REMOVE(wk); 2019 WORKITEM_FREE(indirdep, D_INDIRDEP); 2020 if (LIST_FIRST(&bp->b_dep) != NULL) 2021 panic("indir_trunc: dangling dep"); 2022 FREE_LOCK(&lk); 2023 } else { 2024 FREE_LOCK(&lk); 2025 error = bread(ip->i_devvp, dbn, (int)fs->fs_bsize, NOCRED, &bp); 2026 if (error) 2027 return (error); 2028 } 2029 /* 2030 * Recursively free indirect blocks. 2031 */ 2032 bap = (ufs_daddr_t *)bp->b_data; 2033 nblocks = btodb(fs->fs_bsize); 2034 for (i = NINDIR(fs) - 1; i >= 0; i--) { 2035 if ((nb = bap[i]) == 0) 2036 continue; 2037 if (level != 0) { 2038 if ((error = indir_trunc(ip, fsbtodb(fs, nb), 2039 level - 1, lbn + (i * lbnadd), countp)) != 0) 2040 allerror = error; 2041 } 2042 ffs_blkfree(ip, nb, fs->fs_bsize); 2043 *countp += nblocks; 2044 } 2045 bp->b_flags |= B_INVAL | B_NOCACHE; 2046 bp->b_flags &= ~B_XXX; 2047 brelse(bp); 2048 return (allerror); 2049} 2050 2051/* 2052 * Free an allocindir. 2053 * This routine must be called with splbio interrupts blocked. 2054 */ 2055static void 2056free_allocindir(aip, inodedep) 2057 struct allocindir *aip; 2058 struct inodedep *inodedep; 2059{ 2060 struct freefrag *freefrag; 2061 2062#ifdef DEBUG 2063 if (lk.lkt_held == -1) 2064 panic("free_allocindir: lock not held"); 2065#endif 2066 if ((aip->ai_state & DEPCOMPLETE) == 0) 2067 LIST_REMOVE(aip, ai_deps); 2068 if (aip->ai_state & ONWORKLIST) 2069 WORKLIST_REMOVE(&aip->ai_list); 2070 LIST_REMOVE(aip, ai_next); 2071 if ((freefrag = aip->ai_freefrag) != NULL) { 2072 if (inodedep == NULL) 2073 add_to_worklist(&freefrag->ff_list); 2074 else 2075 WORKLIST_INSERT(&inodedep->id_bufwait, 2076 &freefrag->ff_list); 2077 } 2078 WORKITEM_FREE(aip, D_ALLOCINDIR); 2079} 2080 2081/* 2082 * Directory entry addition dependencies. 2083 * 2084 * When adding a new directory entry, the inode (with its incremented link 2085 * count) must be written to disk before the directory entry's pointer to it. 2086 * Also, if the inode is newly allocated, the corresponding freemap must be 2087 * updated (on disk) before the directory entry's pointer. These requirements 2088 * are met via undo/redo on the directory entry's pointer, which consists 2089 * simply of the inode number. 2090 * 2091 * As directory entries are added and deleted, the free space within a 2092 * directory block can become fragmented. The ufs file system will compact 2093 * a fragmented directory block to make space for a new entry. When this 2094 * occurs, the offsets of previously added entries change. Any "diradd" 2095 * dependency structures corresponding to these entries must be updated with 2096 * the new offsets. 2097 */ 2098 2099/* 2100 * This routine is called after the in-memory inode's link 2101 * count has been incremented, but before the directory entry's 2102 * pointer to the inode has been set. 2103 */ 2104void 2105softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp) 2106 struct buf *bp; /* buffer containing directory block */ 2107 struct inode *dp; /* inode for directory */ 2108 off_t diroffset; /* offset of new entry in directory */ 2109 long newinum; /* inode referenced by new directory entry */ 2110 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 2111{ 2112 int offset; /* offset of new entry within directory block */ 2113 ufs_lbn_t lbn; /* block in directory containing new entry */ 2114 struct fs *fs; 2115 struct diradd *dap; 2116 struct pagedep *pagedep; 2117 struct inodedep *inodedep; 2118 struct mkdir *mkdir1, *mkdir2; 2119 2120 /* 2121 * Whiteouts have no dependencies. 2122 */ 2123 if (newinum == WINO) { 2124 if (newdirbp != NULL) 2125 bdwrite(newdirbp); 2126 return; 2127 } 2128 2129 fs = dp->i_fs; 2130 lbn = lblkno(fs, diroffset); 2131 offset = blkoff(fs, diroffset); 2132 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, M_WAITOK); 2133 bzero(dap, sizeof(struct diradd)); 2134 dap->da_list.wk_type = D_DIRADD; 2135 dap->da_offset = offset; 2136 dap->da_newinum = newinum; 2137 dap->da_state = ATTACHED; 2138 if (newdirbp == NULL) { 2139 dap->da_state |= DEPCOMPLETE; 2140 ACQUIRE_LOCK(&lk); 2141 } else { 2142 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 2143 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2144 M_WAITOK); 2145 mkdir1->md_list.wk_type = D_MKDIR; 2146 mkdir1->md_state = MKDIR_BODY; 2147 mkdir1->md_diradd = dap; 2148 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2149 M_WAITOK); 2150 mkdir2->md_list.wk_type = D_MKDIR; 2151 mkdir2->md_state = MKDIR_PARENT; 2152 mkdir2->md_diradd = dap; 2153 ACQUIRE_LOCK(&lk); 2154 /* 2155 * Dependency on "." and ".." being written to disk. 2156 */ 2157 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 2158 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list); 2159 bdwrite(newdirbp); 2160 /* 2161 * Dependency on link count increase for parent directory 2162 */ 2163 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0 2164 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2165 dap->da_state &= ~MKDIR_PARENT; 2166 WORKITEM_FREE(mkdir2, D_MKDIR); 2167 } else { 2168 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 2169 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 2170 } 2171 } 2172 /* 2173 * Link into parent directory pagedep to await its being written. 2174 */ 2175 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2176 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2177 dap->da_pagedep = pagedep; 2178 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 2179 da_pdlist); 2180 /* 2181 * Link into its inodedep. Put it on the id_bufwait list if the inode 2182 * is not yet written. If it is written, do the post-inode write 2183 * processing to put it on the id_pendinghd list. 2184 */ 2185 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep); 2186 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 2187 diradd_inode_written(dap, inodedep); 2188 else 2189 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2190 FREE_LOCK(&lk); 2191} 2192 2193/* 2194 * This procedure is called to change the offset of a directory 2195 * entry when compacting a directory block which must be owned 2196 * exclusively by the caller. Note that the actual entry movement 2197 * must be done in this procedure to ensure that no I/O completions 2198 * occur while the move is in progress. 2199 */ 2200void 2201softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize) 2202 struct inode *dp; /* inode for directory */ 2203 caddr_t base; /* address of dp->i_offset */ 2204 caddr_t oldloc; /* address of old directory location */ 2205 caddr_t newloc; /* address of new directory location */ 2206 int entrysize; /* size of directory entry */ 2207{ 2208 int offset, oldoffset, newoffset; 2209 struct pagedep *pagedep; 2210 struct diradd *dap; 2211 ufs_lbn_t lbn; 2212 2213 ACQUIRE_LOCK(&lk); 2214 lbn = lblkno(dp->i_fs, dp->i_offset); 2215 offset = blkoff(dp->i_fs, dp->i_offset); 2216 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) 2217 goto done; 2218 oldoffset = offset + (oldloc - base); 2219 newoffset = offset + (newloc - base); 2220 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[DIRADDHASH(oldoffset)]); 2221 dap; dap = LIST_NEXT(dap, da_pdlist)) { 2222 if (dap->da_offset != oldoffset) 2223 continue; 2224 dap->da_offset = newoffset; 2225 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) 2226 break; 2227 LIST_REMOVE(dap, da_pdlist); 2228 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], 2229 dap, da_pdlist); 2230 break; 2231 } 2232done: 2233 bcopy(oldloc, newloc, entrysize); 2234 FREE_LOCK(&lk); 2235} 2236 2237/* 2238 * Free a diradd dependency structure. This routine must be called 2239 * with splbio interrupts blocked. 2240 */ 2241static void 2242free_diradd(dap) 2243 struct diradd *dap; 2244{ 2245 struct dirrem *dirrem; 2246 struct pagedep *pagedep; 2247 struct inodedep *inodedep; 2248 struct mkdir *mkdir, *nextmd; 2249 2250#ifdef DEBUG 2251 if (lk.lkt_held == -1) 2252 panic("free_diradd: lock not held"); 2253#endif 2254 WORKLIST_REMOVE(&dap->da_list); 2255 LIST_REMOVE(dap, da_pdlist); 2256 if ((dap->da_state & DIRCHG) == 0) { 2257 pagedep = dap->da_pagedep; 2258 } else { 2259 dirrem = dap->da_previous; 2260 pagedep = dirrem->dm_pagedep; 2261 add_to_worklist(&dirrem->dm_list); 2262 } 2263 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum, 2264 0, &inodedep) != 0) 2265 (void) free_inodedep(inodedep); 2266 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2267 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 2268 nextmd = LIST_NEXT(mkdir, md_mkdirs); 2269 if (mkdir->md_diradd != dap) 2270 continue; 2271 dap->da_state &= ~mkdir->md_state; 2272 WORKLIST_REMOVE(&mkdir->md_list); 2273 LIST_REMOVE(mkdir, md_mkdirs); 2274 WORKITEM_FREE(mkdir, D_MKDIR); 2275 } 2276 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 2277 panic("free_diradd: unfound ref"); 2278 } 2279 WORKITEM_FREE(dap, D_DIRADD); 2280} 2281 2282/* 2283 * Directory entry removal dependencies. 2284 * 2285 * When removing a directory entry, the entry's inode pointer must be 2286 * zero'ed on disk before the corresponding inode's link count is decremented 2287 * (possibly freeing the inode for re-use). This dependency is handled by 2288 * updating the directory entry but delaying the inode count reduction until 2289 * after the directory block has been written to disk. After this point, the 2290 * inode count can be decremented whenever it is convenient. 2291 */ 2292 2293/* 2294 * This routine should be called immediately after removing 2295 * a directory entry. The inode's link count should not be 2296 * decremented by the calling procedure -- the soft updates 2297 * code will do this task when it is safe. 2298 */ 2299void 2300softdep_setup_remove(bp, dp, ip, isrmdir) 2301 struct buf *bp; /* buffer containing directory block */ 2302 struct inode *dp; /* inode for the directory being modified */ 2303 struct inode *ip; /* inode for directory entry being removed */ 2304 int isrmdir; /* indicates if doing RMDIR */ 2305{ 2306 struct dirrem *dirrem; 2307 2308 /* 2309 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 2310 */ 2311 dirrem = newdirrem(bp, dp, ip, isrmdir); 2312 if ((dirrem->dm_state & COMPLETE) == 0) { 2313 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 2314 dm_next); 2315 } else { 2316 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 2317 add_to_worklist(&dirrem->dm_list); 2318 } 2319 FREE_LOCK(&lk); 2320} 2321 2322/* 2323 * Allocate a new dirrem if appropriate and return it along with 2324 * its associated pagedep. Called without a lock, returns with lock. 2325 */ 2326static struct dirrem * 2327newdirrem(bp, dp, ip, isrmdir) 2328 struct buf *bp; /* buffer containing directory block */ 2329 struct inode *dp; /* inode for the directory being modified */ 2330 struct inode *ip; /* inode for directory entry being removed */ 2331 int isrmdir; /* indicates if doing RMDIR */ 2332{ 2333 int offset; 2334 ufs_lbn_t lbn; 2335 struct diradd *dap; 2336 struct dirrem *dirrem; 2337 struct pagedep *pagedep; 2338 2339 /* 2340 * Whiteouts have no deletion dependencies. 2341 */ 2342 if (ip == NULL) 2343 panic("newdirrem: whiteout"); 2344 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem), 2345 M_DIRREM, M_WAITOK); 2346 bzero(dirrem, sizeof(struct dirrem)); 2347 dirrem->dm_list.wk_type = D_DIRREM; 2348 dirrem->dm_state = isrmdir ? RMDIR : 0; 2349 dirrem->dm_mnt = ITOV(ip)->v_mount; 2350 dirrem->dm_oldinum = ip->i_number; 2351 2352 ACQUIRE_LOCK(&lk); 2353 lbn = lblkno(dp->i_fs, dp->i_offset); 2354 offset = blkoff(dp->i_fs, dp->i_offset); 2355 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2356 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2357 dirrem->dm_pagedep = pagedep; 2358 /* 2359 * Check for a diradd dependency for the same directory entry. 2360 * If present, then both dependencies become obsolete and can 2361 * be de-allocated. Check for an entry on both the pd_dirraddhd 2362 * list and the pd_pendinghd list. 2363 */ 2364 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[DIRADDHASH(offset)]); 2365 dap; dap = LIST_NEXT(dap, da_pdlist)) 2366 if (dap->da_offset == offset) 2367 break; 2368 if (dap == NULL) { 2369 for (dap = LIST_FIRST(&pagedep->pd_pendinghd); 2370 dap; dap = LIST_NEXT(dap, da_pdlist)) 2371 if (dap->da_offset == offset) 2372 break; 2373 if (dap == NULL) 2374 return (dirrem); 2375 } 2376 /* 2377 * Must be ATTACHED at this point, so just delete it. 2378 */ 2379 if ((dap->da_state & ATTACHED) == 0) 2380 panic("newdirrem: not ATTACHED"); 2381 if (dap->da_newinum != ip->i_number) 2382 panic("newdirrem: inum %d should be %d", 2383 ip->i_number, dap->da_newinum); 2384 free_diradd(dap); 2385 dirrem->dm_state |= COMPLETE; 2386 return (dirrem); 2387} 2388 2389/* 2390 * Directory entry change dependencies. 2391 * 2392 * Changing an existing directory entry requires that an add operation 2393 * be completed first followed by a deletion. The semantics for the addition 2394 * are identical to the description of adding a new entry above except 2395 * that the rollback is to the old inode number rather than zero. Once 2396 * the addition dependency is completed, the removal is done as described 2397 * in the removal routine above. 2398 */ 2399 2400/* 2401 * This routine should be called immediately after changing 2402 * a directory entry. The inode's link count should not be 2403 * decremented by the calling procedure -- the soft updates 2404 * code will perform this task when it is safe. 2405 */ 2406void 2407softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 2408 struct buf *bp; /* buffer containing directory block */ 2409 struct inode *dp; /* inode for the directory being modified */ 2410 struct inode *ip; /* inode for directory entry being removed */ 2411 long newinum; /* new inode number for changed entry */ 2412 int isrmdir; /* indicates if doing RMDIR */ 2413{ 2414 int offset; 2415 struct diradd *dap; 2416 struct dirrem *dirrem; 2417 struct inodedep *inodedep; 2418 2419 offset = blkoff(dp->i_fs, dp->i_offset); 2420 2421 /* 2422 * Whiteouts have no addition dependencies. 2423 */ 2424 if (newinum == WINO) { 2425 dap = NULL; 2426 } else { 2427 MALLOC(dap, struct diradd *, sizeof(struct diradd), 2428 M_DIRADD, M_WAITOK); 2429 bzero(dap, sizeof(struct diradd)); 2430 dap->da_list.wk_type = D_DIRADD; 2431 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 2432 dap->da_offset = offset; 2433 dap->da_newinum = newinum; 2434 } 2435 2436 /* 2437 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 2438 */ 2439 dirrem = newdirrem(bp, dp, ip, isrmdir); 2440 2441 /* 2442 * If the inode has already been written, then no addition 2443 * dependency needs to be created. 2444 */ 2445 if (inodedep_lookup(dp->i_fs, newinum, 0, &inodedep) == 0 || 2446 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2447 WORKITEM_FREE(dap, D_DIRADD); 2448 dap = NULL; 2449 } 2450 2451 if (dap) { 2452 dap->da_previous = dirrem; 2453 LIST_INSERT_HEAD( 2454 &dirrem->dm_pagedep->pd_diraddhd[DIRADDHASH(offset)], 2455 dap, da_pdlist); 2456 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2457 } else if ((dirrem->dm_state & COMPLETE) == 0) { 2458 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 2459 dm_next); 2460 } else { 2461 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 2462 add_to_worklist(&dirrem->dm_list); 2463 } 2464 FREE_LOCK(&lk); 2465} 2466 2467/* 2468 * Called whenever the link count on an inode is increased. 2469 * It creates an inode dependency so that the new reference(s) 2470 * to the inode cannot be committed to disk until the updated 2471 * inode has been written. 2472 */ 2473void 2474softdep_increase_linkcnt(ip) 2475 struct inode *ip; /* the inode with the increased link count */ 2476{ 2477 struct inodedep *inodedep; 2478 2479 ACQUIRE_LOCK(&lk); 2480 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 2481 FREE_LOCK(&lk); 2482} 2483 2484/* 2485 * This workitem decrements the inode's link count. 2486 * If the link count reaches zero, the file is removed. 2487 */ 2488static void 2489handle_workitem_remove(dirrem) 2490 struct dirrem *dirrem; 2491{ 2492 struct proc *p = CURPROC; /* XXX */ 2493 struct inodedep *inodedep; 2494 struct vnode *vp; 2495 struct inode *ip; 2496 int error; 2497 2498 if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) { 2499 softdep_error("handle_workitem_remove: vget", error); 2500 return; 2501 } 2502 ip = VTOI(vp); 2503 /* 2504 * Normal file deletion. 2505 */ 2506 if ((dirrem->dm_state & RMDIR) == 0) { 2507 ip->i_nlink--; 2508 if (ip->i_nlink < ip->i_effnlink) 2509 panic("handle_workitem_remove: bad file delta"); 2510 ip->i_flag |= IN_CHANGE; 2511 vput(vp); 2512 WORKITEM_FREE(dirrem, D_DIRREM); 2513 return; 2514 } 2515 /* 2516 * Directory deletion. Decrement reference count for both the 2517 * just deleted parent directory entry and the reference for ".". 2518 * Next truncate the directory to length zero. When the 2519 * truncation completes, arrange to have the reference count on 2520 * the parent decremented to account for the loss of "..". 2521 */ 2522 ip->i_nlink -= 2; 2523 if (ip->i_nlink < ip->i_effnlink) 2524 panic("handle_workitem_remove: bad dir delta"); 2525 ip->i_flag |= IN_CHANGE; 2526 if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, p->p_ucred, p)) != 0) 2527 softdep_error("handle_workitem_remove: truncate", error); 2528 ACQUIRE_LOCK(&lk); 2529 (void) inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, DEPALLOC, 2530 &inodedep); 2531 dirrem->dm_state = 0; 2532 dirrem->dm_oldinum = dirrem->dm_dirinum; 2533 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 2534 FREE_LOCK(&lk); 2535 vput(vp); 2536} 2537 2538/* 2539 * Inode de-allocation dependencies. 2540 * 2541 * When an inode's link count is reduced to zero, it can be de-allocated. We 2542 * found it convenient to postpone de-allocation until after the inode is 2543 * written to disk with its new link count (zero). At this point, all of the 2544 * on-disk inode's block pointers are nullified and, with careful dependency 2545 * list ordering, all dependencies related to the inode will be satisfied and 2546 * the corresponding dependency structures de-allocated. So, if/when the 2547 * inode is reused, there will be no mixing of old dependencies with new 2548 * ones. This artificial dependency is set up by the block de-allocation 2549 * procedure above (softdep_setup_freeblocks) and completed by the 2550 * following procedure. 2551 */ 2552static void 2553handle_workitem_freefile(freefile) 2554 struct freefile *freefile; 2555{ 2556 struct vnode vp; 2557 struct inode tip; 2558 struct inodedep *idp; 2559 int error; 2560 2561#ifdef DEBUG 2562 ACQUIRE_LOCK(&lk); 2563 if (inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp)) 2564 panic("handle_workitem_freefile: inodedep survived"); 2565 FREE_LOCK(&lk); 2566#endif 2567 tip.i_devvp = freefile->fx_devvp; 2568 tip.i_dev = freefile->fx_devvp->v_rdev; 2569 tip.i_fs = freefile->fx_fs; 2570 vp.v_data = &tip; 2571 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0) 2572 softdep_error("handle_workitem_freefile", error); 2573 WORKITEM_FREE(freefile, D_FREEFILE); 2574 num_freefile -= 1; 2575} 2576 2577/* 2578 * Disk writes. 2579 * 2580 * The dependency structures constructed above are most actively used when file 2581 * system blocks are written to disk. No constraints are placed on when a 2582 * block can be written, but unsatisfied update dependencies are made safe by 2583 * modifying (or replacing) the source memory for the duration of the disk 2584 * write. When the disk write completes, the memory block is again brought 2585 * up-to-date. 2586 * 2587 * In-core inode structure reclamation. 2588 * 2589 * Because there are a finite number of "in-core" inode structures, they are 2590 * reused regularly. By transferring all inode-related dependencies to the 2591 * in-memory inode block and indexing them separately (via "inodedep"s), we 2592 * can allow "in-core" inode structures to be reused at any time and avoid 2593 * any increase in contention. 2594 * 2595 * Called just before entering the device driver to initiate a new disk I/O. 2596 * The buffer must be locked, thus, no I/O completion operations can occur 2597 * while we are manipulating its associated dependencies. 2598 */ 2599void 2600softdep_disk_io_initiation(bp) 2601 struct buf *bp; /* structure describing disk write to occur */ 2602{ 2603 struct worklist *wk, *nextwk; 2604 struct indirdep *indirdep; 2605 2606 /* 2607 * We only care about write operations. There should never 2608 * be dependencies for reads. 2609 */ 2610 if (bp->b_flags & B_READ) 2611 panic("softdep_disk_io_initiation: read"); 2612 /* 2613 * Do any necessary pre-I/O processing. 2614 */ 2615 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = nextwk) { 2616 nextwk = LIST_NEXT(wk, wk_list); 2617 switch (wk->wk_type) { 2618 2619 case D_PAGEDEP: 2620 initiate_write_filepage(WK_PAGEDEP(wk), bp); 2621 continue; 2622 2623 case D_INODEDEP: 2624 initiate_write_inodeblock(WK_INODEDEP(wk), bp); 2625 continue; 2626 2627 case D_INDIRDEP: 2628 indirdep = WK_INDIRDEP(wk); 2629 if (indirdep->ir_state & GOINGAWAY) 2630 panic("disk_io_initiation: indirdep gone"); 2631 /* 2632 * If there are no remaining dependencies, this 2633 * will be writing the real pointers, so the 2634 * dependency can be freed. 2635 */ 2636 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) { 2637 indirdep->ir_savebp->b_flags &= ~B_XXX; 2638 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 2639 brelse(indirdep->ir_savebp); 2640 /* inline expand WORKLIST_REMOVE(wk); */ 2641 wk->wk_state &= ~ONWORKLIST; 2642 LIST_REMOVE(wk, wk_list); 2643 WORKITEM_FREE(indirdep, D_INDIRDEP); 2644 continue; 2645 } 2646 /* 2647 * Replace up-to-date version with safe version. 2648 */ 2649 ACQUIRE_LOCK(&lk); 2650 indirdep->ir_state &= ~ATTACHED; 2651 indirdep->ir_state |= UNDONE; 2652 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount, 2653 M_INDIRDEP, M_WAITOK); 2654 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 2655 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 2656 bp->b_bcount); 2657 FREE_LOCK(&lk); 2658 continue; 2659 2660 case D_MKDIR: 2661 case D_BMSAFEMAP: 2662 case D_ALLOCDIRECT: 2663 case D_ALLOCINDIR: 2664 continue; 2665 2666 default: 2667 panic("handle_disk_io_initiation: Unexpected type %s", 2668 TYPENAME(wk->wk_type)); 2669 /* NOTREACHED */ 2670 } 2671 } 2672} 2673 2674/* 2675 * Called from within the procedure above to deal with unsatisfied 2676 * allocation dependencies in a directory. The buffer must be locked, 2677 * thus, no I/O completion operations can occur while we are 2678 * manipulating its associated dependencies. 2679 */ 2680static void 2681initiate_write_filepage(pagedep, bp) 2682 struct pagedep *pagedep; 2683 struct buf *bp; 2684{ 2685 struct diradd *dap; 2686 struct direct *ep; 2687 int i; 2688 2689 if (pagedep->pd_state & IOSTARTED) { 2690 /* 2691 * This can only happen if there is a driver that does not 2692 * understand chaining. Here biodone will reissue the call 2693 * to strategy for the incomplete buffers. 2694 */ 2695 printf("initiate_write_filepage: already started\n"); 2696 return; 2697 } 2698 pagedep->pd_state |= IOSTARTED; 2699 ACQUIRE_LOCK(&lk); 2700 for (i = 0; i < DAHASHSZ; i++) { 2701 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 2702 dap = LIST_NEXT(dap, da_pdlist)) { 2703 ep = (struct direct *) 2704 ((char *)bp->b_data + dap->da_offset); 2705 if (ep->d_ino != dap->da_newinum) 2706 panic("%s: dir inum %d != new %d", 2707 "initiate_write_filepage", 2708 ep->d_ino, dap->da_newinum); 2709 if (dap->da_state & DIRCHG) 2710 ep->d_ino = dap->da_previous->dm_oldinum; 2711 else 2712 ep->d_ino = 0; 2713 dap->da_state &= ~ATTACHED; 2714 dap->da_state |= UNDONE; 2715 } 2716 } 2717 FREE_LOCK(&lk); 2718} 2719 2720/* 2721 * Called from within the procedure above to deal with unsatisfied 2722 * allocation dependencies in an inodeblock. The buffer must be 2723 * locked, thus, no I/O completion operations can occur while we 2724 * are manipulating its associated dependencies. 2725 */ 2726static void 2727initiate_write_inodeblock(inodedep, bp) 2728 struct inodedep *inodedep; 2729 struct buf *bp; /* The inode block */ 2730{ 2731 struct allocdirect *adp, *lastadp; 2732 struct dinode *dp; 2733 struct fs *fs; 2734 ufs_lbn_t prevlbn = 0; 2735 int i, deplist; 2736 2737 if (inodedep->id_state & IOSTARTED) 2738 panic("initiate_write_inodeblock: already started"); 2739 inodedep->id_state |= IOSTARTED; 2740 fs = inodedep->id_fs; 2741 dp = (struct dinode *)bp->b_data + 2742 ino_to_fsbo(fs, inodedep->id_ino); 2743 /* 2744 * If the bitmap is not yet written, then the allocated 2745 * inode cannot be written to disk. 2746 */ 2747 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 2748 if (inodedep->id_savedino != NULL) 2749 panic("initiate_write_inodeblock: already doing I/O"); 2750 MALLOC(inodedep->id_savedino, struct dinode *, 2751 sizeof(struct dinode), M_INODEDEP, M_WAITOK); 2752 *inodedep->id_savedino = *dp; 2753 bzero((caddr_t)dp, sizeof(struct dinode)); 2754 return; 2755 } 2756 /* 2757 * If no dependencies, then there is nothing to roll back. 2758 */ 2759 inodedep->id_savedsize = dp->di_size; 2760 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 2761 return; 2762 /* 2763 * Set the dependencies to busy. 2764 */ 2765 ACQUIRE_LOCK(&lk); 2766 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 2767 adp = TAILQ_NEXT(adp, ad_next)) { 2768#ifdef DIAGNOSTIC 2769 if (deplist != 0 && prevlbn >= adp->ad_lbn) 2770 panic("softdep_write_inodeblock: lbn order"); 2771 prevlbn = adp->ad_lbn; 2772 if (adp->ad_lbn < NDADDR && 2773 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) 2774 panic("%s: direct pointer #%d mismatch %d != %d", 2775 "softdep_write_inodeblock", adp->ad_lbn, 2776 dp->di_db[adp->ad_lbn], adp->ad_newblkno); 2777 if (adp->ad_lbn >= NDADDR && 2778 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) 2779 panic("%s: indirect pointer #%d mismatch %d != %d", 2780 "softdep_write_inodeblock", adp->ad_lbn - NDADDR, 2781 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno); 2782 deplist |= 1 << adp->ad_lbn; 2783 if ((adp->ad_state & ATTACHED) == 0) 2784 panic("softdep_write_inodeblock: Unknown state 0x%x", 2785 adp->ad_state); 2786#endif /* DIAGNOSTIC */ 2787 adp->ad_state &= ~ATTACHED; 2788 adp->ad_state |= UNDONE; 2789 } 2790 /* 2791 * The on-disk inode cannot claim to be any larger than the last 2792 * fragment that has been written. Otherwise, the on-disk inode 2793 * might have fragments that were not the last block in the file 2794 * which would corrupt the filesystem. 2795 */ 2796 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 2797 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 2798 if (adp->ad_lbn >= NDADDR) 2799 break; 2800 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 2801 /* keep going until hitting a rollback to a frag */ 2802 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 2803 continue; 2804 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 2805 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 2806#ifdef DIAGNOSTIC 2807 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 2808 panic("softdep_write_inodeblock: lost dep1"); 2809#endif /* DIAGNOSTIC */ 2810 dp->di_db[i] = 0; 2811 } 2812 for (i = 0; i < NIADDR; i++) { 2813#ifdef DIAGNOSTIC 2814 if (dp->di_ib[i] != 0 && 2815 (deplist & ((1 << NDADDR) << i)) == 0) 2816 panic("softdep_write_inodeblock: lost dep2"); 2817#endif /* DIAGNOSTIC */ 2818 dp->di_ib[i] = 0; 2819 } 2820 FREE_LOCK(&lk); 2821 return; 2822 } 2823 /* 2824 * If we have zero'ed out the last allocated block of the file, 2825 * roll back the size to the last currently allocated block. 2826 * We know that this last allocated block is a full-sized as 2827 * we already checked for fragments in the loop above. 2828 */ 2829 if (lastadp != NULL && 2830 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 2831 for (i = lastadp->ad_lbn; i >= 0; i--) 2832 if (dp->di_db[i] != 0) 2833 break; 2834 dp->di_size = (i + 1) * fs->fs_bsize; 2835 } 2836 /* 2837 * The only dependencies are for indirect blocks. 2838 * 2839 * The file size for indirect block additions is not guaranteed. 2840 * Such a guarantee would be non-trivial to achieve. The conventional 2841 * synchronous write implementation also does not make this guarantee. 2842 * Fsck should catch and fix discrepancies. Arguably, the file size 2843 * can be over-estimated without destroying integrity when the file 2844 * moves into the indirect blocks (i.e., is large). If we want to 2845 * postpone fsck, we are stuck with this argument. 2846 */ 2847 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 2848 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 2849 FREE_LOCK(&lk); 2850} 2851 2852/* 2853 * This routine is called during the completion interrupt 2854 * service routine for a disk write (from the procedure called 2855 * by the device driver to inform the file system caches of 2856 * a request completion). It should be called early in this 2857 * procedure, before the block is made available to other 2858 * processes or other routines are called. 2859 */ 2860void 2861softdep_disk_write_complete(bp) 2862 struct buf *bp; /* describes the completed disk write */ 2863{ 2864 struct worklist *wk; 2865 struct workhead reattach; 2866 struct newblk *newblk; 2867 struct allocindir *aip; 2868 struct allocdirect *adp; 2869 struct indirdep *indirdep; 2870 struct inodedep *inodedep; 2871 struct bmsafemap *bmsafemap; 2872 2873#ifdef DEBUG 2874 if (lk.lkt_held != -1) 2875 panic("softdep_disk_write_complete: lock is held"); 2876 lk.lkt_held = -2; 2877#endif 2878 LIST_INIT(&reattach); 2879 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2880 WORKLIST_REMOVE(wk); 2881 switch (wk->wk_type) { 2882 2883 case D_PAGEDEP: 2884 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 2885 WORKLIST_INSERT(&reattach, wk); 2886 continue; 2887 2888 case D_INODEDEP: 2889 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 2890 WORKLIST_INSERT(&reattach, wk); 2891 continue; 2892 2893 case D_BMSAFEMAP: 2894 bmsafemap = WK_BMSAFEMAP(wk); 2895 while (newblk = LIST_FIRST(&bmsafemap->sm_newblkhd)) { 2896 newblk->nb_state |= DEPCOMPLETE; 2897 newblk->nb_bmsafemap = NULL; 2898 LIST_REMOVE(newblk, nb_deps); 2899 } 2900 while (adp = LIST_FIRST(&bmsafemap->sm_allocdirecthd)) { 2901 adp->ad_state |= DEPCOMPLETE; 2902 adp->ad_buf = NULL; 2903 LIST_REMOVE(adp, ad_deps); 2904 handle_allocdirect_partdone(adp); 2905 } 2906 while (aip = LIST_FIRST(&bmsafemap->sm_allocindirhd)) { 2907 aip->ai_state |= DEPCOMPLETE; 2908 aip->ai_buf = NULL; 2909 LIST_REMOVE(aip, ai_deps); 2910 handle_allocindir_partdone(aip); 2911 } 2912 while ((inodedep = 2913 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { 2914 inodedep->id_state |= DEPCOMPLETE; 2915 LIST_REMOVE(inodedep, id_deps); 2916 inodedep->id_buf = NULL; 2917 } 2918 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 2919 continue; 2920 2921 case D_MKDIR: 2922 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 2923 continue; 2924 2925 case D_ALLOCDIRECT: 2926 adp = WK_ALLOCDIRECT(wk); 2927 adp->ad_state |= COMPLETE; 2928 handle_allocdirect_partdone(adp); 2929 continue; 2930 2931 case D_ALLOCINDIR: 2932 aip = WK_ALLOCINDIR(wk); 2933 aip->ai_state |= COMPLETE; 2934 handle_allocindir_partdone(aip); 2935 continue; 2936 2937 case D_INDIRDEP: 2938 indirdep = WK_INDIRDEP(wk); 2939 if (indirdep->ir_state & GOINGAWAY) 2940 panic("disk_write_complete: indirdep gone"); 2941 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 2942 FREE(indirdep->ir_saveddata, M_INDIRDEP); 2943 indirdep->ir_saveddata = 0; 2944 indirdep->ir_state &= ~UNDONE; 2945 indirdep->ir_state |= ATTACHED; 2946 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 2947 LIST_REMOVE(aip, ai_next); 2948 handle_allocindir_partdone(aip); 2949 } 2950 WORKLIST_INSERT(&reattach, wk); 2951 bdirty(bp); 2952 continue; 2953 2954 default: 2955 panic("handle_disk_write_complete: Unknown type %s", 2956 TYPENAME(wk->wk_type)); 2957 /* NOTREACHED */ 2958 } 2959 } 2960 /* 2961 * Reattach any requests that must be redone. 2962 */ 2963 while ((wk = LIST_FIRST(&reattach)) != NULL) { 2964 WORKLIST_REMOVE(wk); 2965 WORKLIST_INSERT(&bp->b_dep, wk); 2966 } 2967#ifdef DEBUG 2968 if (lk.lkt_held != -2) 2969 panic("softdep_disk_write_complete: lock lost"); 2970 lk.lkt_held = -1; 2971#endif 2972} 2973 2974/* 2975 * Called from within softdep_disk_write_complete above. Note that 2976 * this routine is always called from interrupt level with further 2977 * splbio interrupts blocked. 2978 */ 2979static void 2980handle_allocdirect_partdone(adp) 2981 struct allocdirect *adp; /* the completed allocdirect */ 2982{ 2983 struct allocdirect *listadp; 2984 struct inodedep *inodedep; 2985 long bsize; 2986 2987 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 2988 return; 2989 if (adp->ad_buf != NULL) 2990 panic("handle_allocdirect_partdone: dangling dep"); 2991 /* 2992 * The on-disk inode cannot claim to be any larger than the last 2993 * fragment that has been written. Otherwise, the on-disk inode 2994 * might have fragments that were not the last block in the file 2995 * which would corrupt the filesystem. Thus, we cannot free any 2996 * allocdirects after one whose ad_oldblkno claims a fragment as 2997 * these blocks must be rolled back to zero before writing the inode. 2998 * We check the currently active set of allocdirects in id_inoupdt. 2999 */ 3000 inodedep = adp->ad_inodedep; 3001 bsize = inodedep->id_fs->fs_bsize; 3002 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp; 3003 listadp = TAILQ_NEXT(listadp, ad_next)) { 3004 /* found our block */ 3005 if (listadp == adp) 3006 break; 3007 /* continue if ad_oldlbn is not a fragment */ 3008 if (listadp->ad_oldsize == 0 || 3009 listadp->ad_oldsize == bsize) 3010 continue; 3011 /* hit a fragment */ 3012 return; 3013 } 3014 /* 3015 * If we have reached the end of the current list without 3016 * finding the just finished dependency, then it must be 3017 * on the future dependency list. Future dependencies cannot 3018 * be freed until they are moved to the current list. 3019 */ 3020 if (listadp == NULL) { 3021#ifdef DEBUG 3022 for (listadp = TAILQ_FIRST(&inodedep->id_newinoupdt); listadp; 3023 listadp = TAILQ_NEXT(listadp, ad_next)) 3024 /* found our block */ 3025 if (listadp == adp) 3026 break; 3027 if (listadp == NULL) 3028 panic("handle_allocdirect_partdone: lost dep"); 3029#endif /* DEBUG */ 3030 return; 3031 } 3032 /* 3033 * If we have found the just finished dependency, then free 3034 * it along with anything that follows it that is complete. 3035 */ 3036 for (; adp; adp = listadp) { 3037 listadp = TAILQ_NEXT(adp, ad_next); 3038 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3039 return; 3040 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 3041 } 3042} 3043 3044/* 3045 * Called from within softdep_disk_write_complete above. Note that 3046 * this routine is always called from interrupt level with further 3047 * splbio interrupts blocked. 3048 */ 3049static void 3050handle_allocindir_partdone(aip) 3051 struct allocindir *aip; /* the completed allocindir */ 3052{ 3053 struct indirdep *indirdep; 3054 3055 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 3056 return; 3057 if (aip->ai_buf != NULL) 3058 panic("handle_allocindir_partdone: dangling dependency"); 3059 indirdep = aip->ai_indirdep; 3060 if (indirdep->ir_state & UNDONE) { 3061 LIST_REMOVE(aip, ai_next); 3062 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 3063 return; 3064 } 3065 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 3066 aip->ai_newblkno; 3067 LIST_REMOVE(aip, ai_next); 3068 if (aip->ai_freefrag != NULL) 3069 add_to_worklist(&aip->ai_freefrag->ff_list); 3070 WORKITEM_FREE(aip, D_ALLOCINDIR); 3071} 3072 3073/* 3074 * Called from within softdep_disk_write_complete above to restore 3075 * in-memory inode block contents to their most up-to-date state. Note 3076 * that this routine is always called from interrupt level with further 3077 * splbio interrupts blocked. 3078 */ 3079static int 3080handle_written_inodeblock(inodedep, bp) 3081 struct inodedep *inodedep; 3082 struct buf *bp; /* buffer containing the inode block */ 3083{ 3084 struct pagedep *pagedep; 3085 struct worklist *wk, *filefree; 3086 struct allocdirect *adp, *nextadp; 3087 struct dinode *dp; 3088 struct diradd *dap; 3089 int hadchanges; 3090 3091 if ((inodedep->id_state & IOSTARTED) == 0) 3092 panic("handle_written_inodeblock: not started"); 3093 inodedep->id_state &= ~IOSTARTED; 3094 inodedep->id_state |= COMPLETE; 3095 dp = (struct dinode *)bp->b_data + 3096 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 3097 /* 3098 * If we had to rollback the inode allocation because of 3099 * bitmaps being incomplete, then simply restore it. 3100 * Keep the block dirty so that it will not be reclaimed until 3101 * all associated dependencies have been cleared and the 3102 * corresponding updates written to disk. 3103 */ 3104 if (inodedep->id_savedino != NULL) { 3105 *dp = *inodedep->id_savedino; 3106 FREE(inodedep->id_savedino, M_INODEDEP); 3107 inodedep->id_savedino = NULL; 3108 bdirty(bp); 3109 return (1); 3110 } 3111 /* 3112 * Roll forward anything that had to be rolled back before 3113 * the inode could be updated. 3114 */ 3115 hadchanges = 0; 3116 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 3117 nextadp = TAILQ_NEXT(adp, ad_next); 3118 if (adp->ad_state & ATTACHED) 3119 panic("handle_written_inodeblock: new entry"); 3120 if (adp->ad_lbn < NDADDR) { 3121 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) 3122 panic("%s: %s #%d mismatch %d != %d", 3123 "handle_written_inodeblock", 3124 "direct pointer", adp->ad_lbn, 3125 dp->di_db[adp->ad_lbn], adp->ad_oldblkno); 3126 dp->di_db[adp->ad_lbn] = adp->ad_newblkno; 3127 } else { 3128 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) 3129 panic("%s: %s #%d allocated as %d", 3130 "handle_written_inodeblock", 3131 "indirect pointer", adp->ad_lbn - NDADDR, 3132 dp->di_ib[adp->ad_lbn - NDADDR]); 3133 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno; 3134 } 3135 adp->ad_state &= ~UNDONE; 3136 adp->ad_state |= ATTACHED; 3137 hadchanges = 1; 3138 } 3139 /* 3140 * Reset the file size to its most up-to-date value. 3141 */ 3142 if (inodedep->id_savedsize == -1) 3143 panic("handle_written_inodeblock: bad size"); 3144 if (dp->di_size != inodedep->id_savedsize) { 3145 dp->di_size = inodedep->id_savedsize; 3146 hadchanges = 1; 3147 } 3148 inodedep->id_savedsize = -1; 3149 /* 3150 * If there were any rollbacks in the inode block, then it must be 3151 * marked dirty so that its will eventually get written back in 3152 * its correct form. 3153 */ 3154 if (hadchanges) 3155 bdirty(bp); 3156 /* 3157 * Process any allocdirects that completed during the update. 3158 */ 3159 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 3160 handle_allocdirect_partdone(adp); 3161 /* 3162 * Process deallocations that were held pending until the 3163 * inode had been written to disk. Freeing of the inode 3164 * is delayed until after all blocks have been freed to 3165 * avoid creation of new <vfsid, inum, lbn> triples 3166 * before the old ones have been deleted. 3167 */ 3168 filefree = NULL; 3169 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 3170 WORKLIST_REMOVE(wk); 3171 switch (wk->wk_type) { 3172 3173 case D_FREEFILE: 3174 /* 3175 * We defer adding filefree to the worklist until 3176 * all other additions have been made to ensure 3177 * that it will be done after all the old blocks 3178 * have been freed. 3179 */ 3180 if (filefree != NULL) 3181 panic("handle_written_inodeblock: filefree"); 3182 filefree = wk; 3183 continue; 3184 3185 case D_MKDIR: 3186 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 3187 continue; 3188 3189 case D_DIRADD: 3190 diradd_inode_written(WK_DIRADD(wk), inodedep); 3191 continue; 3192 3193 case D_FREEBLKS: 3194 case D_FREEFRAG: 3195 case D_DIRREM: 3196 add_to_worklist(wk); 3197 continue; 3198 3199 default: 3200 panic("handle_written_inodeblock: Unknown type %s", 3201 TYPENAME(wk->wk_type)); 3202 /* NOTREACHED */ 3203 } 3204 } 3205 if (filefree != NULL) { 3206 if (free_inodedep(inodedep) == 0) 3207 panic("handle_written_inodeblock: live inodedep"); 3208 add_to_worklist(filefree); 3209 return (0); 3210 } 3211 3212 /* 3213 * If no outstanding dependencies, free it. 3214 */ 3215 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0) 3216 return (0); 3217 return (hadchanges); 3218} 3219 3220/* 3221 * Process a diradd entry after its dependent inode has been written. 3222 * This routine must be called with splbio interrupts blocked. 3223 */ 3224static void 3225diradd_inode_written(dap, inodedep) 3226 struct diradd *dap; 3227 struct inodedep *inodedep; 3228{ 3229 struct pagedep *pagedep; 3230 3231 dap->da_state |= COMPLETE; 3232 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3233 if (dap->da_state & DIRCHG) 3234 pagedep = dap->da_previous->dm_pagedep; 3235 else 3236 pagedep = dap->da_pagedep; 3237 LIST_REMOVE(dap, da_pdlist); 3238 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3239 } 3240 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 3241} 3242 3243/* 3244 * Handle the completion of a mkdir dependency. 3245 */ 3246static void 3247handle_written_mkdir(mkdir, type) 3248 struct mkdir *mkdir; 3249 int type; 3250{ 3251 struct diradd *dap; 3252 struct pagedep *pagedep; 3253 3254 if (mkdir->md_state != type) 3255 panic("handle_written_mkdir: bad type"); 3256 dap = mkdir->md_diradd; 3257 dap->da_state &= ~type; 3258 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 3259 dap->da_state |= DEPCOMPLETE; 3260 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3261 if (dap->da_state & DIRCHG) 3262 pagedep = dap->da_previous->dm_pagedep; 3263 else 3264 pagedep = dap->da_pagedep; 3265 LIST_REMOVE(dap, da_pdlist); 3266 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3267 } 3268 LIST_REMOVE(mkdir, md_mkdirs); 3269 WORKITEM_FREE(mkdir, D_MKDIR); 3270} 3271 3272/* 3273 * Called from within softdep_disk_write_complete above. 3274 * A write operation was just completed. Removed inodes can 3275 * now be freed and associated block pointers may be committed. 3276 * Note that this routine is always called from interrupt level 3277 * with further splbio interrupts blocked. 3278 */ 3279static int 3280handle_written_filepage(pagedep, bp) 3281 struct pagedep *pagedep; 3282 struct buf *bp; /* buffer containing the written page */ 3283{ 3284 struct dirrem *dirrem; 3285 struct diradd *dap, *nextdap; 3286 struct direct *ep; 3287 int i, chgs; 3288 3289 if ((pagedep->pd_state & IOSTARTED) == 0) 3290 panic("handle_written_filepage: not started"); 3291 pagedep->pd_state &= ~IOSTARTED; 3292 /* 3293 * Process any directory removals that have been committed. 3294 */ 3295 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 3296 LIST_REMOVE(dirrem, dm_next); 3297 dirrem->dm_dirinum = pagedep->pd_ino; 3298 add_to_worklist(&dirrem->dm_list); 3299 } 3300 /* 3301 * Free any directory additions that have been committed. 3302 */ 3303 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 3304 free_diradd(dap); 3305 /* 3306 * Uncommitted directory entries must be restored. 3307 */ 3308 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 3309 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 3310 dap = nextdap) { 3311 nextdap = LIST_NEXT(dap, da_pdlist); 3312 if (dap->da_state & ATTACHED) 3313 panic("handle_written_filepage: attached"); 3314 ep = (struct direct *) 3315 ((char *)bp->b_data + dap->da_offset); 3316 ep->d_ino = dap->da_newinum; 3317 dap->da_state &= ~UNDONE; 3318 dap->da_state |= ATTACHED; 3319 chgs = 1; 3320 /* 3321 * If the inode referenced by the directory has 3322 * been written out, then the dependency can be 3323 * moved to the pending list. 3324 */ 3325 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3326 LIST_REMOVE(dap, da_pdlist); 3327 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 3328 da_pdlist); 3329 } 3330 } 3331 } 3332 /* 3333 * If there were any rollbacks in the directory, then it must be 3334 * marked dirty so that its will eventually get written back in 3335 * its correct form. 3336 */ 3337 if (chgs) 3338 bdirty(bp); 3339 /* 3340 * If no dependencies remain, the pagedep will be freed. 3341 * Otherwise it will remain to update the page before it 3342 * is written back to disk. 3343 */ 3344 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) { 3345 for (i = 0; i < DAHASHSZ; i++) 3346 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL) 3347 break; 3348 if (i == DAHASHSZ) { 3349 LIST_REMOVE(pagedep, pd_hash); 3350 WORKITEM_FREE(pagedep, D_PAGEDEP); 3351 return (0); 3352 } 3353 } 3354 return (1); 3355} 3356 3357/* 3358 * Writing back in-core inode structures. 3359 * 3360 * The file system only accesses an inode's contents when it occupies an 3361 * "in-core" inode structure. These "in-core" structures are separate from 3362 * the page frames used to cache inode blocks. Only the latter are 3363 * transferred to/from the disk. So, when the updated contents of the 3364 * "in-core" inode structure are copied to the corresponding in-memory inode 3365 * block, the dependencies are also transferred. The following procedure is 3366 * called when copying a dirty "in-core" inode to a cached inode block. 3367 */ 3368 3369/* 3370 * Called when an inode is loaded from disk. If the effective link count 3371 * differed from the actual link count when it was last flushed, then we 3372 * need to ensure that the correct effective link count is put back. 3373 */ 3374void 3375softdep_load_inodeblock(ip) 3376 struct inode *ip; /* the "in_core" copy of the inode */ 3377{ 3378 struct inodedep *inodedep; 3379 int error, gotit; 3380 3381 /* 3382 * Check for alternate nlink count. 3383 */ 3384 ip->i_effnlink = ip->i_nlink; 3385 ACQUIRE_LOCK(&lk); 3386 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3387 FREE_LOCK(&lk); 3388 return; 3389 } 3390 if (inodedep->id_nlinkdelta != 0) { 3391 ip->i_effnlink -= inodedep->id_nlinkdelta; 3392 ip->i_flag |= IN_MODIFIED; 3393 inodedep->id_nlinkdelta = 0; 3394 (void) free_inodedep(inodedep); 3395 } 3396 FREE_LOCK(&lk); 3397} 3398 3399/* 3400 * This routine is called just before the "in-core" inode 3401 * information is to be copied to the in-memory inode block. 3402 * Recall that an inode block contains several inodes. If 3403 * the force flag is set, then the dependencies will be 3404 * cleared so that the update can always be made. Note that 3405 * the buffer is locked when this routine is called, so we 3406 * will never be in the middle of writing the inode block 3407 * to disk. 3408 */ 3409void 3410softdep_update_inodeblock(ip, bp, waitfor) 3411 struct inode *ip; /* the "in_core" copy of the inode */ 3412 struct buf *bp; /* the buffer containing the inode block */ 3413 int waitfor; /* 1 => update must be allowed */ 3414{ 3415 struct inodedep *inodedep; 3416 struct worklist *wk; 3417 int error, gotit; 3418 3419 /* 3420 * If the effective link count is not equal to the actual link 3421 * count, then we must track the difference in an inodedep while 3422 * the inode is (potentially) tossed out of the cache. Otherwise, 3423 * if there is no existing inodedep, then there are no dependencies 3424 * to track. 3425 */ 3426 ACQUIRE_LOCK(&lk); 3427 if (ip->i_effnlink != ip->i_nlink) { 3428 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, 3429 &inodedep); 3430 } else if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3431 FREE_LOCK(&lk); 3432 return; 3433 } 3434 if (ip->i_nlink < ip->i_effnlink) 3435 panic("softdep_update_inodeblock: bad delta"); 3436 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3437 /* 3438 * Changes have been initiated. Anything depending on these 3439 * changes cannot occur until this inode has been written. 3440 */ 3441 inodedep->id_state &= ~COMPLETE; 3442 if ((inodedep->id_state & ONWORKLIST) == 0) 3443 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 3444 /* 3445 * Any new dependencies associated with the incore inode must 3446 * now be moved to the list associated with the buffer holding 3447 * the in-memory copy of the inode. Once merged process any 3448 * allocdirects that are completed by the merger. 3449 */ 3450 merge_inode_lists(inodedep); 3451 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL) 3452 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); 3453 /* 3454 * Now that the inode has been pushed into the buffer, the 3455 * operations dependent on the inode being written to disk 3456 * can be moved to the id_bufwait so that they will be 3457 * processed when the buffer I/O completes. 3458 */ 3459 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 3460 WORKLIST_REMOVE(wk); 3461 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 3462 } 3463 /* 3464 * Newly allocated inodes cannot be written until the bitmap 3465 * that allocates them have been written (indicated by 3466 * DEPCOMPLETE being set in id_state). If we are doing a 3467 * forced sync (e.g., an fsync on a file), we force the bitmap 3468 * to be written so that the update can be done. 3469 */ 3470 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) { 3471 FREE_LOCK(&lk); 3472 return; 3473 } 3474 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 3475 FREE_LOCK(&lk); 3476 if (gotit && (error = VOP_BWRITE(inodedep->id_buf)) != 0) 3477 softdep_error("softdep_update_inodeblock: bwrite", error); 3478 if ((inodedep->id_state & DEPCOMPLETE) == 0) 3479 panic("softdep_update_inodeblock: update failed"); 3480} 3481 3482/* 3483 * Merge the new inode dependency list (id_newinoupdt) into the old 3484 * inode dependency list (id_inoupdt). This routine must be called 3485 * with splbio interrupts blocked. 3486 */ 3487static void 3488merge_inode_lists(inodedep) 3489 struct inodedep *inodedep; 3490{ 3491 struct allocdirect *listadp, *newadp; 3492 3493 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 3494 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) { 3495 if (listadp->ad_lbn < newadp->ad_lbn) { 3496 listadp = TAILQ_NEXT(listadp, ad_next); 3497 continue; 3498 } 3499 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 3500 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 3501 if (listadp->ad_lbn == newadp->ad_lbn) { 3502 allocdirect_merge(&inodedep->id_inoupdt, newadp, 3503 listadp); 3504 listadp = newadp; 3505 } 3506 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 3507 } 3508 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) { 3509 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 3510 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next); 3511 } 3512} 3513 3514/* 3515 * If we are doing an fsync, then we must ensure that any directory 3516 * entries for the inode have been written after the inode gets to disk. 3517 */ 3518int 3519softdep_fsync(vp) 3520 struct vnode *vp; /* the "in_core" copy of the inode */ 3521{ 3522 struct diradd *dap, *olddap; 3523 struct inodedep *inodedep; 3524 struct pagedep *pagedep; 3525 struct worklist *wk; 3526 struct mount *mnt; 3527 struct vnode *pvp; 3528 struct inode *ip; 3529 struct buf *bp; 3530 struct fs *fs; 3531 struct proc *p = CURPROC; /* XXX */ 3532 int error, ret, flushparent; 3533 struct timeval tv; 3534 ino_t parentino; 3535 ufs_lbn_t lbn; 3536 3537 ip = VTOI(vp); 3538 fs = ip->i_fs; 3539 for (error = 0, flushparent = 0, olddap = NULL; ; ) { 3540 ACQUIRE_LOCK(&lk); 3541 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) 3542 break; 3543 if (LIST_FIRST(&inodedep->id_inowait) != NULL || 3544 LIST_FIRST(&inodedep->id_bufwait) != NULL || 3545 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 3546 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) 3547 panic("softdep_fsync: pending ops"); 3548 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 3549 break; 3550 if (wk->wk_type != D_DIRADD) 3551 panic("softdep_fsync: Unexpected type %s", 3552 TYPENAME(wk->wk_type)); 3553 dap = WK_DIRADD(wk); 3554 /* 3555 * If we have failed to get rid of all the dependencies 3556 * then something is seriously wrong. 3557 */ 3558 if (dap == olddap) 3559 panic("softdep_fsync: flush failed"); 3560 olddap = dap; 3561 /* 3562 * Flush our parent if this directory entry 3563 * has a MKDIR_PARENT dependency. 3564 */ 3565 if (dap->da_state & DIRCHG) 3566 pagedep = dap->da_previous->dm_pagedep; 3567 else 3568 pagedep = dap->da_pagedep; 3569 mnt = pagedep->pd_mnt; 3570 parentino = pagedep->pd_ino; 3571 lbn = pagedep->pd_lbn; 3572 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) 3573 panic("softdep_fsync: dirty"); 3574 flushparent = dap->da_state & MKDIR_PARENT; 3575 /* 3576 * If we are being fsync'ed as part of vgone'ing this vnode, 3577 * then we will not be able to release and recover the 3578 * vnode below, so we just have to give up on writing its 3579 * directory entry out. It will eventually be written, just 3580 * not now, but then the user was not asking to have it 3581 * written, so we are not breaking any promises. 3582 */ 3583 if (vp->v_flag & VXLOCK) 3584 break; 3585 /* 3586 * We prevent deadlock by always fetching inodes from the 3587 * root, moving down the directory tree. Thus, when fetching 3588 * our parent directory, we must unlock ourselves before 3589 * requesting the lock on our parent. See the comment in 3590 * ufs_lookup for details on possible races. 3591 */ 3592 FREE_LOCK(&lk); 3593 VOP_UNLOCK(vp, 0, p); 3594 if ((error = VFS_VGET(mnt, parentino, &pvp)) != 0) { 3595 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, p); 3596 return (error); 3597 } 3598 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, p); 3599 if (flushparent) { 3600#ifndef __FreeBSD__ 3601 tv = time; 3602#else 3603 getmicrotime(&tv); 3604#endif /* __FreeBSD__ */ 3605 if (error = UFS_UPDATE(pvp, &tv, &tv, MNT_WAIT)) { 3606 vput(pvp); 3607 return (error); 3608 } 3609 } 3610 /* 3611 * Flush directory page containing the inode's name. 3612 */ 3613 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), p->p_ucred, 3614 &bp); 3615 ret = VOP_BWRITE(bp); 3616 vput(pvp); 3617 if (error != 0) 3618 return (error); 3619 if (ret != 0) 3620 return (ret); 3621 } 3622 FREE_LOCK(&lk); 3623 return (0); 3624} 3625 3626/* 3627 * This routine is called when we are trying to synchronously flush a 3628 * file. This routine must eliminate any filesystem metadata dependencies 3629 * so that the syncing routine can succeed by pushing the dirty blocks 3630 * associated with the file. If any I/O errors occur, they are returned. 3631 */ 3632int 3633softdep_sync_metadata(ap) 3634 struct vop_fsync_args /* { 3635 struct vnode *a_vp; 3636 struct ucred *a_cred; 3637 int a_waitfor; 3638 struct proc *a_p; 3639 } */ *ap; 3640{ 3641 struct vnode *vp = ap->a_vp; 3642 struct pagedep *pagedep; 3643 struct allocdirect *adp; 3644 struct allocindir *aip; 3645 struct buf *bp, *nbp; 3646 struct worklist *wk; 3647 int i, error, waitfor; 3648 3649 /* 3650 * Check whether this vnode is involved in a filesystem 3651 * that is doing soft dependency processing. 3652 */ 3653 if (vp->v_type != VBLK) { 3654 if (!DOINGSOFTDEP(vp)) 3655 return (0); 3656 } else 3657 if (vp->v_specmountpoint == NULL || 3658 (vp->v_specmountpoint->mnt_flag & MNT_SOFTDEP) == 0) 3659 return (0); 3660 /* 3661 * Ensure that any direct block dependencies have been cleared. 3662 */ 3663 ACQUIRE_LOCK(&lk); 3664 if (error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number)) { 3665 FREE_LOCK(&lk); 3666 return (error); 3667 } 3668 /* 3669 * For most files, the only metadata dependencies are the 3670 * cylinder group maps that allocate their inode or blocks. 3671 * The block allocation dependencies can be found by traversing 3672 * the dependency lists for any buffers that remain on their 3673 * dirty buffer list. The inode allocation dependency will 3674 * be resolved when the inode is updated with MNT_WAIT. 3675 * This work is done in two passes. The first pass grabs most 3676 * of the buffers and begins asynchronously writing them. The 3677 * only way to wait for these asynchronous writes is to sleep 3678 * on the filesystem vnode which may stay busy for a long time 3679 * if the filesystem is active. So, instead, we make a second 3680 * pass over the dependencies blocking on each write. In the 3681 * usual case we will be blocking against a write that we 3682 * initiated, so when it is done the dependency will have been 3683 * resolved. Thus the second pass is expected to end quickly. 3684 */ 3685 waitfor = MNT_NOWAIT; 3686top: 3687 if (getdirtybuf(&LIST_FIRST(&vp->v_dirtyblkhd), MNT_WAIT) == 0) { 3688 while (vp->v_numoutput) { 3689 vp->v_flag |= VBWAIT; 3690 FREE_LOCK_INTERLOCKED(&lk); 3691 tsleep((caddr_t)&vp->v_numoutput, PRIBIO + 1, 3692 "sdsynm", 0); 3693 ACQUIRE_LOCK_INTERLOCKED(&lk); 3694 } 3695 FREE_LOCK(&lk); 3696 return (0); 3697 } 3698 bp = LIST_FIRST(&vp->v_dirtyblkhd); 3699loop: 3700 /* 3701 * As we hold the buffer locked, none of its dependencies 3702 * will disappear. 3703 */ 3704 for (wk = LIST_FIRST(&bp->b_dep); wk; 3705 wk = LIST_NEXT(wk, wk_list)) { 3706 switch (wk->wk_type) { 3707 3708 case D_ALLOCDIRECT: 3709 adp = WK_ALLOCDIRECT(wk); 3710 if (adp->ad_state & DEPCOMPLETE) 3711 break; 3712 nbp = adp->ad_buf; 3713 if (getdirtybuf(&nbp, waitfor) == 0) 3714 break; 3715 FREE_LOCK(&lk); 3716 if (waitfor == MNT_NOWAIT) { 3717 bawrite(nbp); 3718 } else if ((error = VOP_BWRITE(nbp)) != 0) { 3719 bawrite(bp); 3720 return (error); 3721 } 3722 ACQUIRE_LOCK(&lk); 3723 break; 3724 3725 case D_ALLOCINDIR: 3726 aip = WK_ALLOCINDIR(wk); 3727 if (aip->ai_state & DEPCOMPLETE) 3728 break; 3729 nbp = aip->ai_buf; 3730 if (getdirtybuf(&nbp, waitfor) == 0) 3731 break; 3732 FREE_LOCK(&lk); 3733 if (waitfor == MNT_NOWAIT) { 3734 bawrite(nbp); 3735 } else if ((error = VOP_BWRITE(nbp)) != 0) { 3736 bawrite(bp); 3737 return (error); 3738 } 3739 ACQUIRE_LOCK(&lk); 3740 break; 3741 3742 case D_INDIRDEP: 3743 restart: 3744 for (aip = LIST_FIRST(&WK_INDIRDEP(wk)->ir_deplisthd); 3745 aip; aip = LIST_NEXT(aip, ai_next)) { 3746 if (aip->ai_state & DEPCOMPLETE) 3747 continue; 3748 nbp = aip->ai_buf; 3749 if (getdirtybuf(&nbp, MNT_WAIT) == 0) 3750 goto restart; 3751 FREE_LOCK(&lk); 3752 if ((error = VOP_BWRITE(nbp)) != 0) { 3753 bawrite(bp); 3754 return (error); 3755 } 3756 ACQUIRE_LOCK(&lk); 3757 goto restart; 3758 } 3759 break; 3760 3761 case D_INODEDEP: 3762 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs, 3763 WK_INODEDEP(wk)->id_ino)) != 0) { 3764 FREE_LOCK(&lk); 3765 bawrite(bp); 3766 return (error); 3767 } 3768 break; 3769 3770 case D_PAGEDEP: 3771 /* 3772 * We are trying to sync a directory that may 3773 * have dependencies on both its own metadata 3774 * and/or dependencies on the inodes of any 3775 * recently allocated files. We walk its diradd 3776 * lists pushing out the associated inode. 3777 */ 3778 pagedep = WK_PAGEDEP(wk); 3779 for (i = 0; i < DAHASHSZ; i++) { 3780 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 3781 continue; 3782 if (error = flush_pagedep_deps(vp, 3783 pagedep->pd_mnt, &pagedep->pd_diraddhd[i])) { 3784 FREE_LOCK(&lk); 3785 bawrite(bp); 3786 return (error); 3787 } 3788 } 3789 break; 3790 3791 default: 3792 panic("softdep_sync_metadata: Unknown type %s", 3793 TYPENAME(wk->wk_type)); 3794 /* NOTREACHED */ 3795 } 3796 } 3797 (void) getdirtybuf(&LIST_NEXT(bp, b_vnbufs), MNT_WAIT); 3798 nbp = LIST_NEXT(bp, b_vnbufs); 3799 FREE_LOCK(&lk); 3800 bawrite(bp); 3801 ACQUIRE_LOCK(&lk); 3802 if (nbp != NULL) { 3803 bp = nbp; 3804 goto loop; 3805 } 3806 /* 3807 * We must wait for any I/O in progress to finish so that 3808 * all potential buffers on the dirty list will be visible. 3809 * Once they are all there, proceed with the second pass 3810 * which will wait for the I/O as per above. 3811 */ 3812 while (vp->v_numoutput) { 3813 vp->v_flag |= VBWAIT; 3814 FREE_LOCK_INTERLOCKED(&lk); 3815 tsleep((caddr_t)&vp->v_numoutput, PRIBIO + 1, "sdsynm", 0); 3816 ACQUIRE_LOCK_INTERLOCKED(&lk); 3817 } 3818 /* 3819 * The brief unlock is to allow any pent up dependency 3820 * processing to be done. 3821 */ 3822 if (waitfor == MNT_NOWAIT) { 3823 waitfor = MNT_WAIT; 3824 FREE_LOCK(&lk); 3825 ACQUIRE_LOCK(&lk); 3826 goto top; 3827 } 3828 3829 /* 3830 * If we have managed to get rid of all the dirty buffers, 3831 * then we are done. For certain directories and block 3832 * devices, we may need to do further work. 3833 */ 3834 if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL) { 3835 FREE_LOCK(&lk); 3836 return (0); 3837 } 3838 3839 FREE_LOCK(&lk); 3840 /* 3841 * If we are trying to sync a block device, some of its buffers may 3842 * contain metadata that cannot be written until the contents of some 3843 * partially written files have been written to disk. The only easy 3844 * way to accomplish this is to sync the entire filesystem (luckily 3845 * this happens rarely). 3846 */ 3847 if (vp->v_type == VBLK && vp->v_specmountpoint && !VOP_ISLOCKED(vp) && 3848 (error = VFS_SYNC(vp->v_specmountpoint, MNT_WAIT, ap->a_cred, 3849 ap->a_p)) != 0) 3850 return (error); 3851 return (0); 3852} 3853 3854/* 3855 * Flush the dependencies associated with an inodedep. 3856 * Called with splbio blocked. 3857 */ 3858static int 3859flush_inodedep_deps(fs, ino) 3860 struct fs *fs; 3861 ino_t ino; 3862{ 3863 struct inodedep *inodedep; 3864 struct allocdirect *adp; 3865 int error, waitfor; 3866 struct buf *bp; 3867 3868 /* 3869 * This work is done in two passes. The first pass grabs most 3870 * of the buffers and begins asynchronously writing them. The 3871 * only way to wait for these asynchronous writes is to sleep 3872 * on the filesystem vnode which may stay busy for a long time 3873 * if the filesystem is active. So, instead, we make a second 3874 * pass over the dependencies blocking on each write. In the 3875 * usual case we will be blocking against a write that we 3876 * initiated, so when it is done the dependency will have been 3877 * resolved. Thus the second pass is expected to end quickly. 3878 * We give a brief window at the top of the loop to allow 3879 * any pending I/O to complete. 3880 */ 3881 for (waitfor = MNT_NOWAIT; ; ) { 3882 FREE_LOCK(&lk); 3883 ACQUIRE_LOCK(&lk); 3884 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 3885 return (0); 3886 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3887 adp = TAILQ_NEXT(adp, ad_next)) { 3888 if (adp->ad_state & DEPCOMPLETE) 3889 continue; 3890 bp = adp->ad_buf; 3891 if (getdirtybuf(&bp, waitfor) == 0) { 3892 if (waitfor == MNT_NOWAIT) 3893 continue; 3894 break; 3895 } 3896 FREE_LOCK(&lk); 3897 if (waitfor == MNT_NOWAIT) { 3898 bawrite(bp); 3899 } else if ((error = VOP_BWRITE(bp)) != 0) { 3900 ACQUIRE_LOCK(&lk); 3901 return (error); 3902 } 3903 ACQUIRE_LOCK(&lk); 3904 break; 3905 } 3906 if (adp != NULL) 3907 continue; 3908 for (adp = TAILQ_FIRST(&inodedep->id_newinoupdt); adp; 3909 adp = TAILQ_NEXT(adp, ad_next)) { 3910 if (adp->ad_state & DEPCOMPLETE) 3911 continue; 3912 bp = adp->ad_buf; 3913 if (getdirtybuf(&bp, waitfor) == 0) { 3914 if (waitfor == MNT_NOWAIT) 3915 continue; 3916 break; 3917 } 3918 FREE_LOCK(&lk); 3919 if (waitfor == MNT_NOWAIT) { 3920 bawrite(bp); 3921 } else if ((error = VOP_BWRITE(bp)) != 0) { 3922 ACQUIRE_LOCK(&lk); 3923 return (error); 3924 } 3925 ACQUIRE_LOCK(&lk); 3926 break; 3927 } 3928 if (adp != NULL) 3929 continue; 3930 /* 3931 * If pass2, we are done, otherwise do pass 2. 3932 */ 3933 if (waitfor == MNT_WAIT) 3934 break; 3935 waitfor = MNT_WAIT; 3936 } 3937 /* 3938 * Try freeing inodedep in case all dependencies have been removed. 3939 */ 3940 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0) 3941 (void) free_inodedep(inodedep); 3942 return (0); 3943} 3944 3945/* 3946 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 3947 * Called with splbio blocked. 3948 */ 3949static int 3950flush_pagedep_deps(pvp, mp, diraddhdp) 3951 struct vnode *pvp; 3952 struct mount *mp; 3953 struct diraddhd *diraddhdp; 3954{ 3955 struct proc *p = CURPROC; /* XXX */ 3956 struct inodedep *inodedep; 3957 struct ufsmount *ump; 3958 struct diradd *dap; 3959 struct timeval tv; 3960 struct vnode *vp; 3961 int gotit, error = 0; 3962 struct buf *bp; 3963 ino_t inum; 3964 3965 ump = VFSTOUFS(mp); 3966 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 3967 /* 3968 * Flush ourselves if this directory entry 3969 * has a MKDIR_PARENT dependency. 3970 */ 3971 if (dap->da_state & MKDIR_PARENT) { 3972#ifndef __FreeBSD__ 3973 tv = time; 3974#else 3975 getmicrotime(&tv); 3976#endif /* __FreeBSD__ */ 3977 FREE_LOCK(&lk); 3978 if (error = UFS_UPDATE(pvp, &tv, &tv, MNT_WAIT)) 3979 break; 3980 ACQUIRE_LOCK(&lk); 3981 /* 3982 * If that cleared dependencies, go on to next. 3983 */ 3984 if (dap != LIST_FIRST(diraddhdp)) 3985 continue; 3986 if (dap->da_state & MKDIR_PARENT) 3987 panic("flush_pagedep_deps: MKDIR"); 3988 } 3989 /* 3990 * Flush the file on which the directory entry depends. 3991 * If the inode has already been pushed out of the cache, 3992 * then all the block dependencies will have been flushed 3993 * leaving only inode dependencies (e.g., bitmaps). Thus, 3994 * we do a ufs_ihashget to check for the vnode in the cache. 3995 * If it is there, we do a full flush. If it is no longer 3996 * there we need only dispose of any remaining bitmap 3997 * dependencies and write the inode to disk. 3998 */ 3999 inum = dap->da_newinum; 4000 FREE_LOCK(&lk); 4001 if ((vp = ufs_ihashget(ump->um_dev, inum)) == NULL) { 4002 ACQUIRE_LOCK(&lk); 4003 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0 4004 && dap == LIST_FIRST(diraddhdp)) 4005 panic("flush_pagedep_deps: flush 1 failed"); 4006 /* 4007 * If the inode still has bitmap dependencies, 4008 * push them to disk. 4009 */ 4010 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4011 gotit = getdirtybuf(&inodedep->id_buf,MNT_WAIT); 4012 FREE_LOCK(&lk); 4013 if (gotit && 4014 (error = VOP_BWRITE(inodedep->id_buf)) != 0) 4015 break; 4016 ACQUIRE_LOCK(&lk); 4017 } 4018 if (dap != LIST_FIRST(diraddhdp)) 4019 continue; 4020 /* 4021 * If the inode is still sitting in a buffer waiting 4022 * to be written, push it to disk. 4023 */ 4024 FREE_LOCK(&lk); 4025 if ((error = bread(ump->um_devvp, 4026 fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), 4027 (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0) 4028 break; 4029 if ((error = VOP_BWRITE(bp)) != 0) 4030 break; 4031 ACQUIRE_LOCK(&lk); 4032 if (dap == LIST_FIRST(diraddhdp)) 4033 panic("flush_pagedep_deps: flush 2 failed"); 4034 continue; 4035 } 4036 if (vp->v_type == VDIR) { 4037 /* 4038 * A newly allocated directory must have its "." and 4039 * ".." entries written out before its name can be 4040 * committed in its parent. We do not want or need 4041 * the full semantics of a synchronous VOP_FSYNC as 4042 * that may end up here again, once for each directory 4043 * level in the filesystem. Instead, we push the blocks 4044 * and wait for them to clear. 4045 */ 4046 if (error = VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p)) { 4047 vput(vp); 4048 break; 4049 } 4050 ACQUIRE_LOCK(&lk); 4051 while (vp->v_numoutput) { 4052 vp->v_flag |= VBWAIT; 4053 FREE_LOCK_INTERLOCKED(&lk); 4054 tsleep((caddr_t)&vp->v_numoutput, PRIBIO + 1, 4055 "sdflpd", 0); 4056 ACQUIRE_LOCK_INTERLOCKED(&lk); 4057 } 4058 FREE_LOCK(&lk); 4059 } 4060#ifndef __FreeBSD__ 4061 tv = time; 4062#else 4063 getmicrotime(&tv); 4064#endif /* __FreeBSD__ */ 4065 error = UFS_UPDATE(vp, &tv, &tv, MNT_WAIT); 4066 vput(vp); 4067 if (error) 4068 break; 4069 /* 4070 * If we have failed to get rid of all the dependencies 4071 * then something is seriously wrong. 4072 */ 4073 if (dap == LIST_FIRST(diraddhdp)) 4074 panic("flush_pagedep_deps: flush 3 failed"); 4075 ACQUIRE_LOCK(&lk); 4076 } 4077 if (error) 4078 ACQUIRE_LOCK(&lk); 4079 return (error); 4080} 4081 4082/* 4083 * Acquire exclusive access to a buffer. 4084 * Must be called with splbio blocked. 4085 * Return 1 if buffer was acquired. 4086 */ 4087static int 4088getdirtybuf(bpp, waitfor) 4089 struct buf **bpp; 4090 int waitfor; 4091{ 4092 struct buf *bp; 4093 4094 for (;;) { 4095 if ((bp = *bpp) == NULL) 4096 return (0); 4097 if ((bp->b_flags & B_BUSY) == 0) 4098 break; 4099 if (waitfor != MNT_WAIT) 4100 return (0); 4101 bp->b_flags |= B_WANTED; 4102 FREE_LOCK_INTERLOCKED(&lk); 4103 tsleep((caddr_t)bp, PRIBIO + 1, "sdsdty", 0); 4104 ACQUIRE_LOCK_INTERLOCKED(&lk); 4105 } 4106 if ((bp->b_flags & B_DELWRI) == 0) 4107 return (0); 4108 bremfree(bp); 4109 bp->b_flags |= B_BUSY; 4110 return (1); 4111} 4112 4113/* 4114 * Called whenever a buffer that is being invalidated or reallocated 4115 * contains dependencies. This should only happen if an I/O error has 4116 * occurred. The routine is called with the buffer locked. 4117 */ 4118void 4119softdep_deallocate_dependencies(bp) 4120 struct buf *bp; 4121{ 4122 struct worklist *wk; 4123 4124 if ((bp->b_flags & B_ERROR) == 0) 4125 panic("softdep_deallocate_dependencies: dangling deps"); 4126 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); 4127 ACQUIRE_LOCK(&lk); 4128 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 4129 WORKLIST_REMOVE(wk); 4130 FREE_LOCK(&lk); 4131 switch (wk->wk_type) { 4132 /* 4133 * XXX - should really clean up, but for now we will 4134 * just leak memory and not worry about it. Also should 4135 * mark the filesystem permanently dirty so that it will 4136 * force fsck to be run (though this would best be done 4137 * in the mainline code). 4138 */ 4139 case D_PAGEDEP: 4140 case D_INODEDEP: 4141 case D_BMSAFEMAP: 4142 case D_ALLOCDIRECT: 4143 case D_INDIRDEP: 4144 case D_ALLOCINDIR: 4145 case D_MKDIR: 4146#ifdef DEBUG 4147 printf("Lost type %s\n", TYPENAME(wk->wk_type)); 4148#endif 4149 break; 4150 default: 4151 panic("%s: Unexpected type %s", 4152 "softdep_deallocate_dependencies", 4153 TYPENAME(wk->wk_type)); 4154 /* NOTREACHED */ 4155 } 4156 ACQUIRE_LOCK(&lk); 4157 } 4158 FREE_LOCK(&lk); 4159} 4160 4161/* 4162 * Function to handle asynchronous write errors in the filesystem. 4163 */ 4164void 4165softdep_error(func, error) 4166 char *func; 4167 int error; 4168{ 4169 4170 /* XXX should do something better! */ 4171 log(LOG_ERR, "%s: got error %d while accessing filesystem\n", 4172 func, error); 4173} 4174 4175