ffs_softdep.c revision 44383
1/* 2 * Copyright 1998 Marshall Kirk McKusick. All Rights Reserved. 3 * 4 * The soft updates code is derived from the appendix of a University 5 * of Michigan technical report (Gregory R. Ganger and Yale N. Patt, 6 * "Soft Updates: A Solution to the Metadata Update Problem in File 7 * Systems", CSE-TR-254-95, August 1995). 8 * 9 * The following are the copyrights and redistribution conditions that 10 * apply to this copy of the soft update software. For a license 11 * to use, redistribute or sell the soft update software under 12 * conditions other than those described here, please contact the 13 * author at one of the following addresses: 14 * 15 * Marshall Kirk McKusick mckusick@mckusick.com 16 * 1614 Oxford Street +1-510-843-9542 17 * Berkeley, CA 94709-1608 18 * USA 19 * 20 * Redistribution and use in source and binary forms, with or without 21 * modification, are permitted provided that the following conditions 22 * are met: 23 * 24 * 1. Redistributions of source code must retain the above copyright 25 * notice, this list of conditions and the following disclaimer. 26 * 2. Redistributions in binary form must reproduce the above copyright 27 * notice, this list of conditions and the following disclaimer in the 28 * documentation and/or other materials provided with the distribution. 29 * 3. None of the names of McKusick, Ganger, Patt, or the University of 30 * Michigan may be used to endorse or promote products derived from 31 * this software without specific prior written permission. 32 * 4. Redistributions in any form must be accompanied by information on 33 * how to obtain complete source code for any accompanying software 34 * that uses this software. This source code must either be included 35 * in the distribution or be available for no more than the cost of 36 * distribution plus a nominal fee, and must be freely redistributable 37 * under reasonable conditions. For an executable file, complete 38 * source code means the source code for all modules it contains. 39 * It does not mean source code for modules or files that typically 40 * accompany the operating system on which the executable file runs, 41 * e.g., standard library modules or system header files. 42 * 43 * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY 44 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 45 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 46 * DISCLAIMED. IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR 47 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 48 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 49 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 50 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 51 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 52 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 53 * SUCH DAMAGE. 54 * 55 * from: @(#)ffs_softdep.c 9.33 (McKusick) 2/25/99 56 * $Id: ffs_softdep.c,v 1.22 1999/02/17 20:01:20 mckusick Exp $ 57 */ 58 59/* 60 * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide. 61 */ 62#ifndef DIAGNOSTIC 63#define DIAGNOSTIC 64#endif 65#ifndef DEBUG 66#define DEBUG 67#endif 68 69#include <sys/param.h> 70#include <sys/buf.h> 71#include <sys/kernel.h> 72#include <sys/malloc.h> 73#include <sys/mount.h> 74#include <sys/proc.h> 75#include <sys/syslog.h> 76#include <sys/systm.h> 77#include <sys/vnode.h> 78#include <miscfs/specfs/specdev.h> 79#include <ufs/ufs/dir.h> 80#include <ufs/ufs/quota.h> 81#include <ufs/ufs/inode.h> 82#include <ufs/ufs/ufsmount.h> 83#include <ufs/ffs/fs.h> 84#include <ufs/ffs/softdep.h> 85#include <ufs/ffs/ffs_extern.h> 86#include <ufs/ufs/ufs_extern.h> 87 88/* 89 * These definitions need to be adapted to the system to which 90 * this file is being ported. 91 */ 92/* 93 * malloc types defined for the softdep system. 94 */ 95MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies"); 96MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies"); 97MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation"); 98MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map"); 99MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode"); 100MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies"); 101MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block"); 102MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode"); 103MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode"); 104MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated"); 105MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry"); 106MALLOC_DEFINE(M_MKDIR, "mkdir","New directory"); 107MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted"); 108 109#define D_PAGEDEP 0 110#define D_INODEDEP 1 111#define D_NEWBLK 2 112#define D_BMSAFEMAP 3 113#define D_ALLOCDIRECT 4 114#define D_INDIRDEP 5 115#define D_ALLOCINDIR 6 116#define D_FREEFRAG 7 117#define D_FREEBLKS 8 118#define D_FREEFILE 9 119#define D_DIRADD 10 120#define D_MKDIR 11 121#define D_DIRREM 12 122#define D_LAST D_DIRREM 123 124/* 125 * translate from workitem type to memory type 126 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 127 */ 128static struct malloc_type *memtype[] = { 129 M_PAGEDEP, 130 M_INODEDEP, 131 M_NEWBLK, 132 M_BMSAFEMAP, 133 M_ALLOCDIRECT, 134 M_INDIRDEP, 135 M_ALLOCINDIR, 136 M_FREEFRAG, 137 M_FREEBLKS, 138 M_FREEFILE, 139 M_DIRADD, 140 M_MKDIR, 141 M_DIRREM 142}; 143 144#define DtoM(type) (memtype[type]) 145 146/* 147 * Names of malloc types. 148 */ 149#define TYPENAME(type) \ 150 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???") 151#define CURPROC curproc 152/* 153 * End system adaptaion definitions. 154 */ 155 156/* 157 * Internal function prototypes. 158 */ 159static void softdep_error __P((char *, int)); 160static int getdirtybuf __P((struct buf **, int)); 161static int flush_pagedep_deps __P((struct vnode *, struct mount *, 162 struct diraddhd *)); 163static int flush_inodedep_deps __P((struct fs *, ino_t)); 164static int handle_written_filepage __P((struct pagedep *, struct buf *)); 165static void diradd_inode_written __P((struct diradd *, struct inodedep *)); 166static int handle_written_inodeblock __P((struct inodedep *, struct buf *)); 167static void handle_allocdirect_partdone __P((struct allocdirect *)); 168static void handle_allocindir_partdone __P((struct allocindir *)); 169static void initiate_write_filepage __P((struct pagedep *, struct buf *)); 170static void handle_written_mkdir __P((struct mkdir *, int)); 171static void initiate_write_inodeblock __P((struct inodedep *, struct buf *)); 172static void handle_workitem_freefile __P((struct freefile *)); 173static void handle_workitem_remove __P((struct dirrem *)); 174static struct dirrem *newdirrem __P((struct buf *, struct inode *, 175 struct inode *, int)); 176static void free_diradd __P((struct diradd *)); 177static void free_allocindir __P((struct allocindir *, struct inodedep *)); 178static int indir_trunc __P((struct inode *, ufs_daddr_t, int, ufs_lbn_t, 179 long *)); 180static void deallocate_dependencies __P((struct buf *, struct inodedep *)); 181static void free_allocdirect __P((struct allocdirectlst *, 182 struct allocdirect *, int)); 183static int free_inodedep __P((struct inodedep *)); 184static void handle_workitem_freeblocks __P((struct freeblks *)); 185static void merge_inode_lists __P((struct inodedep *)); 186static void setup_allocindir_phase2 __P((struct buf *, struct inode *, 187 struct allocindir *)); 188static struct allocindir *newallocindir __P((struct inode *, int, ufs_daddr_t, 189 ufs_daddr_t)); 190static void handle_workitem_freefrag __P((struct freefrag *)); 191static struct freefrag *newfreefrag __P((struct inode *, ufs_daddr_t, long)); 192static void allocdirect_merge __P((struct allocdirectlst *, 193 struct allocdirect *, struct allocdirect *)); 194static struct bmsafemap *bmsafemap_lookup __P((struct buf *)); 195static int newblk_lookup __P((struct fs *, ufs_daddr_t, int, 196 struct newblk **)); 197static int inodedep_lookup __P((struct fs *, ino_t, int, struct inodedep **)); 198static int pagedep_lookup __P((struct inode *, ufs_lbn_t, int, 199 struct pagedep **)); 200static void pause_timer __P((void *)); 201static int checklimit __P((long *, int)); 202static void add_to_worklist __P((struct worklist *)); 203 204/* 205 * Exported softdep operations. 206 */ 207struct bio_ops bioops = { 208 softdep_disk_io_initiation, /* io_start */ 209 softdep_disk_write_complete, /* io_complete */ 210 softdep_deallocate_dependencies, /* io_deallocate */ 211 softdep_fsync, /* io_fsync */ 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 for soft updates\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 %ld >= %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(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT)) { 1667 bp = TAILQ_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 mkdir1->md_buf = newdirbp; 2158 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 2159 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list); 2160 bdwrite(newdirbp); 2161 /* 2162 * Dependency on link count increase for parent directory 2163 */ 2164 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0 2165 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2166 dap->da_state &= ~MKDIR_PARENT; 2167 WORKITEM_FREE(mkdir2, D_MKDIR); 2168 } else { 2169 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 2170 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 2171 } 2172 } 2173 /* 2174 * Link into parent directory pagedep to await its being written. 2175 */ 2176 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2177 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2178 dap->da_pagedep = pagedep; 2179 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 2180 da_pdlist); 2181 /* 2182 * Link into its inodedep. Put it on the id_bufwait list if the inode 2183 * is not yet written. If it is written, do the post-inode write 2184 * processing to put it on the id_pendinghd list. 2185 */ 2186 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep); 2187 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 2188 diradd_inode_written(dap, inodedep); 2189 else 2190 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2191 FREE_LOCK(&lk); 2192} 2193 2194/* 2195 * This procedure is called to change the offset of a directory 2196 * entry when compacting a directory block which must be owned 2197 * exclusively by the caller. Note that the actual entry movement 2198 * must be done in this procedure to ensure that no I/O completions 2199 * occur while the move is in progress. 2200 */ 2201void 2202softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize) 2203 struct inode *dp; /* inode for directory */ 2204 caddr_t base; /* address of dp->i_offset */ 2205 caddr_t oldloc; /* address of old directory location */ 2206 caddr_t newloc; /* address of new directory location */ 2207 int entrysize; /* size of directory entry */ 2208{ 2209 int offset, oldoffset, newoffset; 2210 struct pagedep *pagedep; 2211 struct diradd *dap; 2212 ufs_lbn_t lbn; 2213 2214 ACQUIRE_LOCK(&lk); 2215 lbn = lblkno(dp->i_fs, dp->i_offset); 2216 offset = blkoff(dp->i_fs, dp->i_offset); 2217 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) 2218 goto done; 2219 oldoffset = offset + (oldloc - base); 2220 newoffset = offset + (newloc - base); 2221 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[DIRADDHASH(oldoffset)]); 2222 dap; dap = LIST_NEXT(dap, da_pdlist)) { 2223 if (dap->da_offset != oldoffset) 2224 continue; 2225 dap->da_offset = newoffset; 2226 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) 2227 break; 2228 LIST_REMOVE(dap, da_pdlist); 2229 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], 2230 dap, da_pdlist); 2231 break; 2232 } 2233 if (dap == NULL) { 2234 for (dap = LIST_FIRST(&pagedep->pd_pendinghd); 2235 dap; dap = LIST_NEXT(dap, da_pdlist)) { 2236 if (dap->da_offset == oldoffset) { 2237 dap->da_offset = newoffset; 2238 break; 2239 } 2240 } 2241 } 2242done: 2243 bcopy(oldloc, newloc, entrysize); 2244 FREE_LOCK(&lk); 2245} 2246 2247/* 2248 * Free a diradd dependency structure. This routine must be called 2249 * with splbio interrupts blocked. 2250 */ 2251static void 2252free_diradd(dap) 2253 struct diradd *dap; 2254{ 2255 struct dirrem *dirrem; 2256 struct pagedep *pagedep; 2257 struct inodedep *inodedep; 2258 struct mkdir *mkdir, *nextmd; 2259 2260#ifdef DEBUG 2261 if (lk.lkt_held == -1) 2262 panic("free_diradd: lock not held"); 2263#endif 2264 WORKLIST_REMOVE(&dap->da_list); 2265 LIST_REMOVE(dap, da_pdlist); 2266 if ((dap->da_state & DIRCHG) == 0) { 2267 pagedep = dap->da_pagedep; 2268 } else { 2269 dirrem = dap->da_previous; 2270 pagedep = dirrem->dm_pagedep; 2271 dirrem->dm_dirinum = pagedep->pd_ino; 2272 add_to_worklist(&dirrem->dm_list); 2273 } 2274 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum, 2275 0, &inodedep) != 0) 2276 (void) free_inodedep(inodedep); 2277 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2278 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 2279 nextmd = LIST_NEXT(mkdir, md_mkdirs); 2280 if (mkdir->md_diradd != dap) 2281 continue; 2282 dap->da_state &= ~mkdir->md_state; 2283 WORKLIST_REMOVE(&mkdir->md_list); 2284 LIST_REMOVE(mkdir, md_mkdirs); 2285 WORKITEM_FREE(mkdir, D_MKDIR); 2286 } 2287 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 2288 panic("free_diradd: unfound ref"); 2289 } 2290 WORKITEM_FREE(dap, D_DIRADD); 2291} 2292 2293/* 2294 * Directory entry removal dependencies. 2295 * 2296 * When removing a directory entry, the entry's inode pointer must be 2297 * zero'ed on disk before the corresponding inode's link count is decremented 2298 * (possibly freeing the inode for re-use). This dependency is handled by 2299 * updating the directory entry but delaying the inode count reduction until 2300 * after the directory block has been written to disk. After this point, the 2301 * inode count can be decremented whenever it is convenient. 2302 */ 2303 2304/* 2305 * This routine should be called immediately after removing 2306 * a directory entry. The inode's link count should not be 2307 * decremented by the calling procedure -- the soft updates 2308 * code will do this task when it is safe. 2309 */ 2310void 2311softdep_setup_remove(bp, dp, ip, isrmdir) 2312 struct buf *bp; /* buffer containing directory block */ 2313 struct inode *dp; /* inode for the directory being modified */ 2314 struct inode *ip; /* inode for directory entry being removed */ 2315 int isrmdir; /* indicates if doing RMDIR */ 2316{ 2317 struct dirrem *dirrem; 2318 2319 /* 2320 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 2321 */ 2322 dirrem = newdirrem(bp, dp, ip, isrmdir); 2323 if ((dirrem->dm_state & COMPLETE) == 0) { 2324 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 2325 dm_next); 2326 } else { 2327 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 2328 add_to_worklist(&dirrem->dm_list); 2329 } 2330 FREE_LOCK(&lk); 2331} 2332 2333/* 2334 * Allocate a new dirrem if appropriate and return it along with 2335 * its associated pagedep. Called without a lock, returns with lock. 2336 */ 2337static struct dirrem * 2338newdirrem(bp, dp, ip, isrmdir) 2339 struct buf *bp; /* buffer containing directory block */ 2340 struct inode *dp; /* inode for the directory being modified */ 2341 struct inode *ip; /* inode for directory entry being removed */ 2342 int isrmdir; /* indicates if doing RMDIR */ 2343{ 2344 int offset; 2345 ufs_lbn_t lbn; 2346 struct diradd *dap; 2347 struct dirrem *dirrem; 2348 struct pagedep *pagedep; 2349 2350 /* 2351 * Whiteouts have no deletion dependencies. 2352 */ 2353 if (ip == NULL) 2354 panic("newdirrem: whiteout"); 2355 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem), 2356 M_DIRREM, M_WAITOK); 2357 bzero(dirrem, sizeof(struct dirrem)); 2358 dirrem->dm_list.wk_type = D_DIRREM; 2359 dirrem->dm_state = isrmdir ? RMDIR : 0; 2360 dirrem->dm_mnt = ITOV(ip)->v_mount; 2361 dirrem->dm_oldinum = ip->i_number; 2362 2363 ACQUIRE_LOCK(&lk); 2364 lbn = lblkno(dp->i_fs, dp->i_offset); 2365 offset = blkoff(dp->i_fs, dp->i_offset); 2366 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2367 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2368 dirrem->dm_pagedep = pagedep; 2369 /* 2370 * Check for a diradd dependency for the same directory entry. 2371 * If present, then both dependencies become obsolete and can 2372 * be de-allocated. Check for an entry on both the pd_dirraddhd 2373 * list and the pd_pendinghd list. 2374 */ 2375 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[DIRADDHASH(offset)]); 2376 dap; dap = LIST_NEXT(dap, da_pdlist)) 2377 if (dap->da_offset == offset) 2378 break; 2379 if (dap == NULL) { 2380 for (dap = LIST_FIRST(&pagedep->pd_pendinghd); 2381 dap; dap = LIST_NEXT(dap, da_pdlist)) 2382 if (dap->da_offset == offset) 2383 break; 2384 if (dap == NULL) 2385 return (dirrem); 2386 } 2387 /* 2388 * Must be ATTACHED at this point, so just delete it. 2389 */ 2390 if ((dap->da_state & ATTACHED) == 0) 2391 panic("newdirrem: not ATTACHED"); 2392 if (dap->da_newinum != ip->i_number) 2393 panic("newdirrem: inum %d should be %d", 2394 ip->i_number, dap->da_newinum); 2395 free_diradd(dap); 2396 dirrem->dm_state |= COMPLETE; 2397 return (dirrem); 2398} 2399 2400/* 2401 * Directory entry change dependencies. 2402 * 2403 * Changing an existing directory entry requires that an add operation 2404 * be completed first followed by a deletion. The semantics for the addition 2405 * are identical to the description of adding a new entry above except 2406 * that the rollback is to the old inode number rather than zero. Once 2407 * the addition dependency is completed, the removal is done as described 2408 * in the removal routine above. 2409 */ 2410 2411/* 2412 * This routine should be called immediately after changing 2413 * a directory entry. The inode's link count should not be 2414 * decremented by the calling procedure -- the soft updates 2415 * code will perform this task when it is safe. 2416 */ 2417void 2418softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 2419 struct buf *bp; /* buffer containing directory block */ 2420 struct inode *dp; /* inode for the directory being modified */ 2421 struct inode *ip; /* inode for directory entry being removed */ 2422 long newinum; /* new inode number for changed entry */ 2423 int isrmdir; /* indicates if doing RMDIR */ 2424{ 2425 int offset; 2426 struct diradd *dap = NULL; 2427 struct dirrem *dirrem; 2428 struct pagedep *pagedep; 2429 struct inodedep *inodedep; 2430 2431 offset = blkoff(dp->i_fs, dp->i_offset); 2432 2433 /* 2434 * Whiteouts do not need diradd dependencies. 2435 */ 2436 if (newinum != WINO) { 2437 MALLOC(dap, struct diradd *, sizeof(struct diradd), 2438 M_DIRADD, M_WAITOK); 2439 bzero(dap, sizeof(struct diradd)); 2440 dap->da_list.wk_type = D_DIRADD; 2441 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 2442 dap->da_offset = offset; 2443 dap->da_newinum = newinum; 2444 } 2445 2446 /* 2447 * Allocate a new dirrem and ACQUIRE_LOCK. 2448 */ 2449 dirrem = newdirrem(bp, dp, ip, isrmdir); 2450 pagedep = dirrem->dm_pagedep; 2451 /* 2452 * The possible values for isrmdir: 2453 * 0 - non-directory file rename 2454 * 1 - directory rename within same directory 2455 * inum - directory rename to new directory of given inode number 2456 * When renaming to a new directory, we are both deleting and 2457 * creating a new directory entry, so the link count on the new 2458 * directory should not change. Thus we do not need the followup 2459 * dirrem which is usually done in handle_workitem_remove. We set 2460 * the DIRCHG flag to tell handle_workitem_remove to skip the 2461 * followup dirrem. 2462 */ 2463 if (isrmdir > 1) 2464 dirrem->dm_state |= DIRCHG; 2465 2466 /* 2467 * Whiteouts have no additional dependencies, 2468 * so just put the dirrem on the correct list. 2469 */ 2470 if (newinum == WINO) { 2471 if ((dirrem->dm_state & COMPLETE) == 0) { 2472 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 2473 dm_next); 2474 } else { 2475 dirrem->dm_dirinum = pagedep->pd_ino; 2476 add_to_worklist(&dirrem->dm_list); 2477 } 2478 FREE_LOCK(&lk); 2479 return; 2480 } 2481 2482 /* 2483 * Link into its inodedep. Put it on the id_bufwait list if the inode 2484 * is not yet written. If it is written, do the post-inode write 2485 * processing to put it on the id_pendinghd list. 2486 */ 2487 dap->da_previous = dirrem; 2488 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 || 2489 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2490 dap->da_state |= COMPLETE; 2491 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 2492 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 2493 } else { 2494 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 2495 dap, da_pdlist); 2496 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2497 } 2498 /* 2499 * If the previous inode was never written or its previous directory 2500 * entry was never written, then we do not want to roll back to this 2501 * previous value. Instead we want to roll back to zero and immediately 2502 * free the unwritten or unreferenced inode. 2503 */ 2504 if (dirrem->dm_state & COMPLETE) { 2505 dap->da_state &= ~DIRCHG; 2506 dap->da_pagedep = pagedep; 2507 dirrem->dm_dirinum = pagedep->pd_ino; 2508 add_to_worklist(&dirrem->dm_list); 2509 } 2510 FREE_LOCK(&lk); 2511} 2512 2513/* 2514 * Called whenever the link count on an inode is increased. 2515 * It creates an inode dependency so that the new reference(s) 2516 * to the inode cannot be committed to disk until the updated 2517 * inode has been written. 2518 */ 2519void 2520softdep_increase_linkcnt(ip) 2521 struct inode *ip; /* the inode with the increased link count */ 2522{ 2523 struct inodedep *inodedep; 2524 2525 ACQUIRE_LOCK(&lk); 2526 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 2527 FREE_LOCK(&lk); 2528} 2529 2530/* 2531 * This workitem decrements the inode's link count. 2532 * If the link count reaches zero, the file is removed. 2533 */ 2534static void 2535handle_workitem_remove(dirrem) 2536 struct dirrem *dirrem; 2537{ 2538 struct proc *p = CURPROC; /* XXX */ 2539 struct inodedep *inodedep; 2540 struct vnode *vp; 2541 struct inode *ip; 2542 int error; 2543 2544 if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) { 2545 softdep_error("handle_workitem_remove: vget", error); 2546 return; 2547 } 2548 ip = VTOI(vp); 2549 /* 2550 * Normal file deletion. 2551 */ 2552 if ((dirrem->dm_state & RMDIR) == 0) { 2553 ip->i_nlink--; 2554 if (ip->i_nlink < ip->i_effnlink) 2555 panic("handle_workitem_remove: bad file delta"); 2556 ip->i_flag |= IN_CHANGE; 2557 vput(vp); 2558 WORKITEM_FREE(dirrem, D_DIRREM); 2559 return; 2560 } 2561 /* 2562 * Directory deletion. Decrement reference count for both the 2563 * just deleted parent directory entry and the reference for ".". 2564 * Next truncate the directory to length zero. When the 2565 * truncation completes, arrange to have the reference count on 2566 * the parent decremented to account for the loss of "..". 2567 */ 2568 ip->i_nlink -= 2; 2569 if (ip->i_nlink < ip->i_effnlink) 2570 panic("handle_workitem_remove: bad dir delta"); 2571 ip->i_flag |= IN_CHANGE; 2572 if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, p->p_ucred, p)) != 0) 2573 softdep_error("handle_workitem_remove: truncate", error); 2574 /* 2575 * Rename a directory to a new parent. Since, we are both deleting 2576 * and creating a new directory entry, the link count on the new 2577 * directory should not change. Thus we skip the followup dirrem. 2578 */ 2579 if (dirrem->dm_state & DIRCHG) { 2580 vput(vp); 2581 WORKITEM_FREE(dirrem, D_DIRREM); 2582 return; 2583 } 2584 ACQUIRE_LOCK(&lk); 2585 (void) inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, DEPALLOC, 2586 &inodedep); 2587 dirrem->dm_state = 0; 2588 dirrem->dm_oldinum = dirrem->dm_dirinum; 2589 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 2590 FREE_LOCK(&lk); 2591 vput(vp); 2592} 2593 2594/* 2595 * Inode de-allocation dependencies. 2596 * 2597 * When an inode's link count is reduced to zero, it can be de-allocated. We 2598 * found it convenient to postpone de-allocation until after the inode is 2599 * written to disk with its new link count (zero). At this point, all of the 2600 * on-disk inode's block pointers are nullified and, with careful dependency 2601 * list ordering, all dependencies related to the inode will be satisfied and 2602 * the corresponding dependency structures de-allocated. So, if/when the 2603 * inode is reused, there will be no mixing of old dependencies with new 2604 * ones. This artificial dependency is set up by the block de-allocation 2605 * procedure above (softdep_setup_freeblocks) and completed by the 2606 * following procedure. 2607 */ 2608static void 2609handle_workitem_freefile(freefile) 2610 struct freefile *freefile; 2611{ 2612 struct vnode vp; 2613 struct inode tip; 2614 struct inodedep *idp; 2615 int error; 2616 2617#ifdef DEBUG 2618 ACQUIRE_LOCK(&lk); 2619 if (inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp)) 2620 panic("handle_workitem_freefile: inodedep survived"); 2621 FREE_LOCK(&lk); 2622#endif 2623 tip.i_devvp = freefile->fx_devvp; 2624 tip.i_dev = freefile->fx_devvp->v_rdev; 2625 tip.i_fs = freefile->fx_fs; 2626 vp.v_data = &tip; 2627 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0) 2628 softdep_error("handle_workitem_freefile", error); 2629 WORKITEM_FREE(freefile, D_FREEFILE); 2630 num_freefile -= 1; 2631} 2632 2633/* 2634 * Disk writes. 2635 * 2636 * The dependency structures constructed above are most actively used when file 2637 * system blocks are written to disk. No constraints are placed on when a 2638 * block can be written, but unsatisfied update dependencies are made safe by 2639 * modifying (or replacing) the source memory for the duration of the disk 2640 * write. When the disk write completes, the memory block is again brought 2641 * up-to-date. 2642 * 2643 * In-core inode structure reclamation. 2644 * 2645 * Because there are a finite number of "in-core" inode structures, they are 2646 * reused regularly. By transferring all inode-related dependencies to the 2647 * in-memory inode block and indexing them separately (via "inodedep"s), we 2648 * can allow "in-core" inode structures to be reused at any time and avoid 2649 * any increase in contention. 2650 * 2651 * Called just before entering the device driver to initiate a new disk I/O. 2652 * The buffer must be locked, thus, no I/O completion operations can occur 2653 * while we are manipulating its associated dependencies. 2654 */ 2655void 2656softdep_disk_io_initiation(bp) 2657 struct buf *bp; /* structure describing disk write to occur */ 2658{ 2659 struct worklist *wk, *nextwk; 2660 struct indirdep *indirdep; 2661 2662 /* 2663 * We only care about write operations. There should never 2664 * be dependencies for reads. 2665 */ 2666 if (bp->b_flags & B_READ) 2667 panic("softdep_disk_io_initiation: read"); 2668 /* 2669 * Do any necessary pre-I/O processing. 2670 */ 2671 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = nextwk) { 2672 nextwk = LIST_NEXT(wk, wk_list); 2673 switch (wk->wk_type) { 2674 2675 case D_PAGEDEP: 2676 initiate_write_filepage(WK_PAGEDEP(wk), bp); 2677 continue; 2678 2679 case D_INODEDEP: 2680 initiate_write_inodeblock(WK_INODEDEP(wk), bp); 2681 continue; 2682 2683 case D_INDIRDEP: 2684 indirdep = WK_INDIRDEP(wk); 2685 if (indirdep->ir_state & GOINGAWAY) 2686 panic("disk_io_initiation: indirdep gone"); 2687 /* 2688 * If there are no remaining dependencies, this 2689 * will be writing the real pointers, so the 2690 * dependency can be freed. 2691 */ 2692 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) { 2693 indirdep->ir_savebp->b_flags &= ~B_XXX; 2694 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 2695 brelse(indirdep->ir_savebp); 2696 /* inline expand WORKLIST_REMOVE(wk); */ 2697 wk->wk_state &= ~ONWORKLIST; 2698 LIST_REMOVE(wk, wk_list); 2699 WORKITEM_FREE(indirdep, D_INDIRDEP); 2700 continue; 2701 } 2702 /* 2703 * Replace up-to-date version with safe version. 2704 */ 2705 ACQUIRE_LOCK(&lk); 2706 indirdep->ir_state &= ~ATTACHED; 2707 indirdep->ir_state |= UNDONE; 2708 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount, 2709 M_INDIRDEP, M_WAITOK); 2710 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 2711 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 2712 bp->b_bcount); 2713 FREE_LOCK(&lk); 2714 continue; 2715 2716 case D_MKDIR: 2717 case D_BMSAFEMAP: 2718 case D_ALLOCDIRECT: 2719 case D_ALLOCINDIR: 2720 continue; 2721 2722 default: 2723 panic("handle_disk_io_initiation: Unexpected type %s", 2724 TYPENAME(wk->wk_type)); 2725 /* NOTREACHED */ 2726 } 2727 } 2728} 2729 2730/* 2731 * Called from within the procedure above to deal with unsatisfied 2732 * allocation dependencies in a directory. The buffer must be locked, 2733 * thus, no I/O completion operations can occur while we are 2734 * manipulating its associated dependencies. 2735 */ 2736static void 2737initiate_write_filepage(pagedep, bp) 2738 struct pagedep *pagedep; 2739 struct buf *bp; 2740{ 2741 struct diradd *dap; 2742 struct direct *ep; 2743 int i; 2744 2745 if (pagedep->pd_state & IOSTARTED) { 2746 /* 2747 * This can only happen if there is a driver that does not 2748 * understand chaining. Here biodone will reissue the call 2749 * to strategy for the incomplete buffers. 2750 */ 2751 printf("initiate_write_filepage: already started\n"); 2752 return; 2753 } 2754 pagedep->pd_state |= IOSTARTED; 2755 ACQUIRE_LOCK(&lk); 2756 for (i = 0; i < DAHASHSZ; i++) { 2757 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 2758 dap = LIST_NEXT(dap, da_pdlist)) { 2759 ep = (struct direct *) 2760 ((char *)bp->b_data + dap->da_offset); 2761 if (ep->d_ino != dap->da_newinum) 2762 panic("%s: dir inum %d != new %d", 2763 "initiate_write_filepage", 2764 ep->d_ino, dap->da_newinum); 2765 if (dap->da_state & DIRCHG) 2766 ep->d_ino = dap->da_previous->dm_oldinum; 2767 else 2768 ep->d_ino = 0; 2769 dap->da_state &= ~ATTACHED; 2770 dap->da_state |= UNDONE; 2771 } 2772 } 2773 FREE_LOCK(&lk); 2774} 2775 2776/* 2777 * Called from within the procedure above to deal with unsatisfied 2778 * allocation dependencies in an inodeblock. The buffer must be 2779 * locked, thus, no I/O completion operations can occur while we 2780 * are manipulating its associated dependencies. 2781 */ 2782static void 2783initiate_write_inodeblock(inodedep, bp) 2784 struct inodedep *inodedep; 2785 struct buf *bp; /* The inode block */ 2786{ 2787 struct allocdirect *adp, *lastadp; 2788 struct dinode *dp; 2789 struct fs *fs; 2790 ufs_lbn_t prevlbn = 0; 2791 int i, deplist; 2792 2793 if (inodedep->id_state & IOSTARTED) 2794 panic("initiate_write_inodeblock: already started"); 2795 inodedep->id_state |= IOSTARTED; 2796 fs = inodedep->id_fs; 2797 dp = (struct dinode *)bp->b_data + 2798 ino_to_fsbo(fs, inodedep->id_ino); 2799 /* 2800 * If the bitmap is not yet written, then the allocated 2801 * inode cannot be written to disk. 2802 */ 2803 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 2804 if (inodedep->id_savedino != NULL) 2805 panic("initiate_write_inodeblock: already doing I/O"); 2806 MALLOC(inodedep->id_savedino, struct dinode *, 2807 sizeof(struct dinode), M_INODEDEP, M_WAITOK); 2808 *inodedep->id_savedino = *dp; 2809 bzero((caddr_t)dp, sizeof(struct dinode)); 2810 return; 2811 } 2812 /* 2813 * If no dependencies, then there is nothing to roll back. 2814 */ 2815 inodedep->id_savedsize = dp->di_size; 2816 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 2817 return; 2818 /* 2819 * Set the dependencies to busy. 2820 */ 2821 ACQUIRE_LOCK(&lk); 2822 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 2823 adp = TAILQ_NEXT(adp, ad_next)) { 2824#ifdef DIAGNOSTIC 2825 if (deplist != 0 && prevlbn >= adp->ad_lbn) 2826 panic("softdep_write_inodeblock: lbn order"); 2827 prevlbn = adp->ad_lbn; 2828 if (adp->ad_lbn < NDADDR && 2829 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) 2830 panic("%s: direct pointer #%ld mismatch %d != %d", 2831 "softdep_write_inodeblock", adp->ad_lbn, 2832 dp->di_db[adp->ad_lbn], adp->ad_newblkno); 2833 if (adp->ad_lbn >= NDADDR && 2834 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) 2835 panic("%s: indirect pointer #%ld mismatch %d != %d", 2836 "softdep_write_inodeblock", adp->ad_lbn - NDADDR, 2837 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno); 2838 deplist |= 1 << adp->ad_lbn; 2839 if ((adp->ad_state & ATTACHED) == 0) 2840 panic("softdep_write_inodeblock: Unknown state 0x%x", 2841 adp->ad_state); 2842#endif /* DIAGNOSTIC */ 2843 adp->ad_state &= ~ATTACHED; 2844 adp->ad_state |= UNDONE; 2845 } 2846 /* 2847 * The on-disk inode cannot claim to be any larger than the last 2848 * fragment that has been written. Otherwise, the on-disk inode 2849 * might have fragments that were not the last block in the file 2850 * which would corrupt the filesystem. 2851 */ 2852 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 2853 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 2854 if (adp->ad_lbn >= NDADDR) 2855 break; 2856 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 2857 /* keep going until hitting a rollback to a frag */ 2858 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 2859 continue; 2860 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 2861 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 2862#ifdef DIAGNOSTIC 2863 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 2864 panic("softdep_write_inodeblock: lost dep1"); 2865#endif /* DIAGNOSTIC */ 2866 dp->di_db[i] = 0; 2867 } 2868 for (i = 0; i < NIADDR; i++) { 2869#ifdef DIAGNOSTIC 2870 if (dp->di_ib[i] != 0 && 2871 (deplist & ((1 << NDADDR) << i)) == 0) 2872 panic("softdep_write_inodeblock: lost dep2"); 2873#endif /* DIAGNOSTIC */ 2874 dp->di_ib[i] = 0; 2875 } 2876 FREE_LOCK(&lk); 2877 return; 2878 } 2879 /* 2880 * If we have zero'ed out the last allocated block of the file, 2881 * roll back the size to the last currently allocated block. 2882 * We know that this last allocated block is a full-sized as 2883 * we already checked for fragments in the loop above. 2884 */ 2885 if (lastadp != NULL && 2886 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 2887 for (i = lastadp->ad_lbn; i >= 0; i--) 2888 if (dp->di_db[i] != 0) 2889 break; 2890 dp->di_size = (i + 1) * fs->fs_bsize; 2891 } 2892 /* 2893 * The only dependencies are for indirect blocks. 2894 * 2895 * The file size for indirect block additions is not guaranteed. 2896 * Such a guarantee would be non-trivial to achieve. The conventional 2897 * synchronous write implementation also does not make this guarantee. 2898 * Fsck should catch and fix discrepancies. Arguably, the file size 2899 * can be over-estimated without destroying integrity when the file 2900 * moves into the indirect blocks (i.e., is large). If we want to 2901 * postpone fsck, we are stuck with this argument. 2902 */ 2903 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 2904 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 2905 FREE_LOCK(&lk); 2906} 2907 2908/* 2909 * This routine is called during the completion interrupt 2910 * service routine for a disk write (from the procedure called 2911 * by the device driver to inform the file system caches of 2912 * a request completion). It should be called early in this 2913 * procedure, before the block is made available to other 2914 * processes or other routines are called. 2915 */ 2916void 2917softdep_disk_write_complete(bp) 2918 struct buf *bp; /* describes the completed disk write */ 2919{ 2920 struct worklist *wk; 2921 struct workhead reattach; 2922 struct newblk *newblk; 2923 struct allocindir *aip; 2924 struct allocdirect *adp; 2925 struct indirdep *indirdep; 2926 struct inodedep *inodedep; 2927 struct bmsafemap *bmsafemap; 2928 2929#ifdef DEBUG 2930 if (lk.lkt_held != -1) 2931 panic("softdep_disk_write_complete: lock is held"); 2932 lk.lkt_held = -2; 2933#endif 2934 LIST_INIT(&reattach); 2935 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2936 WORKLIST_REMOVE(wk); 2937 switch (wk->wk_type) { 2938 2939 case D_PAGEDEP: 2940 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 2941 WORKLIST_INSERT(&reattach, wk); 2942 continue; 2943 2944 case D_INODEDEP: 2945 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 2946 WORKLIST_INSERT(&reattach, wk); 2947 continue; 2948 2949 case D_BMSAFEMAP: 2950 bmsafemap = WK_BMSAFEMAP(wk); 2951 while (newblk = LIST_FIRST(&bmsafemap->sm_newblkhd)) { 2952 newblk->nb_state |= DEPCOMPLETE; 2953 newblk->nb_bmsafemap = NULL; 2954 LIST_REMOVE(newblk, nb_deps); 2955 } 2956 while (adp = LIST_FIRST(&bmsafemap->sm_allocdirecthd)) { 2957 adp->ad_state |= DEPCOMPLETE; 2958 adp->ad_buf = NULL; 2959 LIST_REMOVE(adp, ad_deps); 2960 handle_allocdirect_partdone(adp); 2961 } 2962 while (aip = LIST_FIRST(&bmsafemap->sm_allocindirhd)) { 2963 aip->ai_state |= DEPCOMPLETE; 2964 aip->ai_buf = NULL; 2965 LIST_REMOVE(aip, ai_deps); 2966 handle_allocindir_partdone(aip); 2967 } 2968 while ((inodedep = 2969 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { 2970 inodedep->id_state |= DEPCOMPLETE; 2971 LIST_REMOVE(inodedep, id_deps); 2972 inodedep->id_buf = NULL; 2973 } 2974 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 2975 continue; 2976 2977 case D_MKDIR: 2978 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 2979 continue; 2980 2981 case D_ALLOCDIRECT: 2982 adp = WK_ALLOCDIRECT(wk); 2983 adp->ad_state |= COMPLETE; 2984 handle_allocdirect_partdone(adp); 2985 continue; 2986 2987 case D_ALLOCINDIR: 2988 aip = WK_ALLOCINDIR(wk); 2989 aip->ai_state |= COMPLETE; 2990 handle_allocindir_partdone(aip); 2991 continue; 2992 2993 case D_INDIRDEP: 2994 indirdep = WK_INDIRDEP(wk); 2995 if (indirdep->ir_state & GOINGAWAY) 2996 panic("disk_write_complete: indirdep gone"); 2997 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 2998 FREE(indirdep->ir_saveddata, M_INDIRDEP); 2999 indirdep->ir_saveddata = 0; 3000 indirdep->ir_state &= ~UNDONE; 3001 indirdep->ir_state |= ATTACHED; 3002 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 3003 handle_allocindir_partdone(aip); 3004 if (aip == LIST_FIRST(&indirdep->ir_donehd)) 3005 panic("disk_write_complete: not gone"); 3006 } 3007 WORKLIST_INSERT(&reattach, wk); 3008 bdirty(bp); 3009 continue; 3010 3011 default: 3012 panic("handle_disk_write_complete: Unknown type %s", 3013 TYPENAME(wk->wk_type)); 3014 /* NOTREACHED */ 3015 } 3016 } 3017 /* 3018 * Reattach any requests that must be redone. 3019 */ 3020 while ((wk = LIST_FIRST(&reattach)) != NULL) { 3021 WORKLIST_REMOVE(wk); 3022 WORKLIST_INSERT(&bp->b_dep, wk); 3023 } 3024#ifdef DEBUG 3025 if (lk.lkt_held != -2) 3026 panic("softdep_disk_write_complete: lock lost"); 3027 lk.lkt_held = -1; 3028#endif 3029} 3030 3031/* 3032 * Called from within softdep_disk_write_complete above. Note that 3033 * this routine is always called from interrupt level with further 3034 * splbio interrupts blocked. 3035 */ 3036static void 3037handle_allocdirect_partdone(adp) 3038 struct allocdirect *adp; /* the completed allocdirect */ 3039{ 3040 struct allocdirect *listadp; 3041 struct inodedep *inodedep; 3042 long bsize; 3043 3044 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3045 return; 3046 if (adp->ad_buf != NULL) 3047 panic("handle_allocdirect_partdone: dangling dep"); 3048 /* 3049 * The on-disk inode cannot claim to be any larger than the last 3050 * fragment that has been written. Otherwise, the on-disk inode 3051 * might have fragments that were not the last block in the file 3052 * which would corrupt the filesystem. Thus, we cannot free any 3053 * allocdirects after one whose ad_oldblkno claims a fragment as 3054 * these blocks must be rolled back to zero before writing the inode. 3055 * We check the currently active set of allocdirects in id_inoupdt. 3056 */ 3057 inodedep = adp->ad_inodedep; 3058 bsize = inodedep->id_fs->fs_bsize; 3059 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp; 3060 listadp = TAILQ_NEXT(listadp, ad_next)) { 3061 /* found our block */ 3062 if (listadp == adp) 3063 break; 3064 /* continue if ad_oldlbn is not a fragment */ 3065 if (listadp->ad_oldsize == 0 || 3066 listadp->ad_oldsize == bsize) 3067 continue; 3068 /* hit a fragment */ 3069 return; 3070 } 3071 /* 3072 * If we have reached the end of the current list without 3073 * finding the just finished dependency, then it must be 3074 * on the future dependency list. Future dependencies cannot 3075 * be freed until they are moved to the current list. 3076 */ 3077 if (listadp == NULL) { 3078#ifdef DEBUG 3079 for (listadp = TAILQ_FIRST(&inodedep->id_newinoupdt); listadp; 3080 listadp = TAILQ_NEXT(listadp, ad_next)) 3081 /* found our block */ 3082 if (listadp == adp) 3083 break; 3084 if (listadp == NULL) 3085 panic("handle_allocdirect_partdone: lost dep"); 3086#endif /* DEBUG */ 3087 return; 3088 } 3089 /* 3090 * If we have found the just finished dependency, then free 3091 * it along with anything that follows it that is complete. 3092 */ 3093 for (; adp; adp = listadp) { 3094 listadp = TAILQ_NEXT(adp, ad_next); 3095 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3096 return; 3097 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 3098 } 3099} 3100 3101/* 3102 * Called from within softdep_disk_write_complete above. Note that 3103 * this routine is always called from interrupt level with further 3104 * splbio interrupts blocked. 3105 */ 3106static void 3107handle_allocindir_partdone(aip) 3108 struct allocindir *aip; /* the completed allocindir */ 3109{ 3110 struct indirdep *indirdep; 3111 3112 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 3113 return; 3114 if (aip->ai_buf != NULL) 3115 panic("handle_allocindir_partdone: dangling dependency"); 3116 indirdep = aip->ai_indirdep; 3117 if (indirdep->ir_state & UNDONE) { 3118 LIST_REMOVE(aip, ai_next); 3119 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 3120 return; 3121 } 3122 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 3123 aip->ai_newblkno; 3124 LIST_REMOVE(aip, ai_next); 3125 if (aip->ai_freefrag != NULL) 3126 add_to_worklist(&aip->ai_freefrag->ff_list); 3127 WORKITEM_FREE(aip, D_ALLOCINDIR); 3128} 3129 3130/* 3131 * Called from within softdep_disk_write_complete above to restore 3132 * in-memory inode block contents to their most up-to-date state. Note 3133 * that this routine is always called from interrupt level with further 3134 * splbio interrupts blocked. 3135 */ 3136static int 3137handle_written_inodeblock(inodedep, bp) 3138 struct inodedep *inodedep; 3139 struct buf *bp; /* buffer containing the inode block */ 3140{ 3141 struct worklist *wk, *filefree; 3142 struct allocdirect *adp, *nextadp; 3143 struct dinode *dp; 3144 int hadchanges; 3145 3146 if ((inodedep->id_state & IOSTARTED) == 0) 3147 panic("handle_written_inodeblock: not started"); 3148 inodedep->id_state &= ~IOSTARTED; 3149 inodedep->id_state |= COMPLETE; 3150 dp = (struct dinode *)bp->b_data + 3151 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 3152 /* 3153 * If we had to rollback the inode allocation because of 3154 * bitmaps being incomplete, then simply restore it. 3155 * Keep the block dirty so that it will not be reclaimed until 3156 * all associated dependencies have been cleared and the 3157 * corresponding updates written to disk. 3158 */ 3159 if (inodedep->id_savedino != NULL) { 3160 *dp = *inodedep->id_savedino; 3161 FREE(inodedep->id_savedino, M_INODEDEP); 3162 inodedep->id_savedino = NULL; 3163 bdirty(bp); 3164 return (1); 3165 } 3166 /* 3167 * Roll forward anything that had to be rolled back before 3168 * the inode could be updated. 3169 */ 3170 hadchanges = 0; 3171 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 3172 nextadp = TAILQ_NEXT(adp, ad_next); 3173 if (adp->ad_state & ATTACHED) 3174 panic("handle_written_inodeblock: new entry"); 3175 if (adp->ad_lbn < NDADDR) { 3176 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) 3177 panic("%s: %s #%ld mismatch %d != %d", 3178 "handle_written_inodeblock", 3179 "direct pointer", adp->ad_lbn, 3180 dp->di_db[adp->ad_lbn], adp->ad_oldblkno); 3181 dp->di_db[adp->ad_lbn] = adp->ad_newblkno; 3182 } else { 3183 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) 3184 panic("%s: %s #%ld allocated as %d", 3185 "handle_written_inodeblock", 3186 "indirect pointer", adp->ad_lbn - NDADDR, 3187 dp->di_ib[adp->ad_lbn - NDADDR]); 3188 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno; 3189 } 3190 adp->ad_state &= ~UNDONE; 3191 adp->ad_state |= ATTACHED; 3192 hadchanges = 1; 3193 } 3194 /* 3195 * Reset the file size to its most up-to-date value. 3196 */ 3197 if (inodedep->id_savedsize == -1) 3198 panic("handle_written_inodeblock: bad size"); 3199 if (dp->di_size != inodedep->id_savedsize) { 3200 dp->di_size = inodedep->id_savedsize; 3201 hadchanges = 1; 3202 } 3203 inodedep->id_savedsize = -1; 3204 /* 3205 * If there were any rollbacks in the inode block, then it must be 3206 * marked dirty so that its will eventually get written back in 3207 * its correct form. 3208 */ 3209 if (hadchanges) 3210 bdirty(bp); 3211 /* 3212 * Process any allocdirects that completed during the update. 3213 */ 3214 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 3215 handle_allocdirect_partdone(adp); 3216 /* 3217 * Process deallocations that were held pending until the 3218 * inode had been written to disk. Freeing of the inode 3219 * is delayed until after all blocks have been freed to 3220 * avoid creation of new <vfsid, inum, lbn> triples 3221 * before the old ones have been deleted. 3222 */ 3223 filefree = NULL; 3224 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 3225 WORKLIST_REMOVE(wk); 3226 switch (wk->wk_type) { 3227 3228 case D_FREEFILE: 3229 /* 3230 * We defer adding filefree to the worklist until 3231 * all other additions have been made to ensure 3232 * that it will be done after all the old blocks 3233 * have been freed. 3234 */ 3235 if (filefree != NULL) 3236 panic("handle_written_inodeblock: filefree"); 3237 filefree = wk; 3238 continue; 3239 3240 case D_MKDIR: 3241 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 3242 continue; 3243 3244 case D_DIRADD: 3245 diradd_inode_written(WK_DIRADD(wk), inodedep); 3246 continue; 3247 3248 case D_FREEBLKS: 3249 case D_FREEFRAG: 3250 case D_DIRREM: 3251 add_to_worklist(wk); 3252 continue; 3253 3254 default: 3255 panic("handle_written_inodeblock: Unknown type %s", 3256 TYPENAME(wk->wk_type)); 3257 /* NOTREACHED */ 3258 } 3259 } 3260 if (filefree != NULL) { 3261 if (free_inodedep(inodedep) == 0) 3262 panic("handle_written_inodeblock: live inodedep"); 3263 add_to_worklist(filefree); 3264 return (0); 3265 } 3266 3267 /* 3268 * If no outstanding dependencies, free it. 3269 */ 3270 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0) 3271 return (0); 3272 return (hadchanges); 3273} 3274 3275/* 3276 * Process a diradd entry after its dependent inode has been written. 3277 * This routine must be called with splbio interrupts blocked. 3278 */ 3279static void 3280diradd_inode_written(dap, inodedep) 3281 struct diradd *dap; 3282 struct inodedep *inodedep; 3283{ 3284 struct pagedep *pagedep; 3285 3286 dap->da_state |= COMPLETE; 3287 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3288 if (dap->da_state & DIRCHG) 3289 pagedep = dap->da_previous->dm_pagedep; 3290 else 3291 pagedep = dap->da_pagedep; 3292 LIST_REMOVE(dap, da_pdlist); 3293 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3294 } 3295 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 3296} 3297 3298/* 3299 * Handle the completion of a mkdir dependency. 3300 */ 3301static void 3302handle_written_mkdir(mkdir, type) 3303 struct mkdir *mkdir; 3304 int type; 3305{ 3306 struct diradd *dap; 3307 struct pagedep *pagedep; 3308 3309 if (mkdir->md_state != type) 3310 panic("handle_written_mkdir: bad type"); 3311 dap = mkdir->md_diradd; 3312 dap->da_state &= ~type; 3313 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 3314 dap->da_state |= DEPCOMPLETE; 3315 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3316 if (dap->da_state & DIRCHG) 3317 pagedep = dap->da_previous->dm_pagedep; 3318 else 3319 pagedep = dap->da_pagedep; 3320 LIST_REMOVE(dap, da_pdlist); 3321 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3322 } 3323 LIST_REMOVE(mkdir, md_mkdirs); 3324 WORKITEM_FREE(mkdir, D_MKDIR); 3325} 3326 3327/* 3328 * Called from within softdep_disk_write_complete above. 3329 * A write operation was just completed. Removed inodes can 3330 * now be freed and associated block pointers may be committed. 3331 * Note that this routine is always called from interrupt level 3332 * with further splbio interrupts blocked. 3333 */ 3334static int 3335handle_written_filepage(pagedep, bp) 3336 struct pagedep *pagedep; 3337 struct buf *bp; /* buffer containing the written page */ 3338{ 3339 struct dirrem *dirrem; 3340 struct diradd *dap, *nextdap; 3341 struct direct *ep; 3342 int i, chgs; 3343 3344 if ((pagedep->pd_state & IOSTARTED) == 0) 3345 panic("handle_written_filepage: not started"); 3346 pagedep->pd_state &= ~IOSTARTED; 3347 /* 3348 * Process any directory removals that have been committed. 3349 */ 3350 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 3351 LIST_REMOVE(dirrem, dm_next); 3352 dirrem->dm_dirinum = pagedep->pd_ino; 3353 add_to_worklist(&dirrem->dm_list); 3354 } 3355 /* 3356 * Free any directory additions that have been committed. 3357 */ 3358 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 3359 free_diradd(dap); 3360 /* 3361 * Uncommitted directory entries must be restored. 3362 */ 3363 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 3364 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 3365 dap = nextdap) { 3366 nextdap = LIST_NEXT(dap, da_pdlist); 3367 if (dap->da_state & ATTACHED) 3368 panic("handle_written_filepage: attached"); 3369 ep = (struct direct *) 3370 ((char *)bp->b_data + dap->da_offset); 3371 ep->d_ino = dap->da_newinum; 3372 dap->da_state &= ~UNDONE; 3373 dap->da_state |= ATTACHED; 3374 chgs = 1; 3375 /* 3376 * If the inode referenced by the directory has 3377 * been written out, then the dependency can be 3378 * moved to the pending list. 3379 */ 3380 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3381 LIST_REMOVE(dap, da_pdlist); 3382 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 3383 da_pdlist); 3384 } 3385 } 3386 } 3387 /* 3388 * If there were any rollbacks in the directory, then it must be 3389 * marked dirty so that its will eventually get written back in 3390 * its correct form. 3391 */ 3392 if (chgs) 3393 bdirty(bp); 3394 /* 3395 * If no dependencies remain, the pagedep will be freed. 3396 * Otherwise it will remain to update the page before it 3397 * is written back to disk. 3398 */ 3399 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) { 3400 for (i = 0; i < DAHASHSZ; i++) 3401 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL) 3402 break; 3403 if (i == DAHASHSZ) { 3404 LIST_REMOVE(pagedep, pd_hash); 3405 WORKITEM_FREE(pagedep, D_PAGEDEP); 3406 return (0); 3407 } 3408 } 3409 return (1); 3410} 3411 3412/* 3413 * Writing back in-core inode structures. 3414 * 3415 * The file system only accesses an inode's contents when it occupies an 3416 * "in-core" inode structure. These "in-core" structures are separate from 3417 * the page frames used to cache inode blocks. Only the latter are 3418 * transferred to/from the disk. So, when the updated contents of the 3419 * "in-core" inode structure are copied to the corresponding in-memory inode 3420 * block, the dependencies are also transferred. The following procedure is 3421 * called when copying a dirty "in-core" inode to a cached inode block. 3422 */ 3423 3424/* 3425 * Called when an inode is loaded from disk. If the effective link count 3426 * differed from the actual link count when it was last flushed, then we 3427 * need to ensure that the correct effective link count is put back. 3428 */ 3429void 3430softdep_load_inodeblock(ip) 3431 struct inode *ip; /* the "in_core" copy of the inode */ 3432{ 3433 struct inodedep *inodedep; 3434 3435 /* 3436 * Check for alternate nlink count. 3437 */ 3438 ip->i_effnlink = ip->i_nlink; 3439 ACQUIRE_LOCK(&lk); 3440 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3441 FREE_LOCK(&lk); 3442 return; 3443 } 3444 if (inodedep->id_nlinkdelta != 0) { 3445 ip->i_effnlink -= inodedep->id_nlinkdelta; 3446 ip->i_flag |= IN_MODIFIED; 3447 inodedep->id_nlinkdelta = 0; 3448 (void) free_inodedep(inodedep); 3449 } 3450 FREE_LOCK(&lk); 3451} 3452 3453/* 3454 * This routine is called just before the "in-core" inode 3455 * information is to be copied to the in-memory inode block. 3456 * Recall that an inode block contains several inodes. If 3457 * the force flag is set, then the dependencies will be 3458 * cleared so that the update can always be made. Note that 3459 * the buffer is locked when this routine is called, so we 3460 * will never be in the middle of writing the inode block 3461 * to disk. 3462 */ 3463void 3464softdep_update_inodeblock(ip, bp, waitfor) 3465 struct inode *ip; /* the "in_core" copy of the inode */ 3466 struct buf *bp; /* the buffer containing the inode block */ 3467 int waitfor; /* nonzero => update must be allowed */ 3468{ 3469 struct inodedep *inodedep; 3470 struct worklist *wk; 3471 int error, gotit; 3472 3473 /* 3474 * If the effective link count is not equal to the actual link 3475 * count, then we must track the difference in an inodedep while 3476 * the inode is (potentially) tossed out of the cache. Otherwise, 3477 * if there is no existing inodedep, then there are no dependencies 3478 * to track. 3479 */ 3480 ACQUIRE_LOCK(&lk); 3481 if (ip->i_effnlink != ip->i_nlink) { 3482 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, 3483 &inodedep); 3484 } else if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3485 FREE_LOCK(&lk); 3486 return; 3487 } 3488 if (ip->i_nlink < ip->i_effnlink) 3489 panic("softdep_update_inodeblock: bad delta"); 3490 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3491 /* 3492 * Changes have been initiated. Anything depending on these 3493 * changes cannot occur until this inode has been written. 3494 */ 3495 inodedep->id_state &= ~COMPLETE; 3496 if ((inodedep->id_state & ONWORKLIST) == 0) 3497 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 3498 /* 3499 * Any new dependencies associated with the incore inode must 3500 * now be moved to the list associated with the buffer holding 3501 * the in-memory copy of the inode. Once merged process any 3502 * allocdirects that are completed by the merger. 3503 */ 3504 merge_inode_lists(inodedep); 3505 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL) 3506 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); 3507 /* 3508 * Now that the inode has been pushed into the buffer, the 3509 * operations dependent on the inode being written to disk 3510 * can be moved to the id_bufwait so that they will be 3511 * processed when the buffer I/O completes. 3512 */ 3513 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 3514 WORKLIST_REMOVE(wk); 3515 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 3516 } 3517 /* 3518 * Newly allocated inodes cannot be written until the bitmap 3519 * that allocates them have been written (indicated by 3520 * DEPCOMPLETE being set in id_state). If we are doing a 3521 * forced sync (e.g., an fsync on a file), we force the bitmap 3522 * to be written so that the update can be done. 3523 */ 3524 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) { 3525 FREE_LOCK(&lk); 3526 return; 3527 } 3528 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 3529 FREE_LOCK(&lk); 3530 if (gotit && (error = VOP_BWRITE(inodedep->id_buf)) != 0) 3531 softdep_error("softdep_update_inodeblock: bwrite", error); 3532 if ((inodedep->id_state & DEPCOMPLETE) == 0) 3533 panic("softdep_update_inodeblock: update failed"); 3534} 3535 3536/* 3537 * Merge the new inode dependency list (id_newinoupdt) into the old 3538 * inode dependency list (id_inoupdt). This routine must be called 3539 * with splbio interrupts blocked. 3540 */ 3541static void 3542merge_inode_lists(inodedep) 3543 struct inodedep *inodedep; 3544{ 3545 struct allocdirect *listadp, *newadp; 3546 3547 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 3548 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) { 3549 if (listadp->ad_lbn < newadp->ad_lbn) { 3550 listadp = TAILQ_NEXT(listadp, ad_next); 3551 continue; 3552 } 3553 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 3554 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 3555 if (listadp->ad_lbn == newadp->ad_lbn) { 3556 allocdirect_merge(&inodedep->id_inoupdt, newadp, 3557 listadp); 3558 listadp = newadp; 3559 } 3560 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 3561 } 3562 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) { 3563 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 3564 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next); 3565 } 3566} 3567 3568/* 3569 * If we are doing an fsync, then we must ensure that any directory 3570 * entries for the inode have been written after the inode gets to disk. 3571 */ 3572int 3573softdep_fsync(vp) 3574 struct vnode *vp; /* the "in_core" copy of the inode */ 3575{ 3576 struct diradd *dap, *olddap; 3577 struct inodedep *inodedep; 3578 struct pagedep *pagedep; 3579 struct worklist *wk; 3580 struct mount *mnt; 3581 struct vnode *pvp; 3582 struct inode *ip; 3583 struct buf *bp; 3584 struct fs *fs; 3585 struct proc *p = CURPROC; /* XXX */ 3586 int error, ret, flushparent; 3587#ifndef __FreeBSD__ 3588 struct timeval tv; 3589#endif 3590 ino_t parentino; 3591 ufs_lbn_t lbn; 3592 3593 ip = VTOI(vp); 3594 fs = ip->i_fs; 3595 for (error = 0, flushparent = 0, olddap = NULL; ; ) { 3596 ACQUIRE_LOCK(&lk); 3597 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) 3598 break; 3599 if (LIST_FIRST(&inodedep->id_inowait) != NULL || 3600 LIST_FIRST(&inodedep->id_bufwait) != NULL || 3601 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 3602 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) 3603 panic("softdep_fsync: pending ops"); 3604 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 3605 break; 3606 if (wk->wk_type != D_DIRADD) 3607 panic("softdep_fsync: Unexpected type %s", 3608 TYPENAME(wk->wk_type)); 3609 dap = WK_DIRADD(wk); 3610 /* 3611 * If we have failed to get rid of all the dependencies 3612 * then something is seriously wrong. 3613 */ 3614 if (dap == olddap) 3615 panic("softdep_fsync: flush failed"); 3616 olddap = dap; 3617 /* 3618 * Flush our parent if this directory entry 3619 * has a MKDIR_PARENT dependency. 3620 */ 3621 if (dap->da_state & DIRCHG) 3622 pagedep = dap->da_previous->dm_pagedep; 3623 else 3624 pagedep = dap->da_pagedep; 3625 mnt = pagedep->pd_mnt; 3626 parentino = pagedep->pd_ino; 3627 lbn = pagedep->pd_lbn; 3628 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) 3629 panic("softdep_fsync: dirty"); 3630 flushparent = dap->da_state & MKDIR_PARENT; 3631 /* 3632 * If we are being fsync'ed as part of vgone'ing this vnode, 3633 * then we will not be able to release and recover the 3634 * vnode below, so we just have to give up on writing its 3635 * directory entry out. It will eventually be written, just 3636 * not now, but then the user was not asking to have it 3637 * written, so we are not breaking any promises. 3638 */ 3639 if (vp->v_flag & VXLOCK) 3640 break; 3641 /* 3642 * We prevent deadlock by always fetching inodes from the 3643 * root, moving down the directory tree. Thus, when fetching 3644 * our parent directory, we must unlock ourselves before 3645 * requesting the lock on our parent. See the comment in 3646 * ufs_lookup for details on possible races. 3647 */ 3648 FREE_LOCK(&lk); 3649 VOP_UNLOCK(vp, 0, p); 3650 if ((error = VFS_VGET(mnt, parentino, &pvp)) != 0) { 3651 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, p); 3652 return (error); 3653 } 3654 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, p); 3655 if (flushparent) { 3656#ifdef __FreeBSD__ 3657 error = UFS_UPDATE(pvp, 1); 3658#else 3659 tv = time; 3660 error = UFS_UPDATE(pvp, &tv, &tv, 1); 3661#endif 3662 if (error) { 3663 vput(pvp); 3664 return (error); 3665 } 3666 } 3667 /* 3668 * Flush directory page containing the inode's name. 3669 */ 3670 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), p->p_ucred, 3671 &bp); 3672 ret = VOP_BWRITE(bp); 3673 vput(pvp); 3674 if (error != 0) 3675 return (error); 3676 if (ret != 0) 3677 return (ret); 3678 } 3679 FREE_LOCK(&lk); 3680 return (0); 3681} 3682 3683/* 3684 * This routine is called when we are trying to synchronously flush a 3685 * file. This routine must eliminate any filesystem metadata dependencies 3686 * so that the syncing routine can succeed by pushing the dirty blocks 3687 * associated with the file. If any I/O errors occur, they are returned. 3688 */ 3689int 3690softdep_sync_metadata(ap) 3691 struct vop_fsync_args /* { 3692 struct vnode *a_vp; 3693 struct ucred *a_cred; 3694 int a_waitfor; 3695 struct proc *a_p; 3696 } */ *ap; 3697{ 3698 struct vnode *vp = ap->a_vp; 3699 struct pagedep *pagedep; 3700 struct allocdirect *adp; 3701 struct allocindir *aip; 3702 struct buf *bp, *nbp; 3703 struct worklist *wk; 3704 int i, error, waitfor; 3705 3706 /* 3707 * Check whether this vnode is involved in a filesystem 3708 * that is doing soft dependency processing. 3709 */ 3710 if (vp->v_type != VBLK) { 3711 if (!DOINGSOFTDEP(vp)) 3712 return (0); 3713 } else 3714 if (vp->v_specmountpoint == NULL || 3715 (vp->v_specmountpoint->mnt_flag & MNT_SOFTDEP) == 0) 3716 return (0); 3717 /* 3718 * Ensure that any direct block dependencies have been cleared. 3719 */ 3720 ACQUIRE_LOCK(&lk); 3721 if (error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number)) { 3722 FREE_LOCK(&lk); 3723 return (error); 3724 } 3725 /* 3726 * For most files, the only metadata dependencies are the 3727 * cylinder group maps that allocate their inode or blocks. 3728 * The block allocation dependencies can be found by traversing 3729 * the dependency lists for any buffers that remain on their 3730 * dirty buffer list. The inode allocation dependency will 3731 * be resolved when the inode is updated with MNT_WAIT. 3732 * This work is done in two passes. The first pass grabs most 3733 * of the buffers and begins asynchronously writing them. The 3734 * only way to wait for these asynchronous writes is to sleep 3735 * on the filesystem vnode which may stay busy for a long time 3736 * if the filesystem is active. So, instead, we make a second 3737 * pass over the dependencies blocking on each write. In the 3738 * usual case we will be blocking against a write that we 3739 * initiated, so when it is done the dependency will have been 3740 * resolved. Thus the second pass is expected to end quickly. 3741 */ 3742 waitfor = MNT_NOWAIT; 3743top: 3744 if (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT) == 0) { 3745 while (vp->v_numoutput) { 3746 vp->v_flag |= VBWAIT; 3747 FREE_LOCK_INTERLOCKED(&lk); 3748 tsleep((caddr_t)&vp->v_numoutput, PRIBIO + 1, 3749 "sdsynm", 0); 3750 ACQUIRE_LOCK_INTERLOCKED(&lk); 3751 } 3752 FREE_LOCK(&lk); 3753 return (0); 3754 } 3755 bp = TAILQ_FIRST(&vp->v_dirtyblkhd); 3756loop: 3757 /* 3758 * As we hold the buffer locked, none of its dependencies 3759 * will disappear. 3760 */ 3761 for (wk = LIST_FIRST(&bp->b_dep); wk; 3762 wk = LIST_NEXT(wk, wk_list)) { 3763 switch (wk->wk_type) { 3764 3765 case D_ALLOCDIRECT: 3766 adp = WK_ALLOCDIRECT(wk); 3767 if (adp->ad_state & DEPCOMPLETE) 3768 break; 3769 nbp = adp->ad_buf; 3770 if (getdirtybuf(&nbp, waitfor) == 0) 3771 break; 3772 FREE_LOCK(&lk); 3773 if (waitfor == MNT_NOWAIT) { 3774 bawrite(nbp); 3775 } else if ((error = VOP_BWRITE(nbp)) != 0) { 3776 bawrite(bp); 3777 return (error); 3778 } 3779 ACQUIRE_LOCK(&lk); 3780 break; 3781 3782 case D_ALLOCINDIR: 3783 aip = WK_ALLOCINDIR(wk); 3784 if (aip->ai_state & DEPCOMPLETE) 3785 break; 3786 nbp = aip->ai_buf; 3787 if (getdirtybuf(&nbp, waitfor) == 0) 3788 break; 3789 FREE_LOCK(&lk); 3790 if (waitfor == MNT_NOWAIT) { 3791 bawrite(nbp); 3792 } else if ((error = VOP_BWRITE(nbp)) != 0) { 3793 bawrite(bp); 3794 return (error); 3795 } 3796 ACQUIRE_LOCK(&lk); 3797 break; 3798 3799 case D_INDIRDEP: 3800 restart: 3801 for (aip = LIST_FIRST(&WK_INDIRDEP(wk)->ir_deplisthd); 3802 aip; aip = LIST_NEXT(aip, ai_next)) { 3803 if (aip->ai_state & DEPCOMPLETE) 3804 continue; 3805 nbp = aip->ai_buf; 3806 if (getdirtybuf(&nbp, MNT_WAIT) == 0) 3807 goto restart; 3808 FREE_LOCK(&lk); 3809 if ((error = VOP_BWRITE(nbp)) != 0) { 3810 bawrite(bp); 3811 return (error); 3812 } 3813 ACQUIRE_LOCK(&lk); 3814 goto restart; 3815 } 3816 break; 3817 3818 case D_INODEDEP: 3819 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs, 3820 WK_INODEDEP(wk)->id_ino)) != 0) { 3821 FREE_LOCK(&lk); 3822 bawrite(bp); 3823 return (error); 3824 } 3825 break; 3826 3827 case D_PAGEDEP: 3828 /* 3829 * We are trying to sync a directory that may 3830 * have dependencies on both its own metadata 3831 * and/or dependencies on the inodes of any 3832 * recently allocated files. We walk its diradd 3833 * lists pushing out the associated inode. 3834 */ 3835 pagedep = WK_PAGEDEP(wk); 3836 for (i = 0; i < DAHASHSZ; i++) { 3837 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 3838 continue; 3839 if (error = flush_pagedep_deps(vp, 3840 pagedep->pd_mnt, &pagedep->pd_diraddhd[i])) { 3841 FREE_LOCK(&lk); 3842 bawrite(bp); 3843 return (error); 3844 } 3845 } 3846 break; 3847 3848 case D_MKDIR: 3849 /* 3850 * This case should never happen if the vnode has 3851 * been properly sync'ed. However, if this function 3852 * is used at a place where the vnode has not yet 3853 * been sync'ed, this dependency can show up. So, 3854 * rather than panic, just flush it. 3855 */ 3856 nbp = WK_MKDIR(wk)->md_buf; 3857 if (getdirtybuf(&nbp, waitfor) == 0) 3858 break; 3859 FREE_LOCK(&lk); 3860 if (waitfor == MNT_NOWAIT) { 3861 bawrite(nbp); 3862 } else if ((error = VOP_BWRITE(nbp)) != 0) { 3863 bawrite(bp); 3864 return (error); 3865 } 3866 ACQUIRE_LOCK(&lk); 3867 break; 3868 3869 case D_BMSAFEMAP: 3870 /* 3871 * This case should never happen if the vnode has 3872 * been properly sync'ed. However, if this function 3873 * is used at a place where the vnode has not yet 3874 * been sync'ed, this dependency can show up. So, 3875 * rather than panic, just flush it. 3876 */ 3877 nbp = WK_BMSAFEMAP(wk)->sm_buf; 3878 if (getdirtybuf(&nbp, waitfor) == 0) 3879 break; 3880 FREE_LOCK(&lk); 3881 if (waitfor == MNT_NOWAIT) { 3882 bawrite(nbp); 3883 } else if ((error = VOP_BWRITE(nbp)) != 0) { 3884 bawrite(bp); 3885 return (error); 3886 } 3887 ACQUIRE_LOCK(&lk); 3888 break; 3889 3890 default: 3891 panic("softdep_sync_metadata: Unknown type %s", 3892 TYPENAME(wk->wk_type)); 3893 /* NOTREACHED */ 3894 } 3895 } 3896 (void) getdirtybuf(&TAILQ_NEXT(bp, b_vnbufs), MNT_WAIT); 3897 nbp = TAILQ_NEXT(bp, b_vnbufs); 3898 FREE_LOCK(&lk); 3899 bawrite(bp); 3900 ACQUIRE_LOCK(&lk); 3901 if (nbp != NULL) { 3902 bp = nbp; 3903 goto loop; 3904 } 3905 /* 3906 * We must wait for any I/O in progress to finish so that 3907 * all potential buffers on the dirty list will be visible. 3908 * Once they are all there, proceed with the second pass 3909 * which will wait for the I/O as per above. 3910 */ 3911 while (vp->v_numoutput) { 3912 vp->v_flag |= VBWAIT; 3913 FREE_LOCK_INTERLOCKED(&lk); 3914 tsleep((caddr_t)&vp->v_numoutput, PRIBIO + 1, "sdsynm", 0); 3915 ACQUIRE_LOCK_INTERLOCKED(&lk); 3916 } 3917 /* 3918 * The brief unlock is to allow any pent up dependency 3919 * processing to be done. 3920 */ 3921 if (waitfor == MNT_NOWAIT) { 3922 waitfor = MNT_WAIT; 3923 FREE_LOCK(&lk); 3924 ACQUIRE_LOCK(&lk); 3925 goto top; 3926 } 3927 3928 /* 3929 * If we have managed to get rid of all the dirty buffers, 3930 * then we are done. For certain directories and block 3931 * devices, we may need to do further work. 3932 */ 3933 if (TAILQ_FIRST(&vp->v_dirtyblkhd) == NULL) { 3934 FREE_LOCK(&lk); 3935 return (0); 3936 } 3937 3938 FREE_LOCK(&lk); 3939 /* 3940 * If we are trying to sync a block device, some of its buffers may 3941 * contain metadata that cannot be written until the contents of some 3942 * partially written files have been written to disk. The only easy 3943 * way to accomplish this is to sync the entire filesystem (luckily 3944 * this happens rarely). 3945 */ 3946 if (vp->v_type == VBLK && vp->v_specmountpoint && !VOP_ISLOCKED(vp) && 3947 (error = VFS_SYNC(vp->v_specmountpoint, MNT_WAIT, ap->a_cred, 3948 ap->a_p)) != 0) 3949 return (error); 3950 return (0); 3951} 3952 3953/* 3954 * Flush the dependencies associated with an inodedep. 3955 * Called with splbio blocked. 3956 */ 3957static int 3958flush_inodedep_deps(fs, ino) 3959 struct fs *fs; 3960 ino_t ino; 3961{ 3962 struct inodedep *inodedep; 3963 struct allocdirect *adp; 3964 int error, waitfor; 3965 struct buf *bp; 3966 3967 /* 3968 * This work is done in two passes. The first pass grabs most 3969 * of the buffers and begins asynchronously writing them. The 3970 * only way to wait for these asynchronous writes is to sleep 3971 * on the filesystem vnode which may stay busy for a long time 3972 * if the filesystem is active. So, instead, we make a second 3973 * pass over the dependencies blocking on each write. In the 3974 * usual case we will be blocking against a write that we 3975 * initiated, so when it is done the dependency will have been 3976 * resolved. Thus the second pass is expected to end quickly. 3977 * We give a brief window at the top of the loop to allow 3978 * any pending I/O to complete. 3979 */ 3980 for (waitfor = MNT_NOWAIT; ; ) { 3981 FREE_LOCK(&lk); 3982 ACQUIRE_LOCK(&lk); 3983 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 3984 return (0); 3985 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3986 adp = TAILQ_NEXT(adp, ad_next)) { 3987 if (adp->ad_state & DEPCOMPLETE) 3988 continue; 3989 bp = adp->ad_buf; 3990 if (getdirtybuf(&bp, waitfor) == 0) { 3991 if (waitfor == MNT_NOWAIT) 3992 continue; 3993 break; 3994 } 3995 FREE_LOCK(&lk); 3996 if (waitfor == MNT_NOWAIT) { 3997 bawrite(bp); 3998 } else if ((error = VOP_BWRITE(bp)) != 0) { 3999 ACQUIRE_LOCK(&lk); 4000 return (error); 4001 } 4002 ACQUIRE_LOCK(&lk); 4003 break; 4004 } 4005 if (adp != NULL) 4006 continue; 4007 for (adp = TAILQ_FIRST(&inodedep->id_newinoupdt); adp; 4008 adp = TAILQ_NEXT(adp, ad_next)) { 4009 if (adp->ad_state & DEPCOMPLETE) 4010 continue; 4011 bp = adp->ad_buf; 4012 if (getdirtybuf(&bp, waitfor) == 0) { 4013 if (waitfor == MNT_NOWAIT) 4014 continue; 4015 break; 4016 } 4017 FREE_LOCK(&lk); 4018 if (waitfor == MNT_NOWAIT) { 4019 bawrite(bp); 4020 } else if ((error = VOP_BWRITE(bp)) != 0) { 4021 ACQUIRE_LOCK(&lk); 4022 return (error); 4023 } 4024 ACQUIRE_LOCK(&lk); 4025 break; 4026 } 4027 if (adp != NULL) 4028 continue; 4029 /* 4030 * If pass2, we are done, otherwise do pass 2. 4031 */ 4032 if (waitfor == MNT_WAIT) 4033 break; 4034 waitfor = MNT_WAIT; 4035 } 4036 /* 4037 * Try freeing inodedep in case all dependencies have been removed. 4038 */ 4039 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0) 4040 (void) free_inodedep(inodedep); 4041 return (0); 4042} 4043 4044/* 4045 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 4046 * Called with splbio blocked. 4047 */ 4048static int 4049flush_pagedep_deps(pvp, mp, diraddhdp) 4050 struct vnode *pvp; 4051 struct mount *mp; 4052 struct diraddhd *diraddhdp; 4053{ 4054 struct proc *p = CURPROC; /* XXX */ 4055 struct inodedep *inodedep; 4056 struct ufsmount *ump; 4057 struct diradd *dap; 4058#ifndef __FreeBSD__ 4059 struct timeval tv; 4060#endif 4061 struct vnode *vp; 4062 int gotit, error = 0; 4063 struct buf *bp; 4064 ino_t inum; 4065 4066 ump = VFSTOUFS(mp); 4067 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 4068 /* 4069 * Flush ourselves if this directory entry 4070 * has a MKDIR_PARENT dependency. 4071 */ 4072 if (dap->da_state & MKDIR_PARENT) { 4073 FREE_LOCK(&lk); 4074#ifdef __FreeBSD__ 4075 error = UFS_UPDATE(pvp, 1); 4076#else 4077 tv = time; 4078 error = UFS_UPDATE(pvp, &tv, &tv, 1); 4079#endif 4080 if (error) 4081 break; 4082 ACQUIRE_LOCK(&lk); 4083 /* 4084 * If that cleared dependencies, go on to next. 4085 */ 4086 if (dap != LIST_FIRST(diraddhdp)) 4087 continue; 4088 if (dap->da_state & MKDIR_PARENT) 4089 panic("flush_pagedep_deps: MKDIR"); 4090 } 4091 /* 4092 * Flush the file on which the directory entry depends. 4093 * If the inode has already been pushed out of the cache, 4094 * then all the block dependencies will have been flushed 4095 * leaving only inode dependencies (e.g., bitmaps). Thus, 4096 * we do a ufs_ihashget to check for the vnode in the cache. 4097 * If it is there, we do a full flush. If it is no longer 4098 * there we need only dispose of any remaining bitmap 4099 * dependencies and write the inode to disk. 4100 */ 4101 inum = dap->da_newinum; 4102 FREE_LOCK(&lk); 4103 if ((vp = ufs_ihashget(ump->um_dev, inum)) == NULL) { 4104 ACQUIRE_LOCK(&lk); 4105 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0 4106 && dap == LIST_FIRST(diraddhdp)) 4107 panic("flush_pagedep_deps: flush 1 failed"); 4108 /* 4109 * If the inode still has bitmap dependencies, 4110 * push them to disk. 4111 */ 4112 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4113 gotit = getdirtybuf(&inodedep->id_buf,MNT_WAIT); 4114 FREE_LOCK(&lk); 4115 if (gotit && 4116 (error = VOP_BWRITE(inodedep->id_buf)) != 0) 4117 break; 4118 ACQUIRE_LOCK(&lk); 4119 } 4120 if (dap != LIST_FIRST(diraddhdp)) 4121 continue; 4122 /* 4123 * If the inode is still sitting in a buffer waiting 4124 * to be written, push it to disk. 4125 */ 4126 FREE_LOCK(&lk); 4127 if ((error = bread(ump->um_devvp, 4128 fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), 4129 (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0) 4130 break; 4131 if ((error = VOP_BWRITE(bp)) != 0) 4132 break; 4133 ACQUIRE_LOCK(&lk); 4134 if (dap == LIST_FIRST(diraddhdp)) 4135 panic("flush_pagedep_deps: flush 2 failed"); 4136 continue; 4137 } 4138 if (vp->v_type == VDIR) { 4139 /* 4140 * A newly allocated directory must have its "." and 4141 * ".." entries written out before its name can be 4142 * committed in its parent. We do not want or need 4143 * the full semantics of a synchronous VOP_FSYNC as 4144 * that may end up here again, once for each directory 4145 * level in the filesystem. Instead, we push the blocks 4146 * and wait for them to clear. 4147 */ 4148 if (error = VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p)) { 4149 vput(vp); 4150 break; 4151 } 4152 ACQUIRE_LOCK(&lk); 4153 while (vp->v_numoutput) { 4154 vp->v_flag |= VBWAIT; 4155 FREE_LOCK_INTERLOCKED(&lk); 4156 tsleep((caddr_t)&vp->v_numoutput, PRIBIO + 1, 4157 "sdflpd", 0); 4158 ACQUIRE_LOCK_INTERLOCKED(&lk); 4159 } 4160 FREE_LOCK(&lk); 4161 } 4162#ifdef __FreeBSD__ 4163 error = UFS_UPDATE(vp, 1); 4164#else 4165 tv = time; 4166 error = UFS_UPDATE(vp, &tv, &tv, 1); 4167#endif 4168 vput(vp); 4169 if (error) 4170 break; 4171 /* 4172 * If we have failed to get rid of all the dependencies 4173 * then something is seriously wrong. 4174 */ 4175 if (dap == LIST_FIRST(diraddhdp)) 4176 panic("flush_pagedep_deps: flush 3 failed"); 4177 ACQUIRE_LOCK(&lk); 4178 } 4179 if (error) 4180 ACQUIRE_LOCK(&lk); 4181 return (error); 4182} 4183 4184/* 4185 * Acquire exclusive access to a buffer. 4186 * Must be called with splbio blocked. 4187 * Return 1 if buffer was acquired. 4188 */ 4189static int 4190getdirtybuf(bpp, waitfor) 4191 struct buf **bpp; 4192 int waitfor; 4193{ 4194 struct buf *bp; 4195 4196 for (;;) { 4197 if ((bp = *bpp) == NULL) 4198 return (0); 4199 if ((bp->b_flags & B_BUSY) == 0) 4200 break; 4201 if (waitfor != MNT_WAIT) 4202 return (0); 4203 bp->b_flags |= B_WANTED; 4204 FREE_LOCK_INTERLOCKED(&lk); 4205 tsleep((caddr_t)bp, PRIBIO + 1, "sdsdty", 0); 4206 ACQUIRE_LOCK_INTERLOCKED(&lk); 4207 } 4208 if ((bp->b_flags & B_DELWRI) == 0) 4209 return (0); 4210 bremfree(bp); 4211 bp->b_flags |= B_BUSY; 4212 return (1); 4213} 4214 4215/* 4216 * Called whenever a buffer that is being invalidated or reallocated 4217 * contains dependencies. This should only happen if an I/O error has 4218 * occurred. The routine is called with the buffer locked. 4219 */ 4220void 4221softdep_deallocate_dependencies(bp) 4222 struct buf *bp; 4223{ 4224 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); 4225 panic("softdep_deallocate_dependencies: dangling deps"); 4226} 4227 4228/* 4229 * Function to handle asynchronous write errors in the filesystem. 4230 */ 4231void 4232softdep_error(func, error) 4233 char *func; 4234 int error; 4235{ 4236 /* XXX should do something better! */ 4237 printf("%s: got error %d while accessing filesystem\n", func, error); 4238} 4239