ffs_softdep.c revision 48225
195643Smarkm/* 295643Smarkm * Copyright 1998 Marshall Kirk McKusick. All Rights Reserved. 395643Smarkm * 495643Smarkm * The soft updates code is derived from the appendix of a University 595643Smarkm * of Michigan technical report (Gregory R. Ganger and Yale N. Patt, 695643Smarkm * "Soft Updates: A Solution to the Metadata Update Problem in File 795643Smarkm * Systems", CSE-TR-254-95, August 1995). 895643Smarkm * 995643Smarkm * The following are the copyrights and redistribution conditions that 1095643Smarkm * apply to this copy of the soft update software. For a license 1195643Smarkm * to use, redistribute or sell the soft update software under 1295643Smarkm * conditions other than those described here, please contact the 1395643Smarkm * author at one of the following addresses: 1495643Smarkm * 1595643Smarkm * Marshall Kirk McKusick mckusick@mckusick.com 1695643Smarkm * 1614 Oxford Street +1-510-843-9542 1795643Smarkm * Berkeley, CA 94709-1608 1895643Smarkm * USA 1995643Smarkm * 2095643Smarkm * Redistribution and use in source and binary forms, with or without 2195643Smarkm * modification, are permitted provided that the following conditions 2295643Smarkm * are met: 2395643Smarkm * 2495643Smarkm * 1. Redistributions of source code must retain the above copyright 2595643Smarkm * notice, this list of conditions and the following disclaimer. 2695643Smarkm * 2. Redistributions in binary form must reproduce the above copyright 2795643Smarkm * notice, this list of conditions and the following disclaimer in the 2895643Smarkm * documentation and/or other materials provided with the distribution. 2995643Smarkm * 3. None of the names of McKusick, Ganger, Patt, or the University of 3095643Smarkm * Michigan may be used to endorse or promote products derived from 3195643Smarkm * this software without specific prior written permission. 3295643Smarkm * 4. Redistributions in any form must be accompanied by information on 3395643Smarkm * how to obtain complete source code for any accompanying software 3495643Smarkm * that uses this software. This source code must either be included 3595643Smarkm * in the distribution or be available for no more than the cost of 3695643Smarkm * distribution plus a nominal fee, and must be freely redistributable 37 * under reasonable conditions. For an executable file, complete 38 * source code means the source code for all modules it contains. 39 * It does not mean source code for modules or files that typically 40 * accompany the operating system on which the executable file runs, 41 * e.g., standard library modules or system header files. 42 * 43 * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY 44 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 45 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 46 * DISCLAIMED. IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR 47 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 48 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 49 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 50 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 51 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 52 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 53 * SUCH DAMAGE. 54 * 55 * from: @(#)ffs_softdep.c 9.40 (McKusick) 6/15/99 56 * $Id: ffs_softdep.c,v 1.31 1999/06/16 23:27:55 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 void drain_output __P((struct vnode *, int)); 161static int getdirtybuf __P((struct buf **, int)); 162static void clear_remove __P((struct proc *)); 163static void clear_inodedeps __P((struct proc *)); 164static int flush_pagedep_deps __P((struct vnode *, struct mount *, 165 struct diraddhd *)); 166static int flush_inodedep_deps __P((struct fs *, ino_t)); 167static int handle_written_filepage __P((struct pagedep *, struct buf *)); 168static void diradd_inode_written __P((struct diradd *, struct inodedep *)); 169static int handle_written_inodeblock __P((struct inodedep *, struct buf *)); 170static void handle_allocdirect_partdone __P((struct allocdirect *)); 171static void handle_allocindir_partdone __P((struct allocindir *)); 172static void initiate_write_filepage __P((struct pagedep *, struct buf *)); 173static void handle_written_mkdir __P((struct mkdir *, int)); 174static void initiate_write_inodeblock __P((struct inodedep *, struct buf *)); 175static void handle_workitem_freefile __P((struct freefile *)); 176static void handle_workitem_remove __P((struct dirrem *)); 177static struct dirrem *newdirrem __P((struct buf *, struct inode *, 178 struct inode *, int)); 179static void free_diradd __P((struct diradd *)); 180static void free_allocindir __P((struct allocindir *, struct inodedep *)); 181static int indir_trunc __P((struct inode *, ufs_daddr_t, int, ufs_lbn_t, 182 long *)); 183static void deallocate_dependencies __P((struct buf *, struct inodedep *)); 184static void free_allocdirect __P((struct allocdirectlst *, 185 struct allocdirect *, int)); 186static int free_inodedep __P((struct inodedep *)); 187static void handle_workitem_freeblocks __P((struct freeblks *)); 188static void merge_inode_lists __P((struct inodedep *)); 189static void setup_allocindir_phase2 __P((struct buf *, struct inode *, 190 struct allocindir *)); 191static struct allocindir *newallocindir __P((struct inode *, int, ufs_daddr_t, 192 ufs_daddr_t)); 193static void handle_workitem_freefrag __P((struct freefrag *)); 194static struct freefrag *newfreefrag __P((struct inode *, ufs_daddr_t, long)); 195static void allocdirect_merge __P((struct allocdirectlst *, 196 struct allocdirect *, struct allocdirect *)); 197static struct bmsafemap *bmsafemap_lookup __P((struct buf *)); 198static int newblk_lookup __P((struct fs *, ufs_daddr_t, int, 199 struct newblk **)); 200static int inodedep_lookup __P((struct fs *, ino_t, int, struct inodedep **)); 201static int pagedep_lookup __P((struct inode *, ufs_lbn_t, int, 202 struct pagedep **)); 203static void pause_timer __P((void *)); 204static int request_cleanup __P((int, int)); 205static void add_to_worklist __P((struct worklist *)); 206 207/* 208 * Exported softdep operations. 209 */ 210struct bio_ops bioops = { 211 softdep_disk_io_initiation, /* io_start */ 212 softdep_disk_write_complete, /* io_complete */ 213 softdep_deallocate_dependencies, /* io_deallocate */ 214 softdep_fsync, /* io_fsync */ 215 softdep_process_worklist, /* io_sync */ 216}; 217 218/* 219 * Locking primitives. 220 * 221 * For a uniprocessor, all we need to do is protect against disk 222 * interrupts. For a multiprocessor, this lock would have to be 223 * a mutex. A single mutex is used throughout this file, though 224 * finer grain locking could be used if contention warranted it. 225 * 226 * For a multiprocessor, the sleep call would accept a lock and 227 * release it after the sleep processing was complete. In a uniprocessor 228 * implementation there is no such interlock, so we simple mark 229 * the places where it needs to be done with the `interlocked' form 230 * of the lock calls. Since the uniprocessor sleep already interlocks 231 * the spl, there is nothing that really needs to be done. 232 */ 233#ifndef /* NOT */ DEBUG 234static struct lockit { 235 int lkt_spl; 236} lk = { 0 }; 237#define ACQUIRE_LOCK(lk) (lk)->lkt_spl = splbio() 238#define FREE_LOCK(lk) splx((lk)->lkt_spl) 239#define ACQUIRE_LOCK_INTERLOCKED(lk) 240#define FREE_LOCK_INTERLOCKED(lk) 241 242#else /* DEBUG */ 243static struct lockit { 244 int lkt_spl; 245 pid_t lkt_held; 246} lk = { 0, -1 }; 247static int lockcnt; 248 249static void acquire_lock __P((struct lockit *)); 250static void free_lock __P((struct lockit *)); 251static void acquire_lock_interlocked __P((struct lockit *)); 252static void free_lock_interlocked __P((struct lockit *)); 253 254#define ACQUIRE_LOCK(lk) acquire_lock(lk) 255#define FREE_LOCK(lk) free_lock(lk) 256#define ACQUIRE_LOCK_INTERLOCKED(lk) acquire_lock_interlocked(lk) 257#define FREE_LOCK_INTERLOCKED(lk) free_lock_interlocked(lk) 258 259static void 260acquire_lock(lk) 261 struct lockit *lk; 262{ 263 264 if (lk->lkt_held != -1) { 265 if (lk->lkt_held == CURPROC->p_pid) 266 panic("softdep_lock: locking against myself"); 267 else 268 panic("softdep_lock: lock held by %d", lk->lkt_held); 269 } 270 lk->lkt_spl = splbio(); 271 lk->lkt_held = CURPROC->p_pid; 272 lockcnt++; 273} 274 275static void 276free_lock(lk) 277 struct lockit *lk; 278{ 279 280 if (lk->lkt_held == -1) 281 panic("softdep_unlock: lock not held"); 282 lk->lkt_held = -1; 283 splx(lk->lkt_spl); 284} 285 286static void 287acquire_lock_interlocked(lk) 288 struct lockit *lk; 289{ 290 291 if (lk->lkt_held != -1) { 292 if (lk->lkt_held == CURPROC->p_pid) 293 panic("softdep_lock_interlocked: locking against self"); 294 else 295 panic("softdep_lock_interlocked: lock held by %d", 296 lk->lkt_held); 297 } 298 lk->lkt_held = CURPROC->p_pid; 299 lockcnt++; 300} 301 302static void 303free_lock_interlocked(lk) 304 struct lockit *lk; 305{ 306 307 if (lk->lkt_held == -1) 308 panic("softdep_unlock_interlocked: lock not held"); 309 lk->lkt_held = -1; 310} 311#endif /* DEBUG */ 312 313/* 314 * Place holder for real semaphores. 315 */ 316struct sema { 317 int value; 318 pid_t holder; 319 char *name; 320 int prio; 321 int timo; 322}; 323static void sema_init __P((struct sema *, char *, int, int)); 324static int sema_get __P((struct sema *, struct lockit *)); 325static void sema_release __P((struct sema *)); 326 327static void 328sema_init(semap, name, prio, timo) 329 struct sema *semap; 330 char *name; 331 int prio, timo; 332{ 333 334 semap->holder = -1; 335 semap->value = 0; 336 semap->name = name; 337 semap->prio = prio; 338 semap->timo = timo; 339} 340 341static int 342sema_get(semap, interlock) 343 struct sema *semap; 344 struct lockit *interlock; 345{ 346 347 if (semap->value++ > 0) { 348 if (interlock != NULL) 349 FREE_LOCK_INTERLOCKED(interlock); 350 tsleep((caddr_t)semap, semap->prio, semap->name, semap->timo); 351 if (interlock != NULL) { 352 ACQUIRE_LOCK_INTERLOCKED(interlock); 353 FREE_LOCK(interlock); 354 } 355 return (0); 356 } 357 semap->holder = CURPROC->p_pid; 358 if (interlock != NULL) 359 FREE_LOCK(interlock); 360 return (1); 361} 362 363static void 364sema_release(semap) 365 struct sema *semap; 366{ 367 368 if (semap->value <= 0 || semap->holder != CURPROC->p_pid) 369 panic("sema_release: not held"); 370 if (--semap->value > 0) { 371 semap->value = 0; 372 wakeup(semap); 373 } 374 semap->holder = -1; 375} 376 377/* 378 * Worklist queue management. 379 * These routines require that the lock be held. 380 */ 381#ifndef /* NOT */ DEBUG 382#define WORKLIST_INSERT(head, item) do { \ 383 (item)->wk_state |= ONWORKLIST; \ 384 LIST_INSERT_HEAD(head, item, wk_list); \ 385} while (0) 386#define WORKLIST_REMOVE(item) do { \ 387 (item)->wk_state &= ~ONWORKLIST; \ 388 LIST_REMOVE(item, wk_list); \ 389} while (0) 390#define WORKITEM_FREE(item, type) FREE(item, DtoM(type)) 391 392#else /* DEBUG */ 393static void worklist_insert __P((struct workhead *, struct worklist *)); 394static void worklist_remove __P((struct worklist *)); 395static void workitem_free __P((struct worklist *, int)); 396 397#define WORKLIST_INSERT(head, item) worklist_insert(head, item) 398#define WORKLIST_REMOVE(item) worklist_remove(item) 399#define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type) 400 401static void 402worklist_insert(head, item) 403 struct workhead *head; 404 struct worklist *item; 405{ 406 407 if (lk.lkt_held == -1) 408 panic("worklist_insert: lock not held"); 409 if (item->wk_state & ONWORKLIST) 410 panic("worklist_insert: already on list"); 411 item->wk_state |= ONWORKLIST; 412 LIST_INSERT_HEAD(head, item, wk_list); 413} 414 415static void 416worklist_remove(item) 417 struct worklist *item; 418{ 419 420 if (lk.lkt_held == -1) 421 panic("worklist_remove: lock not held"); 422 if ((item->wk_state & ONWORKLIST) == 0) 423 panic("worklist_remove: not on list"); 424 item->wk_state &= ~ONWORKLIST; 425 LIST_REMOVE(item, wk_list); 426} 427 428static void 429workitem_free(item, type) 430 struct worklist *item; 431 int type; 432{ 433 434 if (item->wk_state & ONWORKLIST) 435 panic("workitem_free: still on list"); 436 if (item->wk_type != type) 437 panic("workitem_free: type mismatch"); 438 FREE(item, DtoM(type)); 439} 440#endif /* DEBUG */ 441 442/* 443 * Workitem queue management 444 */ 445static struct workhead softdep_workitem_pending; 446static int softdep_worklist_busy; 447static int max_softdeps; /* maximum number of structs before slowdown */ 448static int tickdelay = 2; /* number of ticks to pause during slowdown */ 449static int proc_waiting; /* tracks whether we have a timeout posted */ 450static struct proc *filesys_syncer; /* proc of filesystem syncer process */ 451static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 452#define FLUSH_INODES 1 453static int req_clear_remove; /* syncer process flush some freeblks */ 454#define FLUSH_REMOVE 2 455/* 456 * runtime statistics 457 */ 458static int stat_blk_limit_push; /* number of times block limit neared */ 459static int stat_ino_limit_push; /* number of times inode limit neared */ 460static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 461static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 462static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 463static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 464static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 465static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 466#ifdef DEBUG 467#include <vm/vm.h> 468#include <sys/sysctl.h> 469#if defined(__FreeBSD__) 470SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, ""); 471SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, ""); 472SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,""); 473SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,""); 474SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, ""); 475SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, ""); 476SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, ""); 477SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, ""); 478SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, ""); 479SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, ""); 480#else /* !__FreeBSD__ */ 481struct ctldebug debug20 = { "max_softdeps", &max_softdeps }; 482struct ctldebug debug21 = { "tickdelay", &tickdelay }; 483struct ctldebug debug23 = { "blk_limit_push", &stat_blk_limit_push }; 484struct ctldebug debug24 = { "ino_limit_push", &stat_ino_limit_push }; 485struct ctldebug debug25 = { "blk_limit_hit", &stat_blk_limit_hit }; 486struct ctldebug debug26 = { "ino_limit_hit", &stat_ino_limit_hit }; 487struct ctldebug debug27 = { "indir_blk_ptrs", &stat_indir_blk_ptrs }; 488struct ctldebug debug28 = { "inode_bitmap", &stat_inode_bitmap }; 489struct ctldebug debug29 = { "direct_blk_ptrs", &stat_direct_blk_ptrs }; 490struct ctldebug debug30 = { "dir_entry", &stat_dir_entry }; 491#endif /* !__FreeBSD__ */ 492 493#endif /* DEBUG */ 494 495/* 496 * Add an item to the end of the work queue. 497 * This routine requires that the lock be held. 498 * This is the only routine that adds items to the list. 499 * The following routine is the only one that removes items 500 * and does so in order from first to last. 501 */ 502static void 503add_to_worklist(wk) 504 struct worklist *wk; 505{ 506 static struct worklist *worklist_tail; 507 508 if (wk->wk_state & ONWORKLIST) 509 panic("add_to_worklist: already on list"); 510 wk->wk_state |= ONWORKLIST; 511 if (LIST_FIRST(&softdep_workitem_pending) == NULL) 512 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list); 513 else 514 LIST_INSERT_AFTER(worklist_tail, wk, wk_list); 515 worklist_tail = wk; 516} 517 518/* 519 * Process that runs once per second to handle items in the background queue. 520 * 521 * Note that we ensure that everything is done in the order in which they 522 * appear in the queue. The code below depends on this property to ensure 523 * that blocks of a file are freed before the inode itself is freed. This 524 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 525 * until all the old ones have been purged from the dependency lists. 526 */ 527int 528softdep_process_worklist(matchmnt) 529 struct mount *matchmnt; 530{ 531 struct proc *p = CURPROC; 532 struct worklist *wk; 533 struct fs *matchfs; 534 int matchcnt; 535 536 /* 537 * Record the process identifier of our caller so that we can give 538 * this process preferential treatment in request_cleanup below. 539 */ 540 filesys_syncer = p; 541 matchcnt = 0; 542 matchfs = NULL; 543 if (matchmnt != NULL) 544 matchfs = VFSTOUFS(matchmnt)->um_fs; 545 /* 546 * There is no danger of having multiple processes run this 547 * code. It is single threaded solely so that softdep_flushfiles 548 * (below) can get an accurate count of the number of items 549 * related to its mount point that are in the list. 550 */ 551 if (softdep_worklist_busy && matchmnt == NULL) 552 return (-1); 553 /* 554 * If requested, try removing inode or removal dependencies. 555 */ 556 if (req_clear_inodedeps) { 557 clear_inodedeps(p); 558 req_clear_inodedeps = 0; 559 wakeup(&proc_waiting); 560 } 561 if (req_clear_remove) { 562 clear_remove(p); 563 req_clear_remove = 0; 564 wakeup(&proc_waiting); 565 } 566 ACQUIRE_LOCK(&lk); 567 while ((wk = LIST_FIRST(&softdep_workitem_pending)) != 0) { 568 WORKLIST_REMOVE(wk); 569 FREE_LOCK(&lk); 570 switch (wk->wk_type) { 571 572 case D_DIRREM: 573 /* removal of a directory entry */ 574 if (WK_DIRREM(wk)->dm_mnt == matchmnt) 575 matchcnt += 1; 576 handle_workitem_remove(WK_DIRREM(wk)); 577 break; 578 579 case D_FREEBLKS: 580 /* releasing blocks and/or fragments from a file */ 581 if (WK_FREEBLKS(wk)->fb_fs == matchfs) 582 matchcnt += 1; 583 handle_workitem_freeblocks(WK_FREEBLKS(wk)); 584 break; 585 586 case D_FREEFRAG: 587 /* releasing a fragment when replaced as a file grows */ 588 if (WK_FREEFRAG(wk)->ff_fs == matchfs) 589 matchcnt += 1; 590 handle_workitem_freefrag(WK_FREEFRAG(wk)); 591 break; 592 593 case D_FREEFILE: 594 /* releasing an inode when its link count drops to 0 */ 595 if (WK_FREEFILE(wk)->fx_fs == matchfs) 596 matchcnt += 1; 597 handle_workitem_freefile(WK_FREEFILE(wk)); 598 break; 599 600 default: 601 panic("%s_process_worklist: Unknown type %s", 602 "softdep", TYPENAME(wk->wk_type)); 603 /* NOTREACHED */ 604 } 605 if (softdep_worklist_busy && matchmnt == NULL) 606 return (-1); 607 /* 608 * If requested, try removing inode or removal dependencies. 609 */ 610 if (req_clear_inodedeps) { 611 clear_inodedeps(p); 612 req_clear_inodedeps = 0; 613 wakeup(&proc_waiting); 614 } 615 if (req_clear_remove) { 616 clear_remove(p); 617 req_clear_remove = 0; 618 wakeup(&proc_waiting); 619 } 620 ACQUIRE_LOCK(&lk); 621 } 622 FREE_LOCK(&lk); 623 return (matchcnt); 624} 625 626/* 627 * Purge the work list of all items associated with a particular mount point. 628 */ 629int 630softdep_flushfiles(oldmnt, flags, p) 631 struct mount *oldmnt; 632 int flags; 633 struct proc *p; 634{ 635 struct vnode *devvp; 636 int error, loopcnt; 637 638 /* 639 * Await our turn to clear out the queue. 640 */ 641 while (softdep_worklist_busy) 642 tsleep(&lbolt, PRIBIO, "softflush", 0); 643 softdep_worklist_busy = 1; 644 if ((error = ffs_flushfiles(oldmnt, flags, p)) != 0) { 645 softdep_worklist_busy = 0; 646 return (error); 647 } 648 /* 649 * Alternately flush the block device associated with the mount 650 * point and process any dependencies that the flushing 651 * creates. In theory, this loop can happen at most twice, 652 * but we give it a few extra just to be sure. 653 */ 654 devvp = VFSTOUFS(oldmnt)->um_devvp; 655 for (loopcnt = 10; loopcnt > 0; loopcnt--) { 656 if (softdep_process_worklist(oldmnt) == 0) { 657 /* 658 * Do another flush in case any vnodes were brought in 659 * as part of the cleanup operations. 660 */ 661 if ((error = ffs_flushfiles(oldmnt, flags, p)) != 0) 662 break; 663 /* 664 * If we still found nothing to do, we are really done. 665 */ 666 if (softdep_process_worklist(oldmnt) == 0) 667 break; 668 } 669 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, p); 670 error = VOP_FSYNC(devvp, p->p_ucred, MNT_WAIT, p); 671 VOP_UNLOCK(devvp, 0, p); 672 if (error) 673 break; 674 } 675 softdep_worklist_busy = 0; 676 /* 677 * If we are unmounting then it is an error to fail. If we 678 * are simply trying to downgrade to read-only, then filesystem 679 * activity can keep us busy forever, so we just fail with EBUSY. 680 */ 681 if (loopcnt == 0) { 682 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 683 panic("softdep_flushfiles: looping"); 684 error = EBUSY; 685 } 686 return (error); 687} 688 689/* 690 * Structure hashing. 691 * 692 * There are three types of structures that can be looked up: 693 * 1) pagedep structures identified by mount point, inode number, 694 * and logical block. 695 * 2) inodedep structures identified by mount point and inode number. 696 * 3) newblk structures identified by mount point and 697 * physical block number. 698 * 699 * The "pagedep" and "inodedep" dependency structures are hashed 700 * separately from the file blocks and inodes to which they correspond. 701 * This separation helps when the in-memory copy of an inode or 702 * file block must be replaced. It also obviates the need to access 703 * an inode or file page when simply updating (or de-allocating) 704 * dependency structures. Lookup of newblk structures is needed to 705 * find newly allocated blocks when trying to associate them with 706 * their allocdirect or allocindir structure. 707 * 708 * The lookup routines optionally create and hash a new instance when 709 * an existing entry is not found. 710 */ 711#define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 712 713/* 714 * Structures and routines associated with pagedep caching. 715 */ 716LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl; 717u_long pagedep_hash; /* size of hash table - 1 */ 718#define PAGEDEP_HASH(mp, inum, lbn) \ 719 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \ 720 pagedep_hash]) 721static struct sema pagedep_in_progress; 722 723/* 724 * Look up a pagedep. Return 1 if found, 0 if not found. 725 * If not found, allocate if DEPALLOC flag is passed. 726 * Found or allocated entry is returned in pagedeppp. 727 * This routine must be called with splbio interrupts blocked. 728 */ 729static int 730pagedep_lookup(ip, lbn, flags, pagedeppp) 731 struct inode *ip; 732 ufs_lbn_t lbn; 733 int flags; 734 struct pagedep **pagedeppp; 735{ 736 struct pagedep *pagedep; 737 struct pagedep_hashhead *pagedephd; 738 struct mount *mp; 739 int i; 740 741#ifdef DEBUG 742 if (lk.lkt_held == -1) 743 panic("pagedep_lookup: lock not held"); 744#endif 745 mp = ITOV(ip)->v_mount; 746 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn); 747top: 748 for (pagedep = LIST_FIRST(pagedephd); pagedep; 749 pagedep = LIST_NEXT(pagedep, pd_hash)) 750 if (ip->i_number == pagedep->pd_ino && 751 lbn == pagedep->pd_lbn && 752 mp == pagedep->pd_mnt) 753 break; 754 if (pagedep) { 755 *pagedeppp = pagedep; 756 return (1); 757 } 758 if ((flags & DEPALLOC) == 0) { 759 *pagedeppp = NULL; 760 return (0); 761 } 762 if (sema_get(&pagedep_in_progress, &lk) == 0) { 763 ACQUIRE_LOCK(&lk); 764 goto top; 765 } 766 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP, 767 M_WAITOK); 768 bzero(pagedep, sizeof(struct pagedep)); 769 pagedep->pd_list.wk_type = D_PAGEDEP; 770 pagedep->pd_mnt = mp; 771 pagedep->pd_ino = ip->i_number; 772 pagedep->pd_lbn = lbn; 773 LIST_INIT(&pagedep->pd_dirremhd); 774 LIST_INIT(&pagedep->pd_pendinghd); 775 for (i = 0; i < DAHASHSZ; i++) 776 LIST_INIT(&pagedep->pd_diraddhd[i]); 777 ACQUIRE_LOCK(&lk); 778 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 779 sema_release(&pagedep_in_progress); 780 *pagedeppp = pagedep; 781 return (0); 782} 783 784/* 785 * Structures and routines associated with inodedep caching. 786 */ 787LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl; 788static u_long inodedep_hash; /* size of hash table - 1 */ 789static long num_inodedep; /* number of inodedep allocated */ 790#define INODEDEP_HASH(fs, inum) \ 791 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash]) 792static struct sema inodedep_in_progress; 793 794/* 795 * Look up a inodedep. Return 1 if found, 0 if not found. 796 * If not found, allocate if DEPALLOC flag is passed. 797 * Found or allocated entry is returned in inodedeppp. 798 * This routine must be called with splbio interrupts blocked. 799 */ 800static int 801inodedep_lookup(fs, inum, flags, inodedeppp) 802 struct fs *fs; 803 ino_t inum; 804 int flags; 805 struct inodedep **inodedeppp; 806{ 807 struct inodedep *inodedep; 808 struct inodedep_hashhead *inodedephd; 809 int firsttry; 810 811#ifdef DEBUG 812 if (lk.lkt_held == -1) 813 panic("inodedep_lookup: lock not held"); 814#endif 815 firsttry = 1; 816 inodedephd = INODEDEP_HASH(fs, inum); 817top: 818 for (inodedep = LIST_FIRST(inodedephd); inodedep; 819 inodedep = LIST_NEXT(inodedep, id_hash)) 820 if (inum == inodedep->id_ino && fs == inodedep->id_fs) 821 break; 822 if (inodedep) { 823 *inodedeppp = inodedep; 824 return (1); 825 } 826 if ((flags & DEPALLOC) == 0) { 827 *inodedeppp = NULL; 828 return (0); 829 } 830 /* 831 * If we are over our limit, try to improve the situation. 832 */ 833 if (num_inodedep > max_softdeps && firsttry && speedup_syncer() == 0 && 834 request_cleanup(FLUSH_INODES, 1)) { 835 firsttry = 0; 836 goto top; 837 } 838 if (sema_get(&inodedep_in_progress, &lk) == 0) { 839 ACQUIRE_LOCK(&lk); 840 goto top; 841 } 842 num_inodedep += 1; 843 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep), 844 M_INODEDEP, M_WAITOK); 845 inodedep->id_list.wk_type = D_INODEDEP; 846 inodedep->id_fs = fs; 847 inodedep->id_ino = inum; 848 inodedep->id_state = ALLCOMPLETE; 849 inodedep->id_nlinkdelta = 0; 850 inodedep->id_savedino = NULL; 851 inodedep->id_savedsize = -1; 852 inodedep->id_buf = NULL; 853 LIST_INIT(&inodedep->id_pendinghd); 854 LIST_INIT(&inodedep->id_inowait); 855 LIST_INIT(&inodedep->id_bufwait); 856 TAILQ_INIT(&inodedep->id_inoupdt); 857 TAILQ_INIT(&inodedep->id_newinoupdt); 858 ACQUIRE_LOCK(&lk); 859 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 860 sema_release(&inodedep_in_progress); 861 *inodedeppp = inodedep; 862 return (0); 863} 864 865/* 866 * Structures and routines associated with newblk caching. 867 */ 868LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl; 869u_long newblk_hash; /* size of hash table - 1 */ 870#define NEWBLK_HASH(fs, inum) \ 871 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash]) 872static struct sema newblk_in_progress; 873 874/* 875 * Look up a newblk. Return 1 if found, 0 if not found. 876 * If not found, allocate if DEPALLOC flag is passed. 877 * Found or allocated entry is returned in newblkpp. 878 */ 879static int 880newblk_lookup(fs, newblkno, flags, newblkpp) 881 struct fs *fs; 882 ufs_daddr_t newblkno; 883 int flags; 884 struct newblk **newblkpp; 885{ 886 struct newblk *newblk; 887 struct newblk_hashhead *newblkhd; 888 889 newblkhd = NEWBLK_HASH(fs, newblkno); 890top: 891 for (newblk = LIST_FIRST(newblkhd); newblk; 892 newblk = LIST_NEXT(newblk, nb_hash)) 893 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs) 894 break; 895 if (newblk) { 896 *newblkpp = newblk; 897 return (1); 898 } 899 if ((flags & DEPALLOC) == 0) { 900 *newblkpp = NULL; 901 return (0); 902 } 903 if (sema_get(&newblk_in_progress, 0) == 0) 904 goto top; 905 MALLOC(newblk, struct newblk *, sizeof(struct newblk), 906 M_NEWBLK, M_WAITOK); 907 newblk->nb_state = 0; 908 newblk->nb_fs = fs; 909 newblk->nb_newblkno = newblkno; 910 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 911 sema_release(&newblk_in_progress); 912 *newblkpp = newblk; 913 return (0); 914} 915 916/* 917 * Executed during filesystem system initialization before 918 * mounting any file systems. 919 */ 920void 921softdep_initialize() 922{ 923 924 LIST_INIT(&mkdirlisthd); 925 LIST_INIT(&softdep_workitem_pending); 926 max_softdeps = desiredvnodes * 8; 927 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 928 &pagedep_hash); 929 sema_init(&pagedep_in_progress, "pagedep", PRIBIO, 0); 930 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash); 931 sema_init(&inodedep_in_progress, "inodedep", PRIBIO, 0); 932 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash); 933 sema_init(&newblk_in_progress, "newblk", PRIBIO, 0); 934} 935 936/* 937 * Called at mount time to notify the dependency code that a 938 * filesystem wishes to use it. 939 */ 940int 941softdep_mount(devvp, mp, fs, cred) 942 struct vnode *devvp; 943 struct mount *mp; 944 struct fs *fs; 945 struct ucred *cred; 946{ 947 struct csum cstotal; 948 struct cg *cgp; 949 struct buf *bp; 950 int error, cyl; 951 952 mp->mnt_flag &= ~MNT_ASYNC; 953 mp->mnt_flag |= MNT_SOFTDEP; 954 /* 955 * When doing soft updates, the counters in the 956 * superblock may have gotten out of sync, so we have 957 * to scan the cylinder groups and recalculate them. 958 */ 959 if (fs->fs_clean != 0) 960 return (0); 961 bzero(&cstotal, sizeof cstotal); 962 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 963 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 964 fs->fs_cgsize, cred, &bp)) != 0) { 965 brelse(bp); 966 return (error); 967 } 968 cgp = (struct cg *)bp->b_data; 969 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 970 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 971 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 972 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 973 fs->fs_cs(fs, cyl) = cgp->cg_cs; 974 brelse(bp); 975 } 976#ifdef DEBUG 977 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 978 printf("ffs_mountfs: superblock updated for soft updates\n"); 979#endif 980 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 981 return (0); 982} 983 984/* 985 * Protecting the freemaps (or bitmaps). 986 * 987 * To eliminate the need to execute fsck before mounting a file system 988 * after a power failure, one must (conservatively) guarantee that the 989 * on-disk copy of the bitmaps never indicate that a live inode or block is 990 * free. So, when a block or inode is allocated, the bitmap should be 991 * updated (on disk) before any new pointers. When a block or inode is 992 * freed, the bitmap should not be updated until all pointers have been 993 * reset. The latter dependency is handled by the delayed de-allocation 994 * approach described below for block and inode de-allocation. The former 995 * dependency is handled by calling the following procedure when a block or 996 * inode is allocated. When an inode is allocated an "inodedep" is created 997 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 998 * Each "inodedep" is also inserted into the hash indexing structure so 999 * that any additional link additions can be made dependent on the inode 1000 * allocation. 1001 * 1002 * The ufs file system maintains a number of free block counts (e.g., per 1003 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 1004 * in addition to the bitmaps. These counts are used to improve efficiency 1005 * during allocation and therefore must be consistent with the bitmaps. 1006 * There is no convenient way to guarantee post-crash consistency of these 1007 * counts with simple update ordering, for two main reasons: (1) The counts 1008 * and bitmaps for a single cylinder group block are not in the same disk 1009 * sector. If a disk write is interrupted (e.g., by power failure), one may 1010 * be written and the other not. (2) Some of the counts are located in the 1011 * superblock rather than the cylinder group block. So, we focus our soft 1012 * updates implementation on protecting the bitmaps. When mounting a 1013 * filesystem, we recompute the auxiliary counts from the bitmaps. 1014 */ 1015 1016/* 1017 * Called just after updating the cylinder group block to allocate an inode. 1018 */ 1019void 1020softdep_setup_inomapdep(bp, ip, newinum) 1021 struct buf *bp; /* buffer for cylgroup block with inode map */ 1022 struct inode *ip; /* inode related to allocation */ 1023 ino_t newinum; /* new inode number being allocated */ 1024{ 1025 struct inodedep *inodedep; 1026 struct bmsafemap *bmsafemap; 1027 1028 /* 1029 * Create a dependency for the newly allocated inode. 1030 * Panic if it already exists as something is seriously wrong. 1031 * Otherwise add it to the dependency list for the buffer holding 1032 * the cylinder group map from which it was allocated. 1033 */ 1034 ACQUIRE_LOCK(&lk); 1035 if (inodedep_lookup(ip->i_fs, newinum, DEPALLOC, &inodedep) != 0) 1036 panic("softdep_setup_inomapdep: found inode"); 1037 inodedep->id_buf = bp; 1038 inodedep->id_state &= ~DEPCOMPLETE; 1039 bmsafemap = bmsafemap_lookup(bp); 1040 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 1041 FREE_LOCK(&lk); 1042} 1043 1044/* 1045 * Called just after updating the cylinder group block to 1046 * allocate block or fragment. 1047 */ 1048void 1049softdep_setup_blkmapdep(bp, fs, newblkno) 1050 struct buf *bp; /* buffer for cylgroup block with block map */ 1051 struct fs *fs; /* filesystem doing allocation */ 1052 ufs_daddr_t newblkno; /* number of newly allocated block */ 1053{ 1054 struct newblk *newblk; 1055 struct bmsafemap *bmsafemap; 1056 1057 /* 1058 * Create a dependency for the newly allocated block. 1059 * Add it to the dependency list for the buffer holding 1060 * the cylinder group map from which it was allocated. 1061 */ 1062 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0) 1063 panic("softdep_setup_blkmapdep: found block"); 1064 ACQUIRE_LOCK(&lk); 1065 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp); 1066 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 1067 FREE_LOCK(&lk); 1068} 1069 1070/* 1071 * Find the bmsafemap associated with a cylinder group buffer. 1072 * If none exists, create one. The buffer must be locked when 1073 * this routine is called and this routine must be called with 1074 * splbio interrupts blocked. 1075 */ 1076static struct bmsafemap * 1077bmsafemap_lookup(bp) 1078 struct buf *bp; 1079{ 1080 struct bmsafemap *bmsafemap; 1081 struct worklist *wk; 1082 1083#ifdef DEBUG 1084 if (lk.lkt_held == -1) 1085 panic("bmsafemap_lookup: lock not held"); 1086#endif 1087 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = LIST_NEXT(wk, wk_list)) 1088 if (wk->wk_type == D_BMSAFEMAP) 1089 return (WK_BMSAFEMAP(wk)); 1090 FREE_LOCK(&lk); 1091 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap), 1092 M_BMSAFEMAP, M_WAITOK); 1093 bmsafemap->sm_list.wk_type = D_BMSAFEMAP; 1094 bmsafemap->sm_list.wk_state = 0; 1095 bmsafemap->sm_buf = bp; 1096 LIST_INIT(&bmsafemap->sm_allocdirecthd); 1097 LIST_INIT(&bmsafemap->sm_allocindirhd); 1098 LIST_INIT(&bmsafemap->sm_inodedephd); 1099 LIST_INIT(&bmsafemap->sm_newblkhd); 1100 ACQUIRE_LOCK(&lk); 1101 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 1102 return (bmsafemap); 1103} 1104 1105/* 1106 * Direct block allocation dependencies. 1107 * 1108 * When a new block is allocated, the corresponding disk locations must be 1109 * initialized (with zeros or new data) before the on-disk inode points to 1110 * them. Also, the freemap from which the block was allocated must be 1111 * updated (on disk) before the inode's pointer. These two dependencies are 1112 * independent of each other and are needed for all file blocks and indirect 1113 * blocks that are pointed to directly by the inode. Just before the 1114 * "in-core" version of the inode is updated with a newly allocated block 1115 * number, a procedure (below) is called to setup allocation dependency 1116 * structures. These structures are removed when the corresponding 1117 * dependencies are satisfied or when the block allocation becomes obsolete 1118 * (i.e., the file is deleted, the block is de-allocated, or the block is a 1119 * fragment that gets upgraded). All of these cases are handled in 1120 * procedures described later. 1121 * 1122 * When a file extension causes a fragment to be upgraded, either to a larger 1123 * fragment or to a full block, the on-disk location may change (if the 1124 * previous fragment could not simply be extended). In this case, the old 1125 * fragment must be de-allocated, but not until after the inode's pointer has 1126 * been updated. In most cases, this is handled by later procedures, which 1127 * will construct a "freefrag" structure to be added to the workitem queue 1128 * when the inode update is complete (or obsolete). The main exception to 1129 * this is when an allocation occurs while a pending allocation dependency 1130 * (for the same block pointer) remains. This case is handled in the main 1131 * allocation dependency setup procedure by immediately freeing the 1132 * unreferenced fragments. 1133 */ 1134void 1135softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 1136 struct inode *ip; /* inode to which block is being added */ 1137 ufs_lbn_t lbn; /* block pointer within inode */ 1138 ufs_daddr_t newblkno; /* disk block number being added */ 1139 ufs_daddr_t oldblkno; /* previous block number, 0 unless frag */ 1140 long newsize; /* size of new block */ 1141 long oldsize; /* size of new block */ 1142 struct buf *bp; /* bp for allocated block */ 1143{ 1144 struct allocdirect *adp, *oldadp; 1145 struct allocdirectlst *adphead; 1146 struct bmsafemap *bmsafemap; 1147 struct inodedep *inodedep; 1148 struct pagedep *pagedep; 1149 struct newblk *newblk; 1150 1151 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect), 1152 M_ALLOCDIRECT, M_WAITOK); 1153 bzero(adp, sizeof(struct allocdirect)); 1154 adp->ad_list.wk_type = D_ALLOCDIRECT; 1155 adp->ad_lbn = lbn; 1156 adp->ad_newblkno = newblkno; 1157 adp->ad_oldblkno = oldblkno; 1158 adp->ad_newsize = newsize; 1159 adp->ad_oldsize = oldsize; 1160 adp->ad_state = ATTACHED; 1161 if (newblkno == oldblkno) 1162 adp->ad_freefrag = NULL; 1163 else 1164 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1165 1166 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1167 panic("softdep_setup_allocdirect: lost block"); 1168 1169 ACQUIRE_LOCK(&lk); 1170 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 1171 adp->ad_inodedep = inodedep; 1172 1173 if (newblk->nb_state == DEPCOMPLETE) { 1174 adp->ad_state |= DEPCOMPLETE; 1175 adp->ad_buf = NULL; 1176 } else { 1177 bmsafemap = newblk->nb_bmsafemap; 1178 adp->ad_buf = bmsafemap->sm_buf; 1179 LIST_REMOVE(newblk, nb_deps); 1180 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1181 } 1182 LIST_REMOVE(newblk, nb_hash); 1183 FREE(newblk, M_NEWBLK); 1184 1185 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list); 1186 if (lbn >= NDADDR) { 1187 /* allocating an indirect block */ 1188 if (oldblkno != 0) 1189 panic("softdep_setup_allocdirect: non-zero indir"); 1190 } else { 1191 /* 1192 * Allocating a direct block. 1193 * 1194 * If we are allocating a directory block, then we must 1195 * allocate an associated pagedep to track additions and 1196 * deletions. 1197 */ 1198 if ((ip->i_mode & IFMT) == IFDIR && 1199 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1200 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 1201 } 1202 /* 1203 * The list of allocdirects must be kept in sorted and ascending 1204 * order so that the rollback routines can quickly determine the 1205 * first uncommitted block (the size of the file stored on disk 1206 * ends at the end of the lowest committed fragment, or if there 1207 * are no fragments, at the end of the highest committed block). 1208 * Since files generally grow, the typical case is that the new 1209 * block is to be added at the end of the list. We speed this 1210 * special case by checking against the last allocdirect in the 1211 * list before laboriously traversing the list looking for the 1212 * insertion point. 1213 */ 1214 adphead = &inodedep->id_newinoupdt; 1215 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1216 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1217 /* insert at end of list */ 1218 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1219 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1220 allocdirect_merge(adphead, adp, oldadp); 1221 FREE_LOCK(&lk); 1222 return; 1223 } 1224 for (oldadp = TAILQ_FIRST(adphead); oldadp; 1225 oldadp = TAILQ_NEXT(oldadp, ad_next)) { 1226 if (oldadp->ad_lbn >= lbn) 1227 break; 1228 } 1229 if (oldadp == NULL) 1230 panic("softdep_setup_allocdirect: lost entry"); 1231 /* insert in middle of list */ 1232 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1233 if (oldadp->ad_lbn == lbn) 1234 allocdirect_merge(adphead, adp, oldadp); 1235 FREE_LOCK(&lk); 1236} 1237 1238/* 1239 * Replace an old allocdirect dependency with a newer one. 1240 * This routine must be called with splbio interrupts blocked. 1241 */ 1242static void 1243allocdirect_merge(adphead, newadp, oldadp) 1244 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 1245 struct allocdirect *newadp; /* allocdirect being added */ 1246 struct allocdirect *oldadp; /* existing allocdirect being checked */ 1247{ 1248 struct freefrag *freefrag; 1249 1250#ifdef DEBUG 1251 if (lk.lkt_held == -1) 1252 panic("allocdirect_merge: lock not held"); 1253#endif 1254 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 1255 newadp->ad_oldsize != oldadp->ad_newsize || 1256 newadp->ad_lbn >= NDADDR) 1257 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d", 1258 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn, 1259 NDADDR); 1260 newadp->ad_oldblkno = oldadp->ad_oldblkno; 1261 newadp->ad_oldsize = oldadp->ad_oldsize; 1262 /* 1263 * If the old dependency had a fragment to free or had never 1264 * previously had a block allocated, then the new dependency 1265 * can immediately post its freefrag and adopt the old freefrag. 1266 * This action is done by swapping the freefrag dependencies. 1267 * The new dependency gains the old one's freefrag, and the 1268 * old one gets the new one and then immediately puts it on 1269 * the worklist when it is freed by free_allocdirect. It is 1270 * not possible to do this swap when the old dependency had a 1271 * non-zero size but no previous fragment to free. This condition 1272 * arises when the new block is an extension of the old block. 1273 * Here, the first part of the fragment allocated to the new 1274 * dependency is part of the block currently claimed on disk by 1275 * the old dependency, so cannot legitimately be freed until the 1276 * conditions for the new dependency are fulfilled. 1277 */ 1278 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 1279 freefrag = newadp->ad_freefrag; 1280 newadp->ad_freefrag = oldadp->ad_freefrag; 1281 oldadp->ad_freefrag = freefrag; 1282 } 1283 free_allocdirect(adphead, oldadp, 0); 1284} 1285 1286/* 1287 * Allocate a new freefrag structure if needed. 1288 */ 1289static struct freefrag * 1290newfreefrag(ip, blkno, size) 1291 struct inode *ip; 1292 ufs_daddr_t blkno; 1293 long size; 1294{ 1295 struct freefrag *freefrag; 1296 struct fs *fs; 1297 1298 if (blkno == 0) 1299 return (NULL); 1300 fs = ip->i_fs; 1301 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 1302 panic("newfreefrag: frag size"); 1303 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag), 1304 M_FREEFRAG, M_WAITOK); 1305 freefrag->ff_list.wk_type = D_FREEFRAG; 1306 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */ 1307 freefrag->ff_inum = ip->i_number; 1308 freefrag->ff_fs = fs; 1309 freefrag->ff_devvp = ip->i_devvp; 1310 freefrag->ff_blkno = blkno; 1311 freefrag->ff_fragsize = size; 1312 return (freefrag); 1313} 1314 1315/* 1316 * This workitem de-allocates fragments that were replaced during 1317 * file block allocation. 1318 */ 1319static void 1320handle_workitem_freefrag(freefrag) 1321 struct freefrag *freefrag; 1322{ 1323 struct inode tip; 1324 1325 tip.i_fs = freefrag->ff_fs; 1326 tip.i_devvp = freefrag->ff_devvp; 1327 tip.i_dev = freefrag->ff_devvp->v_rdev; 1328 tip.i_number = freefrag->ff_inum; 1329 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */ 1330 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize); 1331 FREE(freefrag, M_FREEFRAG); 1332} 1333 1334/* 1335 * Indirect block allocation dependencies. 1336 * 1337 * The same dependencies that exist for a direct block also exist when 1338 * a new block is allocated and pointed to by an entry in a block of 1339 * indirect pointers. The undo/redo states described above are also 1340 * used here. Because an indirect block contains many pointers that 1341 * may have dependencies, a second copy of the entire in-memory indirect 1342 * block is kept. The buffer cache copy is always completely up-to-date. 1343 * The second copy, which is used only as a source for disk writes, 1344 * contains only the safe pointers (i.e., those that have no remaining 1345 * update dependencies). The second copy is freed when all pointers 1346 * are safe. The cache is not allowed to replace indirect blocks with 1347 * pending update dependencies. If a buffer containing an indirect 1348 * block with dependencies is written, these routines will mark it 1349 * dirty again. It can only be successfully written once all the 1350 * dependencies are removed. The ffs_fsync routine in conjunction with 1351 * softdep_sync_metadata work together to get all the dependencies 1352 * removed so that a file can be successfully written to disk. Three 1353 * procedures are used when setting up indirect block pointer 1354 * dependencies. The division is necessary because of the organization 1355 * of the "balloc" routine and because of the distinction between file 1356 * pages and file metadata blocks. 1357 */ 1358 1359/* 1360 * Allocate a new allocindir structure. 1361 */ 1362static struct allocindir * 1363newallocindir(ip, ptrno, newblkno, oldblkno) 1364 struct inode *ip; /* inode for file being extended */ 1365 int ptrno; /* offset of pointer in indirect block */ 1366 ufs_daddr_t newblkno; /* disk block number being added */ 1367 ufs_daddr_t oldblkno; /* previous block number, 0 if none */ 1368{ 1369 struct allocindir *aip; 1370 1371 MALLOC(aip, struct allocindir *, sizeof(struct allocindir), 1372 M_ALLOCINDIR, M_WAITOK); 1373 bzero(aip, sizeof(struct allocindir)); 1374 aip->ai_list.wk_type = D_ALLOCINDIR; 1375 aip->ai_state = ATTACHED; 1376 aip->ai_offset = ptrno; 1377 aip->ai_newblkno = newblkno; 1378 aip->ai_oldblkno = oldblkno; 1379 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize); 1380 return (aip); 1381} 1382 1383/* 1384 * Called just before setting an indirect block pointer 1385 * to a newly allocated file page. 1386 */ 1387void 1388softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 1389 struct inode *ip; /* inode for file being extended */ 1390 ufs_lbn_t lbn; /* allocated block number within file */ 1391 struct buf *bp; /* buffer with indirect blk referencing page */ 1392 int ptrno; /* offset of pointer in indirect block */ 1393 ufs_daddr_t newblkno; /* disk block number being added */ 1394 ufs_daddr_t oldblkno; /* previous block number, 0 if none */ 1395 struct buf *nbp; /* buffer holding allocated page */ 1396{ 1397 struct allocindir *aip; 1398 struct pagedep *pagedep; 1399 1400 aip = newallocindir(ip, ptrno, newblkno, oldblkno); 1401 ACQUIRE_LOCK(&lk); 1402 /* 1403 * If we are allocating a directory page, then we must 1404 * allocate an associated pagedep to track additions and 1405 * deletions. 1406 */ 1407 if ((ip->i_mode & IFMT) == IFDIR && 1408 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1409 WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list); 1410 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1411 FREE_LOCK(&lk); 1412 setup_allocindir_phase2(bp, ip, aip); 1413} 1414 1415/* 1416 * Called just before setting an indirect block pointer to a 1417 * newly allocated indirect block. 1418 */ 1419void 1420softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 1421 struct buf *nbp; /* newly allocated indirect block */ 1422 struct inode *ip; /* inode for file being extended */ 1423 struct buf *bp; /* indirect block referencing allocated block */ 1424 int ptrno; /* offset of pointer in indirect block */ 1425 ufs_daddr_t newblkno; /* disk block number being added */ 1426{ 1427 struct allocindir *aip; 1428 1429 aip = newallocindir(ip, ptrno, newblkno, 0); 1430 ACQUIRE_LOCK(&lk); 1431 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1432 FREE_LOCK(&lk); 1433 setup_allocindir_phase2(bp, ip, aip); 1434} 1435 1436/* 1437 * Called to finish the allocation of the "aip" allocated 1438 * by one of the two routines above. 1439 */ 1440static void 1441setup_allocindir_phase2(bp, ip, aip) 1442 struct buf *bp; /* in-memory copy of the indirect block */ 1443 struct inode *ip; /* inode for file being extended */ 1444 struct allocindir *aip; /* allocindir allocated by the above routines */ 1445{ 1446 struct worklist *wk; 1447 struct indirdep *indirdep, *newindirdep; 1448 struct bmsafemap *bmsafemap; 1449 struct allocindir *oldaip; 1450 struct freefrag *freefrag; 1451 struct newblk *newblk; 1452 1453 if (bp->b_lblkno >= 0) 1454 panic("setup_allocindir_phase2: not indir blk"); 1455 for (indirdep = NULL, newindirdep = NULL; ; ) { 1456 ACQUIRE_LOCK(&lk); 1457 for (wk = LIST_FIRST(&bp->b_dep); wk; 1458 wk = LIST_NEXT(wk, wk_list)) { 1459 if (wk->wk_type != D_INDIRDEP) 1460 continue; 1461 indirdep = WK_INDIRDEP(wk); 1462 break; 1463 } 1464 if (indirdep == NULL && newindirdep) { 1465 indirdep = newindirdep; 1466 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 1467 newindirdep = NULL; 1468 } 1469 FREE_LOCK(&lk); 1470 if (indirdep) { 1471 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0, 1472 &newblk) == 0) 1473 panic("setup_allocindir: lost block"); 1474 ACQUIRE_LOCK(&lk); 1475 if (newblk->nb_state == DEPCOMPLETE) { 1476 aip->ai_state |= DEPCOMPLETE; 1477 aip->ai_buf = NULL; 1478 } else { 1479 bmsafemap = newblk->nb_bmsafemap; 1480 aip->ai_buf = bmsafemap->sm_buf; 1481 LIST_REMOVE(newblk, nb_deps); 1482 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd, 1483 aip, ai_deps); 1484 } 1485 LIST_REMOVE(newblk, nb_hash); 1486 FREE(newblk, M_NEWBLK); 1487 aip->ai_indirdep = indirdep; 1488 /* 1489 * Check to see if there is an existing dependency 1490 * for this block. If there is, merge the old 1491 * dependency into the new one. 1492 */ 1493 if (aip->ai_oldblkno == 0) 1494 oldaip = NULL; 1495 else 1496 for (oldaip=LIST_FIRST(&indirdep->ir_deplisthd); 1497 oldaip; oldaip = LIST_NEXT(oldaip, ai_next)) 1498 if (oldaip->ai_offset == aip->ai_offset) 1499 break; 1500 if (oldaip != NULL) { 1501 if (oldaip->ai_newblkno != aip->ai_oldblkno) 1502 panic("setup_allocindir_phase2: blkno"); 1503 aip->ai_oldblkno = oldaip->ai_oldblkno; 1504 freefrag = oldaip->ai_freefrag; 1505 oldaip->ai_freefrag = aip->ai_freefrag; 1506 aip->ai_freefrag = freefrag; 1507 free_allocindir(oldaip, NULL); 1508 } 1509 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 1510 ((ufs_daddr_t *)indirdep->ir_savebp->b_data) 1511 [aip->ai_offset] = aip->ai_oldblkno; 1512 FREE_LOCK(&lk); 1513 } 1514 if (newindirdep) { 1515 if (indirdep->ir_savebp != NULL) 1516 brelse(newindirdep->ir_savebp); 1517 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP); 1518 } 1519 if (indirdep) 1520 break; 1521 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep), 1522 M_INDIRDEP, M_WAITOK); 1523 newindirdep->ir_list.wk_type = D_INDIRDEP; 1524 newindirdep->ir_state = ATTACHED; 1525 LIST_INIT(&newindirdep->ir_deplisthd); 1526 LIST_INIT(&newindirdep->ir_donehd); 1527 if (bp->b_blkno == bp->b_lblkno) { 1528 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, 1529 NULL, NULL); 1530 } 1531 newindirdep->ir_savebp = 1532 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0); 1533 newindirdep->ir_savebp->b_flags |= B_ASYNC; 1534 BUF_KERNPROC(newindirdep->ir_savebp); 1535 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 1536 } 1537} 1538 1539/* 1540 * Block de-allocation dependencies. 1541 * 1542 * When blocks are de-allocated, the on-disk pointers must be nullified before 1543 * the blocks are made available for use by other files. (The true 1544 * requirement is that old pointers must be nullified before new on-disk 1545 * pointers are set. We chose this slightly more stringent requirement to 1546 * reduce complexity.) Our implementation handles this dependency by updating 1547 * the inode (or indirect block) appropriately but delaying the actual block 1548 * de-allocation (i.e., freemap and free space count manipulation) until 1549 * after the updated versions reach stable storage. After the disk is 1550 * updated, the blocks can be safely de-allocated whenever it is convenient. 1551 * This implementation handles only the common case of reducing a file's 1552 * length to zero. Other cases are handled by the conventional synchronous 1553 * write approach. 1554 * 1555 * The ffs implementation with which we worked double-checks 1556 * the state of the block pointers and file size as it reduces 1557 * a file's length. Some of this code is replicated here in our 1558 * soft updates implementation. The freeblks->fb_chkcnt field is 1559 * used to transfer a part of this information to the procedure 1560 * that eventually de-allocates the blocks. 1561 * 1562 * This routine should be called from the routine that shortens 1563 * a file's length, before the inode's size or block pointers 1564 * are modified. It will save the block pointer information for 1565 * later release and zero the inode so that the calling routine 1566 * can release it. 1567 */ 1568static long num_freeblks; /* number of freeblks allocated */ 1569void 1570softdep_setup_freeblocks(ip, length) 1571 struct inode *ip; /* The inode whose length is to be reduced */ 1572 off_t length; /* The new length for the file */ 1573{ 1574 struct freeblks *freeblks; 1575 struct inodedep *inodedep; 1576 struct allocdirect *adp; 1577 struct vnode *vp; 1578 struct buf *bp; 1579 struct fs *fs; 1580 int i, error; 1581 1582 fs = ip->i_fs; 1583 if (length != 0) 1584 panic("softde_setup_freeblocks: non-zero length"); 1585 /* 1586 * If we are over our limit, try to improve the situation. 1587 */ 1588 if (num_freeblks > max_softdeps / 2 && speedup_syncer() == 0) 1589 (void) request_cleanup(FLUSH_REMOVE, 0); 1590 num_freeblks += 1; 1591 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks), 1592 M_FREEBLKS, M_WAITOK); 1593 bzero(freeblks, sizeof(struct freeblks)); 1594 freeblks->fb_list.wk_type = D_FREEBLKS; 1595 freeblks->fb_uid = ip->i_uid; 1596 freeblks->fb_previousinum = ip->i_number; 1597 freeblks->fb_devvp = ip->i_devvp; 1598 freeblks->fb_fs = fs; 1599 freeblks->fb_oldsize = ip->i_size; 1600 freeblks->fb_newsize = length; 1601 freeblks->fb_chkcnt = ip->i_blocks; 1602 for (i = 0; i < NDADDR; i++) { 1603 freeblks->fb_dblks[i] = ip->i_db[i]; 1604 ip->i_db[i] = 0; 1605 } 1606 for (i = 0; i < NIADDR; i++) { 1607 freeblks->fb_iblks[i] = ip->i_ib[i]; 1608 ip->i_ib[i] = 0; 1609 } 1610 ip->i_blocks = 0; 1611 ip->i_size = 0; 1612 /* 1613 * Push the zero'ed inode to to its disk buffer so that we are free 1614 * to delete its dependencies below. Once the dependencies are gone 1615 * the buffer can be safely released. 1616 */ 1617 if ((error = bread(ip->i_devvp, 1618 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 1619 (int)fs->fs_bsize, NOCRED, &bp)) != 0) 1620 softdep_error("softdep_setup_freeblocks", error); 1621 *((struct dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) = 1622 ip->i_din; 1623 /* 1624 * Find and eliminate any inode dependencies. 1625 */ 1626 ACQUIRE_LOCK(&lk); 1627 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep); 1628 if ((inodedep->id_state & IOSTARTED) != 0) 1629 panic("softdep_setup_freeblocks: inode busy"); 1630 /* 1631 * Add the freeblks structure to the list of operations that 1632 * must await the zero'ed inode being written to disk. 1633 */ 1634 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list); 1635 /* 1636 * Because the file length has been truncated to zero, any 1637 * pending block allocation dependency structures associated 1638 * with this inode are obsolete and can simply be de-allocated. 1639 * We must first merge the two dependency lists to get rid of 1640 * any duplicate freefrag structures, then purge the merged list. 1641 */ 1642 merge_inode_lists(inodedep); 1643 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 1644 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 1645 FREE_LOCK(&lk); 1646 bdwrite(bp); 1647 /* 1648 * We must wait for any I/O in progress to finish so that 1649 * all potential buffers on the dirty list will be visible. 1650 * Once they are all there, walk the list and get rid of 1651 * any dependencies. 1652 */ 1653 vp = ITOV(ip); 1654 ACQUIRE_LOCK(&lk); 1655 drain_output(vp, 1); 1656 while (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT)) { 1657 bp = TAILQ_FIRST(&vp->v_dirtyblkhd); 1658 (void) inodedep_lookup(fs, ip->i_number, 0, &inodedep); 1659 deallocate_dependencies(bp, inodedep); 1660 bp->b_flags |= B_INVAL | B_NOCACHE; 1661 FREE_LOCK(&lk); 1662 brelse(bp); 1663 ACQUIRE_LOCK(&lk); 1664 } 1665 /* 1666 * Try freeing the inodedep in case that was the last dependency. 1667 */ 1668 if ((inodedep_lookup(fs, ip->i_number, 0, &inodedep)) != 0) 1669 (void) free_inodedep(inodedep); 1670 FREE_LOCK(&lk); 1671} 1672 1673/* 1674 * Reclaim any dependency structures from a buffer that is about to 1675 * be reallocated to a new vnode. The buffer must be locked, thus, 1676 * no I/O completion operations can occur while we are manipulating 1677 * its associated dependencies. The mutex is held so that other I/O's 1678 * associated with related dependencies do not occur. 1679 */ 1680static void 1681deallocate_dependencies(bp, inodedep) 1682 struct buf *bp; 1683 struct inodedep *inodedep; 1684{ 1685 struct worklist *wk; 1686 struct indirdep *indirdep; 1687 struct allocindir *aip; 1688 struct pagedep *pagedep; 1689 struct dirrem *dirrem; 1690 struct diradd *dap; 1691 int i; 1692 1693 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 1694 switch (wk->wk_type) { 1695 1696 case D_INDIRDEP: 1697 indirdep = WK_INDIRDEP(wk); 1698 /* 1699 * None of the indirect pointers will ever be visible, 1700 * so they can simply be tossed. GOINGAWAY ensures 1701 * that allocated pointers will be saved in the buffer 1702 * cache until they are freed. Note that they will 1703 * only be able to be found by their physical address 1704 * since the inode mapping the logical address will 1705 * be gone. The save buffer used for the safe copy 1706 * was allocated in setup_allocindir_phase2 using 1707 * the physical address so it could be used for this 1708 * purpose. Hence we swap the safe copy with the real 1709 * copy, allowing the safe copy to be freed and holding 1710 * on to the real copy for later use in indir_trunc. 1711 */ 1712 if (indirdep->ir_state & GOINGAWAY) 1713 panic("deallocate_dependencies: already gone"); 1714 indirdep->ir_state |= GOINGAWAY; 1715 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 1716 free_allocindir(aip, inodedep); 1717 if (bp->b_lblkno >= 0 || 1718 bp->b_blkno != indirdep->ir_savebp->b_lblkno) 1719 panic("deallocate_dependencies: not indir"); 1720 bcopy(bp->b_data, indirdep->ir_savebp->b_data, 1721 bp->b_bcount); 1722 WORKLIST_REMOVE(wk); 1723 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk); 1724 continue; 1725 1726 case D_PAGEDEP: 1727 pagedep = WK_PAGEDEP(wk); 1728 /* 1729 * None of the directory additions will ever be 1730 * visible, so they can simply be tossed. 1731 */ 1732 for (i = 0; i < DAHASHSZ; i++) 1733 while ((dap = 1734 LIST_FIRST(&pagedep->pd_diraddhd[i]))) 1735 free_diradd(dap); 1736 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0) 1737 free_diradd(dap); 1738 /* 1739 * Copy any directory remove dependencies to the list 1740 * to be processed after the zero'ed inode is written. 1741 * If the inode has already been written, then they 1742 * can be dumped directly onto the work list. 1743 */ 1744 for (dirrem = LIST_FIRST(&pagedep->pd_dirremhd); dirrem; 1745 dirrem = LIST_NEXT(dirrem, dm_next)) { 1746 LIST_REMOVE(dirrem, dm_next); 1747 dirrem->dm_dirinum = pagedep->pd_ino; 1748 if (inodedep == NULL) 1749 add_to_worklist(&dirrem->dm_list); 1750 else 1751 WORKLIST_INSERT(&inodedep->id_bufwait, 1752 &dirrem->dm_list); 1753 } 1754 WORKLIST_REMOVE(&pagedep->pd_list); 1755 LIST_REMOVE(pagedep, pd_hash); 1756 WORKITEM_FREE(pagedep, D_PAGEDEP); 1757 continue; 1758 1759 case D_ALLOCINDIR: 1760 free_allocindir(WK_ALLOCINDIR(wk), inodedep); 1761 continue; 1762 1763 case D_ALLOCDIRECT: 1764 case D_INODEDEP: 1765 panic("deallocate_dependencies: Unexpected type %s", 1766 TYPENAME(wk->wk_type)); 1767 /* NOTREACHED */ 1768 1769 default: 1770 panic("deallocate_dependencies: Unknown type %s", 1771 TYPENAME(wk->wk_type)); 1772 /* NOTREACHED */ 1773 } 1774 } 1775} 1776 1777/* 1778 * Free an allocdirect. Generate a new freefrag work request if appropriate. 1779 * This routine must be called with splbio interrupts blocked. 1780 */ 1781static void 1782free_allocdirect(adphead, adp, delay) 1783 struct allocdirectlst *adphead; 1784 struct allocdirect *adp; 1785 int delay; 1786{ 1787 1788#ifdef DEBUG 1789 if (lk.lkt_held == -1) 1790 panic("free_allocdirect: lock not held"); 1791#endif 1792 if ((adp->ad_state & DEPCOMPLETE) == 0) 1793 LIST_REMOVE(adp, ad_deps); 1794 TAILQ_REMOVE(adphead, adp, ad_next); 1795 if ((adp->ad_state & COMPLETE) == 0) 1796 WORKLIST_REMOVE(&adp->ad_list); 1797 if (adp->ad_freefrag != NULL) { 1798 if (delay) 1799 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 1800 &adp->ad_freefrag->ff_list); 1801 else 1802 add_to_worklist(&adp->ad_freefrag->ff_list); 1803 } 1804 WORKITEM_FREE(adp, D_ALLOCDIRECT); 1805} 1806 1807/* 1808 * Prepare an inode to be freed. The actual free operation is not 1809 * done until the zero'ed inode has been written to disk. 1810 */ 1811static long num_freefile; /* number of freefile allocated */ 1812void 1813softdep_freefile(pvp, ino, mode) 1814 struct vnode *pvp; 1815 ino_t ino; 1816 int mode; 1817{ 1818 struct inode *ip = VTOI(pvp); 1819 struct inodedep *inodedep; 1820 struct freefile *freefile; 1821 1822 /* 1823 * If we are over our limit, try to improve the situation. 1824 */ 1825 if (num_freefile > max_softdeps / 2 && speedup_syncer() == 0) 1826 (void) request_cleanup(FLUSH_REMOVE, 0); 1827 /* 1828 * This sets up the inode de-allocation dependency. 1829 */ 1830 num_freefile += 1; 1831 MALLOC(freefile, struct freefile *, sizeof(struct freefile), 1832 M_FREEFILE, M_WAITOK); 1833 freefile->fx_list.wk_type = D_FREEFILE; 1834 freefile->fx_list.wk_state = 0; 1835 freefile->fx_mode = mode; 1836 freefile->fx_oldinum = ino; 1837 freefile->fx_devvp = ip->i_devvp; 1838 freefile->fx_fs = ip->i_fs; 1839 1840 /* 1841 * If the inodedep does not exist, then the zero'ed inode has 1842 * been written to disk and we can free the file immediately. 1843 */ 1844 ACQUIRE_LOCK(&lk); 1845 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0) { 1846 add_to_worklist(&freefile->fx_list); 1847 FREE_LOCK(&lk); 1848 return; 1849 } 1850 1851 /* 1852 * If we still have a bitmap dependency, then the inode has never 1853 * been written to disk. Drop the dependency as it is no longer 1854 * necessary since the inode is being deallocated. We could process 1855 * the freefile immediately, but then we would have to clear the 1856 * id_inowait dependencies here and it is easier just to let the 1857 * zero'ed inode be written and let them be cleaned up in the 1858 * normal followup actions that follow the inode write. 1859 */ 1860 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 1861 inodedep->id_state |= DEPCOMPLETE; 1862 LIST_REMOVE(inodedep, id_deps); 1863 inodedep->id_buf = NULL; 1864 } 1865 /* 1866 * If the inodedep has no dependencies associated with it, 1867 * then we must free it here and free the file immediately. 1868 * This case arises when an early allocation fails (for 1869 * example, the user is over their file quota). 1870 */ 1871 if (free_inodedep(inodedep) == 0) 1872 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 1873 else 1874 add_to_worklist(&freefile->fx_list); 1875 FREE_LOCK(&lk); 1876} 1877 1878/* 1879 * Try to free an inodedep structure. Return 1 if it could be freed. 1880 */ 1881static int 1882free_inodedep(inodedep) 1883 struct inodedep *inodedep; 1884{ 1885 1886 if ((inodedep->id_state & ONWORKLIST) != 0 || 1887 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 1888 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 1889 LIST_FIRST(&inodedep->id_bufwait) != NULL || 1890 LIST_FIRST(&inodedep->id_inowait) != NULL || 1891 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 1892 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 1893 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL) 1894 return (0); 1895 LIST_REMOVE(inodedep, id_hash); 1896 WORKITEM_FREE(inodedep, D_INODEDEP); 1897 num_inodedep -= 1; 1898 return (1); 1899} 1900 1901/* 1902 * This workitem routine performs the block de-allocation. 1903 * The workitem is added to the pending list after the updated 1904 * inode block has been written to disk. As mentioned above, 1905 * checks regarding the number of blocks de-allocated (compared 1906 * to the number of blocks allocated for the file) are also 1907 * performed in this function. 1908 */ 1909static void 1910handle_workitem_freeblocks(freeblks) 1911 struct freeblks *freeblks; 1912{ 1913 struct inode tip; 1914 ufs_daddr_t bn; 1915 struct fs *fs; 1916 int i, level, bsize; 1917 long nblocks, blocksreleased = 0; 1918 int error, allerror = 0; 1919 ufs_lbn_t baselbns[NIADDR], tmpval; 1920 1921 tip.i_number = freeblks->fb_previousinum; 1922 tip.i_devvp = freeblks->fb_devvp; 1923 tip.i_dev = freeblks->fb_devvp->v_rdev; 1924 tip.i_fs = freeblks->fb_fs; 1925 tip.i_size = freeblks->fb_oldsize; 1926 tip.i_uid = freeblks->fb_uid; 1927 fs = freeblks->fb_fs; 1928 tmpval = 1; 1929 baselbns[0] = NDADDR; 1930 for (i = 1; i < NIADDR; i++) { 1931 tmpval *= NINDIR(fs); 1932 baselbns[i] = baselbns[i - 1] + tmpval; 1933 } 1934 nblocks = btodb(fs->fs_bsize); 1935 blocksreleased = 0; 1936 /* 1937 * Indirect blocks first. 1938 */ 1939 for (level = (NIADDR - 1); level >= 0; level--) { 1940 if ((bn = freeblks->fb_iblks[level]) == 0) 1941 continue; 1942 if ((error = indir_trunc(&tip, fsbtodb(fs, bn), level, 1943 baselbns[level], &blocksreleased)) == 0) 1944 allerror = error; 1945 ffs_blkfree(&tip, bn, fs->fs_bsize); 1946 blocksreleased += nblocks; 1947 } 1948 /* 1949 * All direct blocks or frags. 1950 */ 1951 for (i = (NDADDR - 1); i >= 0; i--) { 1952 if ((bn = freeblks->fb_dblks[i]) == 0) 1953 continue; 1954 bsize = blksize(fs, &tip, i); 1955 ffs_blkfree(&tip, bn, bsize); 1956 blocksreleased += btodb(bsize); 1957 } 1958 1959#ifdef DIAGNOSTIC 1960 if (freeblks->fb_chkcnt != blocksreleased) 1961 panic("handle_workitem_freeblocks: block count"); 1962 if (allerror) 1963 softdep_error("handle_workitem_freeblks", allerror); 1964#endif /* DIAGNOSTIC */ 1965 WORKITEM_FREE(freeblks, D_FREEBLKS); 1966 num_freeblks -= 1; 1967} 1968 1969/* 1970 * Release blocks associated with the inode ip and stored in the indirect 1971 * block dbn. If level is greater than SINGLE, the block is an indirect block 1972 * and recursive calls to indirtrunc must be used to cleanse other indirect 1973 * blocks. 1974 */ 1975static int 1976indir_trunc(ip, dbn, level, lbn, countp) 1977 struct inode *ip; 1978 ufs_daddr_t dbn; 1979 int level; 1980 ufs_lbn_t lbn; 1981 long *countp; 1982{ 1983 struct buf *bp; 1984 ufs_daddr_t *bap; 1985 ufs_daddr_t nb; 1986 struct fs *fs; 1987 struct worklist *wk; 1988 struct indirdep *indirdep; 1989 int i, lbnadd, nblocks; 1990 int error, allerror = 0; 1991 1992 fs = ip->i_fs; 1993 lbnadd = 1; 1994 for (i = level; i > 0; i--) 1995 lbnadd *= NINDIR(fs); 1996 /* 1997 * Get buffer of block pointers to be freed. This routine is not 1998 * called until the zero'ed inode has been written, so it is safe 1999 * to free blocks as they are encountered. Because the inode has 2000 * been zero'ed, calls to bmap on these blocks will fail. So, we 2001 * have to use the on-disk address and the block device for the 2002 * filesystem to look them up. If the file was deleted before its 2003 * indirect blocks were all written to disk, the routine that set 2004 * us up (deallocate_dependencies) will have arranged to leave 2005 * a complete copy of the indirect block in memory for our use. 2006 * Otherwise we have to read the blocks in from the disk. 2007 */ 2008 ACQUIRE_LOCK(&lk); 2009 if ((bp = incore(ip->i_devvp, dbn)) != NULL && 2010 (wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2011 if (wk->wk_type != D_INDIRDEP || 2012 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp || 2013 (indirdep->ir_state & GOINGAWAY) == 0) 2014 panic("indir_trunc: lost indirdep"); 2015 WORKLIST_REMOVE(wk); 2016 WORKITEM_FREE(indirdep, D_INDIRDEP); 2017 if (LIST_FIRST(&bp->b_dep) != NULL) 2018 panic("indir_trunc: dangling dep"); 2019 FREE_LOCK(&lk); 2020 } else { 2021 FREE_LOCK(&lk); 2022 error = bread(ip->i_devvp, dbn, (int)fs->fs_bsize, NOCRED, &bp); 2023 if (error) 2024 return (error); 2025 } 2026 /* 2027 * Recursively free indirect blocks. 2028 */ 2029 bap = (ufs_daddr_t *)bp->b_data; 2030 nblocks = btodb(fs->fs_bsize); 2031 for (i = NINDIR(fs) - 1; i >= 0; i--) { 2032 if ((nb = bap[i]) == 0) 2033 continue; 2034 if (level != 0) { 2035 if ((error = indir_trunc(ip, fsbtodb(fs, nb), 2036 level - 1, lbn + (i * lbnadd), countp)) != 0) 2037 allerror = error; 2038 } 2039 ffs_blkfree(ip, nb, fs->fs_bsize); 2040 *countp += nblocks; 2041 } 2042 bp->b_flags |= B_INVAL | B_NOCACHE; 2043 brelse(bp); 2044 return (allerror); 2045} 2046 2047/* 2048 * Free an allocindir. 2049 * This routine must be called with splbio interrupts blocked. 2050 */ 2051static void 2052free_allocindir(aip, inodedep) 2053 struct allocindir *aip; 2054 struct inodedep *inodedep; 2055{ 2056 struct freefrag *freefrag; 2057 2058#ifdef DEBUG 2059 if (lk.lkt_held == -1) 2060 panic("free_allocindir: lock not held"); 2061#endif 2062 if ((aip->ai_state & DEPCOMPLETE) == 0) 2063 LIST_REMOVE(aip, ai_deps); 2064 if (aip->ai_state & ONWORKLIST) 2065 WORKLIST_REMOVE(&aip->ai_list); 2066 LIST_REMOVE(aip, ai_next); 2067 if ((freefrag = aip->ai_freefrag) != NULL) { 2068 if (inodedep == NULL) 2069 add_to_worklist(&freefrag->ff_list); 2070 else 2071 WORKLIST_INSERT(&inodedep->id_bufwait, 2072 &freefrag->ff_list); 2073 } 2074 WORKITEM_FREE(aip, D_ALLOCINDIR); 2075} 2076 2077/* 2078 * Directory entry addition dependencies. 2079 * 2080 * When adding a new directory entry, the inode (with its incremented link 2081 * count) must be written to disk before the directory entry's pointer to it. 2082 * Also, if the inode is newly allocated, the corresponding freemap must be 2083 * updated (on disk) before the directory entry's pointer. These requirements 2084 * are met via undo/redo on the directory entry's pointer, which consists 2085 * simply of the inode number. 2086 * 2087 * As directory entries are added and deleted, the free space within a 2088 * directory block can become fragmented. The ufs file system will compact 2089 * a fragmented directory block to make space for a new entry. When this 2090 * occurs, the offsets of previously added entries change. Any "diradd" 2091 * dependency structures corresponding to these entries must be updated with 2092 * the new offsets. 2093 */ 2094 2095/* 2096 * This routine is called after the in-memory inode's link 2097 * count has been incremented, but before the directory entry's 2098 * pointer to the inode has been set. 2099 */ 2100void 2101softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp) 2102 struct buf *bp; /* buffer containing directory block */ 2103 struct inode *dp; /* inode for directory */ 2104 off_t diroffset; /* offset of new entry in directory */ 2105 long newinum; /* inode referenced by new directory entry */ 2106 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 2107{ 2108 int offset; /* offset of new entry within directory block */ 2109 ufs_lbn_t lbn; /* block in directory containing new entry */ 2110 struct fs *fs; 2111 struct diradd *dap; 2112 struct pagedep *pagedep; 2113 struct inodedep *inodedep; 2114 struct mkdir *mkdir1, *mkdir2; 2115 2116 /* 2117 * Whiteouts have no dependencies. 2118 */ 2119 if (newinum == WINO) { 2120 if (newdirbp != NULL) 2121 bdwrite(newdirbp); 2122 return; 2123 } 2124 2125 fs = dp->i_fs; 2126 lbn = lblkno(fs, diroffset); 2127 offset = blkoff(fs, diroffset); 2128 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, M_WAITOK); 2129 bzero(dap, sizeof(struct diradd)); 2130 dap->da_list.wk_type = D_DIRADD; 2131 dap->da_offset = offset; 2132 dap->da_newinum = newinum; 2133 dap->da_state = ATTACHED; 2134 if (newdirbp == NULL) { 2135 dap->da_state |= DEPCOMPLETE; 2136 ACQUIRE_LOCK(&lk); 2137 } else { 2138 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 2139 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2140 M_WAITOK); 2141 mkdir1->md_list.wk_type = D_MKDIR; 2142 mkdir1->md_state = MKDIR_BODY; 2143 mkdir1->md_diradd = dap; 2144 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2145 M_WAITOK); 2146 mkdir2->md_list.wk_type = D_MKDIR; 2147 mkdir2->md_state = MKDIR_PARENT; 2148 mkdir2->md_diradd = dap; 2149 /* 2150 * Dependency on "." and ".." being written to disk. 2151 */ 2152 mkdir1->md_buf = newdirbp; 2153 ACQUIRE_LOCK(&lk); 2154 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 2155 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list); 2156 FREE_LOCK(&lk); 2157 bdwrite(newdirbp); 2158 /* 2159 * Dependency on link count increase for parent directory 2160 */ 2161 ACQUIRE_LOCK(&lk); 2162 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0 2163 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2164 dap->da_state &= ~MKDIR_PARENT; 2165 WORKITEM_FREE(mkdir2, D_MKDIR); 2166 } else { 2167 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 2168 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 2169 } 2170 } 2171 /* 2172 * Link into parent directory pagedep to await its being written. 2173 */ 2174 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2175 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2176 dap->da_pagedep = pagedep; 2177 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 2178 da_pdlist); 2179 /* 2180 * Link into its inodedep. Put it on the id_bufwait list if the inode 2181 * is not yet written. If it is written, do the post-inode write 2182 * processing to put it on the id_pendinghd list. 2183 */ 2184 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep); 2185 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 2186 diradd_inode_written(dap, inodedep); 2187 else 2188 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2189 FREE_LOCK(&lk); 2190} 2191 2192/* 2193 * This procedure is called to change the offset of a directory 2194 * entry when compacting a directory block which must be owned 2195 * exclusively by the caller. Note that the actual entry movement 2196 * must be done in this procedure to ensure that no I/O completions 2197 * occur while the move is in progress. 2198 */ 2199void 2200softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize) 2201 struct inode *dp; /* inode for directory */ 2202 caddr_t base; /* address of dp->i_offset */ 2203 caddr_t oldloc; /* address of old directory location */ 2204 caddr_t newloc; /* address of new directory location */ 2205 int entrysize; /* size of directory entry */ 2206{ 2207 int offset, oldoffset, newoffset; 2208 struct pagedep *pagedep; 2209 struct diradd *dap; 2210 ufs_lbn_t lbn; 2211 2212 ACQUIRE_LOCK(&lk); 2213 lbn = lblkno(dp->i_fs, dp->i_offset); 2214 offset = blkoff(dp->i_fs, dp->i_offset); 2215 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) 2216 goto done; 2217 oldoffset = offset + (oldloc - base); 2218 newoffset = offset + (newloc - base); 2219 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[DIRADDHASH(oldoffset)]); 2220 dap; dap = LIST_NEXT(dap, da_pdlist)) { 2221 if (dap->da_offset != oldoffset) 2222 continue; 2223 dap->da_offset = newoffset; 2224 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) 2225 break; 2226 LIST_REMOVE(dap, da_pdlist); 2227 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], 2228 dap, da_pdlist); 2229 break; 2230 } 2231 if (dap == NULL) { 2232 for (dap = LIST_FIRST(&pagedep->pd_pendinghd); 2233 dap; dap = LIST_NEXT(dap, da_pdlist)) { 2234 if (dap->da_offset == oldoffset) { 2235 dap->da_offset = newoffset; 2236 break; 2237 } 2238 } 2239 } 2240done: 2241 bcopy(oldloc, newloc, entrysize); 2242 FREE_LOCK(&lk); 2243} 2244 2245/* 2246 * Free a diradd dependency structure. This routine must be called 2247 * with splbio interrupts blocked. 2248 */ 2249static void 2250free_diradd(dap) 2251 struct diradd *dap; 2252{ 2253 struct dirrem *dirrem; 2254 struct pagedep *pagedep; 2255 struct inodedep *inodedep; 2256 struct mkdir *mkdir, *nextmd; 2257 2258#ifdef DEBUG 2259 if (lk.lkt_held == -1) 2260 panic("free_diradd: lock not held"); 2261#endif 2262 WORKLIST_REMOVE(&dap->da_list); 2263 LIST_REMOVE(dap, da_pdlist); 2264 if ((dap->da_state & DIRCHG) == 0) { 2265 pagedep = dap->da_pagedep; 2266 } else { 2267 dirrem = dap->da_previous; 2268 pagedep = dirrem->dm_pagedep; 2269 dirrem->dm_dirinum = pagedep->pd_ino; 2270 add_to_worklist(&dirrem->dm_list); 2271 } 2272 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum, 2273 0, &inodedep) != 0) 2274 (void) free_inodedep(inodedep); 2275 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2276 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 2277 nextmd = LIST_NEXT(mkdir, md_mkdirs); 2278 if (mkdir->md_diradd != dap) 2279 continue; 2280 dap->da_state &= ~mkdir->md_state; 2281 WORKLIST_REMOVE(&mkdir->md_list); 2282 LIST_REMOVE(mkdir, md_mkdirs); 2283 WORKITEM_FREE(mkdir, D_MKDIR); 2284 } 2285 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 2286 panic("free_diradd: unfound ref"); 2287 } 2288 WORKITEM_FREE(dap, D_DIRADD); 2289} 2290 2291/* 2292 * Directory entry removal dependencies. 2293 * 2294 * When removing a directory entry, the entry's inode pointer must be 2295 * zero'ed on disk before the corresponding inode's link count is decremented 2296 * (possibly freeing the inode for re-use). This dependency is handled by 2297 * updating the directory entry but delaying the inode count reduction until 2298 * after the directory block has been written to disk. After this point, the 2299 * inode count can be decremented whenever it is convenient. 2300 */ 2301 2302/* 2303 * This routine should be called immediately after removing 2304 * a directory entry. The inode's link count should not be 2305 * decremented by the calling procedure -- the soft updates 2306 * code will do this task when it is safe. 2307 */ 2308void 2309softdep_setup_remove(bp, dp, ip, isrmdir) 2310 struct buf *bp; /* buffer containing directory block */ 2311 struct inode *dp; /* inode for the directory being modified */ 2312 struct inode *ip; /* inode for directory entry being removed */ 2313 int isrmdir; /* indicates if doing RMDIR */ 2314{ 2315 struct dirrem *dirrem; 2316 2317 /* 2318 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 2319 */ 2320 dirrem = newdirrem(bp, dp, ip, isrmdir); 2321 if ((dirrem->dm_state & COMPLETE) == 0) { 2322 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 2323 dm_next); 2324 } else { 2325 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 2326 add_to_worklist(&dirrem->dm_list); 2327 } 2328 FREE_LOCK(&lk); 2329} 2330 2331/* 2332 * Allocate a new dirrem if appropriate and return it along with 2333 * its associated pagedep. Called without a lock, returns with lock. 2334 */ 2335static struct dirrem * 2336newdirrem(bp, dp, ip, isrmdir) 2337 struct buf *bp; /* buffer containing directory block */ 2338 struct inode *dp; /* inode for the directory being modified */ 2339 struct inode *ip; /* inode for directory entry being removed */ 2340 int isrmdir; /* indicates if doing RMDIR */ 2341{ 2342 int offset; 2343 ufs_lbn_t lbn; 2344 struct diradd *dap; 2345 struct dirrem *dirrem; 2346 struct pagedep *pagedep; 2347 2348 /* 2349 * Whiteouts have no deletion dependencies. 2350 */ 2351 if (ip == NULL) 2352 panic("newdirrem: whiteout"); 2353 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem), 2354 M_DIRREM, M_WAITOK); 2355 bzero(dirrem, sizeof(struct dirrem)); 2356 dirrem->dm_list.wk_type = D_DIRREM; 2357 dirrem->dm_state = isrmdir ? RMDIR : 0; 2358 dirrem->dm_mnt = ITOV(ip)->v_mount; 2359 dirrem->dm_oldinum = ip->i_number; 2360 2361 ACQUIRE_LOCK(&lk); 2362 lbn = lblkno(dp->i_fs, dp->i_offset); 2363 offset = blkoff(dp->i_fs, dp->i_offset); 2364 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2365 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2366 dirrem->dm_pagedep = pagedep; 2367 /* 2368 * Check for a diradd dependency for the same directory entry. 2369 * If present, then both dependencies become obsolete and can 2370 * be de-allocated. Check for an entry on both the pd_dirraddhd 2371 * list and the pd_pendinghd list. 2372 */ 2373 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[DIRADDHASH(offset)]); 2374 dap; dap = LIST_NEXT(dap, da_pdlist)) 2375 if (dap->da_offset == offset) 2376 break; 2377 if (dap == NULL) { 2378 for (dap = LIST_FIRST(&pagedep->pd_pendinghd); 2379 dap; dap = LIST_NEXT(dap, da_pdlist)) 2380 if (dap->da_offset == offset) 2381 break; 2382 if (dap == NULL) 2383 return (dirrem); 2384 } 2385 /* 2386 * Must be ATTACHED at this point, so just delete it. 2387 */ 2388 if ((dap->da_state & ATTACHED) == 0) 2389 panic("newdirrem: not ATTACHED"); 2390 if (dap->da_newinum != ip->i_number) 2391 panic("newdirrem: inum %d should be %d", 2392 ip->i_number, dap->da_newinum); 2393 free_diradd(dap); 2394 dirrem->dm_state |= COMPLETE; 2395 return (dirrem); 2396} 2397 2398/* 2399 * Directory entry change dependencies. 2400 * 2401 * Changing an existing directory entry requires that an add operation 2402 * be completed first followed by a deletion. The semantics for the addition 2403 * are identical to the description of adding a new entry above except 2404 * that the rollback is to the old inode number rather than zero. Once 2405 * the addition dependency is completed, the removal is done as described 2406 * in the removal routine above. 2407 */ 2408 2409/* 2410 * This routine should be called immediately after changing 2411 * a directory entry. The inode's link count should not be 2412 * decremented by the calling procedure -- the soft updates 2413 * code will perform this task when it is safe. 2414 */ 2415void 2416softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 2417 struct buf *bp; /* buffer containing directory block */ 2418 struct inode *dp; /* inode for the directory being modified */ 2419 struct inode *ip; /* inode for directory entry being removed */ 2420 long newinum; /* new inode number for changed entry */ 2421 int isrmdir; /* indicates if doing RMDIR */ 2422{ 2423 int offset; 2424 struct diradd *dap = NULL; 2425 struct dirrem *dirrem; 2426 struct pagedep *pagedep; 2427 struct inodedep *inodedep; 2428 2429 offset = blkoff(dp->i_fs, dp->i_offset); 2430 2431 /* 2432 * Whiteouts do not need diradd dependencies. 2433 */ 2434 if (newinum != WINO) { 2435 MALLOC(dap, struct diradd *, sizeof(struct diradd), 2436 M_DIRADD, M_WAITOK); 2437 bzero(dap, sizeof(struct diradd)); 2438 dap->da_list.wk_type = D_DIRADD; 2439 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 2440 dap->da_offset = offset; 2441 dap->da_newinum = newinum; 2442 } 2443 2444 /* 2445 * Allocate a new dirrem and ACQUIRE_LOCK. 2446 */ 2447 dirrem = newdirrem(bp, dp, ip, isrmdir); 2448 pagedep = dirrem->dm_pagedep; 2449 /* 2450 * The possible values for isrmdir: 2451 * 0 - non-directory file rename 2452 * 1 - directory rename within same directory 2453 * inum - directory rename to new directory of given inode number 2454 * When renaming to a new directory, we are both deleting and 2455 * creating a new directory entry, so the link count on the new 2456 * directory should not change. Thus we do not need the followup 2457 * dirrem which is usually done in handle_workitem_remove. We set 2458 * the DIRCHG flag to tell handle_workitem_remove to skip the 2459 * followup dirrem. 2460 */ 2461 if (isrmdir > 1) 2462 dirrem->dm_state |= DIRCHG; 2463 2464 /* 2465 * Whiteouts have no additional dependencies, 2466 * so just put the dirrem on the correct list. 2467 */ 2468 if (newinum == WINO) { 2469 if ((dirrem->dm_state & COMPLETE) == 0) { 2470 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 2471 dm_next); 2472 } else { 2473 dirrem->dm_dirinum = pagedep->pd_ino; 2474 add_to_worklist(&dirrem->dm_list); 2475 } 2476 FREE_LOCK(&lk); 2477 return; 2478 } 2479 2480 /* 2481 * Link into its inodedep. Put it on the id_bufwait list if the inode 2482 * is not yet written. If it is written, do the post-inode write 2483 * processing to put it on the id_pendinghd list. 2484 */ 2485 dap->da_previous = dirrem; 2486 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 || 2487 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2488 dap->da_state |= COMPLETE; 2489 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 2490 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 2491 } else { 2492 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 2493 dap, da_pdlist); 2494 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2495 } 2496 /* 2497 * If the previous inode was never written or its previous directory 2498 * entry was never written, then we do not want to roll back to this 2499 * previous value. Instead we want to roll back to zero and immediately 2500 * free the unwritten or unreferenced inode. 2501 */ 2502 if (dirrem->dm_state & COMPLETE) { 2503 dap->da_state &= ~DIRCHG; 2504 dap->da_pagedep = pagedep; 2505 dirrem->dm_dirinum = pagedep->pd_ino; 2506 add_to_worklist(&dirrem->dm_list); 2507 } 2508 FREE_LOCK(&lk); 2509} 2510 2511/* 2512 * Called whenever the link count on an inode is increased. 2513 * It creates an inode dependency so that the new reference(s) 2514 * to the inode cannot be committed to disk until the updated 2515 * inode has been written. 2516 */ 2517void 2518softdep_increase_linkcnt(ip) 2519 struct inode *ip; /* the inode with the increased link count */ 2520{ 2521 struct inodedep *inodedep; 2522 2523 ACQUIRE_LOCK(&lk); 2524 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 2525 FREE_LOCK(&lk); 2526} 2527 2528/* 2529 * This workitem decrements the inode's link count. 2530 * If the link count reaches zero, the file is removed. 2531 */ 2532static void 2533handle_workitem_remove(dirrem) 2534 struct dirrem *dirrem; 2535{ 2536 struct proc *p = CURPROC; /* XXX */ 2537 struct inodedep *inodedep; 2538 struct vnode *vp; 2539 struct inode *ip; 2540 int error; 2541 2542 if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) { 2543 softdep_error("handle_workitem_remove: vget", error); 2544 return; 2545 } 2546 ip = VTOI(vp); 2547 /* 2548 * Normal file deletion. 2549 */ 2550 if ((dirrem->dm_state & RMDIR) == 0) { 2551 ip->i_nlink--; 2552 if (ip->i_nlink < ip->i_effnlink) 2553 panic("handle_workitem_remove: bad file delta"); 2554 ip->i_flag |= IN_CHANGE; 2555 vput(vp); 2556 WORKITEM_FREE(dirrem, D_DIRREM); 2557 return; 2558 } 2559 /* 2560 * Directory deletion. Decrement reference count for both the 2561 * just deleted parent directory entry and the reference for ".". 2562 * Next truncate the directory to length zero. When the 2563 * truncation completes, arrange to have the reference count on 2564 * the parent decremented to account for the loss of "..". 2565 */ 2566 ip->i_nlink -= 2; 2567 if (ip->i_nlink < ip->i_effnlink) 2568 panic("handle_workitem_remove: bad dir delta"); 2569 ip->i_flag |= IN_CHANGE; 2570 if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, p->p_ucred, p)) != 0) 2571 softdep_error("handle_workitem_remove: truncate", error); 2572 /* 2573 * Rename a directory to a new parent. Since, we are both deleting 2574 * and creating a new directory entry, the link count on the new 2575 * directory should not change. Thus we skip the followup dirrem. 2576 */ 2577 if (dirrem->dm_state & DIRCHG) { 2578 vput(vp); 2579 WORKITEM_FREE(dirrem, D_DIRREM); 2580 return; 2581 } 2582 ACQUIRE_LOCK(&lk); 2583 (void) inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, DEPALLOC, 2584 &inodedep); 2585 dirrem->dm_state = 0; 2586 dirrem->dm_oldinum = dirrem->dm_dirinum; 2587 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 2588 FREE_LOCK(&lk); 2589 vput(vp); 2590} 2591 2592/* 2593 * Inode de-allocation dependencies. 2594 * 2595 * When an inode's link count is reduced to zero, it can be de-allocated. We 2596 * found it convenient to postpone de-allocation until after the inode is 2597 * written to disk with its new link count (zero). At this point, all of the 2598 * on-disk inode's block pointers are nullified and, with careful dependency 2599 * list ordering, all dependencies related to the inode will be satisfied and 2600 * the corresponding dependency structures de-allocated. So, if/when the 2601 * inode is reused, there will be no mixing of old dependencies with new 2602 * ones. This artificial dependency is set up by the block de-allocation 2603 * procedure above (softdep_setup_freeblocks) and completed by the 2604 * following procedure. 2605 */ 2606static void 2607handle_workitem_freefile(freefile) 2608 struct freefile *freefile; 2609{ 2610 struct vnode vp; 2611 struct inode tip; 2612 struct inodedep *idp; 2613 int error; 2614 2615#ifdef DEBUG 2616 ACQUIRE_LOCK(&lk); 2617 if (inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp)) 2618 panic("handle_workitem_freefile: inodedep survived"); 2619 FREE_LOCK(&lk); 2620#endif 2621 tip.i_devvp = freefile->fx_devvp; 2622 tip.i_dev = freefile->fx_devvp->v_rdev; 2623 tip.i_fs = freefile->fx_fs; 2624 vp.v_data = &tip; 2625 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0) 2626 softdep_error("handle_workitem_freefile", error); 2627 WORKITEM_FREE(freefile, D_FREEFILE); 2628 num_freefile -= 1; 2629} 2630 2631/* 2632 * Disk writes. 2633 * 2634 * The dependency structures constructed above are most actively used when file 2635 * system blocks are written to disk. No constraints are placed on when a 2636 * block can be written, but unsatisfied update dependencies are made safe by 2637 * modifying (or replacing) the source memory for the duration of the disk 2638 * write. When the disk write completes, the memory block is again brought 2639 * up-to-date. 2640 * 2641 * In-core inode structure reclamation. 2642 * 2643 * Because there are a finite number of "in-core" inode structures, they are 2644 * reused regularly. By transferring all inode-related dependencies to the 2645 * in-memory inode block and indexing them separately (via "inodedep"s), we 2646 * can allow "in-core" inode structures to be reused at any time and avoid 2647 * any increase in contention. 2648 * 2649 * Called just before entering the device driver to initiate a new disk I/O. 2650 * The buffer must be locked, thus, no I/O completion operations can occur 2651 * while we are manipulating its associated dependencies. 2652 */ 2653void 2654softdep_disk_io_initiation(bp) 2655 struct buf *bp; /* structure describing disk write to occur */ 2656{ 2657 struct worklist *wk, *nextwk; 2658 struct indirdep *indirdep; 2659 2660 /* 2661 * We only care about write operations. There should never 2662 * be dependencies for reads. 2663 */ 2664 if (bp->b_flags & B_READ) 2665 panic("softdep_disk_io_initiation: read"); 2666 /* 2667 * Do any necessary pre-I/O processing. 2668 */ 2669 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = nextwk) { 2670 nextwk = LIST_NEXT(wk, wk_list); 2671 switch (wk->wk_type) { 2672 2673 case D_PAGEDEP: 2674 initiate_write_filepage(WK_PAGEDEP(wk), bp); 2675 continue; 2676 2677 case D_INODEDEP: 2678 initiate_write_inodeblock(WK_INODEDEP(wk), bp); 2679 continue; 2680 2681 case D_INDIRDEP: 2682 indirdep = WK_INDIRDEP(wk); 2683 if (indirdep->ir_state & GOINGAWAY) 2684 panic("disk_io_initiation: indirdep gone"); 2685 /* 2686 * If there are no remaining dependencies, this 2687 * will be writing the real pointers, so the 2688 * dependency can be freed. 2689 */ 2690 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) { 2691 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 2692 brelse(indirdep->ir_savebp); 2693 /* inline expand WORKLIST_REMOVE(wk); */ 2694 wk->wk_state &= ~ONWORKLIST; 2695 LIST_REMOVE(wk, wk_list); 2696 WORKITEM_FREE(indirdep, D_INDIRDEP); 2697 continue; 2698 } 2699 /* 2700 * Replace up-to-date version with safe version. 2701 */ 2702 ACQUIRE_LOCK(&lk); 2703 indirdep->ir_state &= ~ATTACHED; 2704 indirdep->ir_state |= UNDONE; 2705 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount, 2706 M_INDIRDEP, M_WAITOK); 2707 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 2708 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 2709 bp->b_bcount); 2710 FREE_LOCK(&lk); 2711 continue; 2712 2713 case D_MKDIR: 2714 case D_BMSAFEMAP: 2715 case D_ALLOCDIRECT: 2716 case D_ALLOCINDIR: 2717 continue; 2718 2719 default: 2720 panic("handle_disk_io_initiation: Unexpected type %s", 2721 TYPENAME(wk->wk_type)); 2722 /* NOTREACHED */ 2723 } 2724 } 2725} 2726 2727/* 2728 * Called from within the procedure above to deal with unsatisfied 2729 * allocation dependencies in a directory. The buffer must be locked, 2730 * thus, no I/O completion operations can occur while we are 2731 * manipulating its associated dependencies. 2732 */ 2733static void 2734initiate_write_filepage(pagedep, bp) 2735 struct pagedep *pagedep; 2736 struct buf *bp; 2737{ 2738 struct diradd *dap; 2739 struct direct *ep; 2740 int i; 2741 2742 if (pagedep->pd_state & IOSTARTED) { 2743 /* 2744 * This can only happen if there is a driver that does not 2745 * understand chaining. Here biodone will reissue the call 2746 * to strategy for the incomplete buffers. 2747 */ 2748 printf("initiate_write_filepage: already started\n"); 2749 return; 2750 } 2751 pagedep->pd_state |= IOSTARTED; 2752 ACQUIRE_LOCK(&lk); 2753 for (i = 0; i < DAHASHSZ; i++) { 2754 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 2755 dap = LIST_NEXT(dap, da_pdlist)) { 2756 ep = (struct direct *) 2757 ((char *)bp->b_data + dap->da_offset); 2758 if (ep->d_ino != dap->da_newinum) 2759 panic("%s: dir inum %d != new %d", 2760 "initiate_write_filepage", 2761 ep->d_ino, dap->da_newinum); 2762 if (dap->da_state & DIRCHG) 2763 ep->d_ino = dap->da_previous->dm_oldinum; 2764 else 2765 ep->d_ino = 0; 2766 dap->da_state &= ~ATTACHED; 2767 dap->da_state |= UNDONE; 2768 } 2769 } 2770 FREE_LOCK(&lk); 2771} 2772 2773/* 2774 * Called from within the procedure above to deal with unsatisfied 2775 * allocation dependencies in an inodeblock. The buffer must be 2776 * locked, thus, no I/O completion operations can occur while we 2777 * are manipulating its associated dependencies. 2778 */ 2779static void 2780initiate_write_inodeblock(inodedep, bp) 2781 struct inodedep *inodedep; 2782 struct buf *bp; /* The inode block */ 2783{ 2784 struct allocdirect *adp, *lastadp; 2785 struct dinode *dp; 2786 struct fs *fs; 2787 ufs_lbn_t prevlbn = 0; 2788 int i, deplist; 2789 2790 if (inodedep->id_state & IOSTARTED) 2791 panic("initiate_write_inodeblock: already started"); 2792 inodedep->id_state |= IOSTARTED; 2793 fs = inodedep->id_fs; 2794 dp = (struct dinode *)bp->b_data + 2795 ino_to_fsbo(fs, inodedep->id_ino); 2796 /* 2797 * If the bitmap is not yet written, then the allocated 2798 * inode cannot be written to disk. 2799 */ 2800 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 2801 if (inodedep->id_savedino != NULL) 2802 panic("initiate_write_inodeblock: already doing I/O"); 2803 MALLOC(inodedep->id_savedino, struct dinode *, 2804 sizeof(struct dinode), M_INODEDEP, M_WAITOK); 2805 *inodedep->id_savedino = *dp; 2806 bzero((caddr_t)dp, sizeof(struct dinode)); 2807 return; 2808 } 2809 /* 2810 * If no dependencies, then there is nothing to roll back. 2811 */ 2812 inodedep->id_savedsize = dp->di_size; 2813 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 2814 return; 2815 /* 2816 * Set the dependencies to busy. 2817 */ 2818 ACQUIRE_LOCK(&lk); 2819 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 2820 adp = TAILQ_NEXT(adp, ad_next)) { 2821#ifdef DIAGNOSTIC 2822 if (deplist != 0 && prevlbn >= adp->ad_lbn) 2823 panic("softdep_write_inodeblock: lbn order"); 2824 prevlbn = adp->ad_lbn; 2825 if (adp->ad_lbn < NDADDR && 2826 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) 2827 panic("%s: direct pointer #%ld mismatch %d != %d", 2828 "softdep_write_inodeblock", adp->ad_lbn, 2829 dp->di_db[adp->ad_lbn], adp->ad_newblkno); 2830 if (adp->ad_lbn >= NDADDR && 2831 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) 2832 panic("%s: indirect pointer #%ld mismatch %d != %d", 2833 "softdep_write_inodeblock", adp->ad_lbn - NDADDR, 2834 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno); 2835 deplist |= 1 << adp->ad_lbn; 2836 if ((adp->ad_state & ATTACHED) == 0) 2837 panic("softdep_write_inodeblock: Unknown state 0x%x", 2838 adp->ad_state); 2839#endif /* DIAGNOSTIC */ 2840 adp->ad_state &= ~ATTACHED; 2841 adp->ad_state |= UNDONE; 2842 } 2843 /* 2844 * The on-disk inode cannot claim to be any larger than the last 2845 * fragment that has been written. Otherwise, the on-disk inode 2846 * might have fragments that were not the last block in the file 2847 * which would corrupt the filesystem. 2848 */ 2849 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 2850 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 2851 if (adp->ad_lbn >= NDADDR) 2852 break; 2853 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 2854 /* keep going until hitting a rollback to a frag */ 2855 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 2856 continue; 2857 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 2858 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 2859#ifdef DIAGNOSTIC 2860 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 2861 panic("softdep_write_inodeblock: lost dep1"); 2862#endif /* DIAGNOSTIC */ 2863 dp->di_db[i] = 0; 2864 } 2865 for (i = 0; i < NIADDR; i++) { 2866#ifdef DIAGNOSTIC 2867 if (dp->di_ib[i] != 0 && 2868 (deplist & ((1 << NDADDR) << i)) == 0) 2869 panic("softdep_write_inodeblock: lost dep2"); 2870#endif /* DIAGNOSTIC */ 2871 dp->di_ib[i] = 0; 2872 } 2873 FREE_LOCK(&lk); 2874 return; 2875 } 2876 /* 2877 * If we have zero'ed out the last allocated block of the file, 2878 * roll back the size to the last currently allocated block. 2879 * We know that this last allocated block is a full-sized as 2880 * we already checked for fragments in the loop above. 2881 */ 2882 if (lastadp != NULL && 2883 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 2884 for (i = lastadp->ad_lbn; i >= 0; i--) 2885 if (dp->di_db[i] != 0) 2886 break; 2887 dp->di_size = (i + 1) * fs->fs_bsize; 2888 } 2889 /* 2890 * The only dependencies are for indirect blocks. 2891 * 2892 * The file size for indirect block additions is not guaranteed. 2893 * Such a guarantee would be non-trivial to achieve. The conventional 2894 * synchronous write implementation also does not make this guarantee. 2895 * Fsck should catch and fix discrepancies. Arguably, the file size 2896 * can be over-estimated without destroying integrity when the file 2897 * moves into the indirect blocks (i.e., is large). If we want to 2898 * postpone fsck, we are stuck with this argument. 2899 */ 2900 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 2901 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 2902 FREE_LOCK(&lk); 2903} 2904 2905/* 2906 * This routine is called during the completion interrupt 2907 * service routine for a disk write (from the procedure called 2908 * by the device driver to inform the file system caches of 2909 * a request completion). It should be called early in this 2910 * procedure, before the block is made available to other 2911 * processes or other routines are called. 2912 */ 2913void 2914softdep_disk_write_complete(bp) 2915 struct buf *bp; /* describes the completed disk write */ 2916{ 2917 struct worklist *wk; 2918 struct workhead reattach; 2919 struct newblk *newblk; 2920 struct allocindir *aip; 2921 struct allocdirect *adp; 2922 struct indirdep *indirdep; 2923 struct inodedep *inodedep; 2924 struct bmsafemap *bmsafemap; 2925 2926#ifdef DEBUG 2927 if (lk.lkt_held != -1) 2928 panic("softdep_disk_write_complete: lock is held"); 2929 lk.lkt_held = -2; 2930#endif 2931 LIST_INIT(&reattach); 2932 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2933 WORKLIST_REMOVE(wk); 2934 switch (wk->wk_type) { 2935 2936 case D_PAGEDEP: 2937 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 2938 WORKLIST_INSERT(&reattach, wk); 2939 continue; 2940 2941 case D_INODEDEP: 2942 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 2943 WORKLIST_INSERT(&reattach, wk); 2944 continue; 2945 2946 case D_BMSAFEMAP: 2947 bmsafemap = WK_BMSAFEMAP(wk); 2948 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) { 2949 newblk->nb_state |= DEPCOMPLETE; 2950 newblk->nb_bmsafemap = NULL; 2951 LIST_REMOVE(newblk, nb_deps); 2952 } 2953 while ((adp = 2954 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) { 2955 adp->ad_state |= DEPCOMPLETE; 2956 adp->ad_buf = NULL; 2957 LIST_REMOVE(adp, ad_deps); 2958 handle_allocdirect_partdone(adp); 2959 } 2960 while ((aip = 2961 LIST_FIRST(&bmsafemap->sm_allocindirhd))) { 2962 aip->ai_state |= DEPCOMPLETE; 2963 aip->ai_buf = NULL; 2964 LIST_REMOVE(aip, ai_deps); 2965 handle_allocindir_partdone(aip); 2966 } 2967 while ((inodedep = 2968 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { 2969 inodedep->id_state |= DEPCOMPLETE; 2970 LIST_REMOVE(inodedep, id_deps); 2971 inodedep->id_buf = NULL; 2972 } 2973 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 2974 continue; 2975 2976 case D_MKDIR: 2977 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 2978 continue; 2979 2980 case D_ALLOCDIRECT: 2981 adp = WK_ALLOCDIRECT(wk); 2982 adp->ad_state |= COMPLETE; 2983 handle_allocdirect_partdone(adp); 2984 continue; 2985 2986 case D_ALLOCINDIR: 2987 aip = WK_ALLOCINDIR(wk); 2988 aip->ai_state |= COMPLETE; 2989 handle_allocindir_partdone(aip); 2990 continue; 2991 2992 case D_INDIRDEP: 2993 indirdep = WK_INDIRDEP(wk); 2994 if (indirdep->ir_state & GOINGAWAY) 2995 panic("disk_write_complete: indirdep gone"); 2996 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 2997 FREE(indirdep->ir_saveddata, M_INDIRDEP); 2998 indirdep->ir_saveddata = 0; 2999 indirdep->ir_state &= ~UNDONE; 3000 indirdep->ir_state |= ATTACHED; 3001 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 3002 handle_allocindir_partdone(aip); 3003 if (aip == LIST_FIRST(&indirdep->ir_donehd)) 3004 panic("disk_write_complete: not gone"); 3005 } 3006 WORKLIST_INSERT(&reattach, wk); 3007 if ((bp->b_flags & B_DELWRI) == 0) 3008 stat_indir_blk_ptrs++; 3009 bdirty(bp); 3010 continue; 3011 3012 default: 3013 panic("handle_disk_write_complete: Unknown type %s", 3014 TYPENAME(wk->wk_type)); 3015 /* NOTREACHED */ 3016 } 3017 } 3018 /* 3019 * Reattach any requests that must be redone. 3020 */ 3021 while ((wk = LIST_FIRST(&reattach)) != NULL) { 3022 WORKLIST_REMOVE(wk); 3023 WORKLIST_INSERT(&bp->b_dep, wk); 3024 } 3025#ifdef DEBUG 3026 if (lk.lkt_held != -2) 3027 panic("softdep_disk_write_complete: lock lost"); 3028 lk.lkt_held = -1; 3029#endif 3030} 3031 3032/* 3033 * Called from within softdep_disk_write_complete above. Note that 3034 * this routine is always called from interrupt level with further 3035 * splbio interrupts blocked. 3036 */ 3037static void 3038handle_allocdirect_partdone(adp) 3039 struct allocdirect *adp; /* the completed allocdirect */ 3040{ 3041 struct allocdirect *listadp; 3042 struct inodedep *inodedep; 3043 long bsize; 3044 3045 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3046 return; 3047 if (adp->ad_buf != NULL) 3048 panic("handle_allocdirect_partdone: dangling dep"); 3049 /* 3050 * The on-disk inode cannot claim to be any larger than the last 3051 * fragment that has been written. Otherwise, the on-disk inode 3052 * might have fragments that were not the last block in the file 3053 * which would corrupt the filesystem. Thus, we cannot free any 3054 * allocdirects after one whose ad_oldblkno claims a fragment as 3055 * these blocks must be rolled back to zero before writing the inode. 3056 * We check the currently active set of allocdirects in id_inoupdt. 3057 */ 3058 inodedep = adp->ad_inodedep; 3059 bsize = inodedep->id_fs->fs_bsize; 3060 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp; 3061 listadp = TAILQ_NEXT(listadp, ad_next)) { 3062 /* found our block */ 3063 if (listadp == adp) 3064 break; 3065 /* continue if ad_oldlbn is not a fragment */ 3066 if (listadp->ad_oldsize == 0 || 3067 listadp->ad_oldsize == bsize) 3068 continue; 3069 /* hit a fragment */ 3070 return; 3071 } 3072 /* 3073 * If we have reached the end of the current list without 3074 * finding the just finished dependency, then it must be 3075 * on the future dependency list. Future dependencies cannot 3076 * be freed until they are moved to the current list. 3077 */ 3078 if (listadp == NULL) { 3079#ifdef DEBUG 3080 for (listadp = TAILQ_FIRST(&inodedep->id_newinoupdt); listadp; 3081 listadp = TAILQ_NEXT(listadp, ad_next)) 3082 /* found our block */ 3083 if (listadp == adp) 3084 break; 3085 if (listadp == NULL) 3086 panic("handle_allocdirect_partdone: lost dep"); 3087#endif /* DEBUG */ 3088 return; 3089 } 3090 /* 3091 * If we have found the just finished dependency, then free 3092 * it along with anything that follows it that is complete. 3093 */ 3094 for (; adp; adp = listadp) { 3095 listadp = TAILQ_NEXT(adp, ad_next); 3096 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3097 return; 3098 free_allocdirect(&inodedep->id_inoupdt, adp, 1); 3099 } 3100} 3101 3102/* 3103 * Called from within softdep_disk_write_complete above. Note that 3104 * this routine is always called from interrupt level with further 3105 * splbio interrupts blocked. 3106 */ 3107static void 3108handle_allocindir_partdone(aip) 3109 struct allocindir *aip; /* the completed allocindir */ 3110{ 3111 struct indirdep *indirdep; 3112 3113 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 3114 return; 3115 if (aip->ai_buf != NULL) 3116 panic("handle_allocindir_partdone: dangling dependency"); 3117 indirdep = aip->ai_indirdep; 3118 if (indirdep->ir_state & UNDONE) { 3119 LIST_REMOVE(aip, ai_next); 3120 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 3121 return; 3122 } 3123 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 3124 aip->ai_newblkno; 3125 LIST_REMOVE(aip, ai_next); 3126 if (aip->ai_freefrag != NULL) 3127 add_to_worklist(&aip->ai_freefrag->ff_list); 3128 WORKITEM_FREE(aip, D_ALLOCINDIR); 3129} 3130 3131/* 3132 * Called from within softdep_disk_write_complete above to restore 3133 * in-memory inode block contents to their most up-to-date state. Note 3134 * that this routine is always called from interrupt level with further 3135 * splbio interrupts blocked. 3136 */ 3137static int 3138handle_written_inodeblock(inodedep, bp) 3139 struct inodedep *inodedep; 3140 struct buf *bp; /* buffer containing the inode block */ 3141{ 3142 struct worklist *wk, *filefree; 3143 struct allocdirect *adp, *nextadp; 3144 struct dinode *dp; 3145 int hadchanges; 3146 3147 if ((inodedep->id_state & IOSTARTED) == 0) 3148 panic("handle_written_inodeblock: not started"); 3149 inodedep->id_state &= ~IOSTARTED; 3150 inodedep->id_state |= COMPLETE; 3151 dp = (struct dinode *)bp->b_data + 3152 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 3153 /* 3154 * If we had to rollback the inode allocation because of 3155 * bitmaps being incomplete, then simply restore it. 3156 * Keep the block dirty so that it will not be reclaimed until 3157 * all associated dependencies have been cleared and the 3158 * corresponding updates written to disk. 3159 */ 3160 if (inodedep->id_savedino != NULL) { 3161 *dp = *inodedep->id_savedino; 3162 FREE(inodedep->id_savedino, M_INODEDEP); 3163 inodedep->id_savedino = NULL; 3164 if ((bp->b_flags & B_DELWRI) == 0) 3165 stat_inode_bitmap++; 3166 bdirty(bp); 3167 return (1); 3168 } 3169 /* 3170 * Roll forward anything that had to be rolled back before 3171 * the inode could be updated. 3172 */ 3173 hadchanges = 0; 3174 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 3175 nextadp = TAILQ_NEXT(adp, ad_next); 3176 if (adp->ad_state & ATTACHED) 3177 panic("handle_written_inodeblock: new entry"); 3178 if (adp->ad_lbn < NDADDR) { 3179 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) 3180 panic("%s: %s #%ld mismatch %d != %d", 3181 "handle_written_inodeblock", 3182 "direct pointer", adp->ad_lbn, 3183 dp->di_db[adp->ad_lbn], adp->ad_oldblkno); 3184 dp->di_db[adp->ad_lbn] = adp->ad_newblkno; 3185 } else { 3186 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) 3187 panic("%s: %s #%ld allocated as %d", 3188 "handle_written_inodeblock", 3189 "indirect pointer", adp->ad_lbn - NDADDR, 3190 dp->di_ib[adp->ad_lbn - NDADDR]); 3191 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno; 3192 } 3193 adp->ad_state &= ~UNDONE; 3194 adp->ad_state |= ATTACHED; 3195 hadchanges = 1; 3196 } 3197 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 3198 stat_direct_blk_ptrs++; 3199 /* 3200 * Reset the file size to its most up-to-date value. 3201 */ 3202 if (inodedep->id_savedsize == -1) 3203 panic("handle_written_inodeblock: bad size"); 3204 if (dp->di_size != inodedep->id_savedsize) { 3205 dp->di_size = inodedep->id_savedsize; 3206 hadchanges = 1; 3207 } 3208 inodedep->id_savedsize = -1; 3209 /* 3210 * If there were any rollbacks in the inode block, then it must be 3211 * marked dirty so that its will eventually get written back in 3212 * its correct form. 3213 */ 3214 if (hadchanges) 3215 bdirty(bp); 3216 /* 3217 * Process any allocdirects that completed during the update. 3218 */ 3219 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 3220 handle_allocdirect_partdone(adp); 3221 /* 3222 * Process deallocations that were held pending until the 3223 * inode had been written to disk. Freeing of the inode 3224 * is delayed until after all blocks have been freed to 3225 * avoid creation of new <vfsid, inum, lbn> triples 3226 * before the old ones have been deleted. 3227 */ 3228 filefree = NULL; 3229 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 3230 WORKLIST_REMOVE(wk); 3231 switch (wk->wk_type) { 3232 3233 case D_FREEFILE: 3234 /* 3235 * We defer adding filefree to the worklist until 3236 * all other additions have been made to ensure 3237 * that it will be done after all the old blocks 3238 * have been freed. 3239 */ 3240 if (filefree != NULL) 3241 panic("handle_written_inodeblock: filefree"); 3242 filefree = wk; 3243 continue; 3244 3245 case D_MKDIR: 3246 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 3247 continue; 3248 3249 case D_DIRADD: 3250 diradd_inode_written(WK_DIRADD(wk), inodedep); 3251 continue; 3252 3253 case D_FREEBLKS: 3254 case D_FREEFRAG: 3255 case D_DIRREM: 3256 add_to_worklist(wk); 3257 continue; 3258 3259 default: 3260 panic("handle_written_inodeblock: Unknown type %s", 3261 TYPENAME(wk->wk_type)); 3262 /* NOTREACHED */ 3263 } 3264 } 3265 if (filefree != NULL) { 3266 if (free_inodedep(inodedep) == 0) 3267 panic("handle_written_inodeblock: live inodedep"); 3268 add_to_worklist(filefree); 3269 return (0); 3270 } 3271 3272 /* 3273 * If no outstanding dependencies, free it. 3274 */ 3275 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0) 3276 return (0); 3277 return (hadchanges); 3278} 3279 3280/* 3281 * Process a diradd entry after its dependent inode has been written. 3282 * This routine must be called with splbio interrupts blocked. 3283 */ 3284static void 3285diradd_inode_written(dap, inodedep) 3286 struct diradd *dap; 3287 struct inodedep *inodedep; 3288{ 3289 struct pagedep *pagedep; 3290 3291 dap->da_state |= COMPLETE; 3292 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3293 if (dap->da_state & DIRCHG) 3294 pagedep = dap->da_previous->dm_pagedep; 3295 else 3296 pagedep = dap->da_pagedep; 3297 LIST_REMOVE(dap, da_pdlist); 3298 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3299 } 3300 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 3301} 3302 3303/* 3304 * Handle the completion of a mkdir dependency. 3305 */ 3306static void 3307handle_written_mkdir(mkdir, type) 3308 struct mkdir *mkdir; 3309 int type; 3310{ 3311 struct diradd *dap; 3312 struct pagedep *pagedep; 3313 3314 if (mkdir->md_state != type) 3315 panic("handle_written_mkdir: bad type"); 3316 dap = mkdir->md_diradd; 3317 dap->da_state &= ~type; 3318 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 3319 dap->da_state |= DEPCOMPLETE; 3320 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3321 if (dap->da_state & DIRCHG) 3322 pagedep = dap->da_previous->dm_pagedep; 3323 else 3324 pagedep = dap->da_pagedep; 3325 LIST_REMOVE(dap, da_pdlist); 3326 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3327 } 3328 LIST_REMOVE(mkdir, md_mkdirs); 3329 WORKITEM_FREE(mkdir, D_MKDIR); 3330} 3331 3332/* 3333 * Called from within softdep_disk_write_complete above. 3334 * A write operation was just completed. Removed inodes can 3335 * now be freed and associated block pointers may be committed. 3336 * Note that this routine is always called from interrupt level 3337 * with further splbio interrupts blocked. 3338 */ 3339static int 3340handle_written_filepage(pagedep, bp) 3341 struct pagedep *pagedep; 3342 struct buf *bp; /* buffer containing the written page */ 3343{ 3344 struct dirrem *dirrem; 3345 struct diradd *dap, *nextdap; 3346 struct direct *ep; 3347 int i, chgs; 3348 3349 if ((pagedep->pd_state & IOSTARTED) == 0) 3350 panic("handle_written_filepage: not started"); 3351 pagedep->pd_state &= ~IOSTARTED; 3352 /* 3353 * Process any directory removals that have been committed. 3354 */ 3355 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 3356 LIST_REMOVE(dirrem, dm_next); 3357 dirrem->dm_dirinum = pagedep->pd_ino; 3358 add_to_worklist(&dirrem->dm_list); 3359 } 3360 /* 3361 * Free any directory additions that have been committed. 3362 */ 3363 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 3364 free_diradd(dap); 3365 /* 3366 * Uncommitted directory entries must be restored. 3367 */ 3368 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 3369 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 3370 dap = nextdap) { 3371 nextdap = LIST_NEXT(dap, da_pdlist); 3372 if (dap->da_state & ATTACHED) 3373 panic("handle_written_filepage: attached"); 3374 ep = (struct direct *) 3375 ((char *)bp->b_data + dap->da_offset); 3376 ep->d_ino = dap->da_newinum; 3377 dap->da_state &= ~UNDONE; 3378 dap->da_state |= ATTACHED; 3379 chgs = 1; 3380 /* 3381 * If the inode referenced by the directory has 3382 * been written out, then the dependency can be 3383 * moved to the pending list. 3384 */ 3385 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3386 LIST_REMOVE(dap, da_pdlist); 3387 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 3388 da_pdlist); 3389 } 3390 } 3391 } 3392 /* 3393 * If there were any rollbacks in the directory, then it must be 3394 * marked dirty so that its will eventually get written back in 3395 * its correct form. 3396 */ 3397 if (chgs) { 3398 if ((bp->b_flags & B_DELWRI) == 0) 3399 stat_dir_entry++; 3400 bdirty(bp); 3401 } 3402 /* 3403 * If no dependencies remain, the pagedep will be freed. 3404 * Otherwise it will remain to update the page before it 3405 * is written back to disk. 3406 */ 3407 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) { 3408 for (i = 0; i < DAHASHSZ; i++) 3409 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL) 3410 break; 3411 if (i == DAHASHSZ) { 3412 LIST_REMOVE(pagedep, pd_hash); 3413 WORKITEM_FREE(pagedep, D_PAGEDEP); 3414 return (0); 3415 } 3416 } 3417 return (1); 3418} 3419 3420/* 3421 * Writing back in-core inode structures. 3422 * 3423 * The file system only accesses an inode's contents when it occupies an 3424 * "in-core" inode structure. These "in-core" structures are separate from 3425 * the page frames used to cache inode blocks. Only the latter are 3426 * transferred to/from the disk. So, when the updated contents of the 3427 * "in-core" inode structure are copied to the corresponding in-memory inode 3428 * block, the dependencies are also transferred. The following procedure is 3429 * called when copying a dirty "in-core" inode to a cached inode block. 3430 */ 3431 3432/* 3433 * Called when an inode is loaded from disk. If the effective link count 3434 * differed from the actual link count when it was last flushed, then we 3435 * need to ensure that the correct effective link count is put back. 3436 */ 3437void 3438softdep_load_inodeblock(ip) 3439 struct inode *ip; /* the "in_core" copy of the inode */ 3440{ 3441 struct inodedep *inodedep; 3442 3443 /* 3444 * Check for alternate nlink count. 3445 */ 3446 ip->i_effnlink = ip->i_nlink; 3447 ACQUIRE_LOCK(&lk); 3448 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3449 FREE_LOCK(&lk); 3450 return; 3451 } 3452 if (inodedep->id_nlinkdelta != 0) { 3453 ip->i_effnlink -= inodedep->id_nlinkdelta; 3454 ip->i_flag |= IN_MODIFIED; 3455 inodedep->id_nlinkdelta = 0; 3456 (void) free_inodedep(inodedep); 3457 } 3458 FREE_LOCK(&lk); 3459} 3460 3461/* 3462 * This routine is called just before the "in-core" inode 3463 * information is to be copied to the in-memory inode block. 3464 * Recall that an inode block contains several inodes. If 3465 * the force flag is set, then the dependencies will be 3466 * cleared so that the update can always be made. Note that 3467 * the buffer is locked when this routine is called, so we 3468 * will never be in the middle of writing the inode block 3469 * to disk. 3470 */ 3471void 3472softdep_update_inodeblock(ip, bp, waitfor) 3473 struct inode *ip; /* the "in_core" copy of the inode */ 3474 struct buf *bp; /* the buffer containing the inode block */ 3475 int waitfor; /* nonzero => update must be allowed */ 3476{ 3477 struct inodedep *inodedep; 3478 struct worklist *wk; 3479 int error, gotit; 3480 3481 /* 3482 * If the effective link count is not equal to the actual link 3483 * count, then we must track the difference in an inodedep while 3484 * the inode is (potentially) tossed out of the cache. Otherwise, 3485 * if there is no existing inodedep, then there are no dependencies 3486 * to track. 3487 */ 3488 ACQUIRE_LOCK(&lk); 3489 if (ip->i_effnlink != ip->i_nlink) { 3490 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, 3491 &inodedep); 3492 } else if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3493 FREE_LOCK(&lk); 3494 return; 3495 } 3496 if (ip->i_nlink < ip->i_effnlink) 3497 panic("softdep_update_inodeblock: bad delta"); 3498 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3499 /* 3500 * Changes have been initiated. Anything depending on these 3501 * changes cannot occur until this inode has been written. 3502 */ 3503 inodedep->id_state &= ~COMPLETE; 3504 if ((inodedep->id_state & ONWORKLIST) == 0) 3505 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 3506 /* 3507 * Any new dependencies associated with the incore inode must 3508 * now be moved to the list associated with the buffer holding 3509 * the in-memory copy of the inode. Once merged process any 3510 * allocdirects that are completed by the merger. 3511 */ 3512 merge_inode_lists(inodedep); 3513 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL) 3514 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); 3515 /* 3516 * Now that the inode has been pushed into the buffer, the 3517 * operations dependent on the inode being written to disk 3518 * can be moved to the id_bufwait so that they will be 3519 * processed when the buffer I/O completes. 3520 */ 3521 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 3522 WORKLIST_REMOVE(wk); 3523 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 3524 } 3525 /* 3526 * Newly allocated inodes cannot be written until the bitmap 3527 * that allocates them have been written (indicated by 3528 * DEPCOMPLETE being set in id_state). If we are doing a 3529 * forced sync (e.g., an fsync on a file), we force the bitmap 3530 * to be written so that the update can be done. 3531 */ 3532 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) { 3533 FREE_LOCK(&lk); 3534 return; 3535 } 3536 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 3537 FREE_LOCK(&lk); 3538 if (gotit && 3539 (error = VOP_BWRITE(inodedep->id_buf->b_vp, inodedep->id_buf)) != 0) 3540 softdep_error("softdep_update_inodeblock: bwrite", error); 3541 if ((inodedep->id_state & DEPCOMPLETE) == 0) 3542 panic("softdep_update_inodeblock: update failed"); 3543} 3544 3545/* 3546 * Merge the new inode dependency list (id_newinoupdt) into the old 3547 * inode dependency list (id_inoupdt). This routine must be called 3548 * with splbio interrupts blocked. 3549 */ 3550static void 3551merge_inode_lists(inodedep) 3552 struct inodedep *inodedep; 3553{ 3554 struct allocdirect *listadp, *newadp; 3555 3556 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 3557 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) { 3558 if (listadp->ad_lbn < newadp->ad_lbn) { 3559 listadp = TAILQ_NEXT(listadp, ad_next); 3560 continue; 3561 } 3562 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 3563 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 3564 if (listadp->ad_lbn == newadp->ad_lbn) { 3565 allocdirect_merge(&inodedep->id_inoupdt, newadp, 3566 listadp); 3567 listadp = newadp; 3568 } 3569 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 3570 } 3571 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) { 3572 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 3573 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next); 3574 } 3575} 3576 3577/* 3578 * If we are doing an fsync, then we must ensure that any directory 3579 * entries for the inode have been written after the inode gets to disk. 3580 */ 3581int 3582softdep_fsync(vp) 3583 struct vnode *vp; /* the "in_core" copy of the inode */ 3584{ 3585 struct diradd *dap, *olddap; 3586 struct inodedep *inodedep; 3587 struct pagedep *pagedep; 3588 struct worklist *wk; 3589 struct mount *mnt; 3590 struct vnode *pvp; 3591 struct inode *ip; 3592 struct buf *bp; 3593 struct fs *fs; 3594 struct proc *p = CURPROC; /* XXX */ 3595 int error, ret, flushparent; 3596 ino_t parentino; 3597 ufs_lbn_t lbn; 3598 3599 ip = VTOI(vp); 3600 fs = ip->i_fs; 3601 for (error = 0, flushparent = 0, olddap = NULL; ; ) { 3602 ACQUIRE_LOCK(&lk); 3603 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) 3604 break; 3605 if (LIST_FIRST(&inodedep->id_inowait) != NULL || 3606 LIST_FIRST(&inodedep->id_bufwait) != NULL || 3607 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 3608 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) 3609 panic("softdep_fsync: pending ops"); 3610 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 3611 break; 3612 if (wk->wk_type != D_DIRADD) 3613 panic("softdep_fsync: Unexpected type %s", 3614 TYPENAME(wk->wk_type)); 3615 dap = WK_DIRADD(wk); 3616 /* 3617 * If we have failed to get rid of all the dependencies 3618 * then something is seriously wrong. 3619 */ 3620 if (dap == olddap) 3621 panic("softdep_fsync: flush failed"); 3622 olddap = dap; 3623 /* 3624 * Flush our parent if this directory entry 3625 * has a MKDIR_PARENT dependency. 3626 */ 3627 if (dap->da_state & DIRCHG) 3628 pagedep = dap->da_previous->dm_pagedep; 3629 else 3630 pagedep = dap->da_pagedep; 3631 mnt = pagedep->pd_mnt; 3632 parentino = pagedep->pd_ino; 3633 lbn = pagedep->pd_lbn; 3634 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) 3635 panic("softdep_fsync: dirty"); 3636 flushparent = dap->da_state & MKDIR_PARENT; 3637 /* 3638 * If we are being fsync'ed as part of vgone'ing this vnode, 3639 * then we will not be able to release and recover the 3640 * vnode below, so we just have to give up on writing its 3641 * directory entry out. It will eventually be written, just 3642 * not now, but then the user was not asking to have it 3643 * written, so we are not breaking any promises. 3644 */ 3645 if (vp->v_flag & VXLOCK) 3646 break; 3647 /* 3648 * We prevent deadlock by always fetching inodes from the 3649 * root, moving down the directory tree. Thus, when fetching 3650 * our parent directory, we must unlock ourselves before 3651 * requesting the lock on our parent. See the comment in 3652 * ufs_lookup for details on possible races. 3653 */ 3654 FREE_LOCK(&lk); 3655 VOP_UNLOCK(vp, 0, p); 3656 if ((error = VFS_VGET(mnt, parentino, &pvp)) != 0) { 3657 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, p); 3658 return (error); 3659 } 3660 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, p); 3661 if (flushparent) { 3662 if ((error = UFS_UPDATE(pvp, 1)) != 0) { 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->b_vp, 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 * Flush all the dirty bitmaps associated with the block device 3685 * before flushing the rest of the dirty blocks so as to reduce 3686 * the number of dependencies that will have to be rolled back. 3687 */ 3688void 3689softdep_fsync_mountdev(vp) 3690 struct vnode *vp; 3691{ 3692 struct buf *bp, *nbp; 3693 struct worklist *wk; 3694 3695 if (vp->v_type != VBLK) 3696 panic("softdep_fsync_mountdev: vnode not VBLK"); 3697 ACQUIRE_LOCK(&lk); 3698 for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 3699 nbp = TAILQ_NEXT(bp, b_vnbufs); 3700 /* 3701 * If it is already scheduled, skip to the next buffer. 3702 */ 3703 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) 3704 continue; 3705 if ((bp->b_flags & B_DELWRI) == 0) 3706 panic("softdep_fsync_mountdev: not dirty"); 3707 /* 3708 * We are only interested in bitmaps with outstanding 3709 * dependencies. 3710 */ 3711 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 3712 wk->wk_type != D_BMSAFEMAP) { 3713 BUF_UNLOCK(bp); 3714 continue; 3715 } 3716 bremfree(bp); 3717 FREE_LOCK(&lk); 3718 (void) bawrite(bp); 3719 ACQUIRE_LOCK(&lk); 3720 /* 3721 * Since we may have slept during the I/O, we need 3722 * to start from a known point. 3723 */ 3724 nbp = TAILQ_FIRST(&vp->v_dirtyblkhd); 3725 } 3726 drain_output(vp, 1); 3727 FREE_LOCK(&lk); 3728} 3729 3730/* 3731 * This routine is called when we are trying to synchronously flush a 3732 * file. This routine must eliminate any filesystem metadata dependencies 3733 * so that the syncing routine can succeed by pushing the dirty blocks 3734 * associated with the file. If any I/O errors occur, they are returned. 3735 */ 3736int 3737softdep_sync_metadata(ap) 3738 struct vop_fsync_args /* { 3739 struct vnode *a_vp; 3740 struct ucred *a_cred; 3741 int a_waitfor; 3742 struct proc *a_p; 3743 } */ *ap; 3744{ 3745 struct vnode *vp = ap->a_vp; 3746 struct pagedep *pagedep; 3747 struct allocdirect *adp; 3748 struct allocindir *aip; 3749 struct buf *bp, *nbp; 3750 struct worklist *wk; 3751 int i, error, waitfor; 3752 3753 /* 3754 * Check whether this vnode is involved in a filesystem 3755 * that is doing soft dependency processing. 3756 */ 3757 if (vp->v_type != VBLK) { 3758 if (!DOINGSOFTDEP(vp)) 3759 return (0); 3760 } else 3761 if (vp->v_specmountpoint == NULL || 3762 (vp->v_specmountpoint->mnt_flag & MNT_SOFTDEP) == 0) 3763 return (0); 3764 /* 3765 * Ensure that any direct block dependencies have been cleared. 3766 */ 3767 ACQUIRE_LOCK(&lk); 3768 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) { 3769 FREE_LOCK(&lk); 3770 return (error); 3771 } 3772 /* 3773 * For most files, the only metadata dependencies are the 3774 * cylinder group maps that allocate their inode or blocks. 3775 * The block allocation dependencies can be found by traversing 3776 * the dependency lists for any buffers that remain on their 3777 * dirty buffer list. The inode allocation dependency will 3778 * be resolved when the inode is updated with MNT_WAIT. 3779 * This work is done in two passes. The first pass grabs most 3780 * of the buffers and begins asynchronously writing them. The 3781 * only way to wait for these asynchronous writes is to sleep 3782 * on the filesystem vnode which may stay busy for a long time 3783 * if the filesystem is active. So, instead, we make a second 3784 * pass over the dependencies blocking on each write. In the 3785 * usual case we will be blocking against a write that we 3786 * initiated, so when it is done the dependency will have been 3787 * resolved. Thus the second pass is expected to end quickly. 3788 */ 3789 waitfor = MNT_NOWAIT; 3790top: 3791 if (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT) == 0) { 3792 FREE_LOCK(&lk); 3793 return (0); 3794 } 3795 bp = TAILQ_FIRST(&vp->v_dirtyblkhd); 3796loop: 3797 /* 3798 * As we hold the buffer locked, none of its dependencies 3799 * will disappear. 3800 */ 3801 for (wk = LIST_FIRST(&bp->b_dep); wk; 3802 wk = LIST_NEXT(wk, wk_list)) { 3803 switch (wk->wk_type) { 3804 3805 case D_ALLOCDIRECT: 3806 adp = WK_ALLOCDIRECT(wk); 3807 if (adp->ad_state & DEPCOMPLETE) 3808 break; 3809 nbp = adp->ad_buf; 3810 if (getdirtybuf(&nbp, waitfor) == 0) 3811 break; 3812 FREE_LOCK(&lk); 3813 if (waitfor == MNT_NOWAIT) { 3814 bawrite(nbp); 3815 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 3816 bawrite(bp); 3817 return (error); 3818 } 3819 ACQUIRE_LOCK(&lk); 3820 break; 3821 3822 case D_ALLOCINDIR: 3823 aip = WK_ALLOCINDIR(wk); 3824 if (aip->ai_state & DEPCOMPLETE) 3825 break; 3826 nbp = aip->ai_buf; 3827 if (getdirtybuf(&nbp, waitfor) == 0) 3828 break; 3829 FREE_LOCK(&lk); 3830 if (waitfor == MNT_NOWAIT) { 3831 bawrite(nbp); 3832 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 3833 bawrite(bp); 3834 return (error); 3835 } 3836 ACQUIRE_LOCK(&lk); 3837 break; 3838 3839 case D_INDIRDEP: 3840 restart: 3841 for (aip = LIST_FIRST(&WK_INDIRDEP(wk)->ir_deplisthd); 3842 aip; aip = LIST_NEXT(aip, ai_next)) { 3843 if (aip->ai_state & DEPCOMPLETE) 3844 continue; 3845 nbp = aip->ai_buf; 3846 if (getdirtybuf(&nbp, MNT_WAIT) == 0) 3847 goto restart; 3848 FREE_LOCK(&lk); 3849 if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 3850 bawrite(bp); 3851 return (error); 3852 } 3853 ACQUIRE_LOCK(&lk); 3854 goto restart; 3855 } 3856 break; 3857 3858 case D_INODEDEP: 3859 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs, 3860 WK_INODEDEP(wk)->id_ino)) != 0) { 3861 FREE_LOCK(&lk); 3862 bawrite(bp); 3863 return (error); 3864 } 3865 break; 3866 3867 case D_PAGEDEP: 3868 /* 3869 * We are trying to sync a directory that may 3870 * have dependencies on both its own metadata 3871 * and/or dependencies on the inodes of any 3872 * recently allocated files. We walk its diradd 3873 * lists pushing out the associated inode. 3874 */ 3875 pagedep = WK_PAGEDEP(wk); 3876 for (i = 0; i < DAHASHSZ; i++) { 3877 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 3878 continue; 3879 if ((error = 3880 flush_pagedep_deps(vp, pagedep->pd_mnt, 3881 &pagedep->pd_diraddhd[i]))) { 3882 FREE_LOCK(&lk); 3883 bawrite(bp); 3884 return (error); 3885 } 3886 } 3887 break; 3888 3889 case D_MKDIR: 3890 /* 3891 * This case should never happen if the vnode has 3892 * been properly sync'ed. However, if this function 3893 * is used at a place where the vnode has not yet 3894 * been sync'ed, this dependency can show up. So, 3895 * rather than panic, just flush it. 3896 */ 3897 nbp = WK_MKDIR(wk)->md_buf; 3898 if (getdirtybuf(&nbp, waitfor) == 0) 3899 break; 3900 FREE_LOCK(&lk); 3901 if (waitfor == MNT_NOWAIT) { 3902 bawrite(nbp); 3903 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 3904 bawrite(bp); 3905 return (error); 3906 } 3907 ACQUIRE_LOCK(&lk); 3908 break; 3909 3910 case D_BMSAFEMAP: 3911 /* 3912 * This case should never happen if the vnode has 3913 * been properly sync'ed. However, if this function 3914 * is used at a place where the vnode has not yet 3915 * been sync'ed, this dependency can show up. So, 3916 * rather than panic, just flush it. 3917 */ 3918 nbp = WK_BMSAFEMAP(wk)->sm_buf; 3919 if (getdirtybuf(&nbp, waitfor) == 0) 3920 break; 3921 FREE_LOCK(&lk); 3922 if (waitfor == MNT_NOWAIT) { 3923 bawrite(nbp); 3924 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) { 3925 bawrite(bp); 3926 return (error); 3927 } 3928 ACQUIRE_LOCK(&lk); 3929 break; 3930 3931 default: 3932 panic("softdep_sync_metadata: Unknown type %s", 3933 TYPENAME(wk->wk_type)); 3934 /* NOTREACHED */ 3935 } 3936 } 3937 (void) getdirtybuf(&TAILQ_NEXT(bp, b_vnbufs), MNT_WAIT); 3938 nbp = TAILQ_NEXT(bp, b_vnbufs); 3939 FREE_LOCK(&lk); 3940 bawrite(bp); 3941 ACQUIRE_LOCK(&lk); 3942 if (nbp != NULL) { 3943 bp = nbp; 3944 goto loop; 3945 } 3946 /* 3947 * We must wait for any I/O in progress to finish so that 3948 * all potential buffers on the dirty list will be visible. 3949 * Once they are all there, proceed with the second pass 3950 * which will wait for the I/O as per above. 3951 */ 3952 drain_output(vp, 1); 3953 /* 3954 * The brief unlock is to allow any pent up dependency 3955 * processing to be done. 3956 */ 3957 if (waitfor == MNT_NOWAIT) { 3958 waitfor = MNT_WAIT; 3959 FREE_LOCK(&lk); 3960 ACQUIRE_LOCK(&lk); 3961 goto top; 3962 } 3963 3964 /* 3965 * If we have managed to get rid of all the dirty buffers, 3966 * then we are done. For certain directories and block 3967 * devices, we may need to do further work. 3968 */ 3969 if (TAILQ_FIRST(&vp->v_dirtyblkhd) == NULL) { 3970 FREE_LOCK(&lk); 3971 return (0); 3972 } 3973 3974 FREE_LOCK(&lk); 3975 /* 3976 * If we are trying to sync a block device, some of its buffers may 3977 * contain metadata that cannot be written until the contents of some 3978 * partially written files have been written to disk. The only easy 3979 * way to accomplish this is to sync the entire filesystem (luckily 3980 * this happens rarely). 3981 */ 3982 if (vp->v_type == VBLK && vp->v_specmountpoint && !VOP_ISLOCKED(vp) && 3983 (error = VFS_SYNC(vp->v_specmountpoint, MNT_WAIT, ap->a_cred, 3984 ap->a_p)) != 0) 3985 return (error); 3986 return (0); 3987} 3988 3989/* 3990 * Flush the dependencies associated with an inodedep. 3991 * Called with splbio blocked. 3992 */ 3993static int 3994flush_inodedep_deps(fs, ino) 3995 struct fs *fs; 3996 ino_t ino; 3997{ 3998 struct inodedep *inodedep; 3999 struct allocdirect *adp; 4000 int error, waitfor; 4001 struct buf *bp; 4002 4003 /* 4004 * This work is done in two passes. The first pass grabs most 4005 * of the buffers and begins asynchronously writing them. The 4006 * only way to wait for these asynchronous writes is to sleep 4007 * on the filesystem vnode which may stay busy for a long time 4008 * if the filesystem is active. So, instead, we make a second 4009 * pass over the dependencies blocking on each write. In the 4010 * usual case we will be blocking against a write that we 4011 * initiated, so when it is done the dependency will have been 4012 * resolved. Thus the second pass is expected to end quickly. 4013 * We give a brief window at the top of the loop to allow 4014 * any pending I/O to complete. 4015 */ 4016 for (waitfor = MNT_NOWAIT; ; ) { 4017 FREE_LOCK(&lk); 4018 ACQUIRE_LOCK(&lk); 4019 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4020 return (0); 4021 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 4022 adp = TAILQ_NEXT(adp, ad_next)) { 4023 if (adp->ad_state & DEPCOMPLETE) 4024 continue; 4025 bp = adp->ad_buf; 4026 if (getdirtybuf(&bp, waitfor) == 0) { 4027 if (waitfor == MNT_NOWAIT) 4028 continue; 4029 break; 4030 } 4031 FREE_LOCK(&lk); 4032 if (waitfor == MNT_NOWAIT) { 4033 bawrite(bp); 4034 } else if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) { 4035 ACQUIRE_LOCK(&lk); 4036 return (error); 4037 } 4038 ACQUIRE_LOCK(&lk); 4039 break; 4040 } 4041 if (adp != NULL) 4042 continue; 4043 for (adp = TAILQ_FIRST(&inodedep->id_newinoupdt); adp; 4044 adp = TAILQ_NEXT(adp, ad_next)) { 4045 if (adp->ad_state & DEPCOMPLETE) 4046 continue; 4047 bp = adp->ad_buf; 4048 if (getdirtybuf(&bp, waitfor) == 0) { 4049 if (waitfor == MNT_NOWAIT) 4050 continue; 4051 break; 4052 } 4053 FREE_LOCK(&lk); 4054 if (waitfor == MNT_NOWAIT) { 4055 bawrite(bp); 4056 } else if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) { 4057 ACQUIRE_LOCK(&lk); 4058 return (error); 4059 } 4060 ACQUIRE_LOCK(&lk); 4061 break; 4062 } 4063 if (adp != NULL) 4064 continue; 4065 /* 4066 * If pass2, we are done, otherwise do pass 2. 4067 */ 4068 if (waitfor == MNT_WAIT) 4069 break; 4070 waitfor = MNT_WAIT; 4071 } 4072 /* 4073 * Try freeing inodedep in case all dependencies have been removed. 4074 */ 4075 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0) 4076 (void) free_inodedep(inodedep); 4077 return (0); 4078} 4079 4080/* 4081 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 4082 * Called with splbio blocked. 4083 */ 4084static int 4085flush_pagedep_deps(pvp, mp, diraddhdp) 4086 struct vnode *pvp; 4087 struct mount *mp; 4088 struct diraddhd *diraddhdp; 4089{ 4090 struct proc *p = CURPROC; /* XXX */ 4091 struct inodedep *inodedep; 4092 struct ufsmount *ump; 4093 struct diradd *dap; 4094 struct vnode *vp; 4095 int gotit, error = 0; 4096 struct buf *bp; 4097 ino_t inum; 4098 4099 ump = VFSTOUFS(mp); 4100 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 4101 /* 4102 * Flush ourselves if this directory entry 4103 * has a MKDIR_PARENT dependency. 4104 */ 4105 if (dap->da_state & MKDIR_PARENT) { 4106 FREE_LOCK(&lk); 4107 if ((error = UFS_UPDATE(pvp, 1)) != 0) 4108 break; 4109 ACQUIRE_LOCK(&lk); 4110 /* 4111 * If that cleared dependencies, go on to next. 4112 */ 4113 if (dap != LIST_FIRST(diraddhdp)) 4114 continue; 4115 if (dap->da_state & MKDIR_PARENT) 4116 panic("flush_pagedep_deps: MKDIR"); 4117 } 4118 /* 4119 * Flush the file on which the directory entry depends. 4120 * If the inode has already been pushed out of the cache, 4121 * then all the block dependencies will have been flushed 4122 * leaving only inode dependencies (e.g., bitmaps). Thus, 4123 * we do a ufs_ihashget to check for the vnode in the cache. 4124 * If it is there, we do a full flush. If it is no longer 4125 * there we need only dispose of any remaining bitmap 4126 * dependencies and write the inode to disk. 4127 */ 4128 inum = dap->da_newinum; 4129 FREE_LOCK(&lk); 4130 if ((vp = ufs_ihashget(ump->um_dev, inum)) == NULL) { 4131 ACQUIRE_LOCK(&lk); 4132 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0 4133 && dap == LIST_FIRST(diraddhdp)) 4134 panic("flush_pagedep_deps: flush 1 failed"); 4135 /* 4136 * If the inode still has bitmap dependencies, 4137 * push them to disk. 4138 */ 4139 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4140 gotit = getdirtybuf(&inodedep->id_buf,MNT_WAIT); 4141 FREE_LOCK(&lk); 4142 if (gotit && 4143 (error = VOP_BWRITE(inodedep->id_buf->b_vp, 4144 inodedep->id_buf)) != 0) 4145 break; 4146 ACQUIRE_LOCK(&lk); 4147 } 4148 if (dap != LIST_FIRST(diraddhdp)) 4149 continue; 4150 /* 4151 * If the inode is still sitting in a buffer waiting 4152 * to be written, push it to disk. 4153 */ 4154 FREE_LOCK(&lk); 4155 if ((error = bread(ump->um_devvp, 4156 fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), 4157 (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0) 4158 break; 4159 if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) 4160 break; 4161 ACQUIRE_LOCK(&lk); 4162 if (dap == LIST_FIRST(diraddhdp)) 4163 panic("flush_pagedep_deps: flush 2 failed"); 4164 continue; 4165 } 4166 if (vp->v_type == VDIR) { 4167 /* 4168 * A newly allocated directory must have its "." and 4169 * ".." entries written out before its name can be 4170 * committed in its parent. We do not want or need 4171 * the full semantics of a synchronous VOP_FSYNC as 4172 * that may end up here again, once for each directory 4173 * level in the filesystem. Instead, we push the blocks 4174 * and wait for them to clear. 4175 */ 4176 if ((error = 4177 VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p))) { 4178 vput(vp); 4179 break; 4180 } 4181 drain_output(vp, 0); 4182 } 4183 error = UFS_UPDATE(vp, 1); 4184 vput(vp); 4185 if (error) 4186 break; 4187 /* 4188 * If we have failed to get rid of all the dependencies 4189 * then something is seriously wrong. 4190 */ 4191 if (dap == LIST_FIRST(diraddhdp)) 4192 panic("flush_pagedep_deps: flush 3 failed"); 4193 ACQUIRE_LOCK(&lk); 4194 } 4195 if (error) 4196 ACQUIRE_LOCK(&lk); 4197 return (error); 4198} 4199 4200/* 4201 * A large burst of file addition or deletion activity can drive the 4202 * memory load excessively high. Therefore we deliberately slow things 4203 * down and speed up the I/O processing if we find ourselves with too 4204 * many dependencies in progress. 4205 */ 4206static int 4207request_cleanup(resource, islocked) 4208 int resource; 4209 int islocked; 4210{ 4211 struct callout_handle handle; 4212 struct proc *p = CURPROC; 4213 4214 /* 4215 * We never hold up the filesystem syncer process. 4216 */ 4217 if (p == filesys_syncer) 4218 return (0); 4219 /* 4220 * If we are resource constrained on inode dependencies, try 4221 * flushing some dirty inodes. Otherwise, we are constrained 4222 * by file deletions, so try accelerating flushes of directories 4223 * with removal dependencies. We would like to do the cleanup 4224 * here, but we probably hold an inode locked at this point and 4225 * that might deadlock against one that we try to clean. So, 4226 * the best that we can do is request the syncer daemon to do 4227 * the cleanup for us. 4228 */ 4229 switch (resource) { 4230 4231 case FLUSH_INODES: 4232 stat_ino_limit_push += 1; 4233 req_clear_inodedeps = 1; 4234 break; 4235 4236 case FLUSH_REMOVE: 4237 stat_blk_limit_push += 1; 4238 req_clear_remove = 1; 4239 break; 4240 4241 default: 4242 panic("request_cleanup: unknown type"); 4243 } 4244 /* 4245 * Hopefully the syncer daemon will catch up and awaken us. 4246 * We wait at most tickdelay before proceeding in any case. 4247 */ 4248 if (islocked == 0) 4249 ACQUIRE_LOCK(&lk); 4250 if (proc_waiting == 0) { 4251 proc_waiting = 1; 4252 handle = timeout(pause_timer, NULL, 4253 tickdelay > 2 ? tickdelay : 2); 4254 } 4255 FREE_LOCK_INTERLOCKED(&lk); 4256 (void) tsleep((caddr_t)&proc_waiting, PPAUSE | PCATCH, "softupdate", 0); 4257 ACQUIRE_LOCK_INTERLOCKED(&lk); 4258 if (proc_waiting) { 4259 untimeout(pause_timer, NULL, handle); 4260 proc_waiting = 0; 4261 } else { 4262 switch (resource) { 4263 4264 case FLUSH_INODES: 4265 stat_ino_limit_hit += 1; 4266 break; 4267 4268 case FLUSH_REMOVE: 4269 stat_blk_limit_hit += 1; 4270 break; 4271 } 4272 } 4273 if (islocked == 0) 4274 FREE_LOCK(&lk); 4275 return (1); 4276} 4277 4278/* 4279 * Awaken processes pausing in request_cleanup and clear proc_waiting 4280 * to indicate that there is no longer a timer running. 4281 */ 4282void 4283pause_timer(arg) 4284 void *arg; 4285{ 4286 4287 proc_waiting = 0; 4288 wakeup(&proc_waiting); 4289} 4290 4291/* 4292 * Flush out a directory with at least one removal dependency in an effort 4293 * to reduce the number of freefile and freeblks dependency structures. 4294 */ 4295static void 4296clear_remove(p) 4297 struct proc *p; 4298{ 4299 struct pagedep_hashhead *pagedephd; 4300 struct pagedep *pagedep; 4301 static int next = 0; 4302 struct mount *mp; 4303 struct vnode *vp; 4304 int error, cnt; 4305 ino_t ino; 4306 4307 ACQUIRE_LOCK(&lk); 4308 for (cnt = 0; cnt < pagedep_hash; cnt++) { 4309 pagedephd = &pagedep_hashtbl[next++]; 4310 if (next >= pagedep_hash) 4311 next = 0; 4312 for (pagedep = LIST_FIRST(pagedephd); pagedep; 4313 pagedep = LIST_NEXT(pagedep, pd_hash)) { 4314 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL) 4315 continue; 4316 mp = pagedep->pd_mnt; 4317 ino = pagedep->pd_ino; 4318 FREE_LOCK(&lk); 4319 if ((error = VFS_VGET(mp, ino, &vp)) != 0) { 4320 softdep_error("clear_remove: vget", error); 4321 return; 4322 } 4323 if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p))) 4324 softdep_error("clear_remove: fsync", error); 4325 drain_output(vp, 0); 4326 vput(vp); 4327 return; 4328 } 4329 } 4330 FREE_LOCK(&lk); 4331} 4332 4333/* 4334 * Clear out a block of dirty inodes in an effort to reduce 4335 * the number of inodedep dependency structures. 4336 */ 4337static void 4338clear_inodedeps(p) 4339 struct proc *p; 4340{ 4341 struct inodedep_hashhead *inodedephd; 4342 struct inodedep *inodedep; 4343 static int next = 0; 4344 struct mount *mp; 4345 struct vnode *vp; 4346 struct fs *fs; 4347 int error, cnt; 4348 ino_t firstino, lastino, ino; 4349 4350 ACQUIRE_LOCK(&lk); 4351 /* 4352 * Pick a random inode dependency to be cleared. 4353 * We will then gather up all the inodes in its block 4354 * that have dependencies and flush them out. 4355 */ 4356 for (cnt = 0; cnt < inodedep_hash; cnt++) { 4357 inodedephd = &inodedep_hashtbl[next++]; 4358 if (next >= inodedep_hash) 4359 next = 0; 4360 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 4361 break; 4362 } 4363 /* 4364 * Ugly code to find mount point given pointer to superblock. 4365 */ 4366 fs = inodedep->id_fs; 4367 for (mp = CIRCLEQ_FIRST(&mountlist); mp != (void *)&mountlist; 4368 mp = CIRCLEQ_NEXT(mp, mnt_list)) 4369 if ((mp->mnt_flag & MNT_SOFTDEP) && fs == VFSTOUFS(mp)->um_fs) 4370 break; 4371 /* 4372 * Find the last inode in the block with dependencies. 4373 */ 4374 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 4375 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 4376 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0) 4377 break; 4378 /* 4379 * Asynchronously push all but the last inode with dependencies. 4380 * Synchronously push the last inode with dependencies to ensure 4381 * that the inode block gets written to free up the inodedeps. 4382 */ 4383 for (ino = firstino; ino <= lastino; ino++) { 4384 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4385 continue; 4386 FREE_LOCK(&lk); 4387 if ((error = VFS_VGET(mp, ino, &vp)) != 0) { 4388 softdep_error("clear_inodedeps: vget", error); 4389 return; 4390 } 4391 if (ino == lastino) { 4392 if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_WAIT, p))) 4393 softdep_error("clear_inodedeps: fsync1", error); 4394 } else { 4395 if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p))) 4396 softdep_error("clear_inodedeps: fsync2", error); 4397 drain_output(vp, 0); 4398 } 4399 vput(vp); 4400 ACQUIRE_LOCK(&lk); 4401 } 4402 FREE_LOCK(&lk); 4403} 4404 4405/* 4406 * Acquire exclusive access to a buffer. 4407 * Must be called with splbio blocked. 4408 * Return 1 if buffer was acquired. 4409 */ 4410static int 4411getdirtybuf(bpp, waitfor) 4412 struct buf **bpp; 4413 int waitfor; 4414{ 4415 struct buf *bp; 4416 4417 for (;;) { 4418 if ((bp = *bpp) == NULL) 4419 return (0); 4420 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) 4421 break; 4422 if (waitfor != MNT_WAIT) 4423 return (0); 4424 FREE_LOCK_INTERLOCKED(&lk); 4425 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL) != ENOLCK) 4426 panic("getdirtybuf: inconsistent lock"); 4427 ACQUIRE_LOCK_INTERLOCKED(&lk); 4428 } 4429 if ((bp->b_flags & B_DELWRI) == 0) { 4430 BUF_UNLOCK(bp); 4431 return (0); 4432 } 4433 bremfree(bp); 4434 return (1); 4435} 4436 4437/* 4438 * Wait for pending output on a vnode to complete. 4439 * Must be called with vnode locked. 4440 */ 4441static void 4442drain_output(vp, islocked) 4443 struct vnode *vp; 4444 int islocked; 4445{ 4446 4447 if (!islocked) 4448 ACQUIRE_LOCK(&lk); 4449 while (vp->v_numoutput) { 4450 vp->v_flag |= VBWAIT; 4451 FREE_LOCK_INTERLOCKED(&lk); 4452 tsleep((caddr_t)&vp->v_numoutput, PRIBIO + 1, "drainvp", 0); 4453 ACQUIRE_LOCK_INTERLOCKED(&lk); 4454 } 4455 if (!islocked) 4456 FREE_LOCK(&lk); 4457} 4458 4459/* 4460 * Called whenever a buffer that is being invalidated or reallocated 4461 * contains dependencies. This should only happen if an I/O error has 4462 * occurred. The routine is called with the buffer locked. 4463 */ 4464void 4465softdep_deallocate_dependencies(bp) 4466 struct buf *bp; 4467{ 4468 4469 if ((bp->b_flags & B_ERROR) == 0) 4470 panic("softdep_deallocate_dependencies: dangling deps"); 4471 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); 4472 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 4473} 4474 4475/* 4476 * Function to handle asynchronous write errors in the filesystem. 4477 */ 4478void 4479softdep_error(func, error) 4480 char *func; 4481 int error; 4482{ 4483 4484 /* XXX should do something better! */ 4485 printf("%s: got error %d while accessing filesystem\n", func, error); 4486} 4487