ffs_softdep.c revision 76859
1/* 2 * Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved. 3 * 4 * The soft updates code is derived from the appendix of a University 5 * of Michigan technical report (Gregory R. Ganger and Yale N. Patt, 6 * "Soft Updates: A Solution to the Metadata Update Problem in File 7 * Systems", CSE-TR-254-95, August 1995). 8 * 9 * Further information about soft updates can be obtained from: 10 * 11 * Marshall Kirk McKusick http://www.mckusick.com/softdep/ 12 * 1614 Oxford Street mckusick@mckusick.com 13 * Berkeley, CA 94709-1608 +1-510-843-9542 14 * USA 15 * 16 * Redistribution and use in source and binary forms, with or without 17 * modification, are permitted provided that the following conditions 18 * are met: 19 * 20 * 1. Redistributions of source code must retain the above copyright 21 * notice, this list of conditions and the following disclaimer. 22 * 2. Redistributions in binary form must reproduce the above copyright 23 * notice, this list of conditions and the following disclaimer in the 24 * documentation and/or other materials provided with the distribution. 25 * 26 * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY 27 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 28 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 29 * DISCLAIMED. IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR 30 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * from: @(#)ffs_softdep.c 9.59 (McKusick) 6/21/00 39 * $FreeBSD: head/sys/ufs/ffs/ffs_softdep.c 76859 2001-05-19 19:24:26Z mckusick $ 40 */ 41 42/* 43 * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide. 44 */ 45#ifndef DIAGNOSTIC 46#define DIAGNOSTIC 47#endif 48#ifndef DEBUG 49#define DEBUG 50#endif 51 52#include <sys/param.h> 53#include <sys/kernel.h> 54#include <sys/systm.h> 55#include <sys/bio.h> 56#include <sys/buf.h> 57#include <sys/malloc.h> 58#include <sys/mount.h> 59#include <sys/proc.h> 60#include <sys/stat.h> 61#include <sys/syslog.h> 62#include <sys/vnode.h> 63#include <sys/conf.h> 64#include <ufs/ufs/dir.h> 65#include <ufs/ufs/extattr.h> 66#include <ufs/ufs/quota.h> 67#include <ufs/ufs/inode.h> 68#include <ufs/ufs/ufsmount.h> 69#include <ufs/ffs/fs.h> 70#include <ufs/ffs/softdep.h> 71#include <ufs/ffs/ffs_extern.h> 72#include <ufs/ufs/ufs_extern.h> 73 74/* 75 * These definitions need to be adapted to the system to which 76 * this file is being ported. 77 */ 78/* 79 * malloc types defined for the softdep system. 80 */ 81static MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies"); 82static MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies"); 83static MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation"); 84static MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map"); 85static MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode"); 86static MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies"); 87static MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block"); 88static MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode"); 89static MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode"); 90static MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated"); 91static MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry"); 92static MALLOC_DEFINE(M_MKDIR, "mkdir","New directory"); 93static MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted"); 94static MALLOC_DEFINE(M_NEWDIRBLK, "newdirblk","Unclaimed new directory block"); 95 96#define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE) 97 98#define D_PAGEDEP 0 99#define D_INODEDEP 1 100#define D_NEWBLK 2 101#define D_BMSAFEMAP 3 102#define D_ALLOCDIRECT 4 103#define D_INDIRDEP 5 104#define D_ALLOCINDIR 6 105#define D_FREEFRAG 7 106#define D_FREEBLKS 8 107#define D_FREEFILE 9 108#define D_DIRADD 10 109#define D_MKDIR 11 110#define D_DIRREM 12 111#define D_NEWDIRBLK 13 112#define D_LAST D_NEWDIRBLK 113 114/* 115 * translate from workitem type to memory type 116 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 117 */ 118static struct malloc_type *memtype[] = { 119 M_PAGEDEP, 120 M_INODEDEP, 121 M_NEWBLK, 122 M_BMSAFEMAP, 123 M_ALLOCDIRECT, 124 M_INDIRDEP, 125 M_ALLOCINDIR, 126 M_FREEFRAG, 127 M_FREEBLKS, 128 M_FREEFILE, 129 M_DIRADD, 130 M_MKDIR, 131 M_DIRREM, 132 M_NEWDIRBLK 133}; 134 135#define DtoM(type) (memtype[type]) 136 137/* 138 * Names of malloc types. 139 */ 140#define TYPENAME(type) \ 141 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???") 142/* 143 * End system adaptaion definitions. 144 */ 145 146/* 147 * Internal function prototypes. 148 */ 149static void softdep_error __P((char *, int)); 150static void drain_output __P((struct vnode *, int)); 151static int getdirtybuf __P((struct buf **, int)); 152static void clear_remove __P((struct proc *)); 153static void clear_inodedeps __P((struct proc *)); 154static int flush_pagedep_deps __P((struct vnode *, struct mount *, 155 struct diraddhd *)); 156static int flush_inodedep_deps __P((struct fs *, ino_t)); 157static int handle_written_filepage __P((struct pagedep *, struct buf *)); 158static void diradd_inode_written __P((struct diradd *, struct inodedep *)); 159static int handle_written_inodeblock __P((struct inodedep *, struct buf *)); 160static void handle_allocdirect_partdone __P((struct allocdirect *)); 161static void handle_allocindir_partdone __P((struct allocindir *)); 162static void initiate_write_filepage __P((struct pagedep *, struct buf *)); 163static void handle_written_mkdir __P((struct mkdir *, int)); 164static void initiate_write_inodeblock __P((struct inodedep *, struct buf *)); 165static void handle_workitem_freefile __P((struct freefile *)); 166static void handle_workitem_remove __P((struct dirrem *)); 167static struct dirrem *newdirrem __P((struct buf *, struct inode *, 168 struct inode *, int, struct dirrem **)); 169static void free_diradd __P((struct diradd *)); 170static void free_allocindir __P((struct allocindir *, struct inodedep *)); 171static void free_newdirblk __P((struct newdirblk *)); 172static int indir_trunc __P((struct inode *, ufs_daddr_t, int, ufs_lbn_t, 173 long *)); 174static void deallocate_dependencies __P((struct buf *, struct inodedep *)); 175static void free_allocdirect __P((struct allocdirectlst *, 176 struct allocdirect *, int)); 177static int check_inode_unwritten __P((struct inodedep *)); 178static int free_inodedep __P((struct inodedep *)); 179static void handle_workitem_freeblocks __P((struct freeblks *, int)); 180static void merge_inode_lists __P((struct inodedep *)); 181static void setup_allocindir_phase2 __P((struct buf *, struct inode *, 182 struct allocindir *)); 183static struct allocindir *newallocindir __P((struct inode *, int, ufs_daddr_t, 184 ufs_daddr_t)); 185static void handle_workitem_freefrag __P((struct freefrag *)); 186static struct freefrag *newfreefrag __P((struct inode *, ufs_daddr_t, long)); 187static void allocdirect_merge __P((struct allocdirectlst *, 188 struct allocdirect *, struct allocdirect *)); 189static struct bmsafemap *bmsafemap_lookup __P((struct buf *)); 190static int newblk_lookup __P((struct fs *, ufs_daddr_t, int, 191 struct newblk **)); 192static int inodedep_lookup __P((struct fs *, ino_t, int, struct inodedep **)); 193static int pagedep_lookup __P((struct inode *, ufs_lbn_t, int, 194 struct pagedep **)); 195static void pause_timer __P((void *)); 196static int request_cleanup __P((int, int)); 197static int process_worklist_item __P((struct mount *, int)); 198static void add_to_worklist __P((struct worklist *)); 199 200/* 201 * Exported softdep operations. 202 */ 203static void softdep_disk_io_initiation __P((struct buf *)); 204static void softdep_disk_write_complete __P((struct buf *)); 205static void softdep_deallocate_dependencies __P((struct buf *)); 206static void softdep_move_dependencies __P((struct buf *, struct buf *)); 207static int softdep_count_dependencies __P((struct buf *bp, int)); 208 209struct bio_ops bioops = { 210 softdep_disk_io_initiation, /* io_start */ 211 softdep_disk_write_complete, /* io_complete */ 212 softdep_deallocate_dependencies, /* io_deallocate */ 213 softdep_move_dependencies, /* io_movedeps */ 214 softdep_count_dependencies, /* io_countdeps */ 215}; 216 217/* 218 * Locking primitives. 219 * 220 * For a uniprocessor, all we need to do is protect against disk 221 * interrupts. For a multiprocessor, this lock would have to be 222 * a mutex. A single mutex is used throughout this file, though 223 * finer grain locking could be used if contention warranted it. 224 * 225 * For a multiprocessor, the sleep call would accept a lock and 226 * release it after the sleep processing was complete. In a uniprocessor 227 * implementation there is no such interlock, so we simple mark 228 * the places where it needs to be done with the `interlocked' form 229 * of the lock calls. Since the uniprocessor sleep already interlocks 230 * the spl, there is nothing that really needs to be done. 231 */ 232#ifndef /* NOT */ DEBUG 233static struct lockit { 234 int lkt_spl; 235} lk = { 0 }; 236#define ACQUIRE_LOCK(lk) (lk)->lkt_spl = splbio() 237#define FREE_LOCK(lk) splx((lk)->lkt_spl) 238#define ACQUIRE_LOCK_INTERLOCKED(lk) 239#define FREE_LOCK_INTERLOCKED(lk) 240 241#else /* DEBUG */ 242static struct lockit { 243 int lkt_spl; 244 pid_t lkt_held; 245} lk = { 0, -1 }; 246static int lockcnt; 247 248static void acquire_lock __P((struct lockit *)); 249static void free_lock __P((struct lockit *)); 250static void acquire_lock_interlocked __P((struct lockit *)); 251static void free_lock_interlocked __P((struct lockit *)); 252 253#define ACQUIRE_LOCK(lk) acquire_lock(lk) 254#define FREE_LOCK(lk) free_lock(lk) 255#define ACQUIRE_LOCK_INTERLOCKED(lk) acquire_lock_interlocked(lk) 256#define FREE_LOCK_INTERLOCKED(lk) free_lock_interlocked(lk) 257 258static void 259acquire_lock(lk) 260 struct lockit *lk; 261{ 262 pid_t holder; 263 264 if (lk->lkt_held != -1) { 265 holder = lk->lkt_held; 266 FREE_LOCK(lk); 267 if (holder == CURPROC->p_pid) 268 panic("softdep_lock: locking against myself"); 269 else 270 panic("softdep_lock: lock held by %d", holder); 271 } 272 lk->lkt_spl = splbio(); 273 lk->lkt_held = CURPROC->p_pid; 274 lockcnt++; 275} 276 277static void 278free_lock(lk) 279 struct lockit *lk; 280{ 281 282 if (lk->lkt_held == -1) 283 panic("softdep_unlock: lock not held"); 284 lk->lkt_held = -1; 285 splx(lk->lkt_spl); 286} 287 288static void 289acquire_lock_interlocked(lk) 290 struct lockit *lk; 291{ 292 pid_t holder; 293 294 if (lk->lkt_held != -1) { 295 holder = lk->lkt_held; 296 FREE_LOCK(lk); 297 if (holder == CURPROC->p_pid) 298 panic("softdep_lock_interlocked: locking against self"); 299 else 300 panic("softdep_lock_interlocked: lock held by %d", 301 holder); 302 } 303 lk->lkt_held = CURPROC->p_pid; 304 lockcnt++; 305} 306 307static void 308free_lock_interlocked(lk) 309 struct lockit *lk; 310{ 311 312 if (lk->lkt_held == -1) 313 panic("softdep_unlock_interlocked: lock not held"); 314 lk->lkt_held = -1; 315} 316#endif /* DEBUG */ 317 318/* 319 * Place holder for real semaphores. 320 */ 321struct sema { 322 int value; 323 pid_t holder; 324 char *name; 325 int prio; 326 int timo; 327}; 328static void sema_init __P((struct sema *, char *, int, int)); 329static int sema_get __P((struct sema *, struct lockit *)); 330static void sema_release __P((struct sema *)); 331 332static void 333sema_init(semap, name, prio, timo) 334 struct sema *semap; 335 char *name; 336 int prio, timo; 337{ 338 339 semap->holder = -1; 340 semap->value = 0; 341 semap->name = name; 342 semap->prio = prio; 343 semap->timo = timo; 344} 345 346static int 347sema_get(semap, interlock) 348 struct sema *semap; 349 struct lockit *interlock; 350{ 351 352 if (semap->value++ > 0) { 353 if (interlock != NULL) 354 FREE_LOCK_INTERLOCKED(interlock); 355 tsleep((caddr_t)semap, semap->prio, semap->name, semap->timo); 356 if (interlock != NULL) { 357 ACQUIRE_LOCK_INTERLOCKED(interlock); 358 FREE_LOCK(interlock); 359 } 360 return (0); 361 } 362 semap->holder = CURPROC->p_pid; 363 if (interlock != NULL) 364 FREE_LOCK(interlock); 365 return (1); 366} 367 368static void 369sema_release(semap) 370 struct sema *semap; 371{ 372 373 if (semap->value <= 0 || semap->holder != CURPROC->p_pid) { 374 if (lk.lkt_held != -1) 375 FREE_LOCK(&lk); 376 panic("sema_release: not held"); 377 } 378 if (--semap->value > 0) { 379 semap->value = 0; 380 wakeup(semap); 381 } 382 semap->holder = -1; 383} 384 385/* 386 * Worklist queue management. 387 * These routines require that the lock be held. 388 */ 389#ifndef /* NOT */ DEBUG 390#define WORKLIST_INSERT(head, item) do { \ 391 (item)->wk_state |= ONWORKLIST; \ 392 LIST_INSERT_HEAD(head, item, wk_list); \ 393} while (0) 394#define WORKLIST_REMOVE(item) do { \ 395 (item)->wk_state &= ~ONWORKLIST; \ 396 LIST_REMOVE(item, wk_list); \ 397} while (0) 398#define WORKITEM_FREE(item, type) FREE(item, DtoM(type)) 399 400#else /* DEBUG */ 401static void worklist_insert __P((struct workhead *, struct worklist *)); 402static void worklist_remove __P((struct worklist *)); 403static void workitem_free __P((struct worklist *, int)); 404 405#define WORKLIST_INSERT(head, item) worklist_insert(head, item) 406#define WORKLIST_REMOVE(item) worklist_remove(item) 407#define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type) 408 409static void 410worklist_insert(head, item) 411 struct workhead *head; 412 struct worklist *item; 413{ 414 415 if (lk.lkt_held == -1) 416 panic("worklist_insert: lock not held"); 417 if (item->wk_state & ONWORKLIST) { 418 FREE_LOCK(&lk); 419 panic("worklist_insert: already on list"); 420 } 421 item->wk_state |= ONWORKLIST; 422 LIST_INSERT_HEAD(head, item, wk_list); 423} 424 425static void 426worklist_remove(item) 427 struct worklist *item; 428{ 429 430 if (lk.lkt_held == -1) 431 panic("worklist_remove: lock not held"); 432 if ((item->wk_state & ONWORKLIST) == 0) { 433 FREE_LOCK(&lk); 434 panic("worklist_remove: not on list"); 435 } 436 item->wk_state &= ~ONWORKLIST; 437 LIST_REMOVE(item, wk_list); 438} 439 440static void 441workitem_free(item, type) 442 struct worklist *item; 443 int type; 444{ 445 446 if (item->wk_state & ONWORKLIST) { 447 if (lk.lkt_held != -1) 448 FREE_LOCK(&lk); 449 panic("workitem_free: still on list"); 450 } 451 if (item->wk_type != type) { 452 if (lk.lkt_held != -1) 453 FREE_LOCK(&lk); 454 panic("workitem_free: type mismatch"); 455 } 456 FREE(item, DtoM(type)); 457} 458#endif /* DEBUG */ 459 460/* 461 * Workitem queue management 462 */ 463static struct workhead softdep_workitem_pending; 464static int num_on_worklist; /* number of worklist items to be processed */ 465static int softdep_worklist_busy; /* 1 => trying to do unmount */ 466static int softdep_worklist_req; /* serialized waiters */ 467static int max_softdeps; /* maximum number of structs before slowdown */ 468static int tickdelay = 2; /* number of ticks to pause during slowdown */ 469static int proc_waiting; /* tracks whether we have a timeout posted */ 470static int *stat_countp; /* statistic to count in proc_waiting timeout */ 471static struct callout_handle handle; /* handle on posted proc_waiting timeout */ 472static struct proc *filesys_syncer; /* proc of filesystem syncer process */ 473static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 474#define FLUSH_INODES 1 475static int req_clear_remove; /* syncer process flush some freeblks */ 476#define FLUSH_REMOVE 2 477/* 478 * runtime statistics 479 */ 480static int stat_worklist_push; /* number of worklist cleanups */ 481static int stat_blk_limit_push; /* number of times block limit neared */ 482static int stat_ino_limit_push; /* number of times inode limit neared */ 483static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 484static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 485static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 486static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 487static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 488static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 489static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 490#ifdef DEBUG 491#include <vm/vm.h> 492#include <sys/sysctl.h> 493SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, ""); 494SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, ""); 495SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,""); 496SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,""); 497SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,""); 498SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, ""); 499SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, ""); 500SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, ""); 501SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, ""); 502SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, ""); 503SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, ""); 504SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, ""); 505#endif /* DEBUG */ 506 507/* 508 * Add an item to the end of the work queue. 509 * This routine requires that the lock be held. 510 * This is the only routine that adds items to the list. 511 * The following routine is the only one that removes items 512 * and does so in order from first to last. 513 */ 514static void 515add_to_worklist(wk) 516 struct worklist *wk; 517{ 518 static struct worklist *worklist_tail; 519 520 if (wk->wk_state & ONWORKLIST) { 521 if (lk.lkt_held != -1) 522 FREE_LOCK(&lk); 523 panic("add_to_worklist: already on list"); 524 } 525 wk->wk_state |= ONWORKLIST; 526 if (LIST_FIRST(&softdep_workitem_pending) == NULL) 527 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list); 528 else 529 LIST_INSERT_AFTER(worklist_tail, wk, wk_list); 530 worklist_tail = wk; 531 num_on_worklist += 1; 532} 533 534/* 535 * Process that runs once per second to handle items in the background queue. 536 * 537 * Note that we ensure that everything is done in the order in which they 538 * appear in the queue. The code below depends on this property to ensure 539 * that blocks of a file are freed before the inode itself is freed. This 540 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 541 * until all the old ones have been purged from the dependency lists. 542 */ 543int 544softdep_process_worklist(matchmnt) 545 struct mount *matchmnt; 546{ 547 struct proc *p = CURPROC; 548 int matchcnt, loopcount; 549 long starttime; 550 551 /* 552 * Record the process identifier of our caller so that we can give 553 * this process preferential treatment in request_cleanup below. 554 */ 555 filesys_syncer = p; 556 matchcnt = 0; 557 558 /* 559 * There is no danger of having multiple processes run this 560 * code, but we have to single-thread it when softdep_flushfiles() 561 * is in operation to get an accurate count of the number of items 562 * related to its mount point that are in the list. 563 */ 564 if (matchmnt == NULL) { 565 if (softdep_worklist_busy < 0) 566 return(-1); 567 softdep_worklist_busy += 1; 568 } 569 570 /* 571 * If requested, try removing inode or removal dependencies. 572 */ 573 if (req_clear_inodedeps) { 574 clear_inodedeps(p); 575 req_clear_inodedeps -= 1; 576 wakeup_one(&proc_waiting); 577 } 578 if (req_clear_remove) { 579 clear_remove(p); 580 req_clear_remove -= 1; 581 wakeup_one(&proc_waiting); 582 } 583 loopcount = 1; 584 starttime = time_second; 585 while (num_on_worklist > 0) { 586 matchcnt += process_worklist_item(matchmnt, 0); 587 588 /* 589 * If a umount operation wants to run the worklist 590 * accurately, abort. 591 */ 592 if (softdep_worklist_req && matchmnt == NULL) { 593 matchcnt = -1; 594 break; 595 } 596 597 /* 598 * If requested, try removing inode or removal dependencies. 599 */ 600 if (req_clear_inodedeps) { 601 clear_inodedeps(p); 602 req_clear_inodedeps -= 1; 603 wakeup_one(&proc_waiting); 604 } 605 if (req_clear_remove) { 606 clear_remove(p); 607 req_clear_remove -= 1; 608 wakeup_one(&proc_waiting); 609 } 610 /* 611 * We do not generally want to stop for buffer space, but if 612 * we are really being a buffer hog, we will stop and wait. 613 */ 614 if (loopcount++ % 128 == 0) 615 bwillwrite(); 616 /* 617 * Never allow processing to run for more than one 618 * second. Otherwise the other syncer tasks may get 619 * excessively backlogged. 620 */ 621 if (starttime != time_second && matchmnt == NULL) { 622 matchcnt = -1; 623 break; 624 } 625 } 626 if (matchmnt == NULL) { 627 softdep_worklist_busy -= 1; 628 if (softdep_worklist_req && softdep_worklist_busy == 0) 629 wakeup(&softdep_worklist_req); 630 } 631 return (matchcnt); 632} 633 634/* 635 * Process one item on the worklist. 636 */ 637static int 638process_worklist_item(matchmnt, flags) 639 struct mount *matchmnt; 640 int flags; 641{ 642 struct worklist *wk; 643 struct dirrem *dirrem; 644 struct mount *mp; 645 struct vnode *vp; 646 int matchcnt = 0; 647 648 ACQUIRE_LOCK(&lk); 649 /* 650 * Normally we just process each item on the worklist in order. 651 * However, if we are in a situation where we cannot lock any 652 * inodes, we have to skip over any dirrem requests whose 653 * vnodes are resident and locked. 654 */ 655 LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) { 656 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM) 657 break; 658 dirrem = WK_DIRREM(wk); 659 vp = ufs_ihashlookup(VFSTOUFS(dirrem->dm_mnt)->um_dev, 660 dirrem->dm_oldinum); 661 if (vp == NULL || !VOP_ISLOCKED(vp, CURPROC)) 662 break; 663 } 664 if (wk == 0) { 665 FREE_LOCK(&lk); 666 return (0); 667 } 668 WORKLIST_REMOVE(wk); 669 num_on_worklist -= 1; 670 FREE_LOCK(&lk); 671 switch (wk->wk_type) { 672 673 case D_DIRREM: 674 /* removal of a directory entry */ 675 mp = WK_DIRREM(wk)->dm_mnt; 676 if (vn_write_suspend_wait(NULL, mp, V_NOWAIT)) 677 panic("%s: dirrem on suspended filesystem", 678 "process_worklist_item"); 679 if (mp == matchmnt) 680 matchcnt += 1; 681 handle_workitem_remove(WK_DIRREM(wk)); 682 break; 683 684 case D_FREEBLKS: 685 /* releasing blocks and/or fragments from a file */ 686 mp = WK_FREEBLKS(wk)->fb_mnt; 687 if (vn_write_suspend_wait(NULL, mp, V_NOWAIT)) 688 panic("%s: freeblks on suspended filesystem", 689 "process_worklist_item"); 690 if (mp == matchmnt) 691 matchcnt += 1; 692 handle_workitem_freeblocks(WK_FREEBLKS(wk), flags & LK_NOWAIT); 693 break; 694 695 case D_FREEFRAG: 696 /* releasing a fragment when replaced as a file grows */ 697 mp = WK_FREEFRAG(wk)->ff_mnt; 698 if (vn_write_suspend_wait(NULL, mp, V_NOWAIT)) 699 panic("%s: freefrag on suspended filesystem", 700 "process_worklist_item"); 701 if (mp == matchmnt) 702 matchcnt += 1; 703 handle_workitem_freefrag(WK_FREEFRAG(wk)); 704 break; 705 706 case D_FREEFILE: 707 /* releasing an inode when its link count drops to 0 */ 708 mp = WK_FREEFILE(wk)->fx_mnt; 709 if (vn_write_suspend_wait(NULL, mp, V_NOWAIT)) 710 panic("%s: freefile on suspended filesystem", 711 "process_worklist_item"); 712 if (mp == matchmnt) 713 matchcnt += 1; 714 handle_workitem_freefile(WK_FREEFILE(wk)); 715 break; 716 717 default: 718 panic("%s_process_worklist: Unknown type %s", 719 "softdep", TYPENAME(wk->wk_type)); 720 /* NOTREACHED */ 721 } 722 return (matchcnt); 723} 724 725/* 726 * Move dependencies from one buffer to another. 727 */ 728static void 729softdep_move_dependencies(oldbp, newbp) 730 struct buf *oldbp; 731 struct buf *newbp; 732{ 733 struct worklist *wk, *wktail; 734 735 if (LIST_FIRST(&newbp->b_dep) != NULL) 736 panic("softdep_move_dependencies: need merge code"); 737 wktail = 0; 738 ACQUIRE_LOCK(&lk); 739 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 740 LIST_REMOVE(wk, wk_list); 741 if (wktail == 0) 742 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 743 else 744 LIST_INSERT_AFTER(wktail, wk, wk_list); 745 wktail = wk; 746 } 747 FREE_LOCK(&lk); 748} 749 750/* 751 * Purge the work list of all items associated with a particular mount point. 752 */ 753int 754softdep_flushworklist(oldmnt, countp, p) 755 struct mount *oldmnt; 756 int *countp; 757 struct proc *p; 758{ 759 struct vnode *devvp; 760 int count, error = 0; 761 762 /* 763 * Await our turn to clear out the queue, then serialize access. 764 */ 765 while (softdep_worklist_busy) { 766 softdep_worklist_req += 1; 767 tsleep(&softdep_worklist_req, PRIBIO, "softflush", 0); 768 softdep_worklist_req -= 1; 769 } 770 softdep_worklist_busy = -1; 771 /* 772 * Alternately flush the block device associated with the mount 773 * point and process any dependencies that the flushing 774 * creates. We continue until no more worklist dependencies 775 * are found. 776 */ 777 *countp = 0; 778 devvp = VFSTOUFS(oldmnt)->um_devvp; 779 while ((count = softdep_process_worklist(oldmnt)) > 0) { 780 *countp += count; 781 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, p); 782 error = VOP_FSYNC(devvp, p->p_ucred, MNT_WAIT, p); 783 VOP_UNLOCK(devvp, 0, p); 784 if (error) 785 break; 786 } 787 softdep_worklist_busy = 0; 788 if (softdep_worklist_req) 789 wakeup(&softdep_worklist_req); 790 return (error); 791} 792 793/* 794 * Flush all vnodes and worklist items associated with a specified mount point. 795 */ 796int 797softdep_flushfiles(oldmnt, flags, p) 798 struct mount *oldmnt; 799 int flags; 800 struct proc *p; 801{ 802 int error, count, loopcnt; 803 804 /* 805 * Alternately flush the vnodes associated with the mount 806 * point and process any dependencies that the flushing 807 * creates. In theory, this loop can happen at most twice, 808 * but we give it a few extra just to be sure. 809 */ 810 for (loopcnt = 10; loopcnt > 0; loopcnt--) { 811 /* 812 * Do another flush in case any vnodes were brought in 813 * as part of the cleanup operations. 814 */ 815 if ((error = ffs_flushfiles(oldmnt, flags, p)) != 0) 816 break; 817 if ((error = softdep_flushworklist(oldmnt, &count, p)) != 0 || 818 count == 0) 819 break; 820 } 821 /* 822 * If we are unmounting then it is an error to fail. If we 823 * are simply trying to downgrade to read-only, then filesystem 824 * activity can keep us busy forever, so we just fail with EBUSY. 825 */ 826 if (loopcnt == 0) { 827 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 828 panic("softdep_flushfiles: looping"); 829 error = EBUSY; 830 } 831 return (error); 832} 833 834/* 835 * Structure hashing. 836 * 837 * There are three types of structures that can be looked up: 838 * 1) pagedep structures identified by mount point, inode number, 839 * and logical block. 840 * 2) inodedep structures identified by mount point and inode number. 841 * 3) newblk structures identified by mount point and 842 * physical block number. 843 * 844 * The "pagedep" and "inodedep" dependency structures are hashed 845 * separately from the file blocks and inodes to which they correspond. 846 * This separation helps when the in-memory copy of an inode or 847 * file block must be replaced. It also obviates the need to access 848 * an inode or file page when simply updating (or de-allocating) 849 * dependency structures. Lookup of newblk structures is needed to 850 * find newly allocated blocks when trying to associate them with 851 * their allocdirect or allocindir structure. 852 * 853 * The lookup routines optionally create and hash a new instance when 854 * an existing entry is not found. 855 */ 856#define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 857#define NODELAY 0x0002 /* cannot do background work */ 858 859/* 860 * Structures and routines associated with pagedep caching. 861 */ 862LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl; 863u_long pagedep_hash; /* size of hash table - 1 */ 864#define PAGEDEP_HASH(mp, inum, lbn) \ 865 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \ 866 pagedep_hash]) 867static struct sema pagedep_in_progress; 868 869/* 870 * Look up a pagedep. Return 1 if found, 0 if not found. 871 * If not found, allocate if DEPALLOC flag is passed. 872 * Found or allocated entry is returned in pagedeppp. 873 * This routine must be called with splbio interrupts blocked. 874 */ 875static int 876pagedep_lookup(ip, lbn, flags, pagedeppp) 877 struct inode *ip; 878 ufs_lbn_t lbn; 879 int flags; 880 struct pagedep **pagedeppp; 881{ 882 struct pagedep *pagedep; 883 struct pagedep_hashhead *pagedephd; 884 struct mount *mp; 885 int i; 886 887#ifdef DEBUG 888 if (lk.lkt_held == -1) 889 panic("pagedep_lookup: lock not held"); 890#endif 891 mp = ITOV(ip)->v_mount; 892 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn); 893top: 894 LIST_FOREACH(pagedep, pagedephd, pd_hash) 895 if (ip->i_number == pagedep->pd_ino && 896 lbn == pagedep->pd_lbn && 897 mp == pagedep->pd_mnt) 898 break; 899 if (pagedep) { 900 *pagedeppp = pagedep; 901 return (1); 902 } 903 if ((flags & DEPALLOC) == 0) { 904 *pagedeppp = NULL; 905 return (0); 906 } 907 if (sema_get(&pagedep_in_progress, &lk) == 0) { 908 ACQUIRE_LOCK(&lk); 909 goto top; 910 } 911 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP, 912 M_SOFTDEP_FLAGS|M_ZERO); 913 pagedep->pd_list.wk_type = D_PAGEDEP; 914 pagedep->pd_mnt = mp; 915 pagedep->pd_ino = ip->i_number; 916 pagedep->pd_lbn = lbn; 917 LIST_INIT(&pagedep->pd_dirremhd); 918 LIST_INIT(&pagedep->pd_pendinghd); 919 for (i = 0; i < DAHASHSZ; i++) 920 LIST_INIT(&pagedep->pd_diraddhd[i]); 921 ACQUIRE_LOCK(&lk); 922 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 923 sema_release(&pagedep_in_progress); 924 *pagedeppp = pagedep; 925 return (0); 926} 927 928/* 929 * Structures and routines associated with inodedep caching. 930 */ 931LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl; 932static u_long inodedep_hash; /* size of hash table - 1 */ 933static long num_inodedep; /* number of inodedep allocated */ 934#define INODEDEP_HASH(fs, inum) \ 935 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash]) 936static struct sema inodedep_in_progress; 937 938/* 939 * Look up a inodedep. Return 1 if found, 0 if not found. 940 * If not found, allocate if DEPALLOC flag is passed. 941 * Found or allocated entry is returned in inodedeppp. 942 * This routine must be called with splbio interrupts blocked. 943 */ 944static int 945inodedep_lookup(fs, inum, flags, inodedeppp) 946 struct fs *fs; 947 ino_t inum; 948 int flags; 949 struct inodedep **inodedeppp; 950{ 951 struct inodedep *inodedep; 952 struct inodedep_hashhead *inodedephd; 953 int firsttry; 954 955#ifdef DEBUG 956 if (lk.lkt_held == -1) 957 panic("inodedep_lookup: lock not held"); 958#endif 959 firsttry = 1; 960 inodedephd = INODEDEP_HASH(fs, inum); 961top: 962 LIST_FOREACH(inodedep, inodedephd, id_hash) 963 if (inum == inodedep->id_ino && fs == inodedep->id_fs) 964 break; 965 if (inodedep) { 966 *inodedeppp = inodedep; 967 return (1); 968 } 969 if ((flags & DEPALLOC) == 0) { 970 *inodedeppp = NULL; 971 return (0); 972 } 973 /* 974 * If we are over our limit, try to improve the situation. 975 */ 976 if (num_inodedep > max_softdeps && firsttry && (flags & NODELAY) == 0 && 977 request_cleanup(FLUSH_INODES, 1)) { 978 firsttry = 0; 979 goto top; 980 } 981 if (sema_get(&inodedep_in_progress, &lk) == 0) { 982 ACQUIRE_LOCK(&lk); 983 goto top; 984 } 985 num_inodedep += 1; 986 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep), 987 M_INODEDEP, M_SOFTDEP_FLAGS); 988 inodedep->id_list.wk_type = D_INODEDEP; 989 inodedep->id_fs = fs; 990 inodedep->id_ino = inum; 991 inodedep->id_state = ALLCOMPLETE; 992 inodedep->id_nlinkdelta = 0; 993 inodedep->id_savedino = NULL; 994 inodedep->id_savedsize = -1; 995 inodedep->id_buf = NULL; 996 LIST_INIT(&inodedep->id_pendinghd); 997 LIST_INIT(&inodedep->id_inowait); 998 LIST_INIT(&inodedep->id_bufwait); 999 TAILQ_INIT(&inodedep->id_inoupdt); 1000 TAILQ_INIT(&inodedep->id_newinoupdt); 1001 ACQUIRE_LOCK(&lk); 1002 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 1003 sema_release(&inodedep_in_progress); 1004 *inodedeppp = inodedep; 1005 return (0); 1006} 1007 1008/* 1009 * Structures and routines associated with newblk caching. 1010 */ 1011LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl; 1012u_long newblk_hash; /* size of hash table - 1 */ 1013#define NEWBLK_HASH(fs, inum) \ 1014 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash]) 1015static struct sema newblk_in_progress; 1016 1017/* 1018 * Look up a newblk. Return 1 if found, 0 if not found. 1019 * If not found, allocate if DEPALLOC flag is passed. 1020 * Found or allocated entry is returned in newblkpp. 1021 */ 1022static int 1023newblk_lookup(fs, newblkno, flags, newblkpp) 1024 struct fs *fs; 1025 ufs_daddr_t newblkno; 1026 int flags; 1027 struct newblk **newblkpp; 1028{ 1029 struct newblk *newblk; 1030 struct newblk_hashhead *newblkhd; 1031 1032 newblkhd = NEWBLK_HASH(fs, newblkno); 1033top: 1034 LIST_FOREACH(newblk, newblkhd, nb_hash) 1035 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs) 1036 break; 1037 if (newblk) { 1038 *newblkpp = newblk; 1039 return (1); 1040 } 1041 if ((flags & DEPALLOC) == 0) { 1042 *newblkpp = NULL; 1043 return (0); 1044 } 1045 if (sema_get(&newblk_in_progress, 0) == 0) 1046 goto top; 1047 MALLOC(newblk, struct newblk *, sizeof(struct newblk), 1048 M_NEWBLK, M_SOFTDEP_FLAGS); 1049 newblk->nb_state = 0; 1050 newblk->nb_fs = fs; 1051 newblk->nb_newblkno = newblkno; 1052 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 1053 sema_release(&newblk_in_progress); 1054 *newblkpp = newblk; 1055 return (0); 1056} 1057 1058/* 1059 * Executed during filesystem system initialization before 1060 * mounting any file systems. 1061 */ 1062void 1063softdep_initialize() 1064{ 1065 1066 LIST_INIT(&mkdirlisthd); 1067 LIST_INIT(&softdep_workitem_pending); 1068 max_softdeps = min(desiredvnodes * 8, 1069 M_INODEDEP->ks_limit / (2 * sizeof(struct inodedep))); 1070 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 1071 &pagedep_hash); 1072 sema_init(&pagedep_in_progress, "pagedep", PRIBIO, 0); 1073 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash); 1074 sema_init(&inodedep_in_progress, "inodedep", PRIBIO, 0); 1075 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash); 1076 sema_init(&newblk_in_progress, "newblk", PRIBIO, 0); 1077} 1078 1079/* 1080 * Called at mount time to notify the dependency code that a 1081 * filesystem wishes to use it. 1082 */ 1083int 1084softdep_mount(devvp, mp, fs, cred) 1085 struct vnode *devvp; 1086 struct mount *mp; 1087 struct fs *fs; 1088 struct ucred *cred; 1089{ 1090 struct csum cstotal; 1091 struct cg *cgp; 1092 struct buf *bp; 1093 int error, cyl; 1094 1095 mp->mnt_flag &= ~MNT_ASYNC; 1096 mp->mnt_flag |= MNT_SOFTDEP; 1097 /* 1098 * When doing soft updates, the counters in the 1099 * superblock may have gotten out of sync, so we have 1100 * to scan the cylinder groups and recalculate them. 1101 */ 1102 if (fs->fs_clean != 0) 1103 return (0); 1104 bzero(&cstotal, sizeof cstotal); 1105 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 1106 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 1107 fs->fs_cgsize, cred, &bp)) != 0) { 1108 brelse(bp); 1109 return (error); 1110 } 1111 cgp = (struct cg *)bp->b_data; 1112 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 1113 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 1114 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 1115 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 1116 fs->fs_cs(fs, cyl) = cgp->cg_cs; 1117 brelse(bp); 1118 } 1119#ifdef DEBUG 1120 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 1121 printf("%s: superblock summary recomputed\n", fs->fs_fsmnt); 1122#endif 1123 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 1124 return (0); 1125} 1126 1127/* 1128 * Protecting the freemaps (or bitmaps). 1129 * 1130 * To eliminate the need to execute fsck before mounting a file system 1131 * after a power failure, one must (conservatively) guarantee that the 1132 * on-disk copy of the bitmaps never indicate that a live inode or block is 1133 * free. So, when a block or inode is allocated, the bitmap should be 1134 * updated (on disk) before any new pointers. When a block or inode is 1135 * freed, the bitmap should not be updated until all pointers have been 1136 * reset. The latter dependency is handled by the delayed de-allocation 1137 * approach described below for block and inode de-allocation. The former 1138 * dependency is handled by calling the following procedure when a block or 1139 * inode is allocated. When an inode is allocated an "inodedep" is created 1140 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 1141 * Each "inodedep" is also inserted into the hash indexing structure so 1142 * that any additional link additions can be made dependent on the inode 1143 * allocation. 1144 * 1145 * The ufs file system maintains a number of free block counts (e.g., per 1146 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 1147 * in addition to the bitmaps. These counts are used to improve efficiency 1148 * during allocation and therefore must be consistent with the bitmaps. 1149 * There is no convenient way to guarantee post-crash consistency of these 1150 * counts with simple update ordering, for two main reasons: (1) The counts 1151 * and bitmaps for a single cylinder group block are not in the same disk 1152 * sector. If a disk write is interrupted (e.g., by power failure), one may 1153 * be written and the other not. (2) Some of the counts are located in the 1154 * superblock rather than the cylinder group block. So, we focus our soft 1155 * updates implementation on protecting the bitmaps. When mounting a 1156 * filesystem, we recompute the auxiliary counts from the bitmaps. 1157 */ 1158 1159/* 1160 * Called just after updating the cylinder group block to allocate an inode. 1161 */ 1162void 1163softdep_setup_inomapdep(bp, ip, newinum) 1164 struct buf *bp; /* buffer for cylgroup block with inode map */ 1165 struct inode *ip; /* inode related to allocation */ 1166 ino_t newinum; /* new inode number being allocated */ 1167{ 1168 struct inodedep *inodedep; 1169 struct bmsafemap *bmsafemap; 1170 1171 /* 1172 * Create a dependency for the newly allocated inode. 1173 * Panic if it already exists as something is seriously wrong. 1174 * Otherwise add it to the dependency list for the buffer holding 1175 * the cylinder group map from which it was allocated. 1176 */ 1177 ACQUIRE_LOCK(&lk); 1178 if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) { 1179 FREE_LOCK(&lk); 1180 panic("softdep_setup_inomapdep: found inode"); 1181 } 1182 inodedep->id_buf = bp; 1183 inodedep->id_state &= ~DEPCOMPLETE; 1184 bmsafemap = bmsafemap_lookup(bp); 1185 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 1186 FREE_LOCK(&lk); 1187} 1188 1189/* 1190 * Called just after updating the cylinder group block to 1191 * allocate block or fragment. 1192 */ 1193void 1194softdep_setup_blkmapdep(bp, fs, newblkno) 1195 struct buf *bp; /* buffer for cylgroup block with block map */ 1196 struct fs *fs; /* filesystem doing allocation */ 1197 ufs_daddr_t newblkno; /* number of newly allocated block */ 1198{ 1199 struct newblk *newblk; 1200 struct bmsafemap *bmsafemap; 1201 1202 /* 1203 * Create a dependency for the newly allocated block. 1204 * Add it to the dependency list for the buffer holding 1205 * the cylinder group map from which it was allocated. 1206 */ 1207 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0) 1208 panic("softdep_setup_blkmapdep: found block"); 1209 ACQUIRE_LOCK(&lk); 1210 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp); 1211 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 1212 FREE_LOCK(&lk); 1213} 1214 1215/* 1216 * Find the bmsafemap associated with a cylinder group buffer. 1217 * If none exists, create one. The buffer must be locked when 1218 * this routine is called and this routine must be called with 1219 * splbio interrupts blocked. 1220 */ 1221static struct bmsafemap * 1222bmsafemap_lookup(bp) 1223 struct buf *bp; 1224{ 1225 struct bmsafemap *bmsafemap; 1226 struct worklist *wk; 1227 1228#ifdef DEBUG 1229 if (lk.lkt_held == -1) 1230 panic("bmsafemap_lookup: lock not held"); 1231#endif 1232 LIST_FOREACH(wk, &bp->b_dep, wk_list) 1233 if (wk->wk_type == D_BMSAFEMAP) 1234 return (WK_BMSAFEMAP(wk)); 1235 FREE_LOCK(&lk); 1236 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap), 1237 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 1238 bmsafemap->sm_list.wk_type = D_BMSAFEMAP; 1239 bmsafemap->sm_list.wk_state = 0; 1240 bmsafemap->sm_buf = bp; 1241 LIST_INIT(&bmsafemap->sm_allocdirecthd); 1242 LIST_INIT(&bmsafemap->sm_allocindirhd); 1243 LIST_INIT(&bmsafemap->sm_inodedephd); 1244 LIST_INIT(&bmsafemap->sm_newblkhd); 1245 ACQUIRE_LOCK(&lk); 1246 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 1247 return (bmsafemap); 1248} 1249 1250/* 1251 * Direct block allocation dependencies. 1252 * 1253 * When a new block is allocated, the corresponding disk locations must be 1254 * initialized (with zeros or new data) before the on-disk inode points to 1255 * them. Also, the freemap from which the block was allocated must be 1256 * updated (on disk) before the inode's pointer. These two dependencies are 1257 * independent of each other and are needed for all file blocks and indirect 1258 * blocks that are pointed to directly by the inode. Just before the 1259 * "in-core" version of the inode is updated with a newly allocated block 1260 * number, a procedure (below) is called to setup allocation dependency 1261 * structures. These structures are removed when the corresponding 1262 * dependencies are satisfied or when the block allocation becomes obsolete 1263 * (i.e., the file is deleted, the block is de-allocated, or the block is a 1264 * fragment that gets upgraded). All of these cases are handled in 1265 * procedures described later. 1266 * 1267 * When a file extension causes a fragment to be upgraded, either to a larger 1268 * fragment or to a full block, the on-disk location may change (if the 1269 * previous fragment could not simply be extended). In this case, the old 1270 * fragment must be de-allocated, but not until after the inode's pointer has 1271 * been updated. In most cases, this is handled by later procedures, which 1272 * will construct a "freefrag" structure to be added to the workitem queue 1273 * when the inode update is complete (or obsolete). The main exception to 1274 * this is when an allocation occurs while a pending allocation dependency 1275 * (for the same block pointer) remains. This case is handled in the main 1276 * allocation dependency setup procedure by immediately freeing the 1277 * unreferenced fragments. 1278 */ 1279void 1280softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 1281 struct inode *ip; /* inode to which block is being added */ 1282 ufs_lbn_t lbn; /* block pointer within inode */ 1283 ufs_daddr_t newblkno; /* disk block number being added */ 1284 ufs_daddr_t oldblkno; /* previous block number, 0 unless frag */ 1285 long newsize; /* size of new block */ 1286 long oldsize; /* size of new block */ 1287 struct buf *bp; /* bp for allocated block */ 1288{ 1289 struct allocdirect *adp, *oldadp; 1290 struct allocdirectlst *adphead; 1291 struct bmsafemap *bmsafemap; 1292 struct inodedep *inodedep; 1293 struct pagedep *pagedep; 1294 struct newblk *newblk; 1295 1296 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect), 1297 M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO); 1298 adp->ad_list.wk_type = D_ALLOCDIRECT; 1299 adp->ad_lbn = lbn; 1300 adp->ad_newblkno = newblkno; 1301 adp->ad_oldblkno = oldblkno; 1302 adp->ad_newsize = newsize; 1303 adp->ad_oldsize = oldsize; 1304 adp->ad_state = ATTACHED; 1305 LIST_INIT(&adp->ad_newdirblk); 1306 if (newblkno == oldblkno) 1307 adp->ad_freefrag = NULL; 1308 else 1309 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1310 1311 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1312 panic("softdep_setup_allocdirect: lost block"); 1313 1314 ACQUIRE_LOCK(&lk); 1315 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep); 1316 adp->ad_inodedep = inodedep; 1317 1318 if (newblk->nb_state == DEPCOMPLETE) { 1319 adp->ad_state |= DEPCOMPLETE; 1320 adp->ad_buf = NULL; 1321 } else { 1322 bmsafemap = newblk->nb_bmsafemap; 1323 adp->ad_buf = bmsafemap->sm_buf; 1324 LIST_REMOVE(newblk, nb_deps); 1325 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1326 } 1327 LIST_REMOVE(newblk, nb_hash); 1328 FREE(newblk, M_NEWBLK); 1329 1330 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list); 1331 if (lbn >= NDADDR) { 1332 /* allocating an indirect block */ 1333 if (oldblkno != 0) { 1334 FREE_LOCK(&lk); 1335 panic("softdep_setup_allocdirect: non-zero indir"); 1336 } 1337 } else { 1338 /* 1339 * Allocating a direct block. 1340 * 1341 * If we are allocating a directory block, then we must 1342 * allocate an associated pagedep to track additions and 1343 * deletions. 1344 */ 1345 if ((ip->i_mode & IFMT) == IFDIR && 1346 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1347 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 1348 } 1349 /* 1350 * The list of allocdirects must be kept in sorted and ascending 1351 * order so that the rollback routines can quickly determine the 1352 * first uncommitted block (the size of the file stored on disk 1353 * ends at the end of the lowest committed fragment, or if there 1354 * are no fragments, at the end of the highest committed block). 1355 * Since files generally grow, the typical case is that the new 1356 * block is to be added at the end of the list. We speed this 1357 * special case by checking against the last allocdirect in the 1358 * list before laboriously traversing the list looking for the 1359 * insertion point. 1360 */ 1361 adphead = &inodedep->id_newinoupdt; 1362 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1363 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1364 /* insert at end of list */ 1365 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1366 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1367 allocdirect_merge(adphead, adp, oldadp); 1368 FREE_LOCK(&lk); 1369 return; 1370 } 1371 TAILQ_FOREACH(oldadp, adphead, ad_next) { 1372 if (oldadp->ad_lbn >= lbn) 1373 break; 1374 } 1375 if (oldadp == NULL) { 1376 FREE_LOCK(&lk); 1377 panic("softdep_setup_allocdirect: lost entry"); 1378 } 1379 /* insert in middle of list */ 1380 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1381 if (oldadp->ad_lbn == lbn) 1382 allocdirect_merge(adphead, adp, oldadp); 1383 FREE_LOCK(&lk); 1384} 1385 1386/* 1387 * Replace an old allocdirect dependency with a newer one. 1388 * This routine must be called with splbio interrupts blocked. 1389 */ 1390static void 1391allocdirect_merge(adphead, newadp, oldadp) 1392 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 1393 struct allocdirect *newadp; /* allocdirect being added */ 1394 struct allocdirect *oldadp; /* existing allocdirect being checked */ 1395{ 1396 struct worklist *wk; 1397 struct freefrag *freefrag; 1398 struct newdirblk *newdirblk; 1399 1400#ifdef DEBUG 1401 if (lk.lkt_held == -1) 1402 panic("allocdirect_merge: lock not held"); 1403#endif 1404 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 1405 newadp->ad_oldsize != oldadp->ad_newsize || 1406 newadp->ad_lbn >= NDADDR) { 1407 FREE_LOCK(&lk); 1408 panic("allocdirect_merge: old %d != new %d || lbn %ld >= %d", 1409 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn, 1410 NDADDR); 1411 } 1412 newadp->ad_oldblkno = oldadp->ad_oldblkno; 1413 newadp->ad_oldsize = oldadp->ad_oldsize; 1414 /* 1415 * If the old dependency had a fragment to free or had never 1416 * previously had a block allocated, then the new dependency 1417 * can immediately post its freefrag and adopt the old freefrag. 1418 * This action is done by swapping the freefrag dependencies. 1419 * The new dependency gains the old one's freefrag, and the 1420 * old one gets the new one and then immediately puts it on 1421 * the worklist when it is freed by free_allocdirect. It is 1422 * not possible to do this swap when the old dependency had a 1423 * non-zero size but no previous fragment to free. This condition 1424 * arises when the new block is an extension of the old block. 1425 * Here, the first part of the fragment allocated to the new 1426 * dependency is part of the block currently claimed on disk by 1427 * the old dependency, so cannot legitimately be freed until the 1428 * conditions for the new dependency are fulfilled. 1429 */ 1430 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 1431 freefrag = newadp->ad_freefrag; 1432 newadp->ad_freefrag = oldadp->ad_freefrag; 1433 oldadp->ad_freefrag = freefrag; 1434 } 1435 /* 1436 * If we are tracking a new directory-block allocation, 1437 * move it from the old allocdirect to the new allocdirect. 1438 */ 1439 if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) { 1440 newdirblk = WK_NEWDIRBLK(wk); 1441 WORKLIST_REMOVE(&newdirblk->db_list); 1442 if (LIST_FIRST(&oldadp->ad_newdirblk) != NULL) 1443 panic("allocdirect_merge: extra newdirblk"); 1444 WORKLIST_INSERT(&newadp->ad_newdirblk, &newdirblk->db_list); 1445 } 1446 free_allocdirect(adphead, oldadp, 0); 1447} 1448 1449/* 1450 * Allocate a new freefrag structure if needed. 1451 */ 1452static struct freefrag * 1453newfreefrag(ip, blkno, size) 1454 struct inode *ip; 1455 ufs_daddr_t blkno; 1456 long size; 1457{ 1458 struct freefrag *freefrag; 1459 struct fs *fs; 1460 1461 if (blkno == 0) 1462 return (NULL); 1463 fs = ip->i_fs; 1464 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 1465 panic("newfreefrag: frag size"); 1466 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag), 1467 M_FREEFRAG, M_SOFTDEP_FLAGS); 1468 freefrag->ff_list.wk_type = D_FREEFRAG; 1469 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */ 1470 freefrag->ff_inum = ip->i_number; 1471 freefrag->ff_mnt = ITOV(ip)->v_mount; 1472 freefrag->ff_devvp = ip->i_devvp; 1473 freefrag->ff_blkno = blkno; 1474 freefrag->ff_fragsize = size; 1475 return (freefrag); 1476} 1477 1478/* 1479 * This workitem de-allocates fragments that were replaced during 1480 * file block allocation. 1481 */ 1482static void 1483handle_workitem_freefrag(freefrag) 1484 struct freefrag *freefrag; 1485{ 1486 struct inode tip; 1487 1488 tip.i_vnode = NULL; 1489 tip.i_fs = VFSTOUFS(freefrag->ff_mnt)->um_fs; 1490 tip.i_devvp = freefrag->ff_devvp; 1491 tip.i_dev = freefrag->ff_devvp->v_rdev; 1492 tip.i_number = freefrag->ff_inum; 1493 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */ 1494 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize); 1495 FREE(freefrag, M_FREEFRAG); 1496} 1497 1498/* 1499 * Indirect block allocation dependencies. 1500 * 1501 * The same dependencies that exist for a direct block also exist when 1502 * a new block is allocated and pointed to by an entry in a block of 1503 * indirect pointers. The undo/redo states described above are also 1504 * used here. Because an indirect block contains many pointers that 1505 * may have dependencies, a second copy of the entire in-memory indirect 1506 * block is kept. The buffer cache copy is always completely up-to-date. 1507 * The second copy, which is used only as a source for disk writes, 1508 * contains only the safe pointers (i.e., those that have no remaining 1509 * update dependencies). The second copy is freed when all pointers 1510 * are safe. The cache is not allowed to replace indirect blocks with 1511 * pending update dependencies. If a buffer containing an indirect 1512 * block with dependencies is written, these routines will mark it 1513 * dirty again. It can only be successfully written once all the 1514 * dependencies are removed. The ffs_fsync routine in conjunction with 1515 * softdep_sync_metadata work together to get all the dependencies 1516 * removed so that a file can be successfully written to disk. Three 1517 * procedures are used when setting up indirect block pointer 1518 * dependencies. The division is necessary because of the organization 1519 * of the "balloc" routine and because of the distinction between file 1520 * pages and file metadata blocks. 1521 */ 1522 1523/* 1524 * Allocate a new allocindir structure. 1525 */ 1526static struct allocindir * 1527newallocindir(ip, ptrno, newblkno, oldblkno) 1528 struct inode *ip; /* inode for file being extended */ 1529 int ptrno; /* offset of pointer in indirect block */ 1530 ufs_daddr_t newblkno; /* disk block number being added */ 1531 ufs_daddr_t oldblkno; /* previous block number, 0 if none */ 1532{ 1533 struct allocindir *aip; 1534 1535 MALLOC(aip, struct allocindir *, sizeof(struct allocindir), 1536 M_ALLOCINDIR, M_SOFTDEP_FLAGS|M_ZERO); 1537 aip->ai_list.wk_type = D_ALLOCINDIR; 1538 aip->ai_state = ATTACHED; 1539 aip->ai_offset = ptrno; 1540 aip->ai_newblkno = newblkno; 1541 aip->ai_oldblkno = oldblkno; 1542 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize); 1543 return (aip); 1544} 1545 1546/* 1547 * Called just before setting an indirect block pointer 1548 * to a newly allocated file page. 1549 */ 1550void 1551softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 1552 struct inode *ip; /* inode for file being extended */ 1553 ufs_lbn_t lbn; /* allocated block number within file */ 1554 struct buf *bp; /* buffer with indirect blk referencing page */ 1555 int ptrno; /* offset of pointer in indirect block */ 1556 ufs_daddr_t newblkno; /* disk block number being added */ 1557 ufs_daddr_t oldblkno; /* previous block number, 0 if none */ 1558 struct buf *nbp; /* buffer holding allocated page */ 1559{ 1560 struct allocindir *aip; 1561 struct pagedep *pagedep; 1562 1563 aip = newallocindir(ip, ptrno, newblkno, oldblkno); 1564 ACQUIRE_LOCK(&lk); 1565 /* 1566 * If we are allocating a directory page, then we must 1567 * allocate an associated pagedep to track additions and 1568 * deletions. 1569 */ 1570 if ((ip->i_mode & IFMT) == IFDIR && 1571 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1572 WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list); 1573 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1574 FREE_LOCK(&lk); 1575 setup_allocindir_phase2(bp, ip, aip); 1576} 1577 1578/* 1579 * Called just before setting an indirect block pointer to a 1580 * newly allocated indirect block. 1581 */ 1582void 1583softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 1584 struct buf *nbp; /* newly allocated indirect block */ 1585 struct inode *ip; /* inode for file being extended */ 1586 struct buf *bp; /* indirect block referencing allocated block */ 1587 int ptrno; /* offset of pointer in indirect block */ 1588 ufs_daddr_t newblkno; /* disk block number being added */ 1589{ 1590 struct allocindir *aip; 1591 1592 aip = newallocindir(ip, ptrno, newblkno, 0); 1593 ACQUIRE_LOCK(&lk); 1594 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1595 FREE_LOCK(&lk); 1596 setup_allocindir_phase2(bp, ip, aip); 1597} 1598 1599/* 1600 * Called to finish the allocation of the "aip" allocated 1601 * by one of the two routines above. 1602 */ 1603static void 1604setup_allocindir_phase2(bp, ip, aip) 1605 struct buf *bp; /* in-memory copy of the indirect block */ 1606 struct inode *ip; /* inode for file being extended */ 1607 struct allocindir *aip; /* allocindir allocated by the above routines */ 1608{ 1609 struct worklist *wk; 1610 struct indirdep *indirdep, *newindirdep; 1611 struct bmsafemap *bmsafemap; 1612 struct allocindir *oldaip; 1613 struct freefrag *freefrag; 1614 struct newblk *newblk; 1615 1616 if (bp->b_lblkno >= 0) 1617 panic("setup_allocindir_phase2: not indir blk"); 1618 for (indirdep = NULL, newindirdep = NULL; ; ) { 1619 ACQUIRE_LOCK(&lk); 1620 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 1621 if (wk->wk_type != D_INDIRDEP) 1622 continue; 1623 indirdep = WK_INDIRDEP(wk); 1624 break; 1625 } 1626 if (indirdep == NULL && newindirdep) { 1627 indirdep = newindirdep; 1628 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 1629 newindirdep = NULL; 1630 } 1631 FREE_LOCK(&lk); 1632 if (indirdep) { 1633 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0, 1634 &newblk) == 0) 1635 panic("setup_allocindir: lost block"); 1636 ACQUIRE_LOCK(&lk); 1637 if (newblk->nb_state == DEPCOMPLETE) { 1638 aip->ai_state |= DEPCOMPLETE; 1639 aip->ai_buf = NULL; 1640 } else { 1641 bmsafemap = newblk->nb_bmsafemap; 1642 aip->ai_buf = bmsafemap->sm_buf; 1643 LIST_REMOVE(newblk, nb_deps); 1644 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd, 1645 aip, ai_deps); 1646 } 1647 LIST_REMOVE(newblk, nb_hash); 1648 FREE(newblk, M_NEWBLK); 1649 aip->ai_indirdep = indirdep; 1650 /* 1651 * Check to see if there is an existing dependency 1652 * for this block. If there is, merge the old 1653 * dependency into the new one. 1654 */ 1655 if (aip->ai_oldblkno == 0) 1656 oldaip = NULL; 1657 else 1658 1659 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) 1660 if (oldaip->ai_offset == aip->ai_offset) 1661 break; 1662 freefrag = NULL; 1663 if (oldaip != NULL) { 1664 if (oldaip->ai_newblkno != aip->ai_oldblkno) { 1665 FREE_LOCK(&lk); 1666 panic("setup_allocindir_phase2: blkno"); 1667 } 1668 aip->ai_oldblkno = oldaip->ai_oldblkno; 1669 freefrag = aip->ai_freefrag; 1670 aip->ai_freefrag = oldaip->ai_freefrag; 1671 oldaip->ai_freefrag = NULL; 1672 free_allocindir(oldaip, NULL); 1673 } 1674 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 1675 ((ufs_daddr_t *)indirdep->ir_savebp->b_data) 1676 [aip->ai_offset] = aip->ai_oldblkno; 1677 FREE_LOCK(&lk); 1678 if (freefrag != NULL) 1679 handle_workitem_freefrag(freefrag); 1680 } 1681 if (newindirdep) { 1682 if (indirdep->ir_savebp != NULL) 1683 brelse(newindirdep->ir_savebp); 1684 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP); 1685 } 1686 if (indirdep) 1687 break; 1688 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep), 1689 M_INDIRDEP, M_SOFTDEP_FLAGS); 1690 newindirdep->ir_list.wk_type = D_INDIRDEP; 1691 newindirdep->ir_state = ATTACHED; 1692 LIST_INIT(&newindirdep->ir_deplisthd); 1693 LIST_INIT(&newindirdep->ir_donehd); 1694 if (bp->b_blkno == bp->b_lblkno) 1695 ufs_bmaparray(bp->b_vp, bp->b_lblkno, &bp->b_blkno, NULL, NULL); 1696 newindirdep->ir_savebp = 1697 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0); 1698 BUF_KERNPROC(newindirdep->ir_savebp); 1699 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 1700 } 1701} 1702 1703/* 1704 * Block de-allocation dependencies. 1705 * 1706 * When blocks are de-allocated, the on-disk pointers must be nullified before 1707 * the blocks are made available for use by other files. (The true 1708 * requirement is that old pointers must be nullified before new on-disk 1709 * pointers are set. We chose this slightly more stringent requirement to 1710 * reduce complexity.) Our implementation handles this dependency by updating 1711 * the inode (or indirect block) appropriately but delaying the actual block 1712 * de-allocation (i.e., freemap and free space count manipulation) until 1713 * after the updated versions reach stable storage. After the disk is 1714 * updated, the blocks can be safely de-allocated whenever it is convenient. 1715 * This implementation handles only the common case of reducing a file's 1716 * length to zero. Other cases are handled by the conventional synchronous 1717 * write approach. 1718 * 1719 * The ffs implementation with which we worked double-checks 1720 * the state of the block pointers and file size as it reduces 1721 * a file's length. Some of this code is replicated here in our 1722 * soft updates implementation. The freeblks->fb_chkcnt field is 1723 * used to transfer a part of this information to the procedure 1724 * that eventually de-allocates the blocks. 1725 * 1726 * This routine should be called from the routine that shortens 1727 * a file's length, before the inode's size or block pointers 1728 * are modified. It will save the block pointer information for 1729 * later release and zero the inode so that the calling routine 1730 * can release it. 1731 */ 1732void 1733softdep_setup_freeblocks(ip, length) 1734 struct inode *ip; /* The inode whose length is to be reduced */ 1735 off_t length; /* The new length for the file */ 1736{ 1737 struct freeblks *freeblks; 1738 struct inodedep *inodedep; 1739 struct allocdirect *adp; 1740 struct vnode *vp; 1741 struct buf *bp; 1742 struct fs *fs; 1743 int i, delay, error; 1744 1745 fs = ip->i_fs; 1746 if (length != 0) 1747 panic("softdep_setup_freeblocks: non-zero length"); 1748 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks), 1749 M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO); 1750 freeblks->fb_list.wk_type = D_FREEBLKS; 1751 freeblks->fb_uid = ip->i_uid; 1752 freeblks->fb_previousinum = ip->i_number; 1753 freeblks->fb_devvp = ip->i_devvp; 1754 freeblks->fb_mnt = ITOV(ip)->v_mount; 1755 freeblks->fb_oldsize = ip->i_size; 1756 freeblks->fb_newsize = length; 1757 freeblks->fb_chkcnt = ip->i_blocks; 1758 for (i = 0; i < NDADDR; i++) { 1759 freeblks->fb_dblks[i] = ip->i_db[i]; 1760 ip->i_db[i] = 0; 1761 } 1762 for (i = 0; i < NIADDR; i++) { 1763 freeblks->fb_iblks[i] = ip->i_ib[i]; 1764 ip->i_ib[i] = 0; 1765 } 1766 ip->i_blocks = 0; 1767 ip->i_size = 0; 1768 /* 1769 * If the file was removed, then the space being freed was 1770 * accounted for then (see softdep_filereleased()). If the 1771 * file is merely being truncated, then we account for it now. 1772 */ 1773 if ((ip->i_flag & IN_SPACECOUNTED) == 0) 1774 fs->fs_pendingblocks += freeblks->fb_chkcnt; 1775 /* 1776 * Push the zero'ed inode to to its disk buffer so that we are free 1777 * to delete its dependencies below. Once the dependencies are gone 1778 * the buffer can be safely released. 1779 */ 1780 if ((error = bread(ip->i_devvp, 1781 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 1782 (int)fs->fs_bsize, NOCRED, &bp)) != 0) 1783 softdep_error("softdep_setup_freeblocks", error); 1784 *((struct dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) = 1785 ip->i_din; 1786 /* 1787 * Find and eliminate any inode dependencies. 1788 */ 1789 ACQUIRE_LOCK(&lk); 1790 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep); 1791 if ((inodedep->id_state & IOSTARTED) != 0) { 1792 FREE_LOCK(&lk); 1793 panic("softdep_setup_freeblocks: inode busy"); 1794 } 1795 /* 1796 * Add the freeblks structure to the list of operations that 1797 * must await the zero'ed inode being written to disk. If we 1798 * still have a bitmap dependency (delay == 0), then the inode 1799 * has never been written to disk, so we can process the 1800 * freeblks below once we have deleted the dependencies. 1801 */ 1802 delay = (inodedep->id_state & DEPCOMPLETE); 1803 if (delay) 1804 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list); 1805 /* 1806 * Because the file length has been truncated to zero, any 1807 * pending block allocation dependency structures associated 1808 * with this inode are obsolete and can simply be de-allocated. 1809 * We must first merge the two dependency lists to get rid of 1810 * any duplicate freefrag structures, then purge the merged list. 1811 * If we still have a bitmap dependency, then the inode has never 1812 * been written to disk, so we can free any fragments without delay. 1813 */ 1814 merge_inode_lists(inodedep); 1815 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 1816 free_allocdirect(&inodedep->id_inoupdt, adp, delay); 1817 FREE_LOCK(&lk); 1818 bdwrite(bp); 1819 /* 1820 * We must wait for any I/O in progress to finish so that 1821 * all potential buffers on the dirty list will be visible. 1822 * Once they are all there, walk the list and get rid of 1823 * any dependencies. 1824 */ 1825 vp = ITOV(ip); 1826 ACQUIRE_LOCK(&lk); 1827 drain_output(vp, 1); 1828 while (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT)) { 1829 bp = TAILQ_FIRST(&vp->v_dirtyblkhd); 1830 (void) inodedep_lookup(fs, ip->i_number, 0, &inodedep); 1831 deallocate_dependencies(bp, inodedep); 1832 bp->b_flags |= B_INVAL | B_NOCACHE; 1833 FREE_LOCK(&lk); 1834 brelse(bp); 1835 ACQUIRE_LOCK(&lk); 1836 } 1837 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0) 1838 (void) free_inodedep(inodedep); 1839 FREE_LOCK(&lk); 1840 /* 1841 * If the inode has never been written to disk (delay == 0), 1842 * then we can process the freeblks now that we have deleted 1843 * the dependencies. 1844 */ 1845 if (!delay) 1846 handle_workitem_freeblocks(freeblks, 0); 1847} 1848 1849/* 1850 * Reclaim any dependency structures from a buffer that is about to 1851 * be reallocated to a new vnode. The buffer must be locked, thus, 1852 * no I/O completion operations can occur while we are manipulating 1853 * its associated dependencies. The mutex is held so that other I/O's 1854 * associated with related dependencies do not occur. 1855 */ 1856static void 1857deallocate_dependencies(bp, inodedep) 1858 struct buf *bp; 1859 struct inodedep *inodedep; 1860{ 1861 struct worklist *wk; 1862 struct indirdep *indirdep; 1863 struct allocindir *aip; 1864 struct pagedep *pagedep; 1865 struct dirrem *dirrem; 1866 struct diradd *dap; 1867 int i; 1868 1869 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 1870 switch (wk->wk_type) { 1871 1872 case D_INDIRDEP: 1873 indirdep = WK_INDIRDEP(wk); 1874 /* 1875 * None of the indirect pointers will ever be visible, 1876 * so they can simply be tossed. GOINGAWAY ensures 1877 * that allocated pointers will be saved in the buffer 1878 * cache until they are freed. Note that they will 1879 * only be able to be found by their physical address 1880 * since the inode mapping the logical address will 1881 * be gone. The save buffer used for the safe copy 1882 * was allocated in setup_allocindir_phase2 using 1883 * the physical address so it could be used for this 1884 * purpose. Hence we swap the safe copy with the real 1885 * copy, allowing the safe copy to be freed and holding 1886 * on to the real copy for later use in indir_trunc. 1887 */ 1888 if (indirdep->ir_state & GOINGAWAY) { 1889 FREE_LOCK(&lk); 1890 panic("deallocate_dependencies: already gone"); 1891 } 1892 indirdep->ir_state |= GOINGAWAY; 1893 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 1894 free_allocindir(aip, inodedep); 1895 if (bp->b_lblkno >= 0 || 1896 bp->b_blkno != indirdep->ir_savebp->b_lblkno) { 1897 FREE_LOCK(&lk); 1898 panic("deallocate_dependencies: not indir"); 1899 } 1900 bcopy(bp->b_data, indirdep->ir_savebp->b_data, 1901 bp->b_bcount); 1902 WORKLIST_REMOVE(wk); 1903 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk); 1904 continue; 1905 1906 case D_PAGEDEP: 1907 pagedep = WK_PAGEDEP(wk); 1908 /* 1909 * None of the directory additions will ever be 1910 * visible, so they can simply be tossed. 1911 */ 1912 for (i = 0; i < DAHASHSZ; i++) 1913 while ((dap = 1914 LIST_FIRST(&pagedep->pd_diraddhd[i]))) 1915 free_diradd(dap); 1916 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0) 1917 free_diradd(dap); 1918 /* 1919 * Copy any directory remove dependencies to the list 1920 * to be processed after the zero'ed inode is written. 1921 * If the inode has already been written, then they 1922 * can be dumped directly onto the work list. 1923 */ 1924 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 1925 LIST_REMOVE(dirrem, dm_next); 1926 dirrem->dm_dirinum = pagedep->pd_ino; 1927 if (inodedep == NULL || 1928 (inodedep->id_state & ALLCOMPLETE) == 1929 ALLCOMPLETE) 1930 add_to_worklist(&dirrem->dm_list); 1931 else 1932 WORKLIST_INSERT(&inodedep->id_bufwait, 1933 &dirrem->dm_list); 1934 } 1935 WORKLIST_REMOVE(&pagedep->pd_list); 1936 LIST_REMOVE(pagedep, pd_hash); 1937 WORKITEM_FREE(pagedep, D_PAGEDEP); 1938 continue; 1939 1940 case D_ALLOCINDIR: 1941 free_allocindir(WK_ALLOCINDIR(wk), inodedep); 1942 continue; 1943 1944 case D_ALLOCDIRECT: 1945 case D_INODEDEP: 1946 FREE_LOCK(&lk); 1947 panic("deallocate_dependencies: Unexpected type %s", 1948 TYPENAME(wk->wk_type)); 1949 /* NOTREACHED */ 1950 1951 default: 1952 FREE_LOCK(&lk); 1953 panic("deallocate_dependencies: Unknown type %s", 1954 TYPENAME(wk->wk_type)); 1955 /* NOTREACHED */ 1956 } 1957 } 1958} 1959 1960/* 1961 * Free an allocdirect. Generate a new freefrag work request if appropriate. 1962 * This routine must be called with splbio interrupts blocked. 1963 */ 1964static void 1965free_allocdirect(adphead, adp, delay) 1966 struct allocdirectlst *adphead; 1967 struct allocdirect *adp; 1968 int delay; 1969{ 1970 struct newdirblk *newdirblk; 1971 struct worklist *wk; 1972 1973#ifdef DEBUG 1974 if (lk.lkt_held == -1) 1975 panic("free_allocdirect: lock not held"); 1976#endif 1977 if ((adp->ad_state & DEPCOMPLETE) == 0) 1978 LIST_REMOVE(adp, ad_deps); 1979 TAILQ_REMOVE(adphead, adp, ad_next); 1980 if ((adp->ad_state & COMPLETE) == 0) 1981 WORKLIST_REMOVE(&adp->ad_list); 1982 if (adp->ad_freefrag != NULL) { 1983 if (delay) 1984 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 1985 &adp->ad_freefrag->ff_list); 1986 else 1987 add_to_worklist(&adp->ad_freefrag->ff_list); 1988 } 1989 if ((wk = LIST_FIRST(&adp->ad_newdirblk)) != NULL) { 1990 newdirblk = WK_NEWDIRBLK(wk); 1991 WORKLIST_REMOVE(&newdirblk->db_list); 1992 if (LIST_FIRST(&adp->ad_newdirblk) != NULL) 1993 panic("free_allocdirect: extra newdirblk"); 1994 if (delay) 1995 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 1996 &newdirblk->db_list); 1997 else 1998 free_newdirblk(newdirblk); 1999 } 2000 WORKITEM_FREE(adp, D_ALLOCDIRECT); 2001} 2002 2003/* 2004 * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep. 2005 * This routine must be called with splbio interrupts blocked. 2006 */ 2007static void 2008free_newdirblk(newdirblk) 2009 struct newdirblk *newdirblk; 2010{ 2011 struct pagedep *pagedep; 2012 struct diradd *dap; 2013 int i; 2014 2015#ifdef DEBUG 2016 if (lk.lkt_held == -1) 2017 panic("free_newdirblk: lock not held"); 2018#endif 2019 /* 2020 * If the pagedep is still linked onto the directory buffer 2021 * dependency chain, then some of the entries on the 2022 * pd_pendinghd list may not be committed to disk yet. In 2023 * this case, we will simply clear the NEWBLOCK flag and 2024 * let the pd_pendinghd list be processed when the pagedep 2025 * is next written. If the pagedep is no longer on the buffer 2026 * dependency chain, then all the entries on the pd_pending 2027 * list are committed to disk and we can free them here. 2028 */ 2029 pagedep = newdirblk->db_pagedep; 2030 pagedep->pd_state &= ~NEWBLOCK; 2031 if ((pagedep->pd_state & ONWORKLIST) == 0) 2032 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 2033 free_diradd(dap); 2034 /* 2035 * If no dependencies remain, the pagedep will be freed. 2036 */ 2037 for (i = 0; i < DAHASHSZ; i++) 2038 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL) 2039 break; 2040 if (i == DAHASHSZ && (pagedep->pd_state & ONWORKLIST) == 0) { 2041 LIST_REMOVE(pagedep, pd_hash); 2042 WORKITEM_FREE(pagedep, D_PAGEDEP); 2043 } 2044 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 2045} 2046 2047/* 2048 * Prepare an inode to be freed. The actual free operation is not 2049 * done until the zero'ed inode has been written to disk. 2050 */ 2051void 2052softdep_freefile(pvp, ino, mode) 2053 struct vnode *pvp; 2054 ino_t ino; 2055 int mode; 2056{ 2057 struct inode *ip = VTOI(pvp); 2058 struct inodedep *inodedep; 2059 struct freefile *freefile; 2060 2061 /* 2062 * This sets up the inode de-allocation dependency. 2063 */ 2064 MALLOC(freefile, struct freefile *, sizeof(struct freefile), 2065 M_FREEFILE, M_SOFTDEP_FLAGS); 2066 freefile->fx_list.wk_type = D_FREEFILE; 2067 freefile->fx_list.wk_state = 0; 2068 freefile->fx_mode = mode; 2069 freefile->fx_oldinum = ino; 2070 freefile->fx_devvp = ip->i_devvp; 2071 freefile->fx_mnt = ITOV(ip)->v_mount; 2072 if ((ip->i_flag & IN_SPACECOUNTED) == 0) 2073 ip->i_fs->fs_pendinginodes += 1; 2074 2075 /* 2076 * If the inodedep does not exist, then the zero'ed inode has 2077 * been written to disk. If the allocated inode has never been 2078 * written to disk, then the on-disk inode is zero'ed. In either 2079 * case we can free the file immediately. 2080 */ 2081 ACQUIRE_LOCK(&lk); 2082 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 || 2083 check_inode_unwritten(inodedep)) { 2084 FREE_LOCK(&lk); 2085 handle_workitem_freefile(freefile); 2086 return; 2087 } 2088 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 2089 FREE_LOCK(&lk); 2090} 2091 2092/* 2093 * Check to see if an inode has never been written to disk. If 2094 * so free the inodedep and return success, otherwise return failure. 2095 * This routine must be called with splbio interrupts blocked. 2096 * 2097 * If we still have a bitmap dependency, then the inode has never 2098 * been written to disk. Drop the dependency as it is no longer 2099 * necessary since the inode is being deallocated. We set the 2100 * ALLCOMPLETE flags since the bitmap now properly shows that the 2101 * inode is not allocated. Even if the inode is actively being 2102 * written, it has been rolled back to its zero'ed state, so we 2103 * are ensured that a zero inode is what is on the disk. For short 2104 * lived files, this change will usually result in removing all the 2105 * dependencies from the inode so that it can be freed immediately. 2106 */ 2107static int 2108check_inode_unwritten(inodedep) 2109 struct inodedep *inodedep; 2110{ 2111 2112 if ((inodedep->id_state & DEPCOMPLETE) != 0 || 2113 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2114 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2115 LIST_FIRST(&inodedep->id_inowait) != NULL || 2116 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2117 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2118 inodedep->id_nlinkdelta != 0) 2119 return (0); 2120 inodedep->id_state |= ALLCOMPLETE; 2121 LIST_REMOVE(inodedep, id_deps); 2122 inodedep->id_buf = NULL; 2123 if (inodedep->id_state & ONWORKLIST) 2124 WORKLIST_REMOVE(&inodedep->id_list); 2125 if (inodedep->id_savedino != NULL) { 2126 FREE(inodedep->id_savedino, M_INODEDEP); 2127 inodedep->id_savedino = NULL; 2128 } 2129 if (free_inodedep(inodedep) == 0) { 2130 FREE_LOCK(&lk); 2131 panic("check_inode_unwritten: busy inode"); 2132 } 2133 return (1); 2134} 2135 2136/* 2137 * Try to free an inodedep structure. Return 1 if it could be freed. 2138 */ 2139static int 2140free_inodedep(inodedep) 2141 struct inodedep *inodedep; 2142{ 2143 2144 if ((inodedep->id_state & ONWORKLIST) != 0 || 2145 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 2146 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2147 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2148 LIST_FIRST(&inodedep->id_inowait) != NULL || 2149 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2150 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2151 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL) 2152 return (0); 2153 LIST_REMOVE(inodedep, id_hash); 2154 WORKITEM_FREE(inodedep, D_INODEDEP); 2155 num_inodedep -= 1; 2156 return (1); 2157} 2158 2159/* 2160 * This workitem routine performs the block de-allocation. 2161 * The workitem is added to the pending list after the updated 2162 * inode block has been written to disk. As mentioned above, 2163 * checks regarding the number of blocks de-allocated (compared 2164 * to the number of blocks allocated for the file) are also 2165 * performed in this function. 2166 */ 2167static void 2168handle_workitem_freeblocks(freeblks, flags) 2169 struct freeblks *freeblks; 2170 int flags; 2171{ 2172 struct inode tip, *ip; 2173 struct vnode *vp; 2174 ufs_daddr_t bn; 2175 struct fs *fs; 2176 int i, level, bsize; 2177 long nblocks, blocksreleased = 0; 2178 int error, allerror = 0; 2179 ufs_lbn_t baselbns[NIADDR], tmpval; 2180 2181 tip.i_fs = fs = VFSTOUFS(freeblks->fb_mnt)->um_fs; 2182 tip.i_number = freeblks->fb_previousinum; 2183 tip.i_devvp = freeblks->fb_devvp; 2184 tip.i_dev = freeblks->fb_devvp->v_rdev; 2185 tip.i_size = freeblks->fb_oldsize; 2186 tip.i_uid = freeblks->fb_uid; 2187 tip.i_vnode = NULL; 2188 tmpval = 1; 2189 baselbns[0] = NDADDR; 2190 for (i = 1; i < NIADDR; i++) { 2191 tmpval *= NINDIR(fs); 2192 baselbns[i] = baselbns[i - 1] + tmpval; 2193 } 2194 nblocks = btodb(fs->fs_bsize); 2195 blocksreleased = 0; 2196 /* 2197 * Indirect blocks first. 2198 */ 2199 for (level = (NIADDR - 1); level >= 0; level--) { 2200 if ((bn = freeblks->fb_iblks[level]) == 0) 2201 continue; 2202 if ((error = indir_trunc(&tip, fsbtodb(fs, bn), level, 2203 baselbns[level], &blocksreleased)) == 0) 2204 allerror = error; 2205 ffs_blkfree(&tip, bn, fs->fs_bsize); 2206 fs->fs_pendingblocks -= nblocks; 2207 blocksreleased += nblocks; 2208 } 2209 /* 2210 * All direct blocks or frags. 2211 */ 2212 for (i = (NDADDR - 1); i >= 0; i--) { 2213 if ((bn = freeblks->fb_dblks[i]) == 0) 2214 continue; 2215 bsize = blksize(fs, &tip, i); 2216 ffs_blkfree(&tip, bn, bsize); 2217 fs->fs_pendingblocks -= btodb(bsize); 2218 blocksreleased += btodb(bsize); 2219 } 2220 /* 2221 * If we still have not finished background cleanup, then check 2222 * to see if the block count needs to be adjusted. 2223 */ 2224 if (freeblks->fb_chkcnt != blocksreleased && 2225 (fs->fs_flags & FS_UNCLEAN) != 0 && (flags & LK_NOWAIT) == 0 && 2226 VFS_VGET(freeblks->fb_mnt, freeblks->fb_previousinum, &vp) == 0) { 2227 ip = VTOI(vp); 2228 ip->i_blocks += freeblks->fb_chkcnt - blocksreleased; 2229 ip->i_flag |= IN_CHANGE; 2230 vput(vp); 2231 } 2232 2233#ifdef DIAGNOSTIC 2234 if (freeblks->fb_chkcnt != blocksreleased && 2235 ((fs->fs_flags & FS_UNCLEAN) == 0 || (flags & LK_NOWAIT) != 0)) 2236 printf("handle_workitem_freeblocks: block count"); 2237 if (allerror) 2238 softdep_error("handle_workitem_freeblks", allerror); 2239#endif /* DIAGNOSTIC */ 2240 2241 WORKITEM_FREE(freeblks, D_FREEBLKS); 2242} 2243 2244/* 2245 * Release blocks associated with the inode ip and stored in the indirect 2246 * block dbn. If level is greater than SINGLE, the block is an indirect block 2247 * and recursive calls to indirtrunc must be used to cleanse other indirect 2248 * blocks. 2249 */ 2250static int 2251indir_trunc(ip, dbn, level, lbn, countp) 2252 struct inode *ip; 2253 ufs_daddr_t dbn; 2254 int level; 2255 ufs_lbn_t lbn; 2256 long *countp; 2257{ 2258 struct buf *bp; 2259 ufs_daddr_t *bap; 2260 ufs_daddr_t nb; 2261 struct fs *fs; 2262 struct worklist *wk; 2263 struct indirdep *indirdep; 2264 int i, lbnadd, nblocks; 2265 int error, allerror = 0; 2266 2267 fs = ip->i_fs; 2268 lbnadd = 1; 2269 for (i = level; i > 0; i--) 2270 lbnadd *= NINDIR(fs); 2271 /* 2272 * Get buffer of block pointers to be freed. This routine is not 2273 * called until the zero'ed inode has been written, so it is safe 2274 * to free blocks as they are encountered. Because the inode has 2275 * been zero'ed, calls to bmap on these blocks will fail. So, we 2276 * have to use the on-disk address and the block device for the 2277 * filesystem to look them up. If the file was deleted before its 2278 * indirect blocks were all written to disk, the routine that set 2279 * us up (deallocate_dependencies) will have arranged to leave 2280 * a complete copy of the indirect block in memory for our use. 2281 * Otherwise we have to read the blocks in from the disk. 2282 */ 2283 ACQUIRE_LOCK(&lk); 2284 if ((bp = incore(ip->i_devvp, dbn)) != NULL && 2285 (wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2286 if (wk->wk_type != D_INDIRDEP || 2287 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp || 2288 (indirdep->ir_state & GOINGAWAY) == 0) { 2289 FREE_LOCK(&lk); 2290 panic("indir_trunc: lost indirdep"); 2291 } 2292 WORKLIST_REMOVE(wk); 2293 WORKITEM_FREE(indirdep, D_INDIRDEP); 2294 if (LIST_FIRST(&bp->b_dep) != NULL) { 2295 FREE_LOCK(&lk); 2296 panic("indir_trunc: dangling dep"); 2297 } 2298 FREE_LOCK(&lk); 2299 } else { 2300 FREE_LOCK(&lk); 2301 error = bread(ip->i_devvp, dbn, (int)fs->fs_bsize, NOCRED, &bp); 2302 if (error) 2303 return (error); 2304 } 2305 /* 2306 * Recursively free indirect blocks. 2307 */ 2308 bap = (ufs_daddr_t *)bp->b_data; 2309 nblocks = btodb(fs->fs_bsize); 2310 for (i = NINDIR(fs) - 1; i >= 0; i--) { 2311 if ((nb = bap[i]) == 0) 2312 continue; 2313 if (level != 0) { 2314 if ((error = indir_trunc(ip, fsbtodb(fs, nb), 2315 level - 1, lbn + (i * lbnadd), countp)) != 0) 2316 allerror = error; 2317 } 2318 ffs_blkfree(ip, nb, fs->fs_bsize); 2319 fs->fs_pendingblocks -= nblocks; 2320 *countp += nblocks; 2321 } 2322 bp->b_flags |= B_INVAL | B_NOCACHE; 2323 brelse(bp); 2324 return (allerror); 2325} 2326 2327/* 2328 * Free an allocindir. 2329 * This routine must be called with splbio interrupts blocked. 2330 */ 2331static void 2332free_allocindir(aip, inodedep) 2333 struct allocindir *aip; 2334 struct inodedep *inodedep; 2335{ 2336 struct freefrag *freefrag; 2337 2338#ifdef DEBUG 2339 if (lk.lkt_held == -1) 2340 panic("free_allocindir: lock not held"); 2341#endif 2342 if ((aip->ai_state & DEPCOMPLETE) == 0) 2343 LIST_REMOVE(aip, ai_deps); 2344 if (aip->ai_state & ONWORKLIST) 2345 WORKLIST_REMOVE(&aip->ai_list); 2346 LIST_REMOVE(aip, ai_next); 2347 if ((freefrag = aip->ai_freefrag) != NULL) { 2348 if (inodedep == NULL) 2349 add_to_worklist(&freefrag->ff_list); 2350 else 2351 WORKLIST_INSERT(&inodedep->id_bufwait, 2352 &freefrag->ff_list); 2353 } 2354 WORKITEM_FREE(aip, D_ALLOCINDIR); 2355} 2356 2357/* 2358 * Directory entry addition dependencies. 2359 * 2360 * When adding a new directory entry, the inode (with its incremented link 2361 * count) must be written to disk before the directory entry's pointer to it. 2362 * Also, if the inode is newly allocated, the corresponding freemap must be 2363 * updated (on disk) before the directory entry's pointer. These requirements 2364 * are met via undo/redo on the directory entry's pointer, which consists 2365 * simply of the inode number. 2366 * 2367 * As directory entries are added and deleted, the free space within a 2368 * directory block can become fragmented. The ufs file system will compact 2369 * a fragmented directory block to make space for a new entry. When this 2370 * occurs, the offsets of previously added entries change. Any "diradd" 2371 * dependency structures corresponding to these entries must be updated with 2372 * the new offsets. 2373 */ 2374 2375/* 2376 * This routine is called after the in-memory inode's link 2377 * count has been incremented, but before the directory entry's 2378 * pointer to the inode has been set. 2379 */ 2380int 2381softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 2382 struct buf *bp; /* buffer containing directory block */ 2383 struct inode *dp; /* inode for directory */ 2384 off_t diroffset; /* offset of new entry in directory */ 2385 long newinum; /* inode referenced by new directory entry */ 2386 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 2387 int isnewblk; /* entry is in a newly allocated block */ 2388{ 2389 int offset; /* offset of new entry within directory block */ 2390 ufs_lbn_t lbn; /* block in directory containing new entry */ 2391 struct fs *fs; 2392 struct diradd *dap; 2393 struct allocdirect *adp; 2394 struct pagedep *pagedep; 2395 struct inodedep *inodedep; 2396 struct newdirblk *newdirblk = 0; 2397 struct mkdir *mkdir1, *mkdir2; 2398 2399 /* 2400 * Whiteouts have no dependencies. 2401 */ 2402 if (newinum == WINO) { 2403 if (newdirbp != NULL) 2404 bdwrite(newdirbp); 2405 return (0); 2406 } 2407 2408 fs = dp->i_fs; 2409 lbn = lblkno(fs, diroffset); 2410 offset = blkoff(fs, diroffset); 2411 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, 2412 M_SOFTDEP_FLAGS|M_ZERO); 2413 dap->da_list.wk_type = D_DIRADD; 2414 dap->da_offset = offset; 2415 dap->da_newinum = newinum; 2416 dap->da_state = ATTACHED; 2417 if (isnewblk && lbn < NDADDR && fragoff(fs, diroffset) == 0) { 2418 MALLOC(newdirblk, struct newdirblk *, sizeof(struct newdirblk), 2419 M_NEWDIRBLK, M_SOFTDEP_FLAGS); 2420 newdirblk->db_list.wk_type = D_NEWDIRBLK; 2421 newdirblk->db_state = 0; 2422 } 2423 if (newdirbp == NULL) { 2424 dap->da_state |= DEPCOMPLETE; 2425 ACQUIRE_LOCK(&lk); 2426 } else { 2427 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 2428 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2429 M_SOFTDEP_FLAGS); 2430 mkdir1->md_list.wk_type = D_MKDIR; 2431 mkdir1->md_state = MKDIR_BODY; 2432 mkdir1->md_diradd = dap; 2433 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2434 M_SOFTDEP_FLAGS); 2435 mkdir2->md_list.wk_type = D_MKDIR; 2436 mkdir2->md_state = MKDIR_PARENT; 2437 mkdir2->md_diradd = dap; 2438 /* 2439 * Dependency on "." and ".." being written to disk. 2440 */ 2441 mkdir1->md_buf = newdirbp; 2442 ACQUIRE_LOCK(&lk); 2443 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 2444 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list); 2445 FREE_LOCK(&lk); 2446 bdwrite(newdirbp); 2447 /* 2448 * Dependency on link count increase for parent directory 2449 */ 2450 ACQUIRE_LOCK(&lk); 2451 if (inodedep_lookup(fs, dp->i_number, 0, &inodedep) == 0 2452 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2453 dap->da_state &= ~MKDIR_PARENT; 2454 WORKITEM_FREE(mkdir2, D_MKDIR); 2455 } else { 2456 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 2457 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 2458 } 2459 } 2460 /* 2461 * Link into parent directory pagedep to await its being written. 2462 */ 2463 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2464 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2465 dap->da_pagedep = pagedep; 2466 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 2467 da_pdlist); 2468 /* 2469 * Link into its inodedep. Put it on the id_bufwait list if the inode 2470 * is not yet written. If it is written, do the post-inode write 2471 * processing to put it on the id_pendinghd list. 2472 */ 2473 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep); 2474 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 2475 diradd_inode_written(dap, inodedep); 2476 else 2477 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2478 if (isnewblk) { 2479 /* 2480 * Directories growing into indirect blocks are rare 2481 * enough and the frequency of new block allocation 2482 * in those cases even more rare, that we choose not 2483 * to bother tracking them. Rather we simply force the 2484 * new directory entry to disk. 2485 */ 2486 if (lbn >= NDADDR) { 2487 FREE_LOCK(&lk); 2488 /* 2489 * We only have a new allocation when at the 2490 * beginning of a new block, not when we are 2491 * expanding into an existing block. 2492 */ 2493 if (blkoff(fs, diroffset) == 0) 2494 return (1); 2495 return (0); 2496 } 2497 /* 2498 * We only have a new allocation when at the beginning 2499 * of a new fragment, not when we are expanding into an 2500 * existing fragment. Also, there is nothing to do if we 2501 * are already tracking this block. 2502 */ 2503 if (fragoff(fs, diroffset) != 0) { 2504 FREE_LOCK(&lk); 2505 return (0); 2506 } 2507 if ((pagedep->pd_state & NEWBLOCK) != 0) { 2508 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 2509 FREE_LOCK(&lk); 2510 return (0); 2511 } 2512 /* 2513 * Find our associated allocdirect and have it track us. 2514 */ 2515 if (inodedep_lookup(fs, dp->i_number, 0, &inodedep) == 0) 2516 panic("softdep_setup_directory_add: lost inodedep"); 2517 adp = TAILQ_LAST(&inodedep->id_newinoupdt, allocdirectlst); 2518 if (adp == NULL || adp->ad_lbn != lbn) { 2519 FREE_LOCK(&lk); 2520 panic("softdep_setup_directory_add: lost entry"); 2521 } 2522 pagedep->pd_state |= NEWBLOCK; 2523 newdirblk->db_pagedep = pagedep; 2524 WORKLIST_INSERT(&adp->ad_newdirblk, &newdirblk->db_list); 2525 } 2526 FREE_LOCK(&lk); 2527 return (0); 2528} 2529 2530/* 2531 * This procedure is called to change the offset of a directory 2532 * entry when compacting a directory block which must be owned 2533 * exclusively by the caller. Note that the actual entry movement 2534 * must be done in this procedure to ensure that no I/O completions 2535 * occur while the move is in progress. 2536 */ 2537void 2538softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize) 2539 struct inode *dp; /* inode for directory */ 2540 caddr_t base; /* address of dp->i_offset */ 2541 caddr_t oldloc; /* address of old directory location */ 2542 caddr_t newloc; /* address of new directory location */ 2543 int entrysize; /* size of directory entry */ 2544{ 2545 int offset, oldoffset, newoffset; 2546 struct pagedep *pagedep; 2547 struct diradd *dap; 2548 ufs_lbn_t lbn; 2549 2550 ACQUIRE_LOCK(&lk); 2551 lbn = lblkno(dp->i_fs, dp->i_offset); 2552 offset = blkoff(dp->i_fs, dp->i_offset); 2553 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) 2554 goto done; 2555 oldoffset = offset + (oldloc - base); 2556 newoffset = offset + (newloc - base); 2557 2558 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) { 2559 if (dap->da_offset != oldoffset) 2560 continue; 2561 dap->da_offset = newoffset; 2562 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) 2563 break; 2564 LIST_REMOVE(dap, da_pdlist); 2565 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], 2566 dap, da_pdlist); 2567 break; 2568 } 2569 if (dap == NULL) { 2570 2571 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) { 2572 if (dap->da_offset == oldoffset) { 2573 dap->da_offset = newoffset; 2574 break; 2575 } 2576 } 2577 } 2578done: 2579 bcopy(oldloc, newloc, entrysize); 2580 FREE_LOCK(&lk); 2581} 2582 2583/* 2584 * Free a diradd dependency structure. This routine must be called 2585 * with splbio interrupts blocked. 2586 */ 2587static void 2588free_diradd(dap) 2589 struct diradd *dap; 2590{ 2591 struct dirrem *dirrem; 2592 struct pagedep *pagedep; 2593 struct inodedep *inodedep; 2594 struct mkdir *mkdir, *nextmd; 2595 2596#ifdef DEBUG 2597 if (lk.lkt_held == -1) 2598 panic("free_diradd: lock not held"); 2599#endif 2600 WORKLIST_REMOVE(&dap->da_list); 2601 LIST_REMOVE(dap, da_pdlist); 2602 if ((dap->da_state & DIRCHG) == 0) { 2603 pagedep = dap->da_pagedep; 2604 } else { 2605 dirrem = dap->da_previous; 2606 pagedep = dirrem->dm_pagedep; 2607 dirrem->dm_dirinum = pagedep->pd_ino; 2608 add_to_worklist(&dirrem->dm_list); 2609 } 2610 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum, 2611 0, &inodedep) != 0) 2612 (void) free_inodedep(inodedep); 2613 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2614 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 2615 nextmd = LIST_NEXT(mkdir, md_mkdirs); 2616 if (mkdir->md_diradd != dap) 2617 continue; 2618 dap->da_state &= ~mkdir->md_state; 2619 WORKLIST_REMOVE(&mkdir->md_list); 2620 LIST_REMOVE(mkdir, md_mkdirs); 2621 WORKITEM_FREE(mkdir, D_MKDIR); 2622 } 2623 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 2624 FREE_LOCK(&lk); 2625 panic("free_diradd: unfound ref"); 2626 } 2627 } 2628 WORKITEM_FREE(dap, D_DIRADD); 2629} 2630 2631/* 2632 * Directory entry removal dependencies. 2633 * 2634 * When removing a directory entry, the entry's inode pointer must be 2635 * zero'ed on disk before the corresponding inode's link count is decremented 2636 * (possibly freeing the inode for re-use). This dependency is handled by 2637 * updating the directory entry but delaying the inode count reduction until 2638 * after the directory block has been written to disk. After this point, the 2639 * inode count can be decremented whenever it is convenient. 2640 */ 2641 2642/* 2643 * This routine should be called immediately after removing 2644 * a directory entry. The inode's link count should not be 2645 * decremented by the calling procedure -- the soft updates 2646 * code will do this task when it is safe. 2647 */ 2648void 2649softdep_setup_remove(bp, dp, ip, isrmdir) 2650 struct buf *bp; /* buffer containing directory block */ 2651 struct inode *dp; /* inode for the directory being modified */ 2652 struct inode *ip; /* inode for directory entry being removed */ 2653 int isrmdir; /* indicates if doing RMDIR */ 2654{ 2655 struct dirrem *dirrem, *prevdirrem; 2656 2657 /* 2658 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 2659 */ 2660 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2661 2662 /* 2663 * If the COMPLETE flag is clear, then there were no active 2664 * entries and we want to roll back to a zeroed entry until 2665 * the new inode is committed to disk. If the COMPLETE flag is 2666 * set then we have deleted an entry that never made it to 2667 * disk. If the entry we deleted resulted from a name change, 2668 * then the old name still resides on disk. We cannot delete 2669 * its inode (returned to us in prevdirrem) until the zeroed 2670 * directory entry gets to disk. The new inode has never been 2671 * referenced on the disk, so can be deleted immediately. 2672 */ 2673 if ((dirrem->dm_state & COMPLETE) == 0) { 2674 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 2675 dm_next); 2676 FREE_LOCK(&lk); 2677 } else { 2678 if (prevdirrem != NULL) 2679 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 2680 prevdirrem, dm_next); 2681 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 2682 FREE_LOCK(&lk); 2683 handle_workitem_remove(dirrem); 2684 } 2685} 2686 2687/* 2688 * Allocate a new dirrem if appropriate and return it along with 2689 * its associated pagedep. Called without a lock, returns with lock. 2690 */ 2691static long num_dirrem; /* number of dirrem allocated */ 2692static struct dirrem * 2693newdirrem(bp, dp, ip, isrmdir, prevdirremp) 2694 struct buf *bp; /* buffer containing directory block */ 2695 struct inode *dp; /* inode for the directory being modified */ 2696 struct inode *ip; /* inode for directory entry being removed */ 2697 int isrmdir; /* indicates if doing RMDIR */ 2698 struct dirrem **prevdirremp; /* previously referenced inode, if any */ 2699{ 2700 int offset; 2701 ufs_lbn_t lbn; 2702 struct diradd *dap; 2703 struct dirrem *dirrem; 2704 struct pagedep *pagedep; 2705 2706 /* 2707 * Whiteouts have no deletion dependencies. 2708 */ 2709 if (ip == NULL) 2710 panic("newdirrem: whiteout"); 2711 /* 2712 * If we are over our limit, try to improve the situation. 2713 * Limiting the number of dirrem structures will also limit 2714 * the number of freefile and freeblks structures. 2715 */ 2716 if (num_dirrem > max_softdeps / 2) 2717 (void) request_cleanup(FLUSH_REMOVE, 0); 2718 num_dirrem += 1; 2719 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem), 2720 M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO); 2721 dirrem->dm_list.wk_type = D_DIRREM; 2722 dirrem->dm_state = isrmdir ? RMDIR : 0; 2723 dirrem->dm_mnt = ITOV(ip)->v_mount; 2724 dirrem->dm_oldinum = ip->i_number; 2725 *prevdirremp = NULL; 2726 2727 ACQUIRE_LOCK(&lk); 2728 lbn = lblkno(dp->i_fs, dp->i_offset); 2729 offset = blkoff(dp->i_fs, dp->i_offset); 2730 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2731 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2732 dirrem->dm_pagedep = pagedep; 2733 /* 2734 * Check for a diradd dependency for the same directory entry. 2735 * If present, then both dependencies become obsolete and can 2736 * be de-allocated. Check for an entry on both the pd_dirraddhd 2737 * list and the pd_pendinghd list. 2738 */ 2739 2740 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 2741 if (dap->da_offset == offset) 2742 break; 2743 if (dap == NULL) { 2744 2745 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 2746 if (dap->da_offset == offset) 2747 break; 2748 if (dap == NULL) 2749 return (dirrem); 2750 } 2751 /* 2752 * Must be ATTACHED at this point. 2753 */ 2754 if ((dap->da_state & ATTACHED) == 0) { 2755 FREE_LOCK(&lk); 2756 panic("newdirrem: not ATTACHED"); 2757 } 2758 if (dap->da_newinum != ip->i_number) { 2759 FREE_LOCK(&lk); 2760 panic("newdirrem: inum %d should be %d", 2761 ip->i_number, dap->da_newinum); 2762 } 2763 /* 2764 * If we are deleting a changed name that never made it to disk, 2765 * then return the dirrem describing the previous inode (which 2766 * represents the inode currently referenced from this entry on disk). 2767 */ 2768 if ((dap->da_state & DIRCHG) != 0) { 2769 *prevdirremp = dap->da_previous; 2770 dap->da_state &= ~DIRCHG; 2771 dap->da_pagedep = pagedep; 2772 } 2773 /* 2774 * We are deleting an entry that never made it to disk. 2775 * Mark it COMPLETE so we can delete its inode immediately. 2776 */ 2777 dirrem->dm_state |= COMPLETE; 2778 free_diradd(dap); 2779 return (dirrem); 2780} 2781 2782/* 2783 * Directory entry change dependencies. 2784 * 2785 * Changing an existing directory entry requires that an add operation 2786 * be completed first followed by a deletion. The semantics for the addition 2787 * are identical to the description of adding a new entry above except 2788 * that the rollback is to the old inode number rather than zero. Once 2789 * the addition dependency is completed, the removal is done as described 2790 * in the removal routine above. 2791 */ 2792 2793/* 2794 * This routine should be called immediately after changing 2795 * a directory entry. The inode's link count should not be 2796 * decremented by the calling procedure -- the soft updates 2797 * code will perform this task when it is safe. 2798 */ 2799void 2800softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 2801 struct buf *bp; /* buffer containing directory block */ 2802 struct inode *dp; /* inode for the directory being modified */ 2803 struct inode *ip; /* inode for directory entry being removed */ 2804 long newinum; /* new inode number for changed entry */ 2805 int isrmdir; /* indicates if doing RMDIR */ 2806{ 2807 int offset; 2808 struct diradd *dap = NULL; 2809 struct dirrem *dirrem, *prevdirrem; 2810 struct pagedep *pagedep; 2811 struct inodedep *inodedep; 2812 2813 offset = blkoff(dp->i_fs, dp->i_offset); 2814 2815 /* 2816 * Whiteouts do not need diradd dependencies. 2817 */ 2818 if (newinum != WINO) { 2819 MALLOC(dap, struct diradd *, sizeof(struct diradd), 2820 M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO); 2821 dap->da_list.wk_type = D_DIRADD; 2822 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 2823 dap->da_offset = offset; 2824 dap->da_newinum = newinum; 2825 } 2826 2827 /* 2828 * Allocate a new dirrem and ACQUIRE_LOCK. 2829 */ 2830 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 2831 pagedep = dirrem->dm_pagedep; 2832 /* 2833 * The possible values for isrmdir: 2834 * 0 - non-directory file rename 2835 * 1 - directory rename within same directory 2836 * inum - directory rename to new directory of given inode number 2837 * When renaming to a new directory, we are both deleting and 2838 * creating a new directory entry, so the link count on the new 2839 * directory should not change. Thus we do not need the followup 2840 * dirrem which is usually done in handle_workitem_remove. We set 2841 * the DIRCHG flag to tell handle_workitem_remove to skip the 2842 * followup dirrem. 2843 */ 2844 if (isrmdir > 1) 2845 dirrem->dm_state |= DIRCHG; 2846 2847 /* 2848 * Whiteouts have no additional dependencies, 2849 * so just put the dirrem on the correct list. 2850 */ 2851 if (newinum == WINO) { 2852 if ((dirrem->dm_state & COMPLETE) == 0) { 2853 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 2854 dm_next); 2855 } else { 2856 dirrem->dm_dirinum = pagedep->pd_ino; 2857 add_to_worklist(&dirrem->dm_list); 2858 } 2859 FREE_LOCK(&lk); 2860 return; 2861 } 2862 2863 /* 2864 * If the COMPLETE flag is clear, then there were no active 2865 * entries and we want to roll back to the previous inode until 2866 * the new inode is committed to disk. If the COMPLETE flag is 2867 * set, then we have deleted an entry that never made it to disk. 2868 * If the entry we deleted resulted from a name change, then the old 2869 * inode reference still resides on disk. Any rollback that we do 2870 * needs to be to that old inode (returned to us in prevdirrem). If 2871 * the entry we deleted resulted from a create, then there is 2872 * no entry on the disk, so we want to roll back to zero rather 2873 * than the uncommitted inode. In either of the COMPLETE cases we 2874 * want to immediately free the unwritten and unreferenced inode. 2875 */ 2876 if ((dirrem->dm_state & COMPLETE) == 0) { 2877 dap->da_previous = dirrem; 2878 } else { 2879 if (prevdirrem != NULL) { 2880 dap->da_previous = prevdirrem; 2881 } else { 2882 dap->da_state &= ~DIRCHG; 2883 dap->da_pagedep = pagedep; 2884 } 2885 dirrem->dm_dirinum = pagedep->pd_ino; 2886 add_to_worklist(&dirrem->dm_list); 2887 } 2888 /* 2889 * Link into its inodedep. Put it on the id_bufwait list if the inode 2890 * is not yet written. If it is written, do the post-inode write 2891 * processing to put it on the id_pendinghd list. 2892 */ 2893 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 || 2894 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2895 dap->da_state |= COMPLETE; 2896 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 2897 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 2898 } else { 2899 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 2900 dap, da_pdlist); 2901 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2902 } 2903 FREE_LOCK(&lk); 2904} 2905 2906/* 2907 * Called whenever the link count on an inode is changed. 2908 * It creates an inode dependency so that the new reference(s) 2909 * to the inode cannot be committed to disk until the updated 2910 * inode has been written. 2911 */ 2912void 2913softdep_change_linkcnt(ip) 2914 struct inode *ip; /* the inode with the increased link count */ 2915{ 2916 struct inodedep *inodedep; 2917 2918 ACQUIRE_LOCK(&lk); 2919 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep); 2920 if (ip->i_nlink < ip->i_effnlink) { 2921 FREE_LOCK(&lk); 2922 panic("softdep_change_linkcnt: bad delta"); 2923 } 2924 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 2925 FREE_LOCK(&lk); 2926} 2927 2928/* 2929 * Called when the effective link count and the reference count 2930 * on an inode drops to zero. At this point there are no names 2931 * referencing the file in the filesystem and no active file 2932 * references. The space associated with the file will be freed 2933 * as soon as the necessary soft dependencies are cleared. 2934 */ 2935void 2936softdep_releasefile(ip) 2937 struct inode *ip; /* inode with the zero effective link count */ 2938{ 2939 struct inodedep *inodedep; 2940 2941 if (ip->i_effnlink > 0) 2942 panic("softdep_filerelease: file still referenced"); 2943 /* 2944 * We may be called several times as the real reference count 2945 * drops to zero. We only want to account for the space once. 2946 */ 2947 if (ip->i_flag & IN_SPACECOUNTED) 2948 return; 2949 /* 2950 * We have to deactivate a snapshot otherwise copyonwrites may 2951 * add blocks and the cleanup may remove blocks after we have 2952 * tried to account for them. 2953 */ 2954 if ((ip->i_flags & SF_SNAPSHOT) != 0) 2955 ffs_snapremove(ITOV(ip)); 2956 /* 2957 * If we are tracking an nlinkdelta, we have to also remember 2958 * whether we accounted for the freed space yet. 2959 */ 2960 ACQUIRE_LOCK(&lk); 2961 if ((inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep))) 2962 inodedep->id_state |= SPACECOUNTED; 2963 FREE_LOCK(&lk); 2964 ip->i_fs->fs_pendingblocks += ip->i_blocks; 2965 ip->i_fs->fs_pendinginodes += 1; 2966 ip->i_flag |= IN_SPACECOUNTED; 2967} 2968 2969/* 2970 * This workitem decrements the inode's link count. 2971 * If the link count reaches zero, the file is removed. 2972 */ 2973static void 2974handle_workitem_remove(dirrem) 2975 struct dirrem *dirrem; 2976{ 2977 struct proc *p = CURPROC; /* XXX */ 2978 struct inodedep *inodedep; 2979 struct vnode *vp; 2980 struct inode *ip; 2981 ino_t oldinum; 2982 int error; 2983 2984 if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) { 2985 softdep_error("handle_workitem_remove: vget", error); 2986 return; 2987 } 2988 ip = VTOI(vp); 2989 ACQUIRE_LOCK(&lk); 2990 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){ 2991 FREE_LOCK(&lk); 2992 panic("handle_workitem_remove: lost inodedep"); 2993 } 2994 /* 2995 * Normal file deletion. 2996 */ 2997 if ((dirrem->dm_state & RMDIR) == 0) { 2998 ip->i_nlink--; 2999 ip->i_flag |= IN_CHANGE; 3000 if (ip->i_nlink < ip->i_effnlink) { 3001 FREE_LOCK(&lk); 3002 panic("handle_workitem_remove: bad file delta"); 3003 } 3004 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3005 FREE_LOCK(&lk); 3006 vput(vp); 3007 num_dirrem -= 1; 3008 WORKITEM_FREE(dirrem, D_DIRREM); 3009 return; 3010 } 3011 /* 3012 * Directory deletion. Decrement reference count for both the 3013 * just deleted parent directory entry and the reference for ".". 3014 * Next truncate the directory to length zero. When the 3015 * truncation completes, arrange to have the reference count on 3016 * the parent decremented to account for the loss of "..". 3017 */ 3018 ip->i_nlink -= 2; 3019 ip->i_flag |= IN_CHANGE; 3020 if (ip->i_nlink < ip->i_effnlink) { 3021 FREE_LOCK(&lk); 3022 panic("handle_workitem_remove: bad dir delta"); 3023 } 3024 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3025 FREE_LOCK(&lk); 3026 if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, p->p_ucred, p)) != 0) 3027 softdep_error("handle_workitem_remove: truncate", error); 3028 /* 3029 * Rename a directory to a new parent. Since, we are both deleting 3030 * and creating a new directory entry, the link count on the new 3031 * directory should not change. Thus we skip the followup dirrem. 3032 */ 3033 if (dirrem->dm_state & DIRCHG) { 3034 vput(vp); 3035 num_dirrem -= 1; 3036 WORKITEM_FREE(dirrem, D_DIRREM); 3037 return; 3038 } 3039 /* 3040 * If the inodedep does not exist, then the zero'ed inode has 3041 * been written to disk. If the allocated inode has never been 3042 * written to disk, then the on-disk inode is zero'ed. In either 3043 * case we can remove the file immediately. 3044 */ 3045 ACQUIRE_LOCK(&lk); 3046 dirrem->dm_state = 0; 3047 oldinum = dirrem->dm_oldinum; 3048 dirrem->dm_oldinum = dirrem->dm_dirinum; 3049 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 || 3050 check_inode_unwritten(inodedep)) { 3051 FREE_LOCK(&lk); 3052 vput(vp); 3053 handle_workitem_remove(dirrem); 3054 return; 3055 } 3056 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 3057 FREE_LOCK(&lk); 3058 vput(vp); 3059} 3060 3061/* 3062 * Inode de-allocation dependencies. 3063 * 3064 * When an inode's link count is reduced to zero, it can be de-allocated. We 3065 * found it convenient to postpone de-allocation until after the inode is 3066 * written to disk with its new link count (zero). At this point, all of the 3067 * on-disk inode's block pointers are nullified and, with careful dependency 3068 * list ordering, all dependencies related to the inode will be satisfied and 3069 * the corresponding dependency structures de-allocated. So, if/when the 3070 * inode is reused, there will be no mixing of old dependencies with new 3071 * ones. This artificial dependency is set up by the block de-allocation 3072 * procedure above (softdep_setup_freeblocks) and completed by the 3073 * following procedure. 3074 */ 3075static void 3076handle_workitem_freefile(freefile) 3077 struct freefile *freefile; 3078{ 3079 struct fs *fs; 3080 struct inode tip; 3081 struct inodedep *idp; 3082 int error; 3083 3084 fs = VFSTOUFS(freefile->fx_mnt)->um_fs; 3085#ifdef DEBUG 3086 ACQUIRE_LOCK(&lk); 3087 error = inodedep_lookup(fs, freefile->fx_oldinum, 0, &idp); 3088 FREE_LOCK(&lk); 3089 if (error) 3090 panic("handle_workitem_freefile: inodedep survived"); 3091#endif 3092 tip.i_devvp = freefile->fx_devvp; 3093 tip.i_dev = freefile->fx_devvp->v_rdev; 3094 tip.i_fs = fs; 3095 fs->fs_pendinginodes -= 1; 3096 if ((error = ffs_freefile(&tip, freefile->fx_oldinum, freefile->fx_mode)) != 0) 3097 softdep_error("handle_workitem_freefile", error); 3098 WORKITEM_FREE(freefile, D_FREEFILE); 3099} 3100 3101/* 3102 * Disk writes. 3103 * 3104 * The dependency structures constructed above are most actively used when file 3105 * system blocks are written to disk. No constraints are placed on when a 3106 * block can be written, but unsatisfied update dependencies are made safe by 3107 * modifying (or replacing) the source memory for the duration of the disk 3108 * write. When the disk write completes, the memory block is again brought 3109 * up-to-date. 3110 * 3111 * In-core inode structure reclamation. 3112 * 3113 * Because there are a finite number of "in-core" inode structures, they are 3114 * reused regularly. By transferring all inode-related dependencies to the 3115 * in-memory inode block and indexing them separately (via "inodedep"s), we 3116 * can allow "in-core" inode structures to be reused at any time and avoid 3117 * any increase in contention. 3118 * 3119 * Called just before entering the device driver to initiate a new disk I/O. 3120 * The buffer must be locked, thus, no I/O completion operations can occur 3121 * while we are manipulating its associated dependencies. 3122 */ 3123static void 3124softdep_disk_io_initiation(bp) 3125 struct buf *bp; /* structure describing disk write to occur */ 3126{ 3127 struct worklist *wk, *nextwk; 3128 struct indirdep *indirdep; 3129 3130 /* 3131 * We only care about write operations. There should never 3132 * be dependencies for reads. 3133 */ 3134 if (bp->b_iocmd == BIO_READ) 3135 panic("softdep_disk_io_initiation: read"); 3136 /* 3137 * Do any necessary pre-I/O processing. 3138 */ 3139 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = nextwk) { 3140 nextwk = LIST_NEXT(wk, wk_list); 3141 switch (wk->wk_type) { 3142 3143 case D_PAGEDEP: 3144 initiate_write_filepage(WK_PAGEDEP(wk), bp); 3145 continue; 3146 3147 case D_INODEDEP: 3148 initiate_write_inodeblock(WK_INODEDEP(wk), bp); 3149 continue; 3150 3151 case D_INDIRDEP: 3152 indirdep = WK_INDIRDEP(wk); 3153 if (indirdep->ir_state & GOINGAWAY) 3154 panic("disk_io_initiation: indirdep gone"); 3155 /* 3156 * If there are no remaining dependencies, this 3157 * will be writing the real pointers, so the 3158 * dependency can be freed. 3159 */ 3160 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) { 3161 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 3162 brelse(indirdep->ir_savebp); 3163 /* inline expand WORKLIST_REMOVE(wk); */ 3164 wk->wk_state &= ~ONWORKLIST; 3165 LIST_REMOVE(wk, wk_list); 3166 WORKITEM_FREE(indirdep, D_INDIRDEP); 3167 continue; 3168 } 3169 /* 3170 * Replace up-to-date version with safe version. 3171 */ 3172 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount, 3173 M_INDIRDEP, M_SOFTDEP_FLAGS); 3174 ACQUIRE_LOCK(&lk); 3175 indirdep->ir_state &= ~ATTACHED; 3176 indirdep->ir_state |= UNDONE; 3177 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 3178 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 3179 bp->b_bcount); 3180 FREE_LOCK(&lk); 3181 continue; 3182 3183 case D_MKDIR: 3184 case D_BMSAFEMAP: 3185 case D_ALLOCDIRECT: 3186 case D_ALLOCINDIR: 3187 continue; 3188 3189 default: 3190 panic("handle_disk_io_initiation: Unexpected type %s", 3191 TYPENAME(wk->wk_type)); 3192 /* NOTREACHED */ 3193 } 3194 } 3195} 3196 3197/* 3198 * Called from within the procedure above to deal with unsatisfied 3199 * allocation dependencies in a directory. The buffer must be locked, 3200 * thus, no I/O completion operations can occur while we are 3201 * manipulating its associated dependencies. 3202 */ 3203static void 3204initiate_write_filepage(pagedep, bp) 3205 struct pagedep *pagedep; 3206 struct buf *bp; 3207{ 3208 struct diradd *dap; 3209 struct direct *ep; 3210 int i; 3211 3212 if (pagedep->pd_state & IOSTARTED) { 3213 /* 3214 * This can only happen if there is a driver that does not 3215 * understand chaining. Here biodone will reissue the call 3216 * to strategy for the incomplete buffers. 3217 */ 3218 printf("initiate_write_filepage: already started\n"); 3219 return; 3220 } 3221 pagedep->pd_state |= IOSTARTED; 3222 ACQUIRE_LOCK(&lk); 3223 for (i = 0; i < DAHASHSZ; i++) { 3224 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 3225 ep = (struct direct *) 3226 ((char *)bp->b_data + dap->da_offset); 3227 if (ep->d_ino != dap->da_newinum) { 3228 FREE_LOCK(&lk); 3229 panic("%s: dir inum %d != new %d", 3230 "initiate_write_filepage", 3231 ep->d_ino, dap->da_newinum); 3232 } 3233 if (dap->da_state & DIRCHG) 3234 ep->d_ino = dap->da_previous->dm_oldinum; 3235 else 3236 ep->d_ino = 0; 3237 dap->da_state &= ~ATTACHED; 3238 dap->da_state |= UNDONE; 3239 } 3240 } 3241 FREE_LOCK(&lk); 3242} 3243 3244/* 3245 * Called from within the procedure above to deal with unsatisfied 3246 * allocation dependencies in an inodeblock. The buffer must be 3247 * locked, thus, no I/O completion operations can occur while we 3248 * are manipulating its associated dependencies. 3249 */ 3250static void 3251initiate_write_inodeblock(inodedep, bp) 3252 struct inodedep *inodedep; 3253 struct buf *bp; /* The inode block */ 3254{ 3255 struct allocdirect *adp, *lastadp; 3256 struct dinode *dp; 3257 struct fs *fs; 3258 ufs_lbn_t prevlbn = 0; 3259 int i, deplist; 3260 3261 if (inodedep->id_state & IOSTARTED) 3262 panic("initiate_write_inodeblock: already started"); 3263 inodedep->id_state |= IOSTARTED; 3264 fs = inodedep->id_fs; 3265 dp = (struct dinode *)bp->b_data + 3266 ino_to_fsbo(fs, inodedep->id_ino); 3267 /* 3268 * If the bitmap is not yet written, then the allocated 3269 * inode cannot be written to disk. 3270 */ 3271 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 3272 if (inodedep->id_savedino != NULL) 3273 panic("initiate_write_inodeblock: already doing I/O"); 3274 MALLOC(inodedep->id_savedino, struct dinode *, 3275 sizeof(struct dinode), M_INODEDEP, M_SOFTDEP_FLAGS); 3276 *inodedep->id_savedino = *dp; 3277 bzero((caddr_t)dp, sizeof(struct dinode)); 3278 return; 3279 } 3280 /* 3281 * If no dependencies, then there is nothing to roll back. 3282 */ 3283 inodedep->id_savedsize = dp->di_size; 3284 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 3285 return; 3286 /* 3287 * Set the dependencies to busy. 3288 */ 3289 ACQUIRE_LOCK(&lk); 3290 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3291 adp = TAILQ_NEXT(adp, ad_next)) { 3292#ifdef DIAGNOSTIC 3293 if (deplist != 0 && prevlbn >= adp->ad_lbn) { 3294 FREE_LOCK(&lk); 3295 panic("softdep_write_inodeblock: lbn order"); 3296 } 3297 prevlbn = adp->ad_lbn; 3298 if (adp->ad_lbn < NDADDR && 3299 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) { 3300 FREE_LOCK(&lk); 3301 panic("%s: direct pointer #%ld mismatch %d != %d", 3302 "softdep_write_inodeblock", adp->ad_lbn, 3303 dp->di_db[adp->ad_lbn], adp->ad_newblkno); 3304 } 3305 if (adp->ad_lbn >= NDADDR && 3306 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) { 3307 FREE_LOCK(&lk); 3308 panic("%s: indirect pointer #%ld mismatch %d != %d", 3309 "softdep_write_inodeblock", adp->ad_lbn - NDADDR, 3310 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno); 3311 } 3312 deplist |= 1 << adp->ad_lbn; 3313 if ((adp->ad_state & ATTACHED) == 0) { 3314 FREE_LOCK(&lk); 3315 panic("softdep_write_inodeblock: Unknown state 0x%x", 3316 adp->ad_state); 3317 } 3318#endif /* DIAGNOSTIC */ 3319 adp->ad_state &= ~ATTACHED; 3320 adp->ad_state |= UNDONE; 3321 } 3322 /* 3323 * The on-disk inode cannot claim to be any larger than the last 3324 * fragment that has been written. Otherwise, the on-disk inode 3325 * might have fragments that were not the last block in the file 3326 * which would corrupt the filesystem. 3327 */ 3328 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3329 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 3330 if (adp->ad_lbn >= NDADDR) 3331 break; 3332 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 3333 /* keep going until hitting a rollback to a frag */ 3334 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 3335 continue; 3336 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 3337 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 3338#ifdef DIAGNOSTIC 3339 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) { 3340 FREE_LOCK(&lk); 3341 panic("softdep_write_inodeblock: lost dep1"); 3342 } 3343#endif /* DIAGNOSTIC */ 3344 dp->di_db[i] = 0; 3345 } 3346 for (i = 0; i < NIADDR; i++) { 3347#ifdef DIAGNOSTIC 3348 if (dp->di_ib[i] != 0 && 3349 (deplist & ((1 << NDADDR) << i)) == 0) { 3350 FREE_LOCK(&lk); 3351 panic("softdep_write_inodeblock: lost dep2"); 3352 } 3353#endif /* DIAGNOSTIC */ 3354 dp->di_ib[i] = 0; 3355 } 3356 FREE_LOCK(&lk); 3357 return; 3358 } 3359 /* 3360 * If we have zero'ed out the last allocated block of the file, 3361 * roll back the size to the last currently allocated block. 3362 * We know that this last allocated block is a full-sized as 3363 * we already checked for fragments in the loop above. 3364 */ 3365 if (lastadp != NULL && 3366 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 3367 for (i = lastadp->ad_lbn; i >= 0; i--) 3368 if (dp->di_db[i] != 0) 3369 break; 3370 dp->di_size = (i + 1) * fs->fs_bsize; 3371 } 3372 /* 3373 * The only dependencies are for indirect blocks. 3374 * 3375 * The file size for indirect block additions is not guaranteed. 3376 * Such a guarantee would be non-trivial to achieve. The conventional 3377 * synchronous write implementation also does not make this guarantee. 3378 * Fsck should catch and fix discrepancies. Arguably, the file size 3379 * can be over-estimated without destroying integrity when the file 3380 * moves into the indirect blocks (i.e., is large). If we want to 3381 * postpone fsck, we are stuck with this argument. 3382 */ 3383 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 3384 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 3385 FREE_LOCK(&lk); 3386} 3387 3388/* 3389 * This routine is called during the completion interrupt 3390 * service routine for a disk write (from the procedure called 3391 * by the device driver to inform the file system caches of 3392 * a request completion). It should be called early in this 3393 * procedure, before the block is made available to other 3394 * processes or other routines are called. 3395 */ 3396static void 3397softdep_disk_write_complete(bp) 3398 struct buf *bp; /* describes the completed disk write */ 3399{ 3400 struct worklist *wk; 3401 struct workhead reattach; 3402 struct newblk *newblk; 3403 struct allocindir *aip; 3404 struct allocdirect *adp; 3405 struct indirdep *indirdep; 3406 struct inodedep *inodedep; 3407 struct bmsafemap *bmsafemap; 3408 3409#ifdef DEBUG 3410 if (lk.lkt_held != -1) 3411 panic("softdep_disk_write_complete: lock is held"); 3412 lk.lkt_held = -2; 3413#endif 3414 LIST_INIT(&reattach); 3415 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 3416 WORKLIST_REMOVE(wk); 3417 switch (wk->wk_type) { 3418 3419 case D_PAGEDEP: 3420 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 3421 WORKLIST_INSERT(&reattach, wk); 3422 continue; 3423 3424 case D_INODEDEP: 3425 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 3426 WORKLIST_INSERT(&reattach, wk); 3427 continue; 3428 3429 case D_BMSAFEMAP: 3430 bmsafemap = WK_BMSAFEMAP(wk); 3431 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) { 3432 newblk->nb_state |= DEPCOMPLETE; 3433 newblk->nb_bmsafemap = NULL; 3434 LIST_REMOVE(newblk, nb_deps); 3435 } 3436 while ((adp = 3437 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) { 3438 adp->ad_state |= DEPCOMPLETE; 3439 adp->ad_buf = NULL; 3440 LIST_REMOVE(adp, ad_deps); 3441 handle_allocdirect_partdone(adp); 3442 } 3443 while ((aip = 3444 LIST_FIRST(&bmsafemap->sm_allocindirhd))) { 3445 aip->ai_state |= DEPCOMPLETE; 3446 aip->ai_buf = NULL; 3447 LIST_REMOVE(aip, ai_deps); 3448 handle_allocindir_partdone(aip); 3449 } 3450 while ((inodedep = 3451 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { 3452 inodedep->id_state |= DEPCOMPLETE; 3453 LIST_REMOVE(inodedep, id_deps); 3454 inodedep->id_buf = NULL; 3455 } 3456 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 3457 continue; 3458 3459 case D_MKDIR: 3460 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 3461 continue; 3462 3463 case D_ALLOCDIRECT: 3464 adp = WK_ALLOCDIRECT(wk); 3465 adp->ad_state |= COMPLETE; 3466 handle_allocdirect_partdone(adp); 3467 continue; 3468 3469 case D_ALLOCINDIR: 3470 aip = WK_ALLOCINDIR(wk); 3471 aip->ai_state |= COMPLETE; 3472 handle_allocindir_partdone(aip); 3473 continue; 3474 3475 case D_INDIRDEP: 3476 indirdep = WK_INDIRDEP(wk); 3477 if (indirdep->ir_state & GOINGAWAY) { 3478 lk.lkt_held = -1; 3479 panic("disk_write_complete: indirdep gone"); 3480 } 3481 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 3482 FREE(indirdep->ir_saveddata, M_INDIRDEP); 3483 indirdep->ir_saveddata = 0; 3484 indirdep->ir_state &= ~UNDONE; 3485 indirdep->ir_state |= ATTACHED; 3486 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 3487 handle_allocindir_partdone(aip); 3488 if (aip == LIST_FIRST(&indirdep->ir_donehd)) { 3489 lk.lkt_held = -1; 3490 panic("disk_write_complete: not gone"); 3491 } 3492 } 3493 WORKLIST_INSERT(&reattach, wk); 3494 if ((bp->b_flags & B_DELWRI) == 0) 3495 stat_indir_blk_ptrs++; 3496 bdirty(bp); 3497 continue; 3498 3499 default: 3500 lk.lkt_held = -1; 3501 panic("handle_disk_write_complete: Unknown type %s", 3502 TYPENAME(wk->wk_type)); 3503 /* NOTREACHED */ 3504 } 3505 } 3506 /* 3507 * Reattach any requests that must be redone. 3508 */ 3509 while ((wk = LIST_FIRST(&reattach)) != NULL) { 3510 WORKLIST_REMOVE(wk); 3511 WORKLIST_INSERT(&bp->b_dep, wk); 3512 } 3513#ifdef DEBUG 3514 if (lk.lkt_held != -2) 3515 panic("softdep_disk_write_complete: lock lost"); 3516 lk.lkt_held = -1; 3517#endif 3518} 3519 3520/* 3521 * Called from within softdep_disk_write_complete above. Note that 3522 * this routine is always called from interrupt level with further 3523 * splbio interrupts blocked. 3524 */ 3525static void 3526handle_allocdirect_partdone(adp) 3527 struct allocdirect *adp; /* the completed allocdirect */ 3528{ 3529 struct allocdirect *listadp; 3530 struct inodedep *inodedep; 3531 long bsize, delay; 3532 3533 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3534 return; 3535 if (adp->ad_buf != NULL) { 3536 lk.lkt_held = -1; 3537 panic("handle_allocdirect_partdone: dangling dep"); 3538 } 3539 /* 3540 * The on-disk inode cannot claim to be any larger than the last 3541 * fragment that has been written. Otherwise, the on-disk inode 3542 * might have fragments that were not the last block in the file 3543 * which would corrupt the filesystem. Thus, we cannot free any 3544 * allocdirects after one whose ad_oldblkno claims a fragment as 3545 * these blocks must be rolled back to zero before writing the inode. 3546 * We check the currently active set of allocdirects in id_inoupdt. 3547 */ 3548 inodedep = adp->ad_inodedep; 3549 bsize = inodedep->id_fs->fs_bsize; 3550 TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) { 3551 /* found our block */ 3552 if (listadp == adp) 3553 break; 3554 /* continue if ad_oldlbn is not a fragment */ 3555 if (listadp->ad_oldsize == 0 || 3556 listadp->ad_oldsize == bsize) 3557 continue; 3558 /* hit a fragment */ 3559 return; 3560 } 3561 /* 3562 * If we have reached the end of the current list without 3563 * finding the just finished dependency, then it must be 3564 * on the future dependency list. Future dependencies cannot 3565 * be freed until they are moved to the current list. 3566 */ 3567 if (listadp == NULL) { 3568#ifdef DEBUG 3569 TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next) 3570 /* found our block */ 3571 if (listadp == adp) 3572 break; 3573 if (listadp == NULL) { 3574 lk.lkt_held = -1; 3575 panic("handle_allocdirect_partdone: lost dep"); 3576 } 3577#endif /* DEBUG */ 3578 return; 3579 } 3580 /* 3581 * If we have found the just finished dependency, then free 3582 * it along with anything that follows it that is complete. 3583 * If the inode still has a bitmap dependency, then it has 3584 * never been written to disk, hence the on-disk inode cannot 3585 * reference the old fragment so we can free it without delay. 3586 */ 3587 delay = (inodedep->id_state & DEPCOMPLETE); 3588 for (; adp; adp = listadp) { 3589 listadp = TAILQ_NEXT(adp, ad_next); 3590 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 3591 return; 3592 free_allocdirect(&inodedep->id_inoupdt, adp, delay); 3593 } 3594} 3595 3596/* 3597 * Called from within softdep_disk_write_complete above. Note that 3598 * this routine is always called from interrupt level with further 3599 * splbio interrupts blocked. 3600 */ 3601static void 3602handle_allocindir_partdone(aip) 3603 struct allocindir *aip; /* the completed allocindir */ 3604{ 3605 struct indirdep *indirdep; 3606 3607 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 3608 return; 3609 if (aip->ai_buf != NULL) { 3610 lk.lkt_held = -1; 3611 panic("handle_allocindir_partdone: dangling dependency"); 3612 } 3613 indirdep = aip->ai_indirdep; 3614 if (indirdep->ir_state & UNDONE) { 3615 LIST_REMOVE(aip, ai_next); 3616 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 3617 return; 3618 } 3619 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 3620 aip->ai_newblkno; 3621 LIST_REMOVE(aip, ai_next); 3622 if (aip->ai_freefrag != NULL) 3623 add_to_worklist(&aip->ai_freefrag->ff_list); 3624 WORKITEM_FREE(aip, D_ALLOCINDIR); 3625} 3626 3627/* 3628 * Called from within softdep_disk_write_complete above to restore 3629 * in-memory inode block contents to their most up-to-date state. Note 3630 * that this routine is always called from interrupt level with further 3631 * splbio interrupts blocked. 3632 */ 3633static int 3634handle_written_inodeblock(inodedep, bp) 3635 struct inodedep *inodedep; 3636 struct buf *bp; /* buffer containing the inode block */ 3637{ 3638 struct worklist *wk, *filefree; 3639 struct allocdirect *adp, *nextadp; 3640 struct dinode *dp; 3641 int hadchanges; 3642 3643 if ((inodedep->id_state & IOSTARTED) == 0) { 3644 lk.lkt_held = -1; 3645 panic("handle_written_inodeblock: not started"); 3646 } 3647 inodedep->id_state &= ~IOSTARTED; 3648 inodedep->id_state |= COMPLETE; 3649 dp = (struct dinode *)bp->b_data + 3650 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 3651 /* 3652 * If we had to rollback the inode allocation because of 3653 * bitmaps being incomplete, then simply restore it. 3654 * Keep the block dirty so that it will not be reclaimed until 3655 * all associated dependencies have been cleared and the 3656 * corresponding updates written to disk. 3657 */ 3658 if (inodedep->id_savedino != NULL) { 3659 *dp = *inodedep->id_savedino; 3660 FREE(inodedep->id_savedino, M_INODEDEP); 3661 inodedep->id_savedino = NULL; 3662 if ((bp->b_flags & B_DELWRI) == 0) 3663 stat_inode_bitmap++; 3664 bdirty(bp); 3665 return (1); 3666 } 3667 /* 3668 * Roll forward anything that had to be rolled back before 3669 * the inode could be updated. 3670 */ 3671 hadchanges = 0; 3672 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 3673 nextadp = TAILQ_NEXT(adp, ad_next); 3674 if (adp->ad_state & ATTACHED) { 3675 lk.lkt_held = -1; 3676 panic("handle_written_inodeblock: new entry"); 3677 } 3678 if (adp->ad_lbn < NDADDR) { 3679 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) { 3680 lk.lkt_held = -1; 3681 panic("%s: %s #%ld mismatch %d != %d", 3682 "handle_written_inodeblock", 3683 "direct pointer", adp->ad_lbn, 3684 dp->di_db[adp->ad_lbn], adp->ad_oldblkno); 3685 } 3686 dp->di_db[adp->ad_lbn] = adp->ad_newblkno; 3687 } else { 3688 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) { 3689 lk.lkt_held = -1; 3690 panic("%s: %s #%ld allocated as %d", 3691 "handle_written_inodeblock", 3692 "indirect pointer", adp->ad_lbn - NDADDR, 3693 dp->di_ib[adp->ad_lbn - NDADDR]); 3694 } 3695 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno; 3696 } 3697 adp->ad_state &= ~UNDONE; 3698 adp->ad_state |= ATTACHED; 3699 hadchanges = 1; 3700 } 3701 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 3702 stat_direct_blk_ptrs++; 3703 /* 3704 * Reset the file size to its most up-to-date value. 3705 */ 3706 if (inodedep->id_savedsize == -1) { 3707 lk.lkt_held = -1; 3708 panic("handle_written_inodeblock: bad size"); 3709 } 3710 if (dp->di_size != inodedep->id_savedsize) { 3711 dp->di_size = inodedep->id_savedsize; 3712 hadchanges = 1; 3713 } 3714 inodedep->id_savedsize = -1; 3715 /* 3716 * If there were any rollbacks in the inode block, then it must be 3717 * marked dirty so that its will eventually get written back in 3718 * its correct form. 3719 */ 3720 if (hadchanges) 3721 bdirty(bp); 3722 /* 3723 * Process any allocdirects that completed during the update. 3724 */ 3725 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 3726 handle_allocdirect_partdone(adp); 3727 /* 3728 * Process deallocations that were held pending until the 3729 * inode had been written to disk. Freeing of the inode 3730 * is delayed until after all blocks have been freed to 3731 * avoid creation of new <vfsid, inum, lbn> triples 3732 * before the old ones have been deleted. 3733 */ 3734 filefree = NULL; 3735 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 3736 WORKLIST_REMOVE(wk); 3737 switch (wk->wk_type) { 3738 3739 case D_FREEFILE: 3740 /* 3741 * We defer adding filefree to the worklist until 3742 * all other additions have been made to ensure 3743 * that it will be done after all the old blocks 3744 * have been freed. 3745 */ 3746 if (filefree != NULL) { 3747 lk.lkt_held = -1; 3748 panic("handle_written_inodeblock: filefree"); 3749 } 3750 filefree = wk; 3751 continue; 3752 3753 case D_MKDIR: 3754 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 3755 continue; 3756 3757 case D_DIRADD: 3758 diradd_inode_written(WK_DIRADD(wk), inodedep); 3759 continue; 3760 3761 case D_FREEBLKS: 3762 case D_FREEFRAG: 3763 case D_DIRREM: 3764 add_to_worklist(wk); 3765 continue; 3766 3767 case D_NEWDIRBLK: 3768 free_newdirblk(WK_NEWDIRBLK(wk)); 3769 continue; 3770 3771 default: 3772 lk.lkt_held = -1; 3773 panic("handle_written_inodeblock: Unknown type %s", 3774 TYPENAME(wk->wk_type)); 3775 /* NOTREACHED */ 3776 } 3777 } 3778 if (filefree != NULL) { 3779 if (free_inodedep(inodedep) == 0) { 3780 lk.lkt_held = -1; 3781 panic("handle_written_inodeblock: live inodedep"); 3782 } 3783 add_to_worklist(filefree); 3784 return (0); 3785 } 3786 3787 /* 3788 * If no outstanding dependencies, free it. 3789 */ 3790 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0) 3791 return (0); 3792 return (hadchanges); 3793} 3794 3795/* 3796 * Process a diradd entry after its dependent inode has been written. 3797 * This routine must be called with splbio interrupts blocked. 3798 */ 3799static void 3800diradd_inode_written(dap, inodedep) 3801 struct diradd *dap; 3802 struct inodedep *inodedep; 3803{ 3804 struct pagedep *pagedep; 3805 3806 dap->da_state |= COMPLETE; 3807 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3808 if (dap->da_state & DIRCHG) 3809 pagedep = dap->da_previous->dm_pagedep; 3810 else 3811 pagedep = dap->da_pagedep; 3812 LIST_REMOVE(dap, da_pdlist); 3813 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3814 } 3815 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 3816} 3817 3818/* 3819 * Handle the completion of a mkdir dependency. 3820 */ 3821static void 3822handle_written_mkdir(mkdir, type) 3823 struct mkdir *mkdir; 3824 int type; 3825{ 3826 struct diradd *dap; 3827 struct pagedep *pagedep; 3828 3829 if (mkdir->md_state != type) { 3830 lk.lkt_held = -1; 3831 panic("handle_written_mkdir: bad type"); 3832 } 3833 dap = mkdir->md_diradd; 3834 dap->da_state &= ~type; 3835 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 3836 dap->da_state |= DEPCOMPLETE; 3837 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3838 if (dap->da_state & DIRCHG) 3839 pagedep = dap->da_previous->dm_pagedep; 3840 else 3841 pagedep = dap->da_pagedep; 3842 LIST_REMOVE(dap, da_pdlist); 3843 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3844 } 3845 LIST_REMOVE(mkdir, md_mkdirs); 3846 WORKITEM_FREE(mkdir, D_MKDIR); 3847} 3848 3849/* 3850 * Called from within softdep_disk_write_complete above. 3851 * A write operation was just completed. Removed inodes can 3852 * now be freed and associated block pointers may be committed. 3853 * Note that this routine is always called from interrupt level 3854 * with further splbio interrupts blocked. 3855 */ 3856static int 3857handle_written_filepage(pagedep, bp) 3858 struct pagedep *pagedep; 3859 struct buf *bp; /* buffer containing the written page */ 3860{ 3861 struct dirrem *dirrem; 3862 struct diradd *dap, *nextdap; 3863 struct direct *ep; 3864 int i, chgs; 3865 3866 if ((pagedep->pd_state & IOSTARTED) == 0) { 3867 lk.lkt_held = -1; 3868 panic("handle_written_filepage: not started"); 3869 } 3870 pagedep->pd_state &= ~IOSTARTED; 3871 /* 3872 * Process any directory removals that have been committed. 3873 */ 3874 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 3875 LIST_REMOVE(dirrem, dm_next); 3876 dirrem->dm_dirinum = pagedep->pd_ino; 3877 add_to_worklist(&dirrem->dm_list); 3878 } 3879 /* 3880 * Free any directory additions that have been committed. 3881 * If it is a newly allocated block, we have to wait until 3882 * the on-disk directory inode claims the new block. 3883 */ 3884 if ((pagedep->pd_state & NEWBLOCK) == 0) 3885 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 3886 free_diradd(dap); 3887 /* 3888 * Uncommitted directory entries must be restored. 3889 */ 3890 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 3891 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 3892 dap = nextdap) { 3893 nextdap = LIST_NEXT(dap, da_pdlist); 3894 if (dap->da_state & ATTACHED) { 3895 lk.lkt_held = -1; 3896 panic("handle_written_filepage: attached"); 3897 } 3898 ep = (struct direct *) 3899 ((char *)bp->b_data + dap->da_offset); 3900 ep->d_ino = dap->da_newinum; 3901 dap->da_state &= ~UNDONE; 3902 dap->da_state |= ATTACHED; 3903 chgs = 1; 3904 /* 3905 * If the inode referenced by the directory has 3906 * been written out, then the dependency can be 3907 * moved to the pending list. 3908 */ 3909 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 3910 LIST_REMOVE(dap, da_pdlist); 3911 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 3912 da_pdlist); 3913 } 3914 } 3915 } 3916 /* 3917 * If there were any rollbacks in the directory, then it must be 3918 * marked dirty so that its will eventually get written back in 3919 * its correct form. 3920 */ 3921 if (chgs) { 3922 if ((bp->b_flags & B_DELWRI) == 0) 3923 stat_dir_entry++; 3924 bdirty(bp); 3925 return (1); 3926 } 3927 /* 3928 * If no dependencies remain, the pagedep will be freed. 3929 * Otherwise it will remain to update the page before it 3930 * is written back to disk. 3931 */ 3932 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) { 3933 LIST_REMOVE(pagedep, pd_hash); 3934 WORKITEM_FREE(pagedep, D_PAGEDEP); 3935 } 3936 return (0); 3937} 3938 3939/* 3940 * Writing back in-core inode structures. 3941 * 3942 * The file system only accesses an inode's contents when it occupies an 3943 * "in-core" inode structure. These "in-core" structures are separate from 3944 * the page frames used to cache inode blocks. Only the latter are 3945 * transferred to/from the disk. So, when the updated contents of the 3946 * "in-core" inode structure are copied to the corresponding in-memory inode 3947 * block, the dependencies are also transferred. The following procedure is 3948 * called when copying a dirty "in-core" inode to a cached inode block. 3949 */ 3950 3951/* 3952 * Called when an inode is loaded from disk. If the effective link count 3953 * differed from the actual link count when it was last flushed, then we 3954 * need to ensure that the correct effective link count is put back. 3955 */ 3956void 3957softdep_load_inodeblock(ip) 3958 struct inode *ip; /* the "in_core" copy of the inode */ 3959{ 3960 struct inodedep *inodedep; 3961 3962 /* 3963 * Check for alternate nlink count. 3964 */ 3965 ip->i_effnlink = ip->i_nlink; 3966 ACQUIRE_LOCK(&lk); 3967 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 3968 FREE_LOCK(&lk); 3969 return; 3970 } 3971 ip->i_effnlink -= inodedep->id_nlinkdelta; 3972 if (inodedep->id_state & SPACECOUNTED) 3973 ip->i_flag |= IN_SPACECOUNTED; 3974 FREE_LOCK(&lk); 3975} 3976 3977/* 3978 * This routine is called just before the "in-core" inode 3979 * information is to be copied to the in-memory inode block. 3980 * Recall that an inode block contains several inodes. If 3981 * the force flag is set, then the dependencies will be 3982 * cleared so that the update can always be made. Note that 3983 * the buffer is locked when this routine is called, so we 3984 * will never be in the middle of writing the inode block 3985 * to disk. 3986 */ 3987void 3988softdep_update_inodeblock(ip, bp, waitfor) 3989 struct inode *ip; /* the "in_core" copy of the inode */ 3990 struct buf *bp; /* the buffer containing the inode block */ 3991 int waitfor; /* nonzero => update must be allowed */ 3992{ 3993 struct inodedep *inodedep; 3994 struct worklist *wk; 3995 int error, gotit; 3996 3997 /* 3998 * If the effective link count is not equal to the actual link 3999 * count, then we must track the difference in an inodedep while 4000 * the inode is (potentially) tossed out of the cache. Otherwise, 4001 * if there is no existing inodedep, then there are no dependencies 4002 * to track. 4003 */ 4004 ACQUIRE_LOCK(&lk); 4005 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) { 4006 FREE_LOCK(&lk); 4007 if (ip->i_effnlink != ip->i_nlink) 4008 panic("softdep_update_inodeblock: bad link count"); 4009 return; 4010 } 4011 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) { 4012 FREE_LOCK(&lk); 4013 panic("softdep_update_inodeblock: bad delta"); 4014 } 4015 /* 4016 * Changes have been initiated. Anything depending on these 4017 * changes cannot occur until this inode has been written. 4018 */ 4019 inodedep->id_state &= ~COMPLETE; 4020 if ((inodedep->id_state & ONWORKLIST) == 0) 4021 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 4022 /* 4023 * Any new dependencies associated with the incore inode must 4024 * now be moved to the list associated with the buffer holding 4025 * the in-memory copy of the inode. Once merged process any 4026 * allocdirects that are completed by the merger. 4027 */ 4028 merge_inode_lists(inodedep); 4029 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL) 4030 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); 4031 /* 4032 * Now that the inode has been pushed into the buffer, the 4033 * operations dependent on the inode being written to disk 4034 * can be moved to the id_bufwait so that they will be 4035 * processed when the buffer I/O completes. 4036 */ 4037 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 4038 WORKLIST_REMOVE(wk); 4039 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 4040 } 4041 /* 4042 * Newly allocated inodes cannot be written until the bitmap 4043 * that allocates them have been written (indicated by 4044 * DEPCOMPLETE being set in id_state). If we are doing a 4045 * forced sync (e.g., an fsync on a file), we force the bitmap 4046 * to be written so that the update can be done. 4047 */ 4048 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) { 4049 FREE_LOCK(&lk); 4050 return; 4051 } 4052 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 4053 FREE_LOCK(&lk); 4054 if (gotit && 4055 (error = BUF_WRITE(inodedep->id_buf)) != 0) 4056 softdep_error("softdep_update_inodeblock: bwrite", error); 4057 if ((inodedep->id_state & DEPCOMPLETE) == 0) 4058 panic("softdep_update_inodeblock: update failed"); 4059} 4060 4061/* 4062 * Merge the new inode dependency list (id_newinoupdt) into the old 4063 * inode dependency list (id_inoupdt). This routine must be called 4064 * with splbio interrupts blocked. 4065 */ 4066static void 4067merge_inode_lists(inodedep) 4068 struct inodedep *inodedep; 4069{ 4070 struct allocdirect *listadp, *newadp; 4071 4072 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 4073 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) { 4074 if (listadp->ad_lbn < newadp->ad_lbn) { 4075 listadp = TAILQ_NEXT(listadp, ad_next); 4076 continue; 4077 } 4078 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 4079 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 4080 if (listadp->ad_lbn == newadp->ad_lbn) { 4081 allocdirect_merge(&inodedep->id_inoupdt, newadp, 4082 listadp); 4083 listadp = newadp; 4084 } 4085 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt); 4086 } 4087 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) { 4088 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next); 4089 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next); 4090 } 4091} 4092 4093/* 4094 * If we are doing an fsync, then we must ensure that any directory 4095 * entries for the inode have been written after the inode gets to disk. 4096 */ 4097int 4098softdep_fsync(vp) 4099 struct vnode *vp; /* the "in_core" copy of the inode */ 4100{ 4101 struct inodedep *inodedep; 4102 struct pagedep *pagedep; 4103 struct worklist *wk; 4104 struct diradd *dap; 4105 struct mount *mnt; 4106 struct vnode *pvp; 4107 struct inode *ip; 4108 struct buf *bp; 4109 struct fs *fs; 4110 struct proc *p = CURPROC; /* XXX */ 4111 int error, flushparent; 4112 ino_t parentino; 4113 ufs_lbn_t lbn; 4114 4115 ip = VTOI(vp); 4116 fs = ip->i_fs; 4117 ACQUIRE_LOCK(&lk); 4118 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) { 4119 FREE_LOCK(&lk); 4120 return (0); 4121 } 4122 if (LIST_FIRST(&inodedep->id_inowait) != NULL || 4123 LIST_FIRST(&inodedep->id_bufwait) != NULL || 4124 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 4125 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) { 4126 FREE_LOCK(&lk); 4127 panic("softdep_fsync: pending ops"); 4128 } 4129 for (error = 0, flushparent = 0; ; ) { 4130 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 4131 break; 4132 if (wk->wk_type != D_DIRADD) { 4133 FREE_LOCK(&lk); 4134 panic("softdep_fsync: Unexpected type %s", 4135 TYPENAME(wk->wk_type)); 4136 } 4137 dap = WK_DIRADD(wk); 4138 /* 4139 * Flush our parent if this directory entry has a MKDIR_PARENT 4140 * dependency or is contained in a newly allocated block. 4141 */ 4142 if (dap->da_state & DIRCHG) 4143 pagedep = dap->da_previous->dm_pagedep; 4144 else 4145 pagedep = dap->da_pagedep; 4146 mnt = pagedep->pd_mnt; 4147 parentino = pagedep->pd_ino; 4148 lbn = pagedep->pd_lbn; 4149 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) { 4150 FREE_LOCK(&lk); 4151 panic("softdep_fsync: dirty"); 4152 } 4153 if ((dap->da_state & MKDIR_PARENT) || 4154 (pagedep->pd_state & NEWBLOCK)) 4155 flushparent = 1; 4156 else 4157 flushparent = 0; 4158 /* 4159 * If we are being fsync'ed as part of vgone'ing this vnode, 4160 * then we will not be able to release and recover the 4161 * vnode below, so we just have to give up on writing its 4162 * directory entry out. It will eventually be written, just 4163 * not now, but then the user was not asking to have it 4164 * written, so we are not breaking any promises. 4165 */ 4166 if (vp->v_flag & VXLOCK) 4167 break; 4168 /* 4169 * We prevent deadlock by always fetching inodes from the 4170 * root, moving down the directory tree. Thus, when fetching 4171 * our parent directory, we must unlock ourselves before 4172 * requesting the lock on our parent. See the comment in 4173 * ufs_lookup for details on possible races. 4174 */ 4175 FREE_LOCK(&lk); 4176 VOP_UNLOCK(vp, 0, p); 4177 error = VFS_VGET(mnt, parentino, &pvp); 4178 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, p); 4179 if (error != 0) 4180 return (error); 4181 /* 4182 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps 4183 * that are contained in direct blocks will be resolved by 4184 * doing a UFS_UPDATE. Pagedeps contained in indirect blocks 4185 * may require a complete sync'ing of the directory. So, we 4186 * try the cheap and fast UFS_UPDATE first, and if that fails, 4187 * then we do the slower VOP_FSYNC of the directory. 4188 */ 4189 if (flushparent) { 4190 if ((error = UFS_UPDATE(pvp, 1)) != 0) { 4191 vput(pvp); 4192 return (error); 4193 } 4194 if ((pagedep->pd_state & NEWBLOCK) && 4195 (error = VOP_FSYNC(pvp, p->p_ucred, MNT_WAIT, p))) { 4196 vput(pvp); 4197 return (error); 4198 } 4199 } 4200 /* 4201 * Flush directory page containing the inode's name. 4202 */ 4203 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), p->p_ucred, 4204 &bp); 4205 if (error == 0) 4206 error = BUF_WRITE(bp); 4207 vput(pvp); 4208 if (error != 0) 4209 return (error); 4210 ACQUIRE_LOCK(&lk); 4211 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) 4212 break; 4213 } 4214 FREE_LOCK(&lk); 4215 return (0); 4216} 4217 4218/* 4219 * Flush all the dirty bitmaps associated with the block device 4220 * before flushing the rest of the dirty blocks so as to reduce 4221 * the number of dependencies that will have to be rolled back. 4222 */ 4223void 4224softdep_fsync_mountdev(vp) 4225 struct vnode *vp; 4226{ 4227 struct buf *bp, *nbp; 4228 struct worklist *wk; 4229 4230 if (!vn_isdisk(vp, NULL)) 4231 panic("softdep_fsync_mountdev: vnode not a disk"); 4232 ACQUIRE_LOCK(&lk); 4233 for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) { 4234 nbp = TAILQ_NEXT(bp, b_vnbufs); 4235 /* 4236 * If it is already scheduled, skip to the next buffer. 4237 */ 4238 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) 4239 continue; 4240 if ((bp->b_flags & B_DELWRI) == 0) { 4241 FREE_LOCK(&lk); 4242 panic("softdep_fsync_mountdev: not dirty"); 4243 } 4244 /* 4245 * We are only interested in bitmaps with outstanding 4246 * dependencies. 4247 */ 4248 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 4249 wk->wk_type != D_BMSAFEMAP || 4250 (bp->b_xflags & BX_BKGRDINPROG)) { 4251 BUF_UNLOCK(bp); 4252 continue; 4253 } 4254 bremfree(bp); 4255 FREE_LOCK(&lk); 4256 (void) bawrite(bp); 4257 ACQUIRE_LOCK(&lk); 4258 /* 4259 * Since we may have slept during the I/O, we need 4260 * to start from a known point. 4261 */ 4262 nbp = TAILQ_FIRST(&vp->v_dirtyblkhd); 4263 } 4264 drain_output(vp, 1); 4265 FREE_LOCK(&lk); 4266} 4267 4268/* 4269 * This routine is called when we are trying to synchronously flush a 4270 * file. This routine must eliminate any filesystem metadata dependencies 4271 * so that the syncing routine can succeed by pushing the dirty blocks 4272 * associated with the file. If any I/O errors occur, they are returned. 4273 */ 4274int 4275softdep_sync_metadata(ap) 4276 struct vop_fsync_args /* { 4277 struct vnode *a_vp; 4278 struct ucred *a_cred; 4279 int a_waitfor; 4280 struct proc *a_p; 4281 } */ *ap; 4282{ 4283 struct vnode *vp = ap->a_vp; 4284 struct pagedep *pagedep; 4285 struct allocdirect *adp; 4286 struct allocindir *aip; 4287 struct buf *bp, *nbp; 4288 struct worklist *wk; 4289 int i, error, waitfor; 4290 4291 /* 4292 * Check whether this vnode is involved in a filesystem 4293 * that is doing soft dependency processing. 4294 */ 4295 if (!vn_isdisk(vp, NULL)) { 4296 if (!DOINGSOFTDEP(vp)) 4297 return (0); 4298 } else 4299 if (vp->v_rdev->si_mountpoint == NULL || 4300 (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0) 4301 return (0); 4302 /* 4303 * Ensure that any direct block dependencies have been cleared. 4304 */ 4305 ACQUIRE_LOCK(&lk); 4306 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) { 4307 FREE_LOCK(&lk); 4308 return (error); 4309 } 4310 /* 4311 * For most files, the only metadata dependencies are the 4312 * cylinder group maps that allocate their inode or blocks. 4313 * The block allocation dependencies can be found by traversing 4314 * the dependency lists for any buffers that remain on their 4315 * dirty buffer list. The inode allocation dependency will 4316 * be resolved when the inode is updated with MNT_WAIT. 4317 * This work is done in two passes. The first pass grabs most 4318 * of the buffers and begins asynchronously writing them. The 4319 * only way to wait for these asynchronous writes is to sleep 4320 * on the filesystem vnode which may stay busy for a long time 4321 * if the filesystem is active. So, instead, we make a second 4322 * pass over the dependencies blocking on each write. In the 4323 * usual case we will be blocking against a write that we 4324 * initiated, so when it is done the dependency will have been 4325 * resolved. Thus the second pass is expected to end quickly. 4326 */ 4327 waitfor = MNT_NOWAIT; 4328top: 4329 if (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT) == 0) { 4330 FREE_LOCK(&lk); 4331 return (0); 4332 } 4333 bp = TAILQ_FIRST(&vp->v_dirtyblkhd); 4334 /* While syncing snapshots, we must allow recursive lookups */ 4335 bp->b_lock.lk_flags |= LK_CANRECURSE; 4336loop: 4337 /* 4338 * As we hold the buffer locked, none of its dependencies 4339 * will disappear. 4340 */ 4341 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 4342 switch (wk->wk_type) { 4343 4344 case D_ALLOCDIRECT: 4345 adp = WK_ALLOCDIRECT(wk); 4346 if (adp->ad_state & DEPCOMPLETE) 4347 continue; 4348 nbp = adp->ad_buf; 4349 if (getdirtybuf(&nbp, waitfor) == 0) 4350 continue; 4351 FREE_LOCK(&lk); 4352 if (waitfor == MNT_NOWAIT) { 4353 bawrite(nbp); 4354 } else if ((error = BUF_WRITE(nbp)) != 0) { 4355 break; 4356 } 4357 ACQUIRE_LOCK(&lk); 4358 continue; 4359 4360 case D_ALLOCINDIR: 4361 aip = WK_ALLOCINDIR(wk); 4362 if (aip->ai_state & DEPCOMPLETE) 4363 continue; 4364 nbp = aip->ai_buf; 4365 if (getdirtybuf(&nbp, waitfor) == 0) 4366 continue; 4367 FREE_LOCK(&lk); 4368 if (waitfor == MNT_NOWAIT) { 4369 bawrite(nbp); 4370 } else if ((error = BUF_WRITE(nbp)) != 0) { 4371 break; 4372 } 4373 ACQUIRE_LOCK(&lk); 4374 continue; 4375 4376 case D_INDIRDEP: 4377 restart: 4378 4379 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) { 4380 if (aip->ai_state & DEPCOMPLETE) 4381 continue; 4382 nbp = aip->ai_buf; 4383 if (getdirtybuf(&nbp, MNT_WAIT) == 0) 4384 goto restart; 4385 FREE_LOCK(&lk); 4386 if ((error = BUF_WRITE(nbp)) != 0) { 4387 break; 4388 } 4389 ACQUIRE_LOCK(&lk); 4390 goto restart; 4391 } 4392 continue; 4393 4394 case D_INODEDEP: 4395 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs, 4396 WK_INODEDEP(wk)->id_ino)) != 0) { 4397 FREE_LOCK(&lk); 4398 break; 4399 } 4400 continue; 4401 4402 case D_PAGEDEP: 4403 /* 4404 * We are trying to sync a directory that may 4405 * have dependencies on both its own metadata 4406 * and/or dependencies on the inodes of any 4407 * recently allocated files. We walk its diradd 4408 * lists pushing out the associated inode. 4409 */ 4410 pagedep = WK_PAGEDEP(wk); 4411 for (i = 0; i < DAHASHSZ; i++) { 4412 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 4413 continue; 4414 if ((error = 4415 flush_pagedep_deps(vp, pagedep->pd_mnt, 4416 &pagedep->pd_diraddhd[i]))) { 4417 FREE_LOCK(&lk); 4418 break; 4419 } 4420 } 4421 continue; 4422 4423 case D_MKDIR: 4424 /* 4425 * This case should never happen if the vnode has 4426 * been properly sync'ed. However, if this function 4427 * is used at a place where the vnode has not yet 4428 * been sync'ed, this dependency can show up. So, 4429 * rather than panic, just flush it. 4430 */ 4431 nbp = WK_MKDIR(wk)->md_buf; 4432 if (getdirtybuf(&nbp, waitfor) == 0) 4433 continue; 4434 FREE_LOCK(&lk); 4435 if (waitfor == MNT_NOWAIT) { 4436 bawrite(nbp); 4437 } else if ((error = BUF_WRITE(nbp)) != 0) { 4438 break; 4439 } 4440 ACQUIRE_LOCK(&lk); 4441 continue; 4442 4443 case D_BMSAFEMAP: 4444 /* 4445 * This case should never happen if the vnode has 4446 * been properly sync'ed. However, if this function 4447 * is used at a place where the vnode has not yet 4448 * been sync'ed, this dependency can show up. So, 4449 * rather than panic, just flush it. 4450 */ 4451 nbp = WK_BMSAFEMAP(wk)->sm_buf; 4452 if (getdirtybuf(&nbp, waitfor) == 0) 4453 continue; 4454 FREE_LOCK(&lk); 4455 if (waitfor == MNT_NOWAIT) { 4456 bawrite(nbp); 4457 } else if ((error = BUF_WRITE(nbp)) != 0) { 4458 break; 4459 } 4460 ACQUIRE_LOCK(&lk); 4461 continue; 4462 4463 default: 4464 FREE_LOCK(&lk); 4465 panic("softdep_sync_metadata: Unknown type %s", 4466 TYPENAME(wk->wk_type)); 4467 /* NOTREACHED */ 4468 } 4469 /* We reach here only in error and unlocked */ 4470 if (error == 0) 4471 panic("softdep_sync_metadata: zero error"); 4472 bp->b_lock.lk_flags &= ~LK_CANRECURSE; 4473 bawrite(bp); 4474 return (error); 4475 } 4476 (void) getdirtybuf(&TAILQ_NEXT(bp, b_vnbufs), MNT_WAIT); 4477 nbp = TAILQ_NEXT(bp, b_vnbufs); 4478 FREE_LOCK(&lk); 4479 bp->b_lock.lk_flags &= ~LK_CANRECURSE; 4480 bawrite(bp); 4481 ACQUIRE_LOCK(&lk); 4482 if (nbp != NULL) { 4483 bp = nbp; 4484 goto loop; 4485 } 4486 /* 4487 * We must wait for any I/O in progress to finish so that 4488 * all potential buffers on the dirty list will be visible. 4489 * Once they are all there, proceed with the second pass 4490 * which will wait for the I/O as per above. 4491 */ 4492 drain_output(vp, 1); 4493 /* 4494 * The brief unlock is to allow any pent up dependency 4495 * processing to be done. 4496 */ 4497 if (waitfor == MNT_NOWAIT) { 4498 waitfor = MNT_WAIT; 4499 FREE_LOCK(&lk); 4500 ACQUIRE_LOCK(&lk); 4501 goto top; 4502 } 4503 4504 /* 4505 * If we have managed to get rid of all the dirty buffers, 4506 * then we are done. For certain directories and block 4507 * devices, we may need to do further work. 4508 */ 4509 if (TAILQ_FIRST(&vp->v_dirtyblkhd) == NULL) { 4510 FREE_LOCK(&lk); 4511 return (0); 4512 } 4513 4514 FREE_LOCK(&lk); 4515 /* 4516 * If we are trying to sync a block device, some of its buffers may 4517 * contain metadata that cannot be written until the contents of some 4518 * partially written files have been written to disk. The only easy 4519 * way to accomplish this is to sync the entire filesystem (luckily 4520 * this happens rarely). 4521 */ 4522 if (vn_isdisk(vp, NULL) && 4523 vp->v_rdev->si_mountpoint && !VOP_ISLOCKED(vp, NULL) && 4524 (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT, ap->a_cred, 4525 ap->a_p)) != 0) 4526 return (error); 4527 return (0); 4528} 4529 4530/* 4531 * Flush the dependencies associated with an inodedep. 4532 * Called with splbio blocked. 4533 */ 4534static int 4535flush_inodedep_deps(fs, ino) 4536 struct fs *fs; 4537 ino_t ino; 4538{ 4539 struct inodedep *inodedep; 4540 struct allocdirect *adp; 4541 int error, waitfor; 4542 struct buf *bp; 4543 4544 /* 4545 * This work is done in two passes. The first pass grabs most 4546 * of the buffers and begins asynchronously writing them. The 4547 * only way to wait for these asynchronous writes is to sleep 4548 * on the filesystem vnode which may stay busy for a long time 4549 * if the filesystem is active. So, instead, we make a second 4550 * pass over the dependencies blocking on each write. In the 4551 * usual case we will be blocking against a write that we 4552 * initiated, so when it is done the dependency will have been 4553 * resolved. Thus the second pass is expected to end quickly. 4554 * We give a brief window at the top of the loop to allow 4555 * any pending I/O to complete. 4556 */ 4557 for (waitfor = MNT_NOWAIT; ; ) { 4558 FREE_LOCK(&lk); 4559 ACQUIRE_LOCK(&lk); 4560 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4561 return (0); 4562 TAILQ_FOREACH(adp, &inodedep->id_inoupdt, ad_next) { 4563 if (adp->ad_state & DEPCOMPLETE) 4564 continue; 4565 bp = adp->ad_buf; 4566 if (getdirtybuf(&bp, waitfor) == 0) { 4567 if (waitfor == MNT_NOWAIT) 4568 continue; 4569 break; 4570 } 4571 FREE_LOCK(&lk); 4572 if (waitfor == MNT_NOWAIT) { 4573 bawrite(bp); 4574 } else if ((error = BUF_WRITE(bp)) != 0) { 4575 ACQUIRE_LOCK(&lk); 4576 return (error); 4577 } 4578 ACQUIRE_LOCK(&lk); 4579 break; 4580 } 4581 if (adp != NULL) 4582 continue; 4583 TAILQ_FOREACH(adp, &inodedep->id_newinoupdt, ad_next) { 4584 if (adp->ad_state & DEPCOMPLETE) 4585 continue; 4586 bp = adp->ad_buf; 4587 if (getdirtybuf(&bp, waitfor) == 0) { 4588 if (waitfor == MNT_NOWAIT) 4589 continue; 4590 break; 4591 } 4592 FREE_LOCK(&lk); 4593 if (waitfor == MNT_NOWAIT) { 4594 bawrite(bp); 4595 } else if ((error = BUF_WRITE(bp)) != 0) { 4596 ACQUIRE_LOCK(&lk); 4597 return (error); 4598 } 4599 ACQUIRE_LOCK(&lk); 4600 break; 4601 } 4602 if (adp != NULL) 4603 continue; 4604 /* 4605 * If pass2, we are done, otherwise do pass 2. 4606 */ 4607 if (waitfor == MNT_WAIT) 4608 break; 4609 waitfor = MNT_WAIT; 4610 } 4611 /* 4612 * Try freeing inodedep in case all dependencies have been removed. 4613 */ 4614 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0) 4615 (void) free_inodedep(inodedep); 4616 return (0); 4617} 4618 4619/* 4620 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 4621 * Called with splbio blocked. 4622 */ 4623static int 4624flush_pagedep_deps(pvp, mp, diraddhdp) 4625 struct vnode *pvp; 4626 struct mount *mp; 4627 struct diraddhd *diraddhdp; 4628{ 4629 struct proc *p = CURPROC; /* XXX */ 4630 struct inodedep *inodedep; 4631 struct ufsmount *ump; 4632 struct diradd *dap; 4633 struct vnode *vp; 4634 int gotit, error = 0; 4635 struct buf *bp; 4636 ino_t inum; 4637 4638 ump = VFSTOUFS(mp); 4639 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 4640 /* 4641 * Flush ourselves if this directory entry 4642 * has a MKDIR_PARENT dependency. 4643 */ 4644 if (dap->da_state & MKDIR_PARENT) { 4645 FREE_LOCK(&lk); 4646 if ((error = UFS_UPDATE(pvp, 1)) != 0) 4647 break; 4648 ACQUIRE_LOCK(&lk); 4649 /* 4650 * If that cleared dependencies, go on to next. 4651 */ 4652 if (dap != LIST_FIRST(diraddhdp)) 4653 continue; 4654 if (dap->da_state & MKDIR_PARENT) { 4655 FREE_LOCK(&lk); 4656 panic("flush_pagedep_deps: MKDIR_PARENT"); 4657 } 4658 } 4659 /* 4660 * A newly allocated directory must have its "." and 4661 * ".." entries written out before its name can be 4662 * committed in its parent. We do not want or need 4663 * the full semantics of a synchronous VOP_FSYNC as 4664 * that may end up here again, once for each directory 4665 * level in the filesystem. Instead, we push the blocks 4666 * and wait for them to clear. We have to fsync twice 4667 * because the first call may choose to defer blocks 4668 * that still have dependencies, but deferral will 4669 * happen at most once. 4670 */ 4671 inum = dap->da_newinum; 4672 if (dap->da_state & MKDIR_BODY) { 4673 FREE_LOCK(&lk); 4674 if ((error = VFS_VGET(mp, inum, &vp)) != 0) 4675 break; 4676 if ((error=VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p)) || 4677 (error=VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p))) { 4678 vput(vp); 4679 break; 4680 } 4681 drain_output(vp, 0); 4682 vput(vp); 4683 ACQUIRE_LOCK(&lk); 4684 /* 4685 * If that cleared dependencies, go on to next. 4686 */ 4687 if (dap != LIST_FIRST(diraddhdp)) 4688 continue; 4689 if (dap->da_state & MKDIR_BODY) { 4690 FREE_LOCK(&lk); 4691 panic("flush_pagedep_deps: MKDIR_BODY"); 4692 } 4693 } 4694 /* 4695 * Flush the inode on which the directory entry depends. 4696 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 4697 * the only remaining dependency is that the updated inode 4698 * count must get pushed to disk. The inode has already 4699 * been pushed into its inode buffer (via VOP_UPDATE) at 4700 * the time of the reference count change. So we need only 4701 * locate that buffer, ensure that there will be no rollback 4702 * caused by a bitmap dependency, then write the inode buffer. 4703 */ 4704 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) { 4705 FREE_LOCK(&lk); 4706 panic("flush_pagedep_deps: lost inode"); 4707 } 4708 /* 4709 * If the inode still has bitmap dependencies, 4710 * push them to disk. 4711 */ 4712 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4713 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT); 4714 FREE_LOCK(&lk); 4715 if (gotit && 4716 (error = BUF_WRITE(inodedep->id_buf)) != 0) 4717 break; 4718 ACQUIRE_LOCK(&lk); 4719 if (dap != LIST_FIRST(diraddhdp)) 4720 continue; 4721 } 4722 /* 4723 * If the inode is still sitting in a buffer waiting 4724 * to be written, push it to disk. 4725 */ 4726 FREE_LOCK(&lk); 4727 if ((error = bread(ump->um_devvp, 4728 fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), 4729 (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0) 4730 break; 4731 if ((error = BUF_WRITE(bp)) != 0) 4732 break; 4733 ACQUIRE_LOCK(&lk); 4734 /* 4735 * If we have failed to get rid of all the dependencies 4736 * then something is seriously wrong. 4737 */ 4738 if (dap == LIST_FIRST(diraddhdp)) { 4739 FREE_LOCK(&lk); 4740 panic("flush_pagedep_deps: flush failed"); 4741 } 4742 } 4743 if (error) 4744 ACQUIRE_LOCK(&lk); 4745 return (error); 4746} 4747 4748/* 4749 * A large burst of file addition or deletion activity can drive the 4750 * memory load excessively high. First attempt to slow things down 4751 * using the techniques below. If that fails, this routine requests 4752 * the offending operations to fall back to running synchronously 4753 * until the memory load returns to a reasonable level. 4754 */ 4755int 4756softdep_slowdown(vp) 4757 struct vnode *vp; 4758{ 4759 int max_softdeps_hard; 4760 4761 max_softdeps_hard = max_softdeps * 11 / 10; 4762 if (num_dirrem < max_softdeps_hard / 2 && 4763 num_inodedep < max_softdeps_hard) 4764 return (0); 4765 stat_sync_limit_hit += 1; 4766 return (1); 4767} 4768 4769/* 4770 * If memory utilization has gotten too high, deliberately slow things 4771 * down and speed up the I/O processing. 4772 */ 4773static int 4774request_cleanup(resource, islocked) 4775 int resource; 4776 int islocked; 4777{ 4778 struct proc *p = CURPROC; 4779 4780 /* 4781 * We never hold up the filesystem syncer process. 4782 */ 4783 if (p == filesys_syncer) 4784 return (0); 4785 /* 4786 * First check to see if the work list has gotten backlogged. 4787 * If it has, co-opt this process to help clean up two entries. 4788 * Because this process may hold inodes locked, we cannot 4789 * handle any remove requests that might block on a locked 4790 * inode as that could lead to deadlock. 4791 */ 4792 if (num_on_worklist > max_softdeps / 10) { 4793 if (islocked) 4794 FREE_LOCK(&lk); 4795 process_worklist_item(NULL, LK_NOWAIT); 4796 process_worklist_item(NULL, LK_NOWAIT); 4797 stat_worklist_push += 2; 4798 if (islocked) 4799 ACQUIRE_LOCK(&lk); 4800 return(1); 4801 } 4802 /* 4803 * Next, we attempt to speed up the syncer process. If that 4804 * is successful, then we allow the process to continue. 4805 */ 4806 if (speedup_syncer()) 4807 return(0); 4808 /* 4809 * If we are resource constrained on inode dependencies, try 4810 * flushing some dirty inodes. Otherwise, we are constrained 4811 * by file deletions, so try accelerating flushes of directories 4812 * with removal dependencies. We would like to do the cleanup 4813 * here, but we probably hold an inode locked at this point and 4814 * that might deadlock against one that we try to clean. So, 4815 * the best that we can do is request the syncer daemon to do 4816 * the cleanup for us. 4817 */ 4818 switch (resource) { 4819 4820 case FLUSH_INODES: 4821 stat_ino_limit_push += 1; 4822 req_clear_inodedeps += 1; 4823 stat_countp = &stat_ino_limit_hit; 4824 break; 4825 4826 case FLUSH_REMOVE: 4827 stat_blk_limit_push += 1; 4828 req_clear_remove += 1; 4829 stat_countp = &stat_blk_limit_hit; 4830 break; 4831 4832 default: 4833 if (islocked) 4834 FREE_LOCK(&lk); 4835 panic("request_cleanup: unknown type"); 4836 } 4837 /* 4838 * Hopefully the syncer daemon will catch up and awaken us. 4839 * We wait at most tickdelay before proceeding in any case. 4840 */ 4841 if (islocked == 0) 4842 ACQUIRE_LOCK(&lk); 4843 proc_waiting += 1; 4844 if (handle.callout == NULL) 4845 handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2); 4846 FREE_LOCK_INTERLOCKED(&lk); 4847 (void) tsleep((caddr_t)&proc_waiting, PPAUSE, "softupdate", 0); 4848 ACQUIRE_LOCK_INTERLOCKED(&lk); 4849 proc_waiting -= 1; 4850 if (islocked == 0) 4851 FREE_LOCK(&lk); 4852 return (1); 4853} 4854 4855/* 4856 * Awaken processes pausing in request_cleanup and clear proc_waiting 4857 * to indicate that there is no longer a timer running. 4858 */ 4859void 4860pause_timer(arg) 4861 void *arg; 4862{ 4863 4864 *stat_countp += 1; 4865 wakeup_one(&proc_waiting); 4866 if (proc_waiting > 0) 4867 handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2); 4868 else 4869 handle.callout = NULL; 4870} 4871 4872/* 4873 * Flush out a directory with at least one removal dependency in an effort to 4874 * reduce the number of dirrem, freefile, and freeblks dependency structures. 4875 */ 4876static void 4877clear_remove(p) 4878 struct proc *p; 4879{ 4880 struct pagedep_hashhead *pagedephd; 4881 struct pagedep *pagedep; 4882 static int next = 0; 4883 struct mount *mp; 4884 struct vnode *vp; 4885 int error, cnt; 4886 ino_t ino; 4887 4888 ACQUIRE_LOCK(&lk); 4889 for (cnt = 0; cnt < pagedep_hash; cnt++) { 4890 pagedephd = &pagedep_hashtbl[next++]; 4891 if (next >= pagedep_hash) 4892 next = 0; 4893 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 4894 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL) 4895 continue; 4896 mp = pagedep->pd_mnt; 4897 ino = pagedep->pd_ino; 4898 FREE_LOCK(&lk); 4899 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 4900 continue; 4901 if ((error = VFS_VGET(mp, ino, &vp)) != 0) { 4902 softdep_error("clear_remove: vget", error); 4903 vn_finished_write(mp); 4904 return; 4905 } 4906 if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p))) 4907 softdep_error("clear_remove: fsync", error); 4908 drain_output(vp, 0); 4909 vput(vp); 4910 vn_finished_write(mp); 4911 return; 4912 } 4913 } 4914 FREE_LOCK(&lk); 4915} 4916 4917/* 4918 * Clear out a block of dirty inodes in an effort to reduce 4919 * the number of inodedep dependency structures. 4920 */ 4921static void 4922clear_inodedeps(p) 4923 struct proc *p; 4924{ 4925 struct inodedep_hashhead *inodedephd; 4926 struct inodedep *inodedep; 4927 static int next = 0; 4928 struct mount *mp; 4929 struct vnode *vp; 4930 struct fs *fs; 4931 int error, cnt; 4932 ino_t firstino, lastino, ino; 4933 4934 ACQUIRE_LOCK(&lk); 4935 /* 4936 * Pick a random inode dependency to be cleared. 4937 * We will then gather up all the inodes in its block 4938 * that have dependencies and flush them out. 4939 */ 4940 for (cnt = 0; cnt < inodedep_hash; cnt++) { 4941 inodedephd = &inodedep_hashtbl[next++]; 4942 if (next >= inodedep_hash) 4943 next = 0; 4944 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 4945 break; 4946 } 4947 if (inodedep == NULL) 4948 return; 4949 /* 4950 * Ugly code to find mount point given pointer to superblock. 4951 */ 4952 fs = inodedep->id_fs; 4953 TAILQ_FOREACH(mp, &mountlist, mnt_list) 4954 if ((mp->mnt_flag & MNT_SOFTDEP) && fs == VFSTOUFS(mp)->um_fs) 4955 break; 4956 /* 4957 * Find the last inode in the block with dependencies. 4958 */ 4959 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 4960 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 4961 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0) 4962 break; 4963 /* 4964 * Asynchronously push all but the last inode with dependencies. 4965 * Synchronously push the last inode with dependencies to ensure 4966 * that the inode block gets written to free up the inodedeps. 4967 */ 4968 for (ino = firstino; ino <= lastino; ino++) { 4969 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0) 4970 continue; 4971 FREE_LOCK(&lk); 4972 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 4973 continue; 4974 if ((error = VFS_VGET(mp, ino, &vp)) != 0) { 4975 softdep_error("clear_inodedeps: vget", error); 4976 vn_finished_write(mp); 4977 return; 4978 } 4979 if (ino == lastino) { 4980 if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_WAIT, p))) 4981 softdep_error("clear_inodedeps: fsync1", error); 4982 } else { 4983 if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p))) 4984 softdep_error("clear_inodedeps: fsync2", error); 4985 drain_output(vp, 0); 4986 } 4987 vput(vp); 4988 vn_finished_write(mp); 4989 ACQUIRE_LOCK(&lk); 4990 } 4991 FREE_LOCK(&lk); 4992} 4993 4994/* 4995 * Function to determine if the buffer has outstanding dependencies 4996 * that will cause a roll-back if the buffer is written. If wantcount 4997 * is set, return number of dependencies, otherwise just yes or no. 4998 */ 4999static int 5000softdep_count_dependencies(bp, wantcount) 5001 struct buf *bp; 5002 int wantcount; 5003{ 5004 struct worklist *wk; 5005 struct inodedep *inodedep; 5006 struct indirdep *indirdep; 5007 struct allocindir *aip; 5008 struct pagedep *pagedep; 5009 struct diradd *dap; 5010 int i, retval; 5011 5012 retval = 0; 5013 ACQUIRE_LOCK(&lk); 5014 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5015 switch (wk->wk_type) { 5016 5017 case D_INODEDEP: 5018 inodedep = WK_INODEDEP(wk); 5019 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 5020 /* bitmap allocation dependency */ 5021 retval += 1; 5022 if (!wantcount) 5023 goto out; 5024 } 5025 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 5026 /* direct block pointer dependency */ 5027 retval += 1; 5028 if (!wantcount) 5029 goto out; 5030 } 5031 continue; 5032 5033 case D_INDIRDEP: 5034 indirdep = WK_INDIRDEP(wk); 5035 5036 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 5037 /* indirect block pointer dependency */ 5038 retval += 1; 5039 if (!wantcount) 5040 goto out; 5041 } 5042 continue; 5043 5044 case D_PAGEDEP: 5045 pagedep = WK_PAGEDEP(wk); 5046 for (i = 0; i < DAHASHSZ; i++) { 5047 5048 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 5049 /* directory entry dependency */ 5050 retval += 1; 5051 if (!wantcount) 5052 goto out; 5053 } 5054 } 5055 continue; 5056 5057 case D_BMSAFEMAP: 5058 case D_ALLOCDIRECT: 5059 case D_ALLOCINDIR: 5060 case D_MKDIR: 5061 /* never a dependency on these blocks */ 5062 continue; 5063 5064 default: 5065 FREE_LOCK(&lk); 5066 panic("softdep_check_for_rollback: Unexpected type %s", 5067 TYPENAME(wk->wk_type)); 5068 /* NOTREACHED */ 5069 } 5070 } 5071out: 5072 FREE_LOCK(&lk); 5073 return retval; 5074} 5075 5076/* 5077 * Acquire exclusive access to a buffer. 5078 * Must be called with splbio blocked. 5079 * Return 1 if buffer was acquired. 5080 */ 5081static int 5082getdirtybuf(bpp, waitfor) 5083 struct buf **bpp; 5084 int waitfor; 5085{ 5086 struct buf *bp; 5087 5088 for (;;) { 5089 if ((bp = *bpp) == NULL) 5090 return (0); 5091 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 5092 if ((bp->b_xflags & BX_BKGRDINPROG) == 0) 5093 break; 5094 BUF_UNLOCK(bp); 5095 if (waitfor != MNT_WAIT) 5096 return (0); 5097 bp->b_xflags |= BX_BKGRDWAIT; 5098 FREE_LOCK_INTERLOCKED(&lk); 5099 tsleep(&bp->b_xflags, PRIBIO, "getbuf", 0); 5100 ACQUIRE_LOCK_INTERLOCKED(&lk); 5101 continue; 5102 } 5103 if (waitfor != MNT_WAIT) 5104 return (0); 5105 FREE_LOCK_INTERLOCKED(&lk); 5106 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL) != ENOLCK) 5107 panic("getdirtybuf: inconsistent lock"); 5108 ACQUIRE_LOCK_INTERLOCKED(&lk); 5109 } 5110 if ((bp->b_flags & B_DELWRI) == 0) { 5111 BUF_UNLOCK(bp); 5112 return (0); 5113 } 5114 bremfree(bp); 5115 return (1); 5116} 5117 5118/* 5119 * Wait for pending output on a vnode to complete. 5120 * Must be called with vnode locked. 5121 */ 5122static void 5123drain_output(vp, islocked) 5124 struct vnode *vp; 5125 int islocked; 5126{ 5127 5128 if (!islocked) 5129 ACQUIRE_LOCK(&lk); 5130 while (vp->v_numoutput) { 5131 vp->v_flag |= VBWAIT; 5132 FREE_LOCK_INTERLOCKED(&lk); 5133 tsleep((caddr_t)&vp->v_numoutput, PRIBIO + 1, "drainvp", 0); 5134 ACQUIRE_LOCK_INTERLOCKED(&lk); 5135 } 5136 if (!islocked) 5137 FREE_LOCK(&lk); 5138} 5139 5140/* 5141 * Called whenever a buffer that is being invalidated or reallocated 5142 * contains dependencies. This should only happen if an I/O error has 5143 * occurred. The routine is called with the buffer locked. 5144 */ 5145static void 5146softdep_deallocate_dependencies(bp) 5147 struct buf *bp; 5148{ 5149 5150 if ((bp->b_ioflags & BIO_ERROR) == 0) 5151 panic("softdep_deallocate_dependencies: dangling deps"); 5152 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); 5153 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 5154} 5155 5156/* 5157 * Function to handle asynchronous write errors in the filesystem. 5158 */ 5159void 5160softdep_error(func, error) 5161 char *func; 5162 int error; 5163{ 5164 5165 /* XXX should do something better! */ 5166 printf("%s: got error %d while accessing filesystem\n", func, error); 5167} 5168