1 /*- 2 * Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved. 3 * 4 * The soft updates code is derived from the appendix of a University 5 * of Michigan technical report (Gregory R. Ganger and Yale N. Patt, 6 * "Soft Updates: A Solution to the Metadata Update Problem in File 7 * Systems", CSE-TR-254-95, August 1995). 8 * 9 * Further information about soft updates can be obtained from: 10 * 11 * Marshall Kirk McKusick http://www.mckusick.com/softdep/ 12 * 1614 Oxford Street mckusick@mckusick.com 13 * Berkeley, CA 94709-1608 +1-510-843-9542 14 * USA 15 * 16 * Redistribution and use in source and binary forms, with or without 17 * modification, are permitted provided that the following conditions 18 * are met: 19 * 20 * 1. Redistributions of source code must retain the above copyright 21 * notice, this list of conditions and the following disclaimer. 22 * 2. Redistributions in binary form must reproduce the above copyright 23 * notice, this list of conditions and the following disclaimer in the 24 * documentation and/or other materials provided with the distribution. 25 * 26 * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY 27 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 28 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 29 * DISCLAIMED. IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR 30 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * from: @(#)ffs_softdep.c 9.59 (McKusick) 6/21/00 39 */ 40 41 #include <sys/cdefs.h> 42 __FBSDID("$FreeBSD$"); 43 44 /* 45 * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide. 46 */ 47 #ifndef DIAGNOSTIC 48 #define DIAGNOSTIC 49 #endif 50 #ifndef DEBUG 51 #define DEBUG 52 #endif 53 54 #include <sys/param.h> 55 #include <sys/kernel.h> 56 #include <sys/systm.h> 57 #include <sys/bio.h> 58 #include <sys/buf.h> 59 #include <sys/kdb.h> 60 #include <sys/kthread.h> 61 #include <sys/lock.h> 62 #include <sys/malloc.h> 63 #include <sys/mount.h> 64 #include <sys/mutex.h> 65 #include <sys/proc.h> 66 #include <sys/stat.h> 67 #include <sys/sysctl.h> 68 #include <sys/syslog.h> 69 #include <sys/vnode.h> 70 #include <sys/conf.h> 71 #include <ufs/ufs/dir.h> 72 #include <ufs/ufs/extattr.h> 73 #include <ufs/ufs/quota.h> 74 #include <ufs/ufs/inode.h> 75 #include <ufs/ufs/ufsmount.h> 76 #include <ufs/ffs/fs.h> 77 #include <ufs/ffs/softdep.h> 78 #include <ufs/ffs/ffs_extern.h> 79 #include <ufs/ufs/ufs_extern.h> 80 81 #include <vm/vm.h> 82 83 #include "opt_ffs.h" 84 85 #ifndef SOFTUPDATES 86 87 int 88 softdep_flushfiles(oldmnt, flags, td) 89 struct mount *oldmnt; 90 int flags; 91 struct thread *td; 92 { 93 94 panic("softdep_flushfiles called"); 95 } 96 97 int 98 softdep_mount(devvp, mp, fs, cred) 99 struct vnode *devvp; 100 struct mount *mp; 101 struct fs *fs; 102 struct ucred *cred; 103 { 104 105 return (0); 106 } 107 108 void 109 softdep_initialize() 110 { 111 112 return; 113 } 114 115 void 116 softdep_uninitialize() 117 { 118 119 return; 120 } 121 122 void 123 softdep_setup_inomapdep(bp, ip, newinum) 124 struct buf *bp; 125 struct inode *ip; 126 ino_t newinum; 127 { 128 129 panic("softdep_setup_inomapdep called"); 130 } 131 132 void 133 softdep_setup_blkmapdep(bp, mp, newblkno) 134 struct buf *bp; 135 struct mount *mp; 136 ufs2_daddr_t newblkno; 137 { 138 139 panic("softdep_setup_blkmapdep called"); 140 } 141 142 void 143 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 144 struct inode *ip; 145 ufs_lbn_t lbn; 146 ufs2_daddr_t newblkno; 147 ufs2_daddr_t oldblkno; 148 long newsize; 149 long oldsize; 150 struct buf *bp; 151 { 152 153 panic("softdep_setup_allocdirect called"); 154 } 155 156 void 157 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 158 struct inode *ip; 159 ufs_lbn_t lbn; 160 ufs2_daddr_t newblkno; 161 ufs2_daddr_t oldblkno; 162 long newsize; 163 long oldsize; 164 struct buf *bp; 165 { 166 167 panic("softdep_setup_allocdirect called"); 168 } 169 170 void 171 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 172 struct inode *ip; 173 ufs_lbn_t lbn; 174 struct buf *bp; 175 int ptrno; 176 ufs2_daddr_t newblkno; 177 ufs2_daddr_t oldblkno; 178 struct buf *nbp; 179 { 180 181 panic("softdep_setup_allocindir_page called"); 182 } 183 184 void 185 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 186 struct buf *nbp; 187 struct inode *ip; 188 struct buf *bp; 189 int ptrno; 190 ufs2_daddr_t newblkno; 191 { 192 193 panic("softdep_setup_allocindir_meta called"); 194 } 195 196 void 197 softdep_setup_freeblocks(ip, length, flags) 198 struct inode *ip; 199 off_t length; 200 int flags; 201 { 202 203 panic("softdep_setup_freeblocks called"); 204 } 205 206 void 207 softdep_freefile(pvp, ino, mode) 208 struct vnode *pvp; 209 ino_t ino; 210 int mode; 211 { 212 213 panic("softdep_freefile called"); 214 } 215 216 int 217 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 218 struct buf *bp; 219 struct inode *dp; 220 off_t diroffset; 221 ino_t newinum; 222 struct buf *newdirbp; 223 int isnewblk; 224 { 225 226 panic("softdep_setup_directory_add called"); 227 } 228 229 void 230 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize) 231 struct inode *dp; 232 caddr_t base; 233 caddr_t oldloc; 234 caddr_t newloc; 235 int entrysize; 236 { 237 238 panic("softdep_change_directoryentry_offset called"); 239 } 240 241 void 242 softdep_setup_remove(bp, dp, ip, isrmdir) 243 struct buf *bp; 244 struct inode *dp; 245 struct inode *ip; 246 int isrmdir; 247 { 248 249 panic("softdep_setup_remove called"); 250 } 251 252 void 253 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 254 struct buf *bp; 255 struct inode *dp; 256 struct inode *ip; 257 ino_t newinum; 258 int isrmdir; 259 { 260 261 panic("softdep_setup_directory_change called"); 262 } 263 264 void 265 softdep_change_linkcnt(ip) 266 struct inode *ip; 267 { 268 269 panic("softdep_change_linkcnt called"); 270 } 271 272 void 273 softdep_load_inodeblock(ip) 274 struct inode *ip; 275 { 276 277 panic("softdep_load_inodeblock called"); 278 } 279 280 void 281 softdep_update_inodeblock(ip, bp, waitfor) 282 struct inode *ip; 283 struct buf *bp; 284 int waitfor; 285 { 286 287 panic("softdep_update_inodeblock called"); 288 } 289 290 int 291 softdep_fsync(vp) 292 struct vnode *vp; /* the "in_core" copy of the inode */ 293 { 294 295 return (0); 296 } 297 298 void 299 softdep_fsync_mountdev(vp) 300 struct vnode *vp; 301 { 302 303 return; 304 } 305 306 int 307 softdep_flushworklist(oldmnt, countp, td) 308 struct mount *oldmnt; 309 int *countp; 310 struct thread *td; 311 { 312 313 *countp = 0; 314 return (0); 315 } 316 317 int 318 softdep_sync_metadata(struct vnode *vp) 319 { 320 321 return (0); 322 } 323 324 int 325 softdep_slowdown(vp) 326 struct vnode *vp; 327 { 328 329 panic("softdep_slowdown called"); 330 } 331 332 void 333 softdep_releasefile(ip) 334 struct inode *ip; /* inode with the zero effective link count */ 335 { 336 337 panic("softdep_releasefile called"); 338 } 339 340 int 341 softdep_request_cleanup(fs, vp) 342 struct fs *fs; 343 struct vnode *vp; 344 { 345 346 return (0); 347 } 348 349 int 350 softdep_check_suspend(struct mount *mp, 351 struct vnode *devvp, 352 int softdep_deps, 353 int softdep_accdeps, 354 int secondary_writes, 355 int secondary_accwrites) 356 { 357 struct bufobj *bo; 358 int error; 359 360 (void) softdep_deps, 361 (void) softdep_accdeps; 362 363 ASSERT_VI_LOCKED(devvp, "softdep_check_suspend"); 364 bo = &devvp->v_bufobj; 365 366 for (;;) { 367 if (!MNT_ITRYLOCK(mp)) { 368 VI_UNLOCK(devvp); 369 MNT_ILOCK(mp); 370 MNT_IUNLOCK(mp); 371 VI_LOCK(devvp); 372 continue; 373 } 374 if (mp->mnt_secondary_writes != 0) { 375 VI_UNLOCK(devvp); 376 msleep(&mp->mnt_secondary_writes, 377 MNT_MTX(mp), 378 (PUSER - 1) | PDROP, "secwr", 0); 379 VI_LOCK(devvp); 380 continue; 381 } 382 break; 383 } 384 385 /* 386 * Reasons for needing more work before suspend: 387 * - Dirty buffers on devvp. 388 * - Secondary writes occurred after start of vnode sync loop 389 */ 390 error = 0; 391 if (bo->bo_numoutput > 0 || 392 bo->bo_dirty.bv_cnt > 0 || 393 secondary_writes != 0 || 394 mp->mnt_secondary_writes != 0 || 395 secondary_accwrites != mp->mnt_secondary_accwrites) 396 error = EAGAIN; 397 VI_UNLOCK(devvp); 398 return (error); 399 } 400 401 void 402 softdep_get_depcounts(struct mount *mp, 403 int *softdepactivep, 404 int *softdepactiveaccp) 405 { 406 (void) mp; 407 *softdepactivep = 0; 408 *softdepactiveaccp = 0; 409 } 410 411 #else 412 /* 413 * These definitions need to be adapted to the system to which 414 * this file is being ported. 415 */ 416 /* 417 * malloc types defined for the softdep system. 418 */ 419 static MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies"); 420 static MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies"); 421 static MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation"); 422 static MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map"); 423 static MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode"); 424 static MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies"); 425 static MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block"); 426 static MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode"); 427 static MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode"); 428 static MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated"); 429 static MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry"); 430 static MALLOC_DEFINE(M_MKDIR, "mkdir","New directory"); 431 static MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted"); 432 static MALLOC_DEFINE(M_NEWDIRBLK, "newdirblk","Unclaimed new directory block"); 433 static MALLOC_DEFINE(M_SAVEDINO, "savedino","Saved inodes"); 434 435 #define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE) 436 437 #define D_PAGEDEP 0 438 #define D_INODEDEP 1 439 #define D_NEWBLK 2 440 #define D_BMSAFEMAP 3 441 #define D_ALLOCDIRECT 4 442 #define D_INDIRDEP 5 443 #define D_ALLOCINDIR 6 444 #define D_FREEFRAG 7 445 #define D_FREEBLKS 8 446 #define D_FREEFILE 9 447 #define D_DIRADD 10 448 #define D_MKDIR 11 449 #define D_DIRREM 12 450 #define D_NEWDIRBLK 13 451 #define D_LAST D_NEWDIRBLK 452 453 /* 454 * translate from workitem type to memory type 455 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 456 */ 457 static struct malloc_type *memtype[] = { 458 M_PAGEDEP, 459 M_INODEDEP, 460 M_NEWBLK, 461 M_BMSAFEMAP, 462 M_ALLOCDIRECT, 463 M_INDIRDEP, 464 M_ALLOCINDIR, 465 M_FREEFRAG, 466 M_FREEBLKS, 467 M_FREEFILE, 468 M_DIRADD, 469 M_MKDIR, 470 M_DIRREM, 471 M_NEWDIRBLK 472 }; 473 474 #define DtoM(type) (memtype[type]) 475 476 /* 477 * Names of malloc types. 478 */ 479 #define TYPENAME(type) \ 480 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???") 481 /* 482 * End system adaptation definitions. 483 */ 484 485 /* 486 * Forward declarations. 487 */ 488 struct inodedep_hashhead; 489 struct newblk_hashhead; 490 struct pagedep_hashhead; 491 492 /* 493 * Internal function prototypes. 494 */ 495 static void softdep_error(char *, int); 496 static void drain_output(struct vnode *); 497 static struct buf *getdirtybuf(struct buf *, struct mtx *, int); 498 static void clear_remove(struct thread *); 499 static void clear_inodedeps(struct thread *); 500 static int flush_pagedep_deps(struct vnode *, struct mount *, 501 struct diraddhd *); 502 static int flush_inodedep_deps(struct mount *, ino_t); 503 static int flush_deplist(struct allocdirectlst *, int, int *); 504 static int handle_written_filepage(struct pagedep *, struct buf *); 505 static void diradd_inode_written(struct diradd *, struct inodedep *); 506 static int handle_written_inodeblock(struct inodedep *, struct buf *); 507 static void handle_allocdirect_partdone(struct allocdirect *); 508 static void handle_allocindir_partdone(struct allocindir *); 509 static void initiate_write_filepage(struct pagedep *, struct buf *); 510 static void handle_written_mkdir(struct mkdir *, int); 511 static void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *); 512 static void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *); 513 static void handle_workitem_freefile(struct freefile *); 514 static void handle_workitem_remove(struct dirrem *, struct vnode *); 515 static struct dirrem *newdirrem(struct buf *, struct inode *, 516 struct inode *, int, struct dirrem **); 517 static void free_diradd(struct diradd *); 518 static void free_allocindir(struct allocindir *, struct inodedep *); 519 static void free_newdirblk(struct newdirblk *); 520 static int indir_trunc(struct freeblks *, ufs2_daddr_t, int, ufs_lbn_t, 521 ufs2_daddr_t *); 522 static void deallocate_dependencies(struct buf *, struct inodedep *); 523 static void free_allocdirect(struct allocdirectlst *, 524 struct allocdirect *, int); 525 static int check_inode_unwritten(struct inodedep *); 526 static int free_inodedep(struct inodedep *); 527 static void handle_workitem_freeblocks(struct freeblks *, int); 528 static void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *); 529 static void setup_allocindir_phase2(struct buf *, struct inode *, 530 struct allocindir *); 531 static struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t, 532 ufs2_daddr_t); 533 static void handle_workitem_freefrag(struct freefrag *); 534 static struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long); 535 static void allocdirect_merge(struct allocdirectlst *, 536 struct allocdirect *, struct allocdirect *); 537 static struct bmsafemap *bmsafemap_lookup(struct mount *, struct buf *); 538 static int newblk_find(struct newblk_hashhead *, struct fs *, ufs2_daddr_t, 539 struct newblk **); 540 static int newblk_lookup(struct fs *, ufs2_daddr_t, int, struct newblk **); 541 static int inodedep_find(struct inodedep_hashhead *, struct fs *, ino_t, 542 struct inodedep **); 543 static int inodedep_lookup(struct mount *, ino_t, int, struct inodedep **); 544 static int pagedep_lookup(struct inode *, ufs_lbn_t, int, struct pagedep **); 545 static int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t, 546 struct mount *mp, int, struct pagedep **); 547 static void pause_timer(void *); 548 static int request_cleanup(struct mount *, int); 549 static int process_worklist_item(struct mount *, int); 550 static void add_to_worklist(struct worklist *); 551 static void softdep_flush(void); 552 static int softdep_speedup(void); 553 554 /* 555 * Exported softdep operations. 556 */ 557 static void softdep_disk_io_initiation(struct buf *); 558 static void softdep_disk_write_complete(struct buf *); 559 static void softdep_deallocate_dependencies(struct buf *); 560 static int softdep_count_dependencies(struct buf *bp, int); 561 562 static struct mtx lk; 563 MTX_SYSINIT(softdep_lock, &lk, "Softdep Lock", MTX_DEF); 564 565 #define TRY_ACQUIRE_LOCK(lk) mtx_trylock(lk) 566 #define ACQUIRE_LOCK(lk) mtx_lock(lk) 567 #define FREE_LOCK(lk) mtx_unlock(lk) 568 569 /* 570 * Worklist queue management. 571 * These routines require that the lock be held. 572 */ 573 #ifndef /* NOT */ DEBUG 574 #define WORKLIST_INSERT(head, item) do { \ 575 (item)->wk_state |= ONWORKLIST; \ 576 LIST_INSERT_HEAD(head, item, wk_list); \ 577 } while (0) 578 #define WORKLIST_REMOVE(item) do { \ 579 (item)->wk_state &= ~ONWORKLIST; \ 580 LIST_REMOVE(item, wk_list); \ 581 } while (0) 582 #else /* DEBUG */ 583 static void worklist_insert(struct workhead *, struct worklist *); 584 static void worklist_remove(struct worklist *); 585 586 #define WORKLIST_INSERT(head, item) worklist_insert(head, item) 587 #define WORKLIST_REMOVE(item) worklist_remove(item) 588 589 static void 590 worklist_insert(head, item) 591 struct workhead *head; 592 struct worklist *item; 593 { 594 595 mtx_assert(&lk, MA_OWNED); 596 if (item->wk_state & ONWORKLIST) 597 panic("worklist_insert: already on list"); 598 item->wk_state |= ONWORKLIST; 599 LIST_INSERT_HEAD(head, item, wk_list); 600 } 601 602 static void 603 worklist_remove(item) 604 struct worklist *item; 605 { 606 607 mtx_assert(&lk, MA_OWNED); 608 if ((item->wk_state & ONWORKLIST) == 0) 609 panic("worklist_remove: not on list"); 610 item->wk_state &= ~ONWORKLIST; 611 LIST_REMOVE(item, wk_list); 612 } 613 #endif /* DEBUG */ 614 615 /* 616 * Routines for tracking and managing workitems. 617 */ 618 static void workitem_free(struct worklist *, int); 619 static void workitem_alloc(struct worklist *, int, struct mount *); 620 621 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)(item), (type)) 622 623 static void 624 workitem_free(item, type) 625 struct worklist *item; 626 int type; 627 { 628 struct ufsmount *ump; 629 mtx_assert(&lk, MA_OWNED); 630 631 #ifdef DEBUG 632 if (item->wk_state & ONWORKLIST) 633 panic("workitem_free: still on list"); 634 if (item->wk_type != type) 635 panic("workitem_free: type mismatch"); 636 #endif 637 ump = VFSTOUFS(item->wk_mp); 638 if (--ump->softdep_deps == 0 && ump->softdep_req) 639 wakeup(&ump->softdep_deps); 640 FREE(item, DtoM(type)); 641 } 642 643 static void 644 workitem_alloc(item, type, mp) 645 struct worklist *item; 646 int type; 647 struct mount *mp; 648 { 649 item->wk_type = type; 650 item->wk_mp = mp; 651 item->wk_state = 0; 652 ACQUIRE_LOCK(&lk); 653 VFSTOUFS(mp)->softdep_deps++; 654 VFSTOUFS(mp)->softdep_accdeps++; 655 FREE_LOCK(&lk); 656 } 657 658 /* 659 * Workitem queue management 660 */ 661 static int max_softdeps; /* maximum number of structs before slowdown */ 662 static int maxindirdeps = 50; /* max number of indirdeps before slowdown */ 663 static int tickdelay = 2; /* number of ticks to pause during slowdown */ 664 static int proc_waiting; /* tracks whether we have a timeout posted */ 665 static int *stat_countp; /* statistic to count in proc_waiting timeout */ 666 static struct callout_handle handle; /* handle on posted proc_waiting timeout */ 667 static int req_pending; 668 static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 669 #define FLUSH_INODES 1 670 static int req_clear_remove; /* syncer process flush some freeblks */ 671 #define FLUSH_REMOVE 2 672 #define FLUSH_REMOVE_WAIT 3 673 /* 674 * runtime statistics 675 */ 676 static int stat_worklist_push; /* number of worklist cleanups */ 677 static int stat_blk_limit_push; /* number of times block limit neared */ 678 static int stat_ino_limit_push; /* number of times inode limit neared */ 679 static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 680 static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 681 static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 682 static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 683 static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 684 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 685 static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 686 687 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, ""); 688 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, ""); 689 SYSCTL_INT(_debug, OID_AUTO, maxindirdeps, CTLFLAG_RW, &maxindirdeps, 0, ""); 690 SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,""); 691 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,""); 692 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,""); 693 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, ""); 694 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, ""); 695 SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, ""); 696 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, ""); 697 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, ""); 698 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, ""); 699 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, ""); 700 /* SYSCTL_INT(_debug, OID_AUTO, worklist_num, CTLFLAG_RD, &softdep_on_worklist, 0, ""); */ 701 702 SYSCTL_DECL(_vfs_ffs); 703 704 static int compute_summary_at_mount = 0; /* Whether to recompute the summary at mount time */ 705 SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW, 706 &compute_summary_at_mount, 0, "Recompute summary at mount"); 707 708 static struct proc *softdepproc; 709 static struct kproc_desc softdep_kp = { 710 "softdepflush", 711 softdep_flush, 712 &softdepproc 713 }; 714 SYSINIT(sdproc, SI_SUB_KTHREAD_UPDATE, SI_ORDER_ANY, kproc_start, &softdep_kp) 715 716 static void 717 softdep_flush(void) 718 { 719 struct mount *nmp; 720 struct mount *mp; 721 struct ufsmount *ump; 722 struct thread *td; 723 int remaining; 724 int vfslocked; 725 726 td = curthread; 727 td->td_pflags |= TDP_NORUNNINGBUF; 728 729 for (;;) { 730 kthread_suspend_check(softdepproc); 731 ACQUIRE_LOCK(&lk); 732 /* 733 * If requested, try removing inode or removal dependencies. 734 */ 735 if (req_clear_inodedeps) { 736 clear_inodedeps(td); 737 req_clear_inodedeps -= 1; 738 wakeup_one(&proc_waiting); 739 } 740 if (req_clear_remove) { 741 clear_remove(td); 742 req_clear_remove -= 1; 743 wakeup_one(&proc_waiting); 744 } 745 FREE_LOCK(&lk); 746 remaining = 0; 747 mtx_lock(&mountlist_mtx); 748 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 749 nmp = TAILQ_NEXT(mp, mnt_list); 750 if ((mp->mnt_flag & MNT_SOFTDEP) == 0) 751 continue; 752 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) 753 continue; 754 vfslocked = VFS_LOCK_GIANT(mp); 755 softdep_process_worklist(mp, 0); 756 ump = VFSTOUFS(mp); 757 remaining += ump->softdep_on_worklist - 758 ump->softdep_on_worklist_inprogress; 759 VFS_UNLOCK_GIANT(vfslocked); 760 mtx_lock(&mountlist_mtx); 761 nmp = TAILQ_NEXT(mp, mnt_list); 762 vfs_unbusy(mp, td); 763 } 764 mtx_unlock(&mountlist_mtx); 765 if (remaining) 766 continue; 767 ACQUIRE_LOCK(&lk); 768 if (!req_pending) 769 msleep(&req_pending, &lk, PVM, "sdflush", hz); 770 req_pending = 0; 771 FREE_LOCK(&lk); 772 } 773 } 774 775 static int 776 softdep_speedup(void) 777 { 778 779 mtx_assert(&lk, MA_OWNED); 780 if (req_pending == 0) { 781 req_pending = 1; 782 wakeup(&req_pending); 783 } 784 785 return speedup_syncer(); 786 } 787 788 /* 789 * Add an item to the end of the work queue. 790 * This routine requires that the lock be held. 791 * This is the only routine that adds items to the list. 792 * The following routine is the only one that removes items 793 * and does so in order from first to last. 794 */ 795 static void 796 add_to_worklist(wk) 797 struct worklist *wk; 798 { 799 struct ufsmount *ump; 800 801 mtx_assert(&lk, MA_OWNED); 802 ump = VFSTOUFS(wk->wk_mp); 803 if (wk->wk_state & ONWORKLIST) 804 panic("add_to_worklist: already on list"); 805 wk->wk_state |= ONWORKLIST; 806 if (LIST_FIRST(&ump->softdep_workitem_pending) == NULL) 807 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); 808 else 809 LIST_INSERT_AFTER(ump->softdep_worklist_tail, wk, wk_list); 810 ump->softdep_worklist_tail = wk; 811 ump->softdep_on_worklist += 1; 812 } 813 814 /* 815 * Process that runs once per second to handle items in the background queue. 816 * 817 * Note that we ensure that everything is done in the order in which they 818 * appear in the queue. The code below depends on this property to ensure 819 * that blocks of a file are freed before the inode itself is freed. This 820 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 821 * until all the old ones have been purged from the dependency lists. 822 */ 823 int 824 softdep_process_worklist(mp, full) 825 struct mount *mp; 826 int full; 827 { 828 struct thread *td = curthread; 829 int cnt, matchcnt, loopcount; 830 struct ufsmount *ump; 831 long starttime; 832 833 KASSERT(mp != NULL, ("softdep_process_worklist: NULL mp")); 834 /* 835 * Record the process identifier of our caller so that we can give 836 * this process preferential treatment in request_cleanup below. 837 */ 838 matchcnt = 0; 839 ump = VFSTOUFS(mp); 840 ACQUIRE_LOCK(&lk); 841 loopcount = 1; 842 starttime = time_second; 843 while (ump->softdep_on_worklist > 0) { 844 if ((cnt = process_worklist_item(mp, 0)) == -1) 845 break; 846 else 847 matchcnt += cnt; 848 /* 849 * If requested, try removing inode or removal dependencies. 850 */ 851 if (req_clear_inodedeps) { 852 clear_inodedeps(td); 853 req_clear_inodedeps -= 1; 854 wakeup_one(&proc_waiting); 855 } 856 if (req_clear_remove) { 857 clear_remove(td); 858 req_clear_remove -= 1; 859 wakeup_one(&proc_waiting); 860 } 861 /* 862 * We do not generally want to stop for buffer space, but if 863 * we are really being a buffer hog, we will stop and wait. 864 */ 865 if (loopcount++ % 128 == 0) { 866 FREE_LOCK(&lk); 867 bwillwrite(); 868 ACQUIRE_LOCK(&lk); 869 } 870 /* 871 * Never allow processing to run for more than one 872 * second. Otherwise the other mountpoints may get 873 * excessively backlogged. 874 */ 875 if (!full && starttime != time_second) { 876 matchcnt = -1; 877 break; 878 } 879 } 880 FREE_LOCK(&lk); 881 return (matchcnt); 882 } 883 884 /* 885 * Process one item on the worklist. 886 */ 887 static int 888 process_worklist_item(mp, flags) 889 struct mount *mp; 890 int flags; 891 { 892 struct worklist *wk, *wkend; 893 struct ufsmount *ump; 894 struct vnode *vp; 895 int matchcnt = 0; 896 897 mtx_assert(&lk, MA_OWNED); 898 KASSERT(mp != NULL, ("process_worklist_item: NULL mp")); 899 /* 900 * If we are being called because of a process doing a 901 * copy-on-write, then it is not safe to write as we may 902 * recurse into the copy-on-write routine. 903 */ 904 if (curthread->td_pflags & TDP_COWINPROGRESS) 905 return (-1); 906 /* 907 * Normally we just process each item on the worklist in order. 908 * However, if we are in a situation where we cannot lock any 909 * inodes, we have to skip over any dirrem requests whose 910 * vnodes are resident and locked. 911 */ 912 ump = VFSTOUFS(mp); 913 vp = NULL; 914 LIST_FOREACH(wk, &ump->softdep_workitem_pending, wk_list) { 915 if (wk->wk_state & INPROGRESS) 916 continue; 917 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM) 918 break; 919 wk->wk_state |= INPROGRESS; 920 ump->softdep_on_worklist_inprogress++; 921 FREE_LOCK(&lk); 922 ffs_vget(mp, WK_DIRREM(wk)->dm_oldinum, 923 LK_NOWAIT | LK_EXCLUSIVE, &vp); 924 ACQUIRE_LOCK(&lk); 925 wk->wk_state &= ~INPROGRESS; 926 ump->softdep_on_worklist_inprogress--; 927 if (vp != NULL) 928 break; 929 } 930 if (wk == 0) 931 return (-1); 932 /* 933 * Remove the item to be processed. If we are removing the last 934 * item on the list, we need to recalculate the tail pointer. 935 * As this happens rarely and usually when the list is short, 936 * we just run down the list to find it rather than tracking it 937 * in the above loop. 938 */ 939 WORKLIST_REMOVE(wk); 940 if (wk == ump->softdep_worklist_tail) { 941 LIST_FOREACH(wkend, &ump->softdep_workitem_pending, wk_list) 942 if (LIST_NEXT(wkend, wk_list) == NULL) 943 break; 944 ump->softdep_worklist_tail = wkend; 945 } 946 ump->softdep_on_worklist -= 1; 947 FREE_LOCK(&lk); 948 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT)) 949 panic("process_worklist_item: suspended filesystem"); 950 matchcnt++; 951 switch (wk->wk_type) { 952 953 case D_DIRREM: 954 /* removal of a directory entry */ 955 handle_workitem_remove(WK_DIRREM(wk), vp); 956 break; 957 958 case D_FREEBLKS: 959 /* releasing blocks and/or fragments from a file */ 960 handle_workitem_freeblocks(WK_FREEBLKS(wk), flags & LK_NOWAIT); 961 break; 962 963 case D_FREEFRAG: 964 /* releasing a fragment when replaced as a file grows */ 965 handle_workitem_freefrag(WK_FREEFRAG(wk)); 966 break; 967 968 case D_FREEFILE: 969 /* releasing an inode when its link count drops to 0 */ 970 handle_workitem_freefile(WK_FREEFILE(wk)); 971 break; 972 973 default: 974 panic("%s_process_worklist: Unknown type %s", 975 "softdep", TYPENAME(wk->wk_type)); 976 /* NOTREACHED */ 977 } 978 vn_finished_secondary_write(mp); 979 ACQUIRE_LOCK(&lk); 980 return (matchcnt); 981 } 982 983 /* 984 * Move dependencies from one buffer to another. 985 */ 986 void 987 softdep_move_dependencies(oldbp, newbp) 988 struct buf *oldbp; 989 struct buf *newbp; 990 { 991 struct worklist *wk, *wktail; 992 993 if (LIST_FIRST(&newbp->b_dep) != NULL) 994 panic("softdep_move_dependencies: need merge code"); 995 wktail = 0; 996 ACQUIRE_LOCK(&lk); 997 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 998 LIST_REMOVE(wk, wk_list); 999 if (wktail == 0) 1000 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 1001 else 1002 LIST_INSERT_AFTER(wktail, wk, wk_list); 1003 wktail = wk; 1004 } 1005 FREE_LOCK(&lk); 1006 } 1007 1008 /* 1009 * Purge the work list of all items associated with a particular mount point. 1010 */ 1011 int 1012 softdep_flushworklist(oldmnt, countp, td) 1013 struct mount *oldmnt; 1014 int *countp; 1015 struct thread *td; 1016 { 1017 struct vnode *devvp; 1018 int count, error = 0; 1019 struct ufsmount *ump; 1020 1021 /* 1022 * Alternately flush the block device associated with the mount 1023 * point and process any dependencies that the flushing 1024 * creates. We continue until no more worklist dependencies 1025 * are found. 1026 */ 1027 *countp = 0; 1028 ump = VFSTOUFS(oldmnt); 1029 devvp = ump->um_devvp; 1030 while ((count = softdep_process_worklist(oldmnt, 1)) > 0) { 1031 *countp += count; 1032 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, td); 1033 error = VOP_FSYNC(devvp, MNT_WAIT, td); 1034 VOP_UNLOCK(devvp, 0, td); 1035 if (error) 1036 break; 1037 } 1038 return (error); 1039 } 1040 1041 int 1042 softdep_waitidle(struct mount *mp) 1043 { 1044 struct ufsmount *ump; 1045 int error; 1046 int i; 1047 1048 ump = VFSTOUFS(mp); 1049 ACQUIRE_LOCK(&lk); 1050 for (i = 0; i < 10 && ump->softdep_deps; i++) { 1051 ump->softdep_req = 1; 1052 if (ump->softdep_on_worklist) 1053 panic("softdep_waitidle: work added after flush."); 1054 msleep(&ump->softdep_deps, &lk, PVM, "softdeps", 1); 1055 } 1056 ump->softdep_req = 0; 1057 FREE_LOCK(&lk); 1058 error = 0; 1059 if (i == 10) { 1060 error = EBUSY; 1061 printf("softdep_waitidle: Failed to flush worklist for %p", 1062 mp); 1063 } 1064 1065 return (error); 1066 } 1067 1068 /* 1069 * Flush all vnodes and worklist items associated with a specified mount point. 1070 */ 1071 int 1072 softdep_flushfiles(oldmnt, flags, td) 1073 struct mount *oldmnt; 1074 int flags; 1075 struct thread *td; 1076 { 1077 int error, count, loopcnt; 1078 1079 error = 0; 1080 1081 /* 1082 * Alternately flush the vnodes associated with the mount 1083 * point and process any dependencies that the flushing 1084 * creates. In theory, this loop can happen at most twice, 1085 * but we give it a few extra just to be sure. 1086 */ 1087 for (loopcnt = 10; loopcnt > 0; loopcnt--) { 1088 /* 1089 * Do another flush in case any vnodes were brought in 1090 * as part of the cleanup operations. 1091 */ 1092 if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0) 1093 break; 1094 if ((error = softdep_flushworklist(oldmnt, &count, td)) != 0 || 1095 count == 0) 1096 break; 1097 } 1098 /* 1099 * If we are unmounting then it is an error to fail. If we 1100 * are simply trying to downgrade to read-only, then filesystem 1101 * activity can keep us busy forever, so we just fail with EBUSY. 1102 */ 1103 if (loopcnt == 0) { 1104 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 1105 panic("softdep_flushfiles: looping"); 1106 error = EBUSY; 1107 } 1108 if (!error) 1109 error = softdep_waitidle(oldmnt); 1110 return (error); 1111 } 1112 1113 /* 1114 * Structure hashing. 1115 * 1116 * There are three types of structures that can be looked up: 1117 * 1) pagedep structures identified by mount point, inode number, 1118 * and logical block. 1119 * 2) inodedep structures identified by mount point and inode number. 1120 * 3) newblk structures identified by mount point and 1121 * physical block number. 1122 * 1123 * The "pagedep" and "inodedep" dependency structures are hashed 1124 * separately from the file blocks and inodes to which they correspond. 1125 * This separation helps when the in-memory copy of an inode or 1126 * file block must be replaced. It also obviates the need to access 1127 * an inode or file page when simply updating (or de-allocating) 1128 * dependency structures. Lookup of newblk structures is needed to 1129 * find newly allocated blocks when trying to associate them with 1130 * their allocdirect or allocindir structure. 1131 * 1132 * The lookup routines optionally create and hash a new instance when 1133 * an existing entry is not found. 1134 */ 1135 #define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 1136 #define NODELAY 0x0002 /* cannot do background work */ 1137 1138 /* 1139 * Structures and routines associated with pagedep caching. 1140 */ 1141 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl; 1142 u_long pagedep_hash; /* size of hash table - 1 */ 1143 #define PAGEDEP_HASH(mp, inum, lbn) \ 1144 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \ 1145 pagedep_hash]) 1146 1147 static int 1148 pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp) 1149 struct pagedep_hashhead *pagedephd; 1150 ino_t ino; 1151 ufs_lbn_t lbn; 1152 struct mount *mp; 1153 int flags; 1154 struct pagedep **pagedeppp; 1155 { 1156 struct pagedep *pagedep; 1157 1158 LIST_FOREACH(pagedep, pagedephd, pd_hash) 1159 if (ino == pagedep->pd_ino && 1160 lbn == pagedep->pd_lbn && 1161 mp == pagedep->pd_list.wk_mp) 1162 break; 1163 if (pagedep) { 1164 *pagedeppp = pagedep; 1165 if ((flags & DEPALLOC) != 0 && 1166 (pagedep->pd_state & ONWORKLIST) == 0) 1167 return (0); 1168 return (1); 1169 } 1170 *pagedeppp = NULL; 1171 return (0); 1172 } 1173 /* 1174 * Look up a pagedep. Return 1 if found, 0 if not found or found 1175 * when asked to allocate but not associated with any buffer. 1176 * If not found, allocate if DEPALLOC flag is passed. 1177 * Found or allocated entry is returned in pagedeppp. 1178 * This routine must be called with splbio interrupts blocked. 1179 */ 1180 static int 1181 pagedep_lookup(ip, lbn, flags, pagedeppp) 1182 struct inode *ip; 1183 ufs_lbn_t lbn; 1184 int flags; 1185 struct pagedep **pagedeppp; 1186 { 1187 struct pagedep *pagedep; 1188 struct pagedep_hashhead *pagedephd; 1189 struct mount *mp; 1190 int ret; 1191 int i; 1192 1193 mtx_assert(&lk, MA_OWNED); 1194 mp = ITOV(ip)->v_mount; 1195 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn); 1196 1197 ret = pagedep_find(pagedephd, ip->i_number, lbn, mp, flags, pagedeppp); 1198 if (*pagedeppp || (flags & DEPALLOC) == 0) 1199 return (ret); 1200 FREE_LOCK(&lk); 1201 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), 1202 M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO); 1203 workitem_alloc(&pagedep->pd_list, D_PAGEDEP, mp); 1204 ACQUIRE_LOCK(&lk); 1205 ret = pagedep_find(pagedephd, ip->i_number, lbn, mp, flags, pagedeppp); 1206 if (*pagedeppp) { 1207 WORKITEM_FREE(pagedep, D_PAGEDEP); 1208 return (ret); 1209 } 1210 pagedep->pd_ino = ip->i_number; 1211 pagedep->pd_lbn = lbn; 1212 LIST_INIT(&pagedep->pd_dirremhd); 1213 LIST_INIT(&pagedep->pd_pendinghd); 1214 for (i = 0; i < DAHASHSZ; i++) 1215 LIST_INIT(&pagedep->pd_diraddhd[i]); 1216 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 1217 *pagedeppp = pagedep; 1218 return (0); 1219 } 1220 1221 /* 1222 * Structures and routines associated with inodedep caching. 1223 */ 1224 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl; 1225 static u_long inodedep_hash; /* size of hash table - 1 */ 1226 static long num_inodedep; /* number of inodedep allocated */ 1227 #define INODEDEP_HASH(fs, inum) \ 1228 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash]) 1229 1230 static int 1231 inodedep_find(inodedephd, fs, inum, inodedeppp) 1232 struct inodedep_hashhead *inodedephd; 1233 struct fs *fs; 1234 ino_t inum; 1235 struct inodedep **inodedeppp; 1236 { 1237 struct inodedep *inodedep; 1238 1239 LIST_FOREACH(inodedep, inodedephd, id_hash) 1240 if (inum == inodedep->id_ino && fs == inodedep->id_fs) 1241 break; 1242 if (inodedep) { 1243 *inodedeppp = inodedep; 1244 return (1); 1245 } 1246 *inodedeppp = NULL; 1247 1248 return (0); 1249 } 1250 /* 1251 * Look up an inodedep. Return 1 if found, 0 if not found. 1252 * If not found, allocate if DEPALLOC flag is passed. 1253 * Found or allocated entry is returned in inodedeppp. 1254 * This routine must be called with splbio interrupts blocked. 1255 */ 1256 static int 1257 inodedep_lookup(mp, inum, flags, inodedeppp) 1258 struct mount *mp; 1259 ino_t inum; 1260 int flags; 1261 struct inodedep **inodedeppp; 1262 { 1263 struct inodedep *inodedep; 1264 struct inodedep_hashhead *inodedephd; 1265 struct fs *fs; 1266 1267 mtx_assert(&lk, MA_OWNED); 1268 fs = VFSTOUFS(mp)->um_fs; 1269 inodedephd = INODEDEP_HASH(fs, inum); 1270 1271 if (inodedep_find(inodedephd, fs, inum, inodedeppp)) 1272 return (1); 1273 if ((flags & DEPALLOC) == 0) 1274 return (0); 1275 /* 1276 * If we are over our limit, try to improve the situation. 1277 */ 1278 if (num_inodedep > max_softdeps && (flags & NODELAY) == 0) 1279 request_cleanup(mp, FLUSH_INODES); 1280 FREE_LOCK(&lk); 1281 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep), 1282 M_INODEDEP, M_SOFTDEP_FLAGS); 1283 workitem_alloc(&inodedep->id_list, D_INODEDEP, mp); 1284 ACQUIRE_LOCK(&lk); 1285 if (inodedep_find(inodedephd, fs, inum, inodedeppp)) { 1286 WORKITEM_FREE(inodedep, D_INODEDEP); 1287 return (1); 1288 } 1289 num_inodedep += 1; 1290 inodedep->id_fs = fs; 1291 inodedep->id_ino = inum; 1292 inodedep->id_state = ALLCOMPLETE; 1293 inodedep->id_nlinkdelta = 0; 1294 inodedep->id_savedino1 = NULL; 1295 inodedep->id_savedsize = -1; 1296 inodedep->id_savedextsize = -1; 1297 inodedep->id_buf = NULL; 1298 LIST_INIT(&inodedep->id_pendinghd); 1299 LIST_INIT(&inodedep->id_inowait); 1300 LIST_INIT(&inodedep->id_bufwait); 1301 TAILQ_INIT(&inodedep->id_inoupdt); 1302 TAILQ_INIT(&inodedep->id_newinoupdt); 1303 TAILQ_INIT(&inodedep->id_extupdt); 1304 TAILQ_INIT(&inodedep->id_newextupdt); 1305 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 1306 *inodedeppp = inodedep; 1307 return (0); 1308 } 1309 1310 /* 1311 * Structures and routines associated with newblk caching. 1312 */ 1313 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl; 1314 u_long newblk_hash; /* size of hash table - 1 */ 1315 #define NEWBLK_HASH(fs, inum) \ 1316 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash]) 1317 1318 static int 1319 newblk_find(newblkhd, fs, newblkno, newblkpp) 1320 struct newblk_hashhead *newblkhd; 1321 struct fs *fs; 1322 ufs2_daddr_t newblkno; 1323 struct newblk **newblkpp; 1324 { 1325 struct newblk *newblk; 1326 1327 LIST_FOREACH(newblk, newblkhd, nb_hash) 1328 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs) 1329 break; 1330 if (newblk) { 1331 *newblkpp = newblk; 1332 return (1); 1333 } 1334 *newblkpp = NULL; 1335 return (0); 1336 } 1337 1338 /* 1339 * Look up a newblk. Return 1 if found, 0 if not found. 1340 * If not found, allocate if DEPALLOC flag is passed. 1341 * Found or allocated entry is returned in newblkpp. 1342 */ 1343 static int 1344 newblk_lookup(fs, newblkno, flags, newblkpp) 1345 struct fs *fs; 1346 ufs2_daddr_t newblkno; 1347 int flags; 1348 struct newblk **newblkpp; 1349 { 1350 struct newblk *newblk; 1351 struct newblk_hashhead *newblkhd; 1352 1353 newblkhd = NEWBLK_HASH(fs, newblkno); 1354 if (newblk_find(newblkhd, fs, newblkno, newblkpp)) 1355 return (1); 1356 if ((flags & DEPALLOC) == 0) 1357 return (0); 1358 FREE_LOCK(&lk); 1359 MALLOC(newblk, struct newblk *, sizeof(struct newblk), 1360 M_NEWBLK, M_SOFTDEP_FLAGS); 1361 ACQUIRE_LOCK(&lk); 1362 if (newblk_find(newblkhd, fs, newblkno, newblkpp)) { 1363 FREE(newblk, M_NEWBLK); 1364 return (1); 1365 } 1366 newblk->nb_state = 0; 1367 newblk->nb_fs = fs; 1368 newblk->nb_newblkno = newblkno; 1369 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 1370 *newblkpp = newblk; 1371 return (0); 1372 } 1373 1374 /* 1375 * Executed during filesystem system initialization before 1376 * mounting any filesystems. 1377 */ 1378 void 1379 softdep_initialize() 1380 { 1381 1382 LIST_INIT(&mkdirlisthd); 1383 max_softdeps = desiredvnodes * 4; 1384 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 1385 &pagedep_hash); 1386 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash); 1387 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash); 1388 1389 /* initialise bioops hack */ 1390 bioops.io_start = softdep_disk_io_initiation; 1391 bioops.io_complete = softdep_disk_write_complete; 1392 bioops.io_deallocate = softdep_deallocate_dependencies; 1393 bioops.io_countdeps = softdep_count_dependencies; 1394 } 1395 1396 /* 1397 * Executed after all filesystems have been unmounted during 1398 * filesystem module unload. 1399 */ 1400 void 1401 softdep_uninitialize() 1402 { 1403 1404 hashdestroy(pagedep_hashtbl, M_PAGEDEP, pagedep_hash); 1405 hashdestroy(inodedep_hashtbl, M_INODEDEP, inodedep_hash); 1406 hashdestroy(newblk_hashtbl, M_NEWBLK, newblk_hash); 1407 } 1408 1409 /* 1410 * Called at mount time to notify the dependency code that a 1411 * filesystem wishes to use it. 1412 */ 1413 int 1414 softdep_mount(devvp, mp, fs, cred) 1415 struct vnode *devvp; 1416 struct mount *mp; 1417 struct fs *fs; 1418 struct ucred *cred; 1419 { 1420 struct csum_total cstotal; 1421 struct ufsmount *ump; 1422 struct cg *cgp; 1423 struct buf *bp; 1424 int error, cyl; 1425 1426 mp->mnt_flag &= ~MNT_ASYNC; 1427 mp->mnt_flag |= MNT_SOFTDEP; 1428 ump = VFSTOUFS(mp); 1429 LIST_INIT(&ump->softdep_workitem_pending); 1430 ump->softdep_worklist_tail = NULL; 1431 ump->softdep_on_worklist = 0; 1432 ump->softdep_deps = 0; 1433 /* 1434 * When doing soft updates, the counters in the 1435 * superblock may have gotten out of sync. Recomputation 1436 * can take a long time and can be deferred for background 1437 * fsck. However, the old behavior of scanning the cylinder 1438 * groups and recalculating them at mount time is available 1439 * by setting vfs.ffs.compute_summary_at_mount to one. 1440 */ 1441 if (compute_summary_at_mount == 0 || fs->fs_clean != 0) 1442 return (0); 1443 bzero(&cstotal, sizeof cstotal); 1444 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 1445 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 1446 fs->fs_cgsize, cred, &bp)) != 0) { 1447 brelse(bp); 1448 return (error); 1449 } 1450 cgp = (struct cg *)bp->b_data; 1451 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 1452 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 1453 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 1454 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 1455 fs->fs_cs(fs, cyl) = cgp->cg_cs; 1456 brelse(bp); 1457 } 1458 #ifdef DEBUG 1459 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 1460 printf("%s: superblock summary recomputed\n", fs->fs_fsmnt); 1461 #endif 1462 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 1463 return (0); 1464 } 1465 1466 /* 1467 * Protecting the freemaps (or bitmaps). 1468 * 1469 * To eliminate the need to execute fsck before mounting a filesystem 1470 * after a power failure, one must (conservatively) guarantee that the 1471 * on-disk copy of the bitmaps never indicate that a live inode or block is 1472 * free. So, when a block or inode is allocated, the bitmap should be 1473 * updated (on disk) before any new pointers. When a block or inode is 1474 * freed, the bitmap should not be updated until all pointers have been 1475 * reset. The latter dependency is handled by the delayed de-allocation 1476 * approach described below for block and inode de-allocation. The former 1477 * dependency is handled by calling the following procedure when a block or 1478 * inode is allocated. When an inode is allocated an "inodedep" is created 1479 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 1480 * Each "inodedep" is also inserted into the hash indexing structure so 1481 * that any additional link additions can be made dependent on the inode 1482 * allocation. 1483 * 1484 * The ufs filesystem maintains a number of free block counts (e.g., per 1485 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 1486 * in addition to the bitmaps. These counts are used to improve efficiency 1487 * during allocation and therefore must be consistent with the bitmaps. 1488 * There is no convenient way to guarantee post-crash consistency of these 1489 * counts with simple update ordering, for two main reasons: (1) The counts 1490 * and bitmaps for a single cylinder group block are not in the same disk 1491 * sector. If a disk write is interrupted (e.g., by power failure), one may 1492 * be written and the other not. (2) Some of the counts are located in the 1493 * superblock rather than the cylinder group block. So, we focus our soft 1494 * updates implementation on protecting the bitmaps. When mounting a 1495 * filesystem, we recompute the auxiliary counts from the bitmaps. 1496 */ 1497 1498 /* 1499 * Called just after updating the cylinder group block to allocate an inode. 1500 */ 1501 void 1502 softdep_setup_inomapdep(bp, ip, newinum) 1503 struct buf *bp; /* buffer for cylgroup block with inode map */ 1504 struct inode *ip; /* inode related to allocation */ 1505 ino_t newinum; /* new inode number being allocated */ 1506 { 1507 struct inodedep *inodedep; 1508 struct bmsafemap *bmsafemap; 1509 1510 /* 1511 * Create a dependency for the newly allocated inode. 1512 * Panic if it already exists as something is seriously wrong. 1513 * Otherwise add it to the dependency list for the buffer holding 1514 * the cylinder group map from which it was allocated. 1515 */ 1516 ACQUIRE_LOCK(&lk); 1517 if ((inodedep_lookup(UFSTOVFS(ip->i_ump), newinum, DEPALLOC|NODELAY, 1518 &inodedep))) 1519 panic("softdep_setup_inomapdep: found inode"); 1520 inodedep->id_buf = bp; 1521 inodedep->id_state &= ~DEPCOMPLETE; 1522 bmsafemap = bmsafemap_lookup(inodedep->id_list.wk_mp, bp); 1523 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 1524 FREE_LOCK(&lk); 1525 } 1526 1527 /* 1528 * Called just after updating the cylinder group block to 1529 * allocate block or fragment. 1530 */ 1531 void 1532 softdep_setup_blkmapdep(bp, mp, newblkno) 1533 struct buf *bp; /* buffer for cylgroup block with block map */ 1534 struct mount *mp; /* filesystem doing allocation */ 1535 ufs2_daddr_t newblkno; /* number of newly allocated block */ 1536 { 1537 struct newblk *newblk; 1538 struct bmsafemap *bmsafemap; 1539 struct fs *fs; 1540 1541 fs = VFSTOUFS(mp)->um_fs; 1542 /* 1543 * Create a dependency for the newly allocated block. 1544 * Add it to the dependency list for the buffer holding 1545 * the cylinder group map from which it was allocated. 1546 */ 1547 ACQUIRE_LOCK(&lk); 1548 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0) 1549 panic("softdep_setup_blkmapdep: found block"); 1550 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp); 1551 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 1552 FREE_LOCK(&lk); 1553 } 1554 1555 /* 1556 * Find the bmsafemap associated with a cylinder group buffer. 1557 * If none exists, create one. The buffer must be locked when 1558 * this routine is called and this routine must be called with 1559 * splbio interrupts blocked. 1560 */ 1561 static struct bmsafemap * 1562 bmsafemap_lookup(mp, bp) 1563 struct mount *mp; 1564 struct buf *bp; 1565 { 1566 struct bmsafemap *bmsafemap; 1567 struct worklist *wk; 1568 1569 mtx_assert(&lk, MA_OWNED); 1570 LIST_FOREACH(wk, &bp->b_dep, wk_list) 1571 if (wk->wk_type == D_BMSAFEMAP) 1572 return (WK_BMSAFEMAP(wk)); 1573 FREE_LOCK(&lk); 1574 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap), 1575 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 1576 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); 1577 bmsafemap->sm_buf = bp; 1578 LIST_INIT(&bmsafemap->sm_allocdirecthd); 1579 LIST_INIT(&bmsafemap->sm_allocindirhd); 1580 LIST_INIT(&bmsafemap->sm_inodedephd); 1581 LIST_INIT(&bmsafemap->sm_newblkhd); 1582 ACQUIRE_LOCK(&lk); 1583 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 1584 return (bmsafemap); 1585 } 1586 1587 /* 1588 * Direct block allocation dependencies. 1589 * 1590 * When a new block is allocated, the corresponding disk locations must be 1591 * initialized (with zeros or new data) before the on-disk inode points to 1592 * them. Also, the freemap from which the block was allocated must be 1593 * updated (on disk) before the inode's pointer. These two dependencies are 1594 * independent of each other and are needed for all file blocks and indirect 1595 * blocks that are pointed to directly by the inode. Just before the 1596 * "in-core" version of the inode is updated with a newly allocated block 1597 * number, a procedure (below) is called to setup allocation dependency 1598 * structures. These structures are removed when the corresponding 1599 * dependencies are satisfied or when the block allocation becomes obsolete 1600 * (i.e., the file is deleted, the block is de-allocated, or the block is a 1601 * fragment that gets upgraded). All of these cases are handled in 1602 * procedures described later. 1603 * 1604 * When a file extension causes a fragment to be upgraded, either to a larger 1605 * fragment or to a full block, the on-disk location may change (if the 1606 * previous fragment could not simply be extended). In this case, the old 1607 * fragment must be de-allocated, but not until after the inode's pointer has 1608 * been updated. In most cases, this is handled by later procedures, which 1609 * will construct a "freefrag" structure to be added to the workitem queue 1610 * when the inode update is complete (or obsolete). The main exception to 1611 * this is when an allocation occurs while a pending allocation dependency 1612 * (for the same block pointer) remains. This case is handled in the main 1613 * allocation dependency setup procedure by immediately freeing the 1614 * unreferenced fragments. 1615 */ 1616 void 1617 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 1618 struct inode *ip; /* inode to which block is being added */ 1619 ufs_lbn_t lbn; /* block pointer within inode */ 1620 ufs2_daddr_t newblkno; /* disk block number being added */ 1621 ufs2_daddr_t oldblkno; /* previous block number, 0 unless frag */ 1622 long newsize; /* size of new block */ 1623 long oldsize; /* size of new block */ 1624 struct buf *bp; /* bp for allocated block */ 1625 { 1626 struct allocdirect *adp, *oldadp; 1627 struct allocdirectlst *adphead; 1628 struct bmsafemap *bmsafemap; 1629 struct inodedep *inodedep; 1630 struct pagedep *pagedep; 1631 struct newblk *newblk; 1632 struct mount *mp; 1633 1634 mp = UFSTOVFS(ip->i_ump); 1635 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect), 1636 M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO); 1637 workitem_alloc(&adp->ad_list, D_ALLOCDIRECT, mp); 1638 adp->ad_lbn = lbn; 1639 adp->ad_newblkno = newblkno; 1640 adp->ad_oldblkno = oldblkno; 1641 adp->ad_newsize = newsize; 1642 adp->ad_oldsize = oldsize; 1643 adp->ad_state = ATTACHED; 1644 LIST_INIT(&adp->ad_newdirblk); 1645 if (newblkno == oldblkno) 1646 adp->ad_freefrag = NULL; 1647 else 1648 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1649 1650 ACQUIRE_LOCK(&lk); 1651 if (lbn >= NDADDR) { 1652 /* allocating an indirect block */ 1653 if (oldblkno != 0) 1654 panic("softdep_setup_allocdirect: non-zero indir"); 1655 } else { 1656 /* 1657 * Allocating a direct block. 1658 * 1659 * If we are allocating a directory block, then we must 1660 * allocate an associated pagedep to track additions and 1661 * deletions. 1662 */ 1663 if ((ip->i_mode & IFMT) == IFDIR && 1664 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1665 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 1666 } 1667 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1668 panic("softdep_setup_allocdirect: lost block"); 1669 if (newblk->nb_state == DEPCOMPLETE) { 1670 adp->ad_state |= DEPCOMPLETE; 1671 adp->ad_buf = NULL; 1672 } else { 1673 bmsafemap = newblk->nb_bmsafemap; 1674 adp->ad_buf = bmsafemap->sm_buf; 1675 LIST_REMOVE(newblk, nb_deps); 1676 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1677 } 1678 LIST_REMOVE(newblk, nb_hash); 1679 FREE(newblk, M_NEWBLK); 1680 1681 inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep); 1682 adp->ad_inodedep = inodedep; 1683 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list); 1684 /* 1685 * The list of allocdirects must be kept in sorted and ascending 1686 * order so that the rollback routines can quickly determine the 1687 * first uncommitted block (the size of the file stored on disk 1688 * ends at the end of the lowest committed fragment, or if there 1689 * are no fragments, at the end of the highest committed block). 1690 * Since files generally grow, the typical case is that the new 1691 * block is to be added at the end of the list. We speed this 1692 * special case by checking against the last allocdirect in the 1693 * list before laboriously traversing the list looking for the 1694 * insertion point. 1695 */ 1696 adphead = &inodedep->id_newinoupdt; 1697 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1698 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1699 /* insert at end of list */ 1700 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1701 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1702 allocdirect_merge(adphead, adp, oldadp); 1703 FREE_LOCK(&lk); 1704 return; 1705 } 1706 TAILQ_FOREACH(oldadp, adphead, ad_next) { 1707 if (oldadp->ad_lbn >= lbn) 1708 break; 1709 } 1710 if (oldadp == NULL) 1711 panic("softdep_setup_allocdirect: lost entry"); 1712 /* insert in middle of list */ 1713 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1714 if (oldadp->ad_lbn == lbn) 1715 allocdirect_merge(adphead, adp, oldadp); 1716 FREE_LOCK(&lk); 1717 } 1718 1719 /* 1720 * Replace an old allocdirect dependency with a newer one. 1721 * This routine must be called with splbio interrupts blocked. 1722 */ 1723 static void 1724 allocdirect_merge(adphead, newadp, oldadp) 1725 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 1726 struct allocdirect *newadp; /* allocdirect being added */ 1727 struct allocdirect *oldadp; /* existing allocdirect being checked */ 1728 { 1729 struct worklist *wk; 1730 struct freefrag *freefrag; 1731 struct newdirblk *newdirblk; 1732 1733 mtx_assert(&lk, MA_OWNED); 1734 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 1735 newadp->ad_oldsize != oldadp->ad_newsize || 1736 newadp->ad_lbn >= NDADDR) 1737 panic("%s %jd != new %jd || old size %ld != new %ld", 1738 "allocdirect_merge: old blkno", 1739 (intmax_t)newadp->ad_oldblkno, 1740 (intmax_t)oldadp->ad_newblkno, 1741 newadp->ad_oldsize, oldadp->ad_newsize); 1742 newadp->ad_oldblkno = oldadp->ad_oldblkno; 1743 newadp->ad_oldsize = oldadp->ad_oldsize; 1744 /* 1745 * If the old dependency had a fragment to free or had never 1746 * previously had a block allocated, then the new dependency 1747 * can immediately post its freefrag and adopt the old freefrag. 1748 * This action is done by swapping the freefrag dependencies. 1749 * The new dependency gains the old one's freefrag, and the 1750 * old one gets the new one and then immediately puts it on 1751 * the worklist when it is freed by free_allocdirect. It is 1752 * not possible to do this swap when the old dependency had a 1753 * non-zero size but no previous fragment to free. This condition 1754 * arises when the new block is an extension of the old block. 1755 * Here, the first part of the fragment allocated to the new 1756 * dependency is part of the block currently claimed on disk by 1757 * the old dependency, so cannot legitimately be freed until the 1758 * conditions for the new dependency are fulfilled. 1759 */ 1760 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 1761 freefrag = newadp->ad_freefrag; 1762 newadp->ad_freefrag = oldadp->ad_freefrag; 1763 oldadp->ad_freefrag = freefrag; 1764 } 1765 /* 1766 * If we are tracking a new directory-block allocation, 1767 * move it from the old allocdirect to the new allocdirect. 1768 */ 1769 if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) { 1770 newdirblk = WK_NEWDIRBLK(wk); 1771 WORKLIST_REMOVE(&newdirblk->db_list); 1772 if (LIST_FIRST(&oldadp->ad_newdirblk) != NULL) 1773 panic("allocdirect_merge: extra newdirblk"); 1774 WORKLIST_INSERT(&newadp->ad_newdirblk, &newdirblk->db_list); 1775 } 1776 free_allocdirect(adphead, oldadp, 0); 1777 } 1778 1779 /* 1780 * Allocate a new freefrag structure if needed. 1781 */ 1782 static struct freefrag * 1783 newfreefrag(ip, blkno, size) 1784 struct inode *ip; 1785 ufs2_daddr_t blkno; 1786 long size; 1787 { 1788 struct freefrag *freefrag; 1789 struct fs *fs; 1790 1791 if (blkno == 0) 1792 return (NULL); 1793 fs = ip->i_fs; 1794 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 1795 panic("newfreefrag: frag size"); 1796 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag), 1797 M_FREEFRAG, M_SOFTDEP_FLAGS); 1798 workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ip->i_ump)); 1799 freefrag->ff_inum = ip->i_number; 1800 freefrag->ff_blkno = blkno; 1801 freefrag->ff_fragsize = size; 1802 return (freefrag); 1803 } 1804 1805 /* 1806 * This workitem de-allocates fragments that were replaced during 1807 * file block allocation. 1808 */ 1809 static void 1810 handle_workitem_freefrag(freefrag) 1811 struct freefrag *freefrag; 1812 { 1813 struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp); 1814 1815 ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno, 1816 freefrag->ff_fragsize, freefrag->ff_inum); 1817 ACQUIRE_LOCK(&lk); 1818 WORKITEM_FREE(freefrag, D_FREEFRAG); 1819 FREE_LOCK(&lk); 1820 } 1821 1822 /* 1823 * Set up a dependency structure for an external attributes data block. 1824 * This routine follows much of the structure of softdep_setup_allocdirect. 1825 * See the description of softdep_setup_allocdirect above for details. 1826 */ 1827 void 1828 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 1829 struct inode *ip; 1830 ufs_lbn_t lbn; 1831 ufs2_daddr_t newblkno; 1832 ufs2_daddr_t oldblkno; 1833 long newsize; 1834 long oldsize; 1835 struct buf *bp; 1836 { 1837 struct allocdirect *adp, *oldadp; 1838 struct allocdirectlst *adphead; 1839 struct bmsafemap *bmsafemap; 1840 struct inodedep *inodedep; 1841 struct newblk *newblk; 1842 struct mount *mp; 1843 1844 mp = UFSTOVFS(ip->i_ump); 1845 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect), 1846 M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO); 1847 workitem_alloc(&adp->ad_list, D_ALLOCDIRECT, mp); 1848 adp->ad_lbn = lbn; 1849 adp->ad_newblkno = newblkno; 1850 adp->ad_oldblkno = oldblkno; 1851 adp->ad_newsize = newsize; 1852 adp->ad_oldsize = oldsize; 1853 adp->ad_state = ATTACHED | EXTDATA; 1854 LIST_INIT(&adp->ad_newdirblk); 1855 if (newblkno == oldblkno) 1856 adp->ad_freefrag = NULL; 1857 else 1858 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1859 1860 ACQUIRE_LOCK(&lk); 1861 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1862 panic("softdep_setup_allocext: lost block"); 1863 1864 inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep); 1865 adp->ad_inodedep = inodedep; 1866 1867 if (newblk->nb_state == DEPCOMPLETE) { 1868 adp->ad_state |= DEPCOMPLETE; 1869 adp->ad_buf = NULL; 1870 } else { 1871 bmsafemap = newblk->nb_bmsafemap; 1872 adp->ad_buf = bmsafemap->sm_buf; 1873 LIST_REMOVE(newblk, nb_deps); 1874 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1875 } 1876 LIST_REMOVE(newblk, nb_hash); 1877 FREE(newblk, M_NEWBLK); 1878 1879 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list); 1880 if (lbn >= NXADDR) 1881 panic("softdep_setup_allocext: lbn %lld > NXADDR", 1882 (long long)lbn); 1883 /* 1884 * The list of allocdirects must be kept in sorted and ascending 1885 * order so that the rollback routines can quickly determine the 1886 * first uncommitted block (the size of the file stored on disk 1887 * ends at the end of the lowest committed fragment, or if there 1888 * are no fragments, at the end of the highest committed block). 1889 * Since files generally grow, the typical case is that the new 1890 * block is to be added at the end of the list. We speed this 1891 * special case by checking against the last allocdirect in the 1892 * list before laboriously traversing the list looking for the 1893 * insertion point. 1894 */ 1895 adphead = &inodedep->id_newextupdt; 1896 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1897 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1898 /* insert at end of list */ 1899 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1900 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1901 allocdirect_merge(adphead, adp, oldadp); 1902 FREE_LOCK(&lk); 1903 return; 1904 } 1905 TAILQ_FOREACH(oldadp, adphead, ad_next) { 1906 if (oldadp->ad_lbn >= lbn) 1907 break; 1908 } 1909 if (oldadp == NULL) 1910 panic("softdep_setup_allocext: lost entry"); 1911 /* insert in middle of list */ 1912 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1913 if (oldadp->ad_lbn == lbn) 1914 allocdirect_merge(adphead, adp, oldadp); 1915 FREE_LOCK(&lk); 1916 } 1917 1918 /* 1919 * Indirect block allocation dependencies. 1920 * 1921 * The same dependencies that exist for a direct block also exist when 1922 * a new block is allocated and pointed to by an entry in a block of 1923 * indirect pointers. The undo/redo states described above are also 1924 * used here. Because an indirect block contains many pointers that 1925 * may have dependencies, a second copy of the entire in-memory indirect 1926 * block is kept. The buffer cache copy is always completely up-to-date. 1927 * The second copy, which is used only as a source for disk writes, 1928 * contains only the safe pointers (i.e., those that have no remaining 1929 * update dependencies). The second copy is freed when all pointers 1930 * are safe. The cache is not allowed to replace indirect blocks with 1931 * pending update dependencies. If a buffer containing an indirect 1932 * block with dependencies is written, these routines will mark it 1933 * dirty again. It can only be successfully written once all the 1934 * dependencies are removed. The ffs_fsync routine in conjunction with 1935 * softdep_sync_metadata work together to get all the dependencies 1936 * removed so that a file can be successfully written to disk. Three 1937 * procedures are used when setting up indirect block pointer 1938 * dependencies. The division is necessary because of the organization 1939 * of the "balloc" routine and because of the distinction between file 1940 * pages and file metadata blocks. 1941 */ 1942 1943 /* 1944 * Allocate a new allocindir structure. 1945 */ 1946 static struct allocindir * 1947 newallocindir(ip, ptrno, newblkno, oldblkno) 1948 struct inode *ip; /* inode for file being extended */ 1949 int ptrno; /* offset of pointer in indirect block */ 1950 ufs2_daddr_t newblkno; /* disk block number being added */ 1951 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 1952 { 1953 struct allocindir *aip; 1954 1955 MALLOC(aip, struct allocindir *, sizeof(struct allocindir), 1956 M_ALLOCINDIR, M_SOFTDEP_FLAGS|M_ZERO); 1957 workitem_alloc(&aip->ai_list, D_ALLOCINDIR, UFSTOVFS(ip->i_ump)); 1958 aip->ai_state = ATTACHED; 1959 aip->ai_offset = ptrno; 1960 aip->ai_newblkno = newblkno; 1961 aip->ai_oldblkno = oldblkno; 1962 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize); 1963 return (aip); 1964 } 1965 1966 /* 1967 * Called just before setting an indirect block pointer 1968 * to a newly allocated file page. 1969 */ 1970 void 1971 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 1972 struct inode *ip; /* inode for file being extended */ 1973 ufs_lbn_t lbn; /* allocated block number within file */ 1974 struct buf *bp; /* buffer with indirect blk referencing page */ 1975 int ptrno; /* offset of pointer in indirect block */ 1976 ufs2_daddr_t newblkno; /* disk block number being added */ 1977 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 1978 struct buf *nbp; /* buffer holding allocated page */ 1979 { 1980 struct allocindir *aip; 1981 struct pagedep *pagedep; 1982 1983 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page"); 1984 aip = newallocindir(ip, ptrno, newblkno, oldblkno); 1985 ACQUIRE_LOCK(&lk); 1986 /* 1987 * If we are allocating a directory page, then we must 1988 * allocate an associated pagedep to track additions and 1989 * deletions. 1990 */ 1991 if ((ip->i_mode & IFMT) == IFDIR && 1992 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1993 WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list); 1994 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1995 setup_allocindir_phase2(bp, ip, aip); 1996 FREE_LOCK(&lk); 1997 } 1998 1999 /* 2000 * Called just before setting an indirect block pointer to a 2001 * newly allocated indirect block. 2002 */ 2003 void 2004 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 2005 struct buf *nbp; /* newly allocated indirect block */ 2006 struct inode *ip; /* inode for file being extended */ 2007 struct buf *bp; /* indirect block referencing allocated block */ 2008 int ptrno; /* offset of pointer in indirect block */ 2009 ufs2_daddr_t newblkno; /* disk block number being added */ 2010 { 2011 struct allocindir *aip; 2012 2013 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta"); 2014 aip = newallocindir(ip, ptrno, newblkno, 0); 2015 ACQUIRE_LOCK(&lk); 2016 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 2017 setup_allocindir_phase2(bp, ip, aip); 2018 FREE_LOCK(&lk); 2019 } 2020 2021 /* 2022 * Called to finish the allocation of the "aip" allocated 2023 * by one of the two routines above. 2024 */ 2025 static void 2026 setup_allocindir_phase2(bp, ip, aip) 2027 struct buf *bp; /* in-memory copy of the indirect block */ 2028 struct inode *ip; /* inode for file being extended */ 2029 struct allocindir *aip; /* allocindir allocated by the above routines */ 2030 { 2031 struct worklist *wk; 2032 struct indirdep *indirdep, *newindirdep; 2033 struct bmsafemap *bmsafemap; 2034 struct allocindir *oldaip; 2035 struct freefrag *freefrag; 2036 struct newblk *newblk; 2037 ufs2_daddr_t blkno; 2038 2039 mtx_assert(&lk, MA_OWNED); 2040 if (bp->b_lblkno >= 0) 2041 panic("setup_allocindir_phase2: not indir blk"); 2042 for (indirdep = NULL, newindirdep = NULL; ; ) { 2043 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 2044 if (wk->wk_type != D_INDIRDEP) 2045 continue; 2046 indirdep = WK_INDIRDEP(wk); 2047 break; 2048 } 2049 if (indirdep == NULL && newindirdep) { 2050 indirdep = newindirdep; 2051 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 2052 newindirdep = NULL; 2053 } 2054 if (indirdep) { 2055 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0, 2056 &newblk) == 0) 2057 panic("setup_allocindir: lost block"); 2058 if (newblk->nb_state == DEPCOMPLETE) { 2059 aip->ai_state |= DEPCOMPLETE; 2060 aip->ai_buf = NULL; 2061 } else { 2062 bmsafemap = newblk->nb_bmsafemap; 2063 aip->ai_buf = bmsafemap->sm_buf; 2064 LIST_REMOVE(newblk, nb_deps); 2065 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd, 2066 aip, ai_deps); 2067 } 2068 LIST_REMOVE(newblk, nb_hash); 2069 FREE(newblk, M_NEWBLK); 2070 aip->ai_indirdep = indirdep; 2071 /* 2072 * Check to see if there is an existing dependency 2073 * for this block. If there is, merge the old 2074 * dependency into the new one. 2075 */ 2076 if (aip->ai_oldblkno == 0) 2077 oldaip = NULL; 2078 else 2079 2080 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) 2081 if (oldaip->ai_offset == aip->ai_offset) 2082 break; 2083 freefrag = NULL; 2084 if (oldaip != NULL) { 2085 if (oldaip->ai_newblkno != aip->ai_oldblkno) 2086 panic("setup_allocindir_phase2: blkno"); 2087 aip->ai_oldblkno = oldaip->ai_oldblkno; 2088 freefrag = aip->ai_freefrag; 2089 aip->ai_freefrag = oldaip->ai_freefrag; 2090 oldaip->ai_freefrag = NULL; 2091 free_allocindir(oldaip, NULL); 2092 } 2093 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 2094 if (ip->i_ump->um_fstype == UFS1) 2095 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data) 2096 [aip->ai_offset] = aip->ai_oldblkno; 2097 else 2098 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data) 2099 [aip->ai_offset] = aip->ai_oldblkno; 2100 FREE_LOCK(&lk); 2101 if (freefrag != NULL) 2102 handle_workitem_freefrag(freefrag); 2103 } else 2104 FREE_LOCK(&lk); 2105 if (newindirdep) { 2106 newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 2107 brelse(newindirdep->ir_savebp); 2108 ACQUIRE_LOCK(&lk); 2109 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP); 2110 if (indirdep) 2111 break; 2112 FREE_LOCK(&lk); 2113 } 2114 if (indirdep) { 2115 ACQUIRE_LOCK(&lk); 2116 break; 2117 } 2118 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep), 2119 M_INDIRDEP, M_SOFTDEP_FLAGS); 2120 workitem_alloc(&newindirdep->ir_list, D_INDIRDEP, 2121 UFSTOVFS(ip->i_ump)); 2122 newindirdep->ir_state = ATTACHED; 2123 if (ip->i_ump->um_fstype == UFS1) 2124 newindirdep->ir_state |= UFS1FMT; 2125 LIST_INIT(&newindirdep->ir_deplisthd); 2126 LIST_INIT(&newindirdep->ir_donehd); 2127 if (bp->b_blkno == bp->b_lblkno) { 2128 ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp, 2129 NULL, NULL); 2130 bp->b_blkno = blkno; 2131 } 2132 newindirdep->ir_savebp = 2133 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0); 2134 BUF_KERNPROC(newindirdep->ir_savebp); 2135 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 2136 ACQUIRE_LOCK(&lk); 2137 } 2138 } 2139 2140 /* 2141 * Block de-allocation dependencies. 2142 * 2143 * When blocks are de-allocated, the on-disk pointers must be nullified before 2144 * the blocks are made available for use by other files. (The true 2145 * requirement is that old pointers must be nullified before new on-disk 2146 * pointers are set. We chose this slightly more stringent requirement to 2147 * reduce complexity.) Our implementation handles this dependency by updating 2148 * the inode (or indirect block) appropriately but delaying the actual block 2149 * de-allocation (i.e., freemap and free space count manipulation) until 2150 * after the updated versions reach stable storage. After the disk is 2151 * updated, the blocks can be safely de-allocated whenever it is convenient. 2152 * This implementation handles only the common case of reducing a file's 2153 * length to zero. Other cases are handled by the conventional synchronous 2154 * write approach. 2155 * 2156 * The ffs implementation with which we worked double-checks 2157 * the state of the block pointers and file size as it reduces 2158 * a file's length. Some of this code is replicated here in our 2159 * soft updates implementation. The freeblks->fb_chkcnt field is 2160 * used to transfer a part of this information to the procedure 2161 * that eventually de-allocates the blocks. 2162 * 2163 * This routine should be called from the routine that shortens 2164 * a file's length, before the inode's size or block pointers 2165 * are modified. It will save the block pointer information for 2166 * later release and zero the inode so that the calling routine 2167 * can release it. 2168 */ 2169 void 2170 softdep_setup_freeblocks(ip, length, flags) 2171 struct inode *ip; /* The inode whose length is to be reduced */ 2172 off_t length; /* The new length for the file */ 2173 int flags; /* IO_EXT and/or IO_NORMAL */ 2174 { 2175 struct freeblks *freeblks; 2176 struct inodedep *inodedep; 2177 struct allocdirect *adp; 2178 struct vnode *vp; 2179 struct buf *bp; 2180 struct fs *fs; 2181 ufs2_daddr_t extblocks, datablocks; 2182 struct mount *mp; 2183 int i, delay, error; 2184 2185 fs = ip->i_fs; 2186 mp = UFSTOVFS(ip->i_ump); 2187 if (length != 0) 2188 panic("softdep_setup_freeblocks: non-zero length"); 2189 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks), 2190 M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO); 2191 workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp); 2192 freeblks->fb_state = ATTACHED; 2193 freeblks->fb_uid = ip->i_uid; 2194 freeblks->fb_previousinum = ip->i_number; 2195 freeblks->fb_devvp = ip->i_devvp; 2196 extblocks = 0; 2197 if (fs->fs_magic == FS_UFS2_MAGIC) 2198 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 2199 datablocks = DIP(ip, i_blocks) - extblocks; 2200 if ((flags & IO_NORMAL) == 0) { 2201 freeblks->fb_oldsize = 0; 2202 freeblks->fb_chkcnt = 0; 2203 } else { 2204 freeblks->fb_oldsize = ip->i_size; 2205 ip->i_size = 0; 2206 DIP_SET(ip, i_size, 0); 2207 freeblks->fb_chkcnt = datablocks; 2208 for (i = 0; i < NDADDR; i++) { 2209 freeblks->fb_dblks[i] = DIP(ip, i_db[i]); 2210 DIP_SET(ip, i_db[i], 0); 2211 } 2212 for (i = 0; i < NIADDR; i++) { 2213 freeblks->fb_iblks[i] = DIP(ip, i_ib[i]); 2214 DIP_SET(ip, i_ib[i], 0); 2215 } 2216 /* 2217 * If the file was removed, then the space being freed was 2218 * accounted for then (see softdep_filereleased()). If the 2219 * file is merely being truncated, then we account for it now. 2220 */ 2221 if ((ip->i_flag & IN_SPACECOUNTED) == 0) { 2222 UFS_LOCK(ip->i_ump); 2223 fs->fs_pendingblocks += datablocks; 2224 UFS_UNLOCK(ip->i_ump); 2225 } 2226 } 2227 if ((flags & IO_EXT) == 0) { 2228 freeblks->fb_oldextsize = 0; 2229 } else { 2230 freeblks->fb_oldextsize = ip->i_din2->di_extsize; 2231 ip->i_din2->di_extsize = 0; 2232 freeblks->fb_chkcnt += extblocks; 2233 for (i = 0; i < NXADDR; i++) { 2234 freeblks->fb_eblks[i] = ip->i_din2->di_extb[i]; 2235 ip->i_din2->di_extb[i] = 0; 2236 } 2237 } 2238 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - freeblks->fb_chkcnt); 2239 /* 2240 * Push the zero'ed inode to to its disk buffer so that we are free 2241 * to delete its dependencies below. Once the dependencies are gone 2242 * the buffer can be safely released. 2243 */ 2244 if ((error = bread(ip->i_devvp, 2245 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 2246 (int)fs->fs_bsize, NOCRED, &bp)) != 0) { 2247 brelse(bp); 2248 softdep_error("softdep_setup_freeblocks", error); 2249 } 2250 if (ip->i_ump->um_fstype == UFS1) 2251 *((struct ufs1_dinode *)bp->b_data + 2252 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1; 2253 else 2254 *((struct ufs2_dinode *)bp->b_data + 2255 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2; 2256 /* 2257 * Find and eliminate any inode dependencies. 2258 */ 2259 ACQUIRE_LOCK(&lk); 2260 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 2261 if ((inodedep->id_state & IOSTARTED) != 0) 2262 panic("softdep_setup_freeblocks: inode busy"); 2263 /* 2264 * Add the freeblks structure to the list of operations that 2265 * must await the zero'ed inode being written to disk. If we 2266 * still have a bitmap dependency (delay == 0), then the inode 2267 * has never been written to disk, so we can process the 2268 * freeblks below once we have deleted the dependencies. 2269 */ 2270 delay = (inodedep->id_state & DEPCOMPLETE); 2271 if (delay) 2272 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list); 2273 /* 2274 * Because the file length has been truncated to zero, any 2275 * pending block allocation dependency structures associated 2276 * with this inode are obsolete and can simply be de-allocated. 2277 * We must first merge the two dependency lists to get rid of 2278 * any duplicate freefrag structures, then purge the merged list. 2279 * If we still have a bitmap dependency, then the inode has never 2280 * been written to disk, so we can free any fragments without delay. 2281 */ 2282 if (flags & IO_NORMAL) { 2283 merge_inode_lists(&inodedep->id_newinoupdt, 2284 &inodedep->id_inoupdt); 2285 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 2286 free_allocdirect(&inodedep->id_inoupdt, adp, delay); 2287 } 2288 if (flags & IO_EXT) { 2289 merge_inode_lists(&inodedep->id_newextupdt, 2290 &inodedep->id_extupdt); 2291 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0) 2292 free_allocdirect(&inodedep->id_extupdt, adp, delay); 2293 } 2294 FREE_LOCK(&lk); 2295 bdwrite(bp); 2296 /* 2297 * We must wait for any I/O in progress to finish so that 2298 * all potential buffers on the dirty list will be visible. 2299 * Once they are all there, walk the list and get rid of 2300 * any dependencies. 2301 */ 2302 vp = ITOV(ip); 2303 VI_LOCK(vp); 2304 drain_output(vp); 2305 restart: 2306 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 2307 if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) || 2308 ((flags & IO_NORMAL) == 0 && 2309 (bp->b_xflags & BX_ALTDATA) == 0)) 2310 continue; 2311 if ((bp = getdirtybuf(bp, VI_MTX(vp), MNT_WAIT)) == NULL) 2312 goto restart; 2313 VI_UNLOCK(vp); 2314 ACQUIRE_LOCK(&lk); 2315 (void) inodedep_lookup(mp, ip->i_number, 0, &inodedep); 2316 deallocate_dependencies(bp, inodedep); 2317 FREE_LOCK(&lk); 2318 bp->b_flags |= B_INVAL | B_NOCACHE; 2319 brelse(bp); 2320 VI_LOCK(vp); 2321 goto restart; 2322 } 2323 VI_UNLOCK(vp); 2324 ACQUIRE_LOCK(&lk); 2325 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 2326 (void) free_inodedep(inodedep); 2327 2328 if(delay) { 2329 freeblks->fb_state |= DEPCOMPLETE; 2330 /* 2331 * If the inode with zeroed block pointers is now on disk 2332 * we can start freeing blocks. Add freeblks to the worklist 2333 * instead of calling handle_workitem_freeblocks directly as 2334 * it is more likely that additional IO is needed to complete 2335 * the request here than in the !delay case. 2336 */ 2337 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 2338 add_to_worklist(&freeblks->fb_list); 2339 } 2340 2341 FREE_LOCK(&lk); 2342 /* 2343 * If the inode has never been written to disk (delay == 0), 2344 * then we can process the freeblks now that we have deleted 2345 * the dependencies. 2346 */ 2347 if (!delay) 2348 handle_workitem_freeblocks(freeblks, 0); 2349 } 2350 2351 /* 2352 * Reclaim any dependency structures from a buffer that is about to 2353 * be reallocated to a new vnode. The buffer must be locked, thus, 2354 * no I/O completion operations can occur while we are manipulating 2355 * its associated dependencies. The mutex is held so that other I/O's 2356 * associated with related dependencies do not occur. 2357 */ 2358 static void 2359 deallocate_dependencies(bp, inodedep) 2360 struct buf *bp; 2361 struct inodedep *inodedep; 2362 { 2363 struct worklist *wk; 2364 struct indirdep *indirdep; 2365 struct allocindir *aip; 2366 struct pagedep *pagedep; 2367 struct dirrem *dirrem; 2368 struct diradd *dap; 2369 int i; 2370 2371 mtx_assert(&lk, MA_OWNED); 2372 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2373 switch (wk->wk_type) { 2374 2375 case D_INDIRDEP: 2376 indirdep = WK_INDIRDEP(wk); 2377 /* 2378 * None of the indirect pointers will ever be visible, 2379 * so they can simply be tossed. GOINGAWAY ensures 2380 * that allocated pointers will be saved in the buffer 2381 * cache until they are freed. Note that they will 2382 * only be able to be found by their physical address 2383 * since the inode mapping the logical address will 2384 * be gone. The save buffer used for the safe copy 2385 * was allocated in setup_allocindir_phase2 using 2386 * the physical address so it could be used for this 2387 * purpose. Hence we swap the safe copy with the real 2388 * copy, allowing the safe copy to be freed and holding 2389 * on to the real copy for later use in indir_trunc. 2390 */ 2391 if (indirdep->ir_state & GOINGAWAY) 2392 panic("deallocate_dependencies: already gone"); 2393 indirdep->ir_state |= GOINGAWAY; 2394 VFSTOUFS(bp->b_vp->v_mount)->um_numindirdeps += 1; 2395 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 2396 free_allocindir(aip, inodedep); 2397 if (bp->b_lblkno >= 0 || 2398 bp->b_blkno != indirdep->ir_savebp->b_lblkno) 2399 panic("deallocate_dependencies: not indir"); 2400 bcopy(bp->b_data, indirdep->ir_savebp->b_data, 2401 bp->b_bcount); 2402 WORKLIST_REMOVE(wk); 2403 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk); 2404 continue; 2405 2406 case D_PAGEDEP: 2407 pagedep = WK_PAGEDEP(wk); 2408 /* 2409 * None of the directory additions will ever be 2410 * visible, so they can simply be tossed. 2411 */ 2412 for (i = 0; i < DAHASHSZ; i++) 2413 while ((dap = 2414 LIST_FIRST(&pagedep->pd_diraddhd[i]))) 2415 free_diradd(dap); 2416 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0) 2417 free_diradd(dap); 2418 /* 2419 * Copy any directory remove dependencies to the list 2420 * to be processed after the zero'ed inode is written. 2421 * If the inode has already been written, then they 2422 * can be dumped directly onto the work list. 2423 */ 2424 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 2425 LIST_REMOVE(dirrem, dm_next); 2426 dirrem->dm_dirinum = pagedep->pd_ino; 2427 if (inodedep == NULL || 2428 (inodedep->id_state & ALLCOMPLETE) == 2429 ALLCOMPLETE) 2430 add_to_worklist(&dirrem->dm_list); 2431 else 2432 WORKLIST_INSERT(&inodedep->id_bufwait, 2433 &dirrem->dm_list); 2434 } 2435 if ((pagedep->pd_state & NEWBLOCK) != 0) { 2436 LIST_FOREACH(wk, &inodedep->id_bufwait, wk_list) 2437 if (wk->wk_type == D_NEWDIRBLK && 2438 WK_NEWDIRBLK(wk)->db_pagedep == 2439 pagedep) 2440 break; 2441 if (wk != NULL) { 2442 WORKLIST_REMOVE(wk); 2443 free_newdirblk(WK_NEWDIRBLK(wk)); 2444 } else 2445 panic("deallocate_dependencies: " 2446 "lost pagedep"); 2447 } 2448 WORKLIST_REMOVE(&pagedep->pd_list); 2449 LIST_REMOVE(pagedep, pd_hash); 2450 WORKITEM_FREE(pagedep, D_PAGEDEP); 2451 continue; 2452 2453 case D_ALLOCINDIR: 2454 free_allocindir(WK_ALLOCINDIR(wk), inodedep); 2455 continue; 2456 2457 case D_ALLOCDIRECT: 2458 case D_INODEDEP: 2459 panic("deallocate_dependencies: Unexpected type %s", 2460 TYPENAME(wk->wk_type)); 2461 /* NOTREACHED */ 2462 2463 default: 2464 panic("deallocate_dependencies: Unknown type %s", 2465 TYPENAME(wk->wk_type)); 2466 /* NOTREACHED */ 2467 } 2468 } 2469 } 2470 2471 /* 2472 * Free an allocdirect. Generate a new freefrag work request if appropriate. 2473 * This routine must be called with splbio interrupts blocked. 2474 */ 2475 static void 2476 free_allocdirect(adphead, adp, delay) 2477 struct allocdirectlst *adphead; 2478 struct allocdirect *adp; 2479 int delay; 2480 { 2481 struct newdirblk *newdirblk; 2482 struct worklist *wk; 2483 2484 mtx_assert(&lk, MA_OWNED); 2485 if ((adp->ad_state & DEPCOMPLETE) == 0) 2486 LIST_REMOVE(adp, ad_deps); 2487 TAILQ_REMOVE(adphead, adp, ad_next); 2488 if ((adp->ad_state & COMPLETE) == 0) 2489 WORKLIST_REMOVE(&adp->ad_list); 2490 if (adp->ad_freefrag != NULL) { 2491 if (delay) 2492 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 2493 &adp->ad_freefrag->ff_list); 2494 else 2495 add_to_worklist(&adp->ad_freefrag->ff_list); 2496 } 2497 if ((wk = LIST_FIRST(&adp->ad_newdirblk)) != NULL) { 2498 newdirblk = WK_NEWDIRBLK(wk); 2499 WORKLIST_REMOVE(&newdirblk->db_list); 2500 if (LIST_FIRST(&adp->ad_newdirblk) != NULL) 2501 panic("free_allocdirect: extra newdirblk"); 2502 if (delay) 2503 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 2504 &newdirblk->db_list); 2505 else 2506 free_newdirblk(newdirblk); 2507 } 2508 WORKITEM_FREE(adp, D_ALLOCDIRECT); 2509 } 2510 2511 /* 2512 * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep. 2513 * This routine must be called with splbio interrupts blocked. 2514 */ 2515 static void 2516 free_newdirblk(newdirblk) 2517 struct newdirblk *newdirblk; 2518 { 2519 struct pagedep *pagedep; 2520 struct diradd *dap; 2521 int i; 2522 2523 mtx_assert(&lk, MA_OWNED); 2524 /* 2525 * If the pagedep is still linked onto the directory buffer 2526 * dependency chain, then some of the entries on the 2527 * pd_pendinghd list may not be committed to disk yet. In 2528 * this case, we will simply clear the NEWBLOCK flag and 2529 * let the pd_pendinghd list be processed when the pagedep 2530 * is next written. If the pagedep is no longer on the buffer 2531 * dependency chain, then all the entries on the pd_pending 2532 * list are committed to disk and we can free them here. 2533 */ 2534 pagedep = newdirblk->db_pagedep; 2535 pagedep->pd_state &= ~NEWBLOCK; 2536 if ((pagedep->pd_state & ONWORKLIST) == 0) 2537 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 2538 free_diradd(dap); 2539 /* 2540 * If no dependencies remain, the pagedep will be freed. 2541 */ 2542 for (i = 0; i < DAHASHSZ; i++) 2543 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL) 2544 break; 2545 if (i == DAHASHSZ && (pagedep->pd_state & ONWORKLIST) == 0) { 2546 LIST_REMOVE(pagedep, pd_hash); 2547 WORKITEM_FREE(pagedep, D_PAGEDEP); 2548 } 2549 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 2550 } 2551 2552 /* 2553 * Prepare an inode to be freed. The actual free operation is not 2554 * done until the zero'ed inode has been written to disk. 2555 */ 2556 void 2557 softdep_freefile(pvp, ino, mode) 2558 struct vnode *pvp; 2559 ino_t ino; 2560 int mode; 2561 { 2562 struct inode *ip = VTOI(pvp); 2563 struct inodedep *inodedep; 2564 struct freefile *freefile; 2565 2566 /* 2567 * This sets up the inode de-allocation dependency. 2568 */ 2569 MALLOC(freefile, struct freefile *, sizeof(struct freefile), 2570 M_FREEFILE, M_SOFTDEP_FLAGS); 2571 workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount); 2572 freefile->fx_mode = mode; 2573 freefile->fx_oldinum = ino; 2574 freefile->fx_devvp = ip->i_devvp; 2575 if ((ip->i_flag & IN_SPACECOUNTED) == 0) { 2576 UFS_LOCK(ip->i_ump); 2577 ip->i_fs->fs_pendinginodes += 1; 2578 UFS_UNLOCK(ip->i_ump); 2579 } 2580 2581 /* 2582 * If the inodedep does not exist, then the zero'ed inode has 2583 * been written to disk. If the allocated inode has never been 2584 * written to disk, then the on-disk inode is zero'ed. In either 2585 * case we can free the file immediately. 2586 */ 2587 ACQUIRE_LOCK(&lk); 2588 if (inodedep_lookup(pvp->v_mount, ino, 0, &inodedep) == 0 || 2589 check_inode_unwritten(inodedep)) { 2590 FREE_LOCK(&lk); 2591 handle_workitem_freefile(freefile); 2592 return; 2593 } 2594 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 2595 FREE_LOCK(&lk); 2596 } 2597 2598 /* 2599 * Check to see if an inode has never been written to disk. If 2600 * so free the inodedep and return success, otherwise return failure. 2601 * This routine must be called with splbio interrupts blocked. 2602 * 2603 * If we still have a bitmap dependency, then the inode has never 2604 * been written to disk. Drop the dependency as it is no longer 2605 * necessary since the inode is being deallocated. We set the 2606 * ALLCOMPLETE flags since the bitmap now properly shows that the 2607 * inode is not allocated. Even if the inode is actively being 2608 * written, it has been rolled back to its zero'ed state, so we 2609 * are ensured that a zero inode is what is on the disk. For short 2610 * lived files, this change will usually result in removing all the 2611 * dependencies from the inode so that it can be freed immediately. 2612 */ 2613 static int 2614 check_inode_unwritten(inodedep) 2615 struct inodedep *inodedep; 2616 { 2617 2618 mtx_assert(&lk, MA_OWNED); 2619 if ((inodedep->id_state & DEPCOMPLETE) != 0 || 2620 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2621 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2622 LIST_FIRST(&inodedep->id_inowait) != NULL || 2623 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2624 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2625 TAILQ_FIRST(&inodedep->id_extupdt) != NULL || 2626 TAILQ_FIRST(&inodedep->id_newextupdt) != NULL || 2627 inodedep->id_nlinkdelta != 0) 2628 return (0); 2629 2630 /* 2631 * Another process might be in initiate_write_inodeblock_ufs[12] 2632 * trying to allocate memory without holding "Softdep Lock". 2633 */ 2634 if ((inodedep->id_state & IOSTARTED) != 0 && 2635 inodedep->id_savedino1 == NULL) 2636 return (0); 2637 2638 inodedep->id_state |= ALLCOMPLETE; 2639 LIST_REMOVE(inodedep, id_deps); 2640 inodedep->id_buf = NULL; 2641 if (inodedep->id_state & ONWORKLIST) 2642 WORKLIST_REMOVE(&inodedep->id_list); 2643 if (inodedep->id_savedino1 != NULL) { 2644 FREE(inodedep->id_savedino1, M_SAVEDINO); 2645 inodedep->id_savedino1 = NULL; 2646 } 2647 if (free_inodedep(inodedep) == 0) 2648 panic("check_inode_unwritten: busy inode"); 2649 return (1); 2650 } 2651 2652 /* 2653 * Try to free an inodedep structure. Return 1 if it could be freed. 2654 */ 2655 static int 2656 free_inodedep(inodedep) 2657 struct inodedep *inodedep; 2658 { 2659 2660 mtx_assert(&lk, MA_OWNED); 2661 if ((inodedep->id_state & ONWORKLIST) != 0 || 2662 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 2663 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2664 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2665 LIST_FIRST(&inodedep->id_inowait) != NULL || 2666 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2667 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2668 TAILQ_FIRST(&inodedep->id_extupdt) != NULL || 2669 TAILQ_FIRST(&inodedep->id_newextupdt) != NULL || 2670 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino1 != NULL) 2671 return (0); 2672 LIST_REMOVE(inodedep, id_hash); 2673 WORKITEM_FREE(inodedep, D_INODEDEP); 2674 num_inodedep -= 1; 2675 return (1); 2676 } 2677 2678 /* 2679 * This workitem routine performs the block de-allocation. 2680 * The workitem is added to the pending list after the updated 2681 * inode block has been written to disk. As mentioned above, 2682 * checks regarding the number of blocks de-allocated (compared 2683 * to the number of blocks allocated for the file) are also 2684 * performed in this function. 2685 */ 2686 static void 2687 handle_workitem_freeblocks(freeblks, flags) 2688 struct freeblks *freeblks; 2689 int flags; 2690 { 2691 struct inode *ip; 2692 struct vnode *vp; 2693 struct fs *fs; 2694 struct ufsmount *ump; 2695 int i, nblocks, level, bsize; 2696 ufs2_daddr_t bn, blocksreleased = 0; 2697 int error, allerror = 0; 2698 ufs_lbn_t baselbns[NIADDR], tmpval; 2699 int fs_pendingblocks; 2700 2701 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 2702 fs = ump->um_fs; 2703 fs_pendingblocks = 0; 2704 tmpval = 1; 2705 baselbns[0] = NDADDR; 2706 for (i = 1; i < NIADDR; i++) { 2707 tmpval *= NINDIR(fs); 2708 baselbns[i] = baselbns[i - 1] + tmpval; 2709 } 2710 nblocks = btodb(fs->fs_bsize); 2711 blocksreleased = 0; 2712 /* 2713 * Release all extended attribute blocks or frags. 2714 */ 2715 if (freeblks->fb_oldextsize > 0) { 2716 for (i = (NXADDR - 1); i >= 0; i--) { 2717 if ((bn = freeblks->fb_eblks[i]) == 0) 2718 continue; 2719 bsize = sblksize(fs, freeblks->fb_oldextsize, i); 2720 ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, bsize, 2721 freeblks->fb_previousinum); 2722 blocksreleased += btodb(bsize); 2723 } 2724 } 2725 /* 2726 * Release all data blocks or frags. 2727 */ 2728 if (freeblks->fb_oldsize > 0) { 2729 /* 2730 * Indirect blocks first. 2731 */ 2732 for (level = (NIADDR - 1); level >= 0; level--) { 2733 if ((bn = freeblks->fb_iblks[level]) == 0) 2734 continue; 2735 if ((error = indir_trunc(freeblks, fsbtodb(fs, bn), 2736 level, baselbns[level], &blocksreleased)) == 0) 2737 allerror = error; 2738 ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, 2739 fs->fs_bsize, freeblks->fb_previousinum); 2740 fs_pendingblocks += nblocks; 2741 blocksreleased += nblocks; 2742 } 2743 /* 2744 * All direct blocks or frags. 2745 */ 2746 for (i = (NDADDR - 1); i >= 0; i--) { 2747 if ((bn = freeblks->fb_dblks[i]) == 0) 2748 continue; 2749 bsize = sblksize(fs, freeblks->fb_oldsize, i); 2750 ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, bsize, 2751 freeblks->fb_previousinum); 2752 fs_pendingblocks += btodb(bsize); 2753 blocksreleased += btodb(bsize); 2754 } 2755 } 2756 UFS_LOCK(ump); 2757 fs->fs_pendingblocks -= fs_pendingblocks; 2758 UFS_UNLOCK(ump); 2759 /* 2760 * If we still have not finished background cleanup, then check 2761 * to see if the block count needs to be adjusted. 2762 */ 2763 if (freeblks->fb_chkcnt != blocksreleased && 2764 (fs->fs_flags & FS_UNCLEAN) != 0 && 2765 ffs_vget(freeblks->fb_list.wk_mp, freeblks->fb_previousinum, 2766 (flags & LK_NOWAIT) | LK_EXCLUSIVE, &vp) == 0) { 2767 ip = VTOI(vp); 2768 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + \ 2769 freeblks->fb_chkcnt - blocksreleased); 2770 ip->i_flag |= IN_CHANGE; 2771 vput(vp); 2772 } 2773 2774 #ifdef DIAGNOSTIC 2775 if (freeblks->fb_chkcnt != blocksreleased && 2776 ((fs->fs_flags & FS_UNCLEAN) == 0 || (flags & LK_NOWAIT) != 0)) 2777 printf("handle_workitem_freeblocks: block count\n"); 2778 if (allerror) 2779 softdep_error("handle_workitem_freeblks", allerror); 2780 #endif /* DIAGNOSTIC */ 2781 2782 ACQUIRE_LOCK(&lk); 2783 WORKITEM_FREE(freeblks, D_FREEBLKS); 2784 FREE_LOCK(&lk); 2785 } 2786 2787 /* 2788 * Release blocks associated with the inode ip and stored in the indirect 2789 * block dbn. If level is greater than SINGLE, the block is an indirect block 2790 * and recursive calls to indirtrunc must be used to cleanse other indirect 2791 * blocks. 2792 */ 2793 static int 2794 indir_trunc(freeblks, dbn, level, lbn, countp) 2795 struct freeblks *freeblks; 2796 ufs2_daddr_t dbn; 2797 int level; 2798 ufs_lbn_t lbn; 2799 ufs2_daddr_t *countp; 2800 { 2801 struct buf *bp; 2802 struct fs *fs; 2803 struct worklist *wk; 2804 struct indirdep *indirdep; 2805 struct ufsmount *ump; 2806 ufs1_daddr_t *bap1 = 0; 2807 ufs2_daddr_t nb, *bap2 = 0; 2808 ufs_lbn_t lbnadd; 2809 int i, nblocks, ufs1fmt; 2810 int error, allerror = 0; 2811 int fs_pendingblocks; 2812 2813 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 2814 fs = ump->um_fs; 2815 fs_pendingblocks = 0; 2816 lbnadd = 1; 2817 for (i = level; i > 0; i--) 2818 lbnadd *= NINDIR(fs); 2819 /* 2820 * Get buffer of block pointers to be freed. This routine is not 2821 * called until the zero'ed inode has been written, so it is safe 2822 * to free blocks as they are encountered. Because the inode has 2823 * been zero'ed, calls to bmap on these blocks will fail. So, we 2824 * have to use the on-disk address and the block device for the 2825 * filesystem to look them up. If the file was deleted before its 2826 * indirect blocks were all written to disk, the routine that set 2827 * us up (deallocate_dependencies) will have arranged to leave 2828 * a complete copy of the indirect block in memory for our use. 2829 * Otherwise we have to read the blocks in from the disk. 2830 */ 2831 #ifdef notyet 2832 bp = getblk(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 0, 0, 2833 GB_NOCREAT); 2834 #else 2835 bp = incore(&freeblks->fb_devvp->v_bufobj, dbn); 2836 #endif 2837 ACQUIRE_LOCK(&lk); 2838 if (bp != NULL && (wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2839 if (wk->wk_type != D_INDIRDEP || 2840 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp || 2841 (indirdep->ir_state & GOINGAWAY) == 0) 2842 panic("indir_trunc: lost indirdep"); 2843 WORKLIST_REMOVE(wk); 2844 WORKITEM_FREE(indirdep, D_INDIRDEP); 2845 if (LIST_FIRST(&bp->b_dep) != NULL) 2846 panic("indir_trunc: dangling dep"); 2847 ump->um_numindirdeps -= 1; 2848 FREE_LOCK(&lk); 2849 } else { 2850 #ifdef notyet 2851 if (bp) 2852 brelse(bp); 2853 #endif 2854 FREE_LOCK(&lk); 2855 error = bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 2856 NOCRED, &bp); 2857 if (error) { 2858 brelse(bp); 2859 return (error); 2860 } 2861 } 2862 /* 2863 * Recursively free indirect blocks. 2864 */ 2865 if (ump->um_fstype == UFS1) { 2866 ufs1fmt = 1; 2867 bap1 = (ufs1_daddr_t *)bp->b_data; 2868 } else { 2869 ufs1fmt = 0; 2870 bap2 = (ufs2_daddr_t *)bp->b_data; 2871 } 2872 nblocks = btodb(fs->fs_bsize); 2873 for (i = NINDIR(fs) - 1; i >= 0; i--) { 2874 if (ufs1fmt) 2875 nb = bap1[i]; 2876 else 2877 nb = bap2[i]; 2878 if (nb == 0) 2879 continue; 2880 if (level != 0) { 2881 if ((error = indir_trunc(freeblks, fsbtodb(fs, nb), 2882 level - 1, lbn + (i * lbnadd), countp)) != 0) 2883 allerror = error; 2884 } 2885 ffs_blkfree(ump, fs, freeblks->fb_devvp, nb, fs->fs_bsize, 2886 freeblks->fb_previousinum); 2887 fs_pendingblocks += nblocks; 2888 *countp += nblocks; 2889 } 2890 UFS_LOCK(ump); 2891 fs->fs_pendingblocks -= fs_pendingblocks; 2892 UFS_UNLOCK(ump); 2893 bp->b_flags |= B_INVAL | B_NOCACHE; 2894 brelse(bp); 2895 return (allerror); 2896 } 2897 2898 /* 2899 * Free an allocindir. 2900 * This routine must be called with splbio interrupts blocked. 2901 */ 2902 static void 2903 free_allocindir(aip, inodedep) 2904 struct allocindir *aip; 2905 struct inodedep *inodedep; 2906 { 2907 struct freefrag *freefrag; 2908 2909 mtx_assert(&lk, MA_OWNED); 2910 if ((aip->ai_state & DEPCOMPLETE) == 0) 2911 LIST_REMOVE(aip, ai_deps); 2912 if (aip->ai_state & ONWORKLIST) 2913 WORKLIST_REMOVE(&aip->ai_list); 2914 LIST_REMOVE(aip, ai_next); 2915 if ((freefrag = aip->ai_freefrag) != NULL) { 2916 if (inodedep == NULL) 2917 add_to_worklist(&freefrag->ff_list); 2918 else 2919 WORKLIST_INSERT(&inodedep->id_bufwait, 2920 &freefrag->ff_list); 2921 } 2922 WORKITEM_FREE(aip, D_ALLOCINDIR); 2923 } 2924 2925 /* 2926 * Directory entry addition dependencies. 2927 * 2928 * When adding a new directory entry, the inode (with its incremented link 2929 * count) must be written to disk before the directory entry's pointer to it. 2930 * Also, if the inode is newly allocated, the corresponding freemap must be 2931 * updated (on disk) before the directory entry's pointer. These requirements 2932 * are met via undo/redo on the directory entry's pointer, which consists 2933 * simply of the inode number. 2934 * 2935 * As directory entries are added and deleted, the free space within a 2936 * directory block can become fragmented. The ufs filesystem will compact 2937 * a fragmented directory block to make space for a new entry. When this 2938 * occurs, the offsets of previously added entries change. Any "diradd" 2939 * dependency structures corresponding to these entries must be updated with 2940 * the new offsets. 2941 */ 2942 2943 /* 2944 * This routine is called after the in-memory inode's link 2945 * count has been incremented, but before the directory entry's 2946 * pointer to the inode has been set. 2947 */ 2948 int 2949 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 2950 struct buf *bp; /* buffer containing directory block */ 2951 struct inode *dp; /* inode for directory */ 2952 off_t diroffset; /* offset of new entry in directory */ 2953 ino_t newinum; /* inode referenced by new directory entry */ 2954 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 2955 int isnewblk; /* entry is in a newly allocated block */ 2956 { 2957 int offset; /* offset of new entry within directory block */ 2958 ufs_lbn_t lbn; /* block in directory containing new entry */ 2959 struct fs *fs; 2960 struct diradd *dap; 2961 struct allocdirect *adp; 2962 struct pagedep *pagedep; 2963 struct inodedep *inodedep; 2964 struct newdirblk *newdirblk = 0; 2965 struct mkdir *mkdir1, *mkdir2; 2966 struct mount *mp; 2967 2968 /* 2969 * Whiteouts have no dependencies. 2970 */ 2971 if (newinum == WINO) { 2972 if (newdirbp != NULL) 2973 bdwrite(newdirbp); 2974 return (0); 2975 } 2976 mp = UFSTOVFS(dp->i_ump); 2977 fs = dp->i_fs; 2978 lbn = lblkno(fs, diroffset); 2979 offset = blkoff(fs, diroffset); 2980 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, 2981 M_SOFTDEP_FLAGS|M_ZERO); 2982 workitem_alloc(&dap->da_list, D_DIRADD, mp); 2983 dap->da_offset = offset; 2984 dap->da_newinum = newinum; 2985 dap->da_state = ATTACHED; 2986 if (isnewblk && lbn < NDADDR && fragoff(fs, diroffset) == 0) { 2987 MALLOC(newdirblk, struct newdirblk *, sizeof(struct newdirblk), 2988 M_NEWDIRBLK, M_SOFTDEP_FLAGS); 2989 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 2990 } 2991 if (newdirbp == NULL) { 2992 dap->da_state |= DEPCOMPLETE; 2993 ACQUIRE_LOCK(&lk); 2994 } else { 2995 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 2996 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2997 M_SOFTDEP_FLAGS); 2998 workitem_alloc(&mkdir1->md_list, D_MKDIR, mp); 2999 mkdir1->md_state = MKDIR_BODY; 3000 mkdir1->md_diradd = dap; 3001 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 3002 M_SOFTDEP_FLAGS); 3003 workitem_alloc(&mkdir2->md_list, D_MKDIR, mp); 3004 mkdir2->md_state = MKDIR_PARENT; 3005 mkdir2->md_diradd = dap; 3006 /* 3007 * Dependency on "." and ".." being written to disk. 3008 */ 3009 mkdir1->md_buf = newdirbp; 3010 ACQUIRE_LOCK(&lk); 3011 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 3012 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list); 3013 FREE_LOCK(&lk); 3014 bdwrite(newdirbp); 3015 /* 3016 * Dependency on link count increase for parent directory 3017 */ 3018 ACQUIRE_LOCK(&lk); 3019 if (inodedep_lookup(mp, dp->i_number, 0, &inodedep) == 0 3020 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 3021 dap->da_state &= ~MKDIR_PARENT; 3022 WORKITEM_FREE(mkdir2, D_MKDIR); 3023 } else { 3024 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 3025 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 3026 } 3027 } 3028 /* 3029 * Link into parent directory pagedep to await its being written. 3030 */ 3031 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 3032 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 3033 dap->da_pagedep = pagedep; 3034 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 3035 da_pdlist); 3036 /* 3037 * Link into its inodedep. Put it on the id_bufwait list if the inode 3038 * is not yet written. If it is written, do the post-inode write 3039 * processing to put it on the id_pendinghd list. 3040 */ 3041 (void) inodedep_lookup(mp, newinum, DEPALLOC, &inodedep); 3042 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 3043 diradd_inode_written(dap, inodedep); 3044 else 3045 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 3046 if (isnewblk) { 3047 /* 3048 * Directories growing into indirect blocks are rare 3049 * enough and the frequency of new block allocation 3050 * in those cases even more rare, that we choose not 3051 * to bother tracking them. Rather we simply force the 3052 * new directory entry to disk. 3053 */ 3054 if (lbn >= NDADDR) { 3055 FREE_LOCK(&lk); 3056 /* 3057 * We only have a new allocation when at the 3058 * beginning of a new block, not when we are 3059 * expanding into an existing block. 3060 */ 3061 if (blkoff(fs, diroffset) == 0) 3062 return (1); 3063 return (0); 3064 } 3065 /* 3066 * We only have a new allocation when at the beginning 3067 * of a new fragment, not when we are expanding into an 3068 * existing fragment. Also, there is nothing to do if we 3069 * are already tracking this block. 3070 */ 3071 if (fragoff(fs, diroffset) != 0) { 3072 FREE_LOCK(&lk); 3073 return (0); 3074 } 3075 if ((pagedep->pd_state & NEWBLOCK) != 0) { 3076 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 3077 FREE_LOCK(&lk); 3078 return (0); 3079 } 3080 /* 3081 * Find our associated allocdirect and have it track us. 3082 */ 3083 if (inodedep_lookup(mp, dp->i_number, 0, &inodedep) == 0) 3084 panic("softdep_setup_directory_add: lost inodedep"); 3085 adp = TAILQ_LAST(&inodedep->id_newinoupdt, allocdirectlst); 3086 if (adp == NULL || adp->ad_lbn != lbn) 3087 panic("softdep_setup_directory_add: lost entry"); 3088 pagedep->pd_state |= NEWBLOCK; 3089 newdirblk->db_pagedep = pagedep; 3090 WORKLIST_INSERT(&adp->ad_newdirblk, &newdirblk->db_list); 3091 } 3092 FREE_LOCK(&lk); 3093 return (0); 3094 } 3095 3096 /* 3097 * This procedure is called to change the offset of a directory 3098 * entry when compacting a directory block which must be owned 3099 * exclusively by the caller. Note that the actual entry movement 3100 * must be done in this procedure to ensure that no I/O completions 3101 * occur while the move is in progress. 3102 */ 3103 void 3104 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize) 3105 struct inode *dp; /* inode for directory */ 3106 caddr_t base; /* address of dp->i_offset */ 3107 caddr_t oldloc; /* address of old directory location */ 3108 caddr_t newloc; /* address of new directory location */ 3109 int entrysize; /* size of directory entry */ 3110 { 3111 int offset, oldoffset, newoffset; 3112 struct pagedep *pagedep; 3113 struct diradd *dap; 3114 ufs_lbn_t lbn; 3115 3116 ACQUIRE_LOCK(&lk); 3117 lbn = lblkno(dp->i_fs, dp->i_offset); 3118 offset = blkoff(dp->i_fs, dp->i_offset); 3119 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) 3120 goto done; 3121 oldoffset = offset + (oldloc - base); 3122 newoffset = offset + (newloc - base); 3123 3124 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) { 3125 if (dap->da_offset != oldoffset) 3126 continue; 3127 dap->da_offset = newoffset; 3128 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) 3129 break; 3130 LIST_REMOVE(dap, da_pdlist); 3131 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], 3132 dap, da_pdlist); 3133 break; 3134 } 3135 if (dap == NULL) { 3136 3137 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) { 3138 if (dap->da_offset == oldoffset) { 3139 dap->da_offset = newoffset; 3140 break; 3141 } 3142 } 3143 } 3144 done: 3145 bcopy(oldloc, newloc, entrysize); 3146 FREE_LOCK(&lk); 3147 } 3148 3149 /* 3150 * Free a diradd dependency structure. This routine must be called 3151 * with splbio interrupts blocked. 3152 */ 3153 static void 3154 free_diradd(dap) 3155 struct diradd *dap; 3156 { 3157 struct dirrem *dirrem; 3158 struct pagedep *pagedep; 3159 struct inodedep *inodedep; 3160 struct mkdir *mkdir, *nextmd; 3161 3162 mtx_assert(&lk, MA_OWNED); 3163 WORKLIST_REMOVE(&dap->da_list); 3164 LIST_REMOVE(dap, da_pdlist); 3165 if ((dap->da_state & DIRCHG) == 0) { 3166 pagedep = dap->da_pagedep; 3167 } else { 3168 dirrem = dap->da_previous; 3169 pagedep = dirrem->dm_pagedep; 3170 dirrem->dm_dirinum = pagedep->pd_ino; 3171 add_to_worklist(&dirrem->dm_list); 3172 } 3173 if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum, 3174 0, &inodedep) != 0) 3175 (void) free_inodedep(inodedep); 3176 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 3177 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 3178 nextmd = LIST_NEXT(mkdir, md_mkdirs); 3179 if (mkdir->md_diradd != dap) 3180 continue; 3181 dap->da_state &= ~mkdir->md_state; 3182 WORKLIST_REMOVE(&mkdir->md_list); 3183 LIST_REMOVE(mkdir, md_mkdirs); 3184 WORKITEM_FREE(mkdir, D_MKDIR); 3185 } 3186 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 3187 panic("free_diradd: unfound ref"); 3188 } 3189 WORKITEM_FREE(dap, D_DIRADD); 3190 } 3191 3192 /* 3193 * Directory entry removal dependencies. 3194 * 3195 * When removing a directory entry, the entry's inode pointer must be 3196 * zero'ed on disk before the corresponding inode's link count is decremented 3197 * (possibly freeing the inode for re-use). This dependency is handled by 3198 * updating the directory entry but delaying the inode count reduction until 3199 * after the directory block has been written to disk. After this point, the 3200 * inode count can be decremented whenever it is convenient. 3201 */ 3202 3203 /* 3204 * This routine should be called immediately after removing 3205 * a directory entry. The inode's link count should not be 3206 * decremented by the calling procedure -- the soft updates 3207 * code will do this task when it is safe. 3208 */ 3209 void 3210 softdep_setup_remove(bp, dp, ip, isrmdir) 3211 struct buf *bp; /* buffer containing directory block */ 3212 struct inode *dp; /* inode for the directory being modified */ 3213 struct inode *ip; /* inode for directory entry being removed */ 3214 int isrmdir; /* indicates if doing RMDIR */ 3215 { 3216 struct dirrem *dirrem, *prevdirrem; 3217 3218 /* 3219 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 3220 */ 3221 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 3222 3223 /* 3224 * If the COMPLETE flag is clear, then there were no active 3225 * entries and we want to roll back to a zeroed entry until 3226 * the new inode is committed to disk. If the COMPLETE flag is 3227 * set then we have deleted an entry that never made it to 3228 * disk. If the entry we deleted resulted from a name change, 3229 * then the old name still resides on disk. We cannot delete 3230 * its inode (returned to us in prevdirrem) until the zeroed 3231 * directory entry gets to disk. The new inode has never been 3232 * referenced on the disk, so can be deleted immediately. 3233 */ 3234 if ((dirrem->dm_state & COMPLETE) == 0) { 3235 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 3236 dm_next); 3237 FREE_LOCK(&lk); 3238 } else { 3239 if (prevdirrem != NULL) 3240 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 3241 prevdirrem, dm_next); 3242 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 3243 FREE_LOCK(&lk); 3244 handle_workitem_remove(dirrem, NULL); 3245 } 3246 } 3247 3248 /* 3249 * Allocate a new dirrem if appropriate and return it along with 3250 * its associated pagedep. Called without a lock, returns with lock. 3251 */ 3252 static long num_dirrem; /* number of dirrem allocated */ 3253 static struct dirrem * 3254 newdirrem(bp, dp, ip, isrmdir, prevdirremp) 3255 struct buf *bp; /* buffer containing directory block */ 3256 struct inode *dp; /* inode for the directory being modified */ 3257 struct inode *ip; /* inode for directory entry being removed */ 3258 int isrmdir; /* indicates if doing RMDIR */ 3259 struct dirrem **prevdirremp; /* previously referenced inode, if any */ 3260 { 3261 int offset; 3262 ufs_lbn_t lbn; 3263 struct diradd *dap; 3264 struct dirrem *dirrem; 3265 struct pagedep *pagedep; 3266 3267 /* 3268 * Whiteouts have no deletion dependencies. 3269 */ 3270 if (ip == NULL) 3271 panic("newdirrem: whiteout"); 3272 /* 3273 * If we are over our limit, try to improve the situation. 3274 * Limiting the number of dirrem structures will also limit 3275 * the number of freefile and freeblks structures. 3276 */ 3277 ACQUIRE_LOCK(&lk); 3278 if (num_dirrem > max_softdeps / 2) 3279 (void) request_cleanup(ITOV(dp)->v_mount, FLUSH_REMOVE); 3280 num_dirrem += 1; 3281 FREE_LOCK(&lk); 3282 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem), 3283 M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO); 3284 workitem_alloc(&dirrem->dm_list, D_DIRREM, ITOV(dp)->v_mount); 3285 dirrem->dm_state = isrmdir ? RMDIR : 0; 3286 dirrem->dm_oldinum = ip->i_number; 3287 *prevdirremp = NULL; 3288 3289 ACQUIRE_LOCK(&lk); 3290 lbn = lblkno(dp->i_fs, dp->i_offset); 3291 offset = blkoff(dp->i_fs, dp->i_offset); 3292 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 3293 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 3294 dirrem->dm_pagedep = pagedep; 3295 /* 3296 * Check for a diradd dependency for the same directory entry. 3297 * If present, then both dependencies become obsolete and can 3298 * be de-allocated. Check for an entry on both the pd_dirraddhd 3299 * list and the pd_pendinghd list. 3300 */ 3301 3302 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 3303 if (dap->da_offset == offset) 3304 break; 3305 if (dap == NULL) { 3306 3307 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 3308 if (dap->da_offset == offset) 3309 break; 3310 if (dap == NULL) 3311 return (dirrem); 3312 } 3313 /* 3314 * Must be ATTACHED at this point. 3315 */ 3316 if ((dap->da_state & ATTACHED) == 0) 3317 panic("newdirrem: not ATTACHED"); 3318 if (dap->da_newinum != ip->i_number) 3319 panic("newdirrem: inum %d should be %d", 3320 ip->i_number, dap->da_newinum); 3321 /* 3322 * If we are deleting a changed name that never made it to disk, 3323 * then return the dirrem describing the previous inode (which 3324 * represents the inode currently referenced from this entry on disk). 3325 */ 3326 if ((dap->da_state & DIRCHG) != 0) { 3327 *prevdirremp = dap->da_previous; 3328 dap->da_state &= ~DIRCHG; 3329 dap->da_pagedep = pagedep; 3330 } 3331 /* 3332 * We are deleting an entry that never made it to disk. 3333 * Mark it COMPLETE so we can delete its inode immediately. 3334 */ 3335 dirrem->dm_state |= COMPLETE; 3336 free_diradd(dap); 3337 return (dirrem); 3338 } 3339 3340 /* 3341 * Directory entry change dependencies. 3342 * 3343 * Changing an existing directory entry requires that an add operation 3344 * be completed first followed by a deletion. The semantics for the addition 3345 * are identical to the description of adding a new entry above except 3346 * that the rollback is to the old inode number rather than zero. Once 3347 * the addition dependency is completed, the removal is done as described 3348 * in the removal routine above. 3349 */ 3350 3351 /* 3352 * This routine should be called immediately after changing 3353 * a directory entry. The inode's link count should not be 3354 * decremented by the calling procedure -- the soft updates 3355 * code will perform this task when it is safe. 3356 */ 3357 void 3358 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 3359 struct buf *bp; /* buffer containing directory block */ 3360 struct inode *dp; /* inode for the directory being modified */ 3361 struct inode *ip; /* inode for directory entry being removed */ 3362 ino_t newinum; /* new inode number for changed entry */ 3363 int isrmdir; /* indicates if doing RMDIR */ 3364 { 3365 int offset; 3366 struct diradd *dap = NULL; 3367 struct dirrem *dirrem, *prevdirrem; 3368 struct pagedep *pagedep; 3369 struct inodedep *inodedep; 3370 struct mount *mp; 3371 3372 offset = blkoff(dp->i_fs, dp->i_offset); 3373 mp = UFSTOVFS(dp->i_ump); 3374 3375 /* 3376 * Whiteouts do not need diradd dependencies. 3377 */ 3378 if (newinum != WINO) { 3379 MALLOC(dap, struct diradd *, sizeof(struct diradd), 3380 M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO); 3381 workitem_alloc(&dap->da_list, D_DIRADD, mp); 3382 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 3383 dap->da_offset = offset; 3384 dap->da_newinum = newinum; 3385 } 3386 3387 /* 3388 * Allocate a new dirrem and ACQUIRE_LOCK. 3389 */ 3390 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 3391 pagedep = dirrem->dm_pagedep; 3392 /* 3393 * The possible values for isrmdir: 3394 * 0 - non-directory file rename 3395 * 1 - directory rename within same directory 3396 * inum - directory rename to new directory of given inode number 3397 * When renaming to a new directory, we are both deleting and 3398 * creating a new directory entry, so the link count on the new 3399 * directory should not change. Thus we do not need the followup 3400 * dirrem which is usually done in handle_workitem_remove. We set 3401 * the DIRCHG flag to tell handle_workitem_remove to skip the 3402 * followup dirrem. 3403 */ 3404 if (isrmdir > 1) 3405 dirrem->dm_state |= DIRCHG; 3406 3407 /* 3408 * Whiteouts have no additional dependencies, 3409 * so just put the dirrem on the correct list. 3410 */ 3411 if (newinum == WINO) { 3412 if ((dirrem->dm_state & COMPLETE) == 0) { 3413 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 3414 dm_next); 3415 } else { 3416 dirrem->dm_dirinum = pagedep->pd_ino; 3417 add_to_worklist(&dirrem->dm_list); 3418 } 3419 FREE_LOCK(&lk); 3420 return; 3421 } 3422 3423 /* 3424 * If the COMPLETE flag is clear, then there were no active 3425 * entries and we want to roll back to the previous inode until 3426 * the new inode is committed to disk. If the COMPLETE flag is 3427 * set, then we have deleted an entry that never made it to disk. 3428 * If the entry we deleted resulted from a name change, then the old 3429 * inode reference still resides on disk. Any rollback that we do 3430 * needs to be to that old inode (returned to us in prevdirrem). If 3431 * the entry we deleted resulted from a create, then there is 3432 * no entry on the disk, so we want to roll back to zero rather 3433 * than the uncommitted inode. In either of the COMPLETE cases we 3434 * want to immediately free the unwritten and unreferenced inode. 3435 */ 3436 if ((dirrem->dm_state & COMPLETE) == 0) { 3437 dap->da_previous = dirrem; 3438 } else { 3439 if (prevdirrem != NULL) { 3440 dap->da_previous = prevdirrem; 3441 } else { 3442 dap->da_state &= ~DIRCHG; 3443 dap->da_pagedep = pagedep; 3444 } 3445 dirrem->dm_dirinum = pagedep->pd_ino; 3446 add_to_worklist(&dirrem->dm_list); 3447 } 3448 /* 3449 * Link into its inodedep. Put it on the id_bufwait list if the inode 3450 * is not yet written. If it is written, do the post-inode write 3451 * processing to put it on the id_pendinghd list. 3452 */ 3453 if (inodedep_lookup(mp, newinum, DEPALLOC, &inodedep) == 0 || 3454 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 3455 dap->da_state |= COMPLETE; 3456 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3457 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 3458 } else { 3459 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 3460 dap, da_pdlist); 3461 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 3462 } 3463 FREE_LOCK(&lk); 3464 } 3465 3466 /* 3467 * Called whenever the link count on an inode is changed. 3468 * It creates an inode dependency so that the new reference(s) 3469 * to the inode cannot be committed to disk until the updated 3470 * inode has been written. 3471 */ 3472 void 3473 softdep_change_linkcnt(ip) 3474 struct inode *ip; /* the inode with the increased link count */ 3475 { 3476 struct inodedep *inodedep; 3477 3478 ACQUIRE_LOCK(&lk); 3479 (void) inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 3480 DEPALLOC, &inodedep); 3481 if (ip->i_nlink < ip->i_effnlink) 3482 panic("softdep_change_linkcnt: bad delta"); 3483 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3484 FREE_LOCK(&lk); 3485 } 3486 3487 /* 3488 * Called when the effective link count and the reference count 3489 * on an inode drops to zero. At this point there are no names 3490 * referencing the file in the filesystem and no active file 3491 * references. The space associated with the file will be freed 3492 * as soon as the necessary soft dependencies are cleared. 3493 */ 3494 void 3495 softdep_releasefile(ip) 3496 struct inode *ip; /* inode with the zero effective link count */ 3497 { 3498 struct inodedep *inodedep; 3499 struct fs *fs; 3500 int extblocks; 3501 3502 if (ip->i_effnlink > 0) 3503 panic("softdep_filerelease: file still referenced"); 3504 /* 3505 * We may be called several times as the real reference count 3506 * drops to zero. We only want to account for the space once. 3507 */ 3508 if (ip->i_flag & IN_SPACECOUNTED) 3509 return; 3510 /* 3511 * We have to deactivate a snapshot otherwise copyonwrites may 3512 * add blocks and the cleanup may remove blocks after we have 3513 * tried to account for them. 3514 */ 3515 if ((ip->i_flags & SF_SNAPSHOT) != 0) 3516 ffs_snapremove(ITOV(ip)); 3517 /* 3518 * If we are tracking an nlinkdelta, we have to also remember 3519 * whether we accounted for the freed space yet. 3520 */ 3521 ACQUIRE_LOCK(&lk); 3522 if ((inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, &inodedep))) 3523 inodedep->id_state |= SPACECOUNTED; 3524 FREE_LOCK(&lk); 3525 fs = ip->i_fs; 3526 extblocks = 0; 3527 if (fs->fs_magic == FS_UFS2_MAGIC) 3528 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 3529 UFS_LOCK(ip->i_ump); 3530 ip->i_fs->fs_pendingblocks += DIP(ip, i_blocks) - extblocks; 3531 ip->i_fs->fs_pendinginodes += 1; 3532 UFS_UNLOCK(ip->i_ump); 3533 ip->i_flag |= IN_SPACECOUNTED; 3534 } 3535 3536 /* 3537 * This workitem decrements the inode's link count. 3538 * If the link count reaches zero, the file is removed. 3539 */ 3540 static void 3541 handle_workitem_remove(dirrem, xp) 3542 struct dirrem *dirrem; 3543 struct vnode *xp; 3544 { 3545 struct thread *td = curthread; 3546 struct inodedep *inodedep; 3547 struct vnode *vp; 3548 struct inode *ip; 3549 ino_t oldinum; 3550 int error; 3551 3552 if ((vp = xp) == NULL && 3553 (error = ffs_vget(dirrem->dm_list.wk_mp, 3554 dirrem->dm_oldinum, LK_EXCLUSIVE, &vp)) != 0) { 3555 softdep_error("handle_workitem_remove: vget", error); 3556 return; 3557 } 3558 ip = VTOI(vp); 3559 ACQUIRE_LOCK(&lk); 3560 if ((inodedep_lookup(dirrem->dm_list.wk_mp, 3561 dirrem->dm_oldinum, 0, &inodedep)) == 0) 3562 panic("handle_workitem_remove: lost inodedep"); 3563 /* 3564 * Normal file deletion. 3565 */ 3566 if ((dirrem->dm_state & RMDIR) == 0) { 3567 ip->i_nlink--; 3568 DIP_SET(ip, i_nlink, ip->i_nlink); 3569 ip->i_flag |= IN_CHANGE; 3570 if (ip->i_nlink < ip->i_effnlink) 3571 panic("handle_workitem_remove: bad file delta"); 3572 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3573 num_dirrem -= 1; 3574 WORKITEM_FREE(dirrem, D_DIRREM); 3575 FREE_LOCK(&lk); 3576 vput(vp); 3577 return; 3578 } 3579 /* 3580 * Directory deletion. Decrement reference count for both the 3581 * just deleted parent directory entry and the reference for ".". 3582 * Next truncate the directory to length zero. When the 3583 * truncation completes, arrange to have the reference count on 3584 * the parent decremented to account for the loss of "..". 3585 */ 3586 ip->i_nlink -= 2; 3587 DIP_SET(ip, i_nlink, ip->i_nlink); 3588 ip->i_flag |= IN_CHANGE; 3589 if (ip->i_nlink < ip->i_effnlink) 3590 panic("handle_workitem_remove: bad dir delta"); 3591 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3592 FREE_LOCK(&lk); 3593 if ((error = ffs_truncate(vp, (off_t)0, 0, td->td_ucred, td)) != 0) 3594 softdep_error("handle_workitem_remove: truncate", error); 3595 ACQUIRE_LOCK(&lk); 3596 /* 3597 * Rename a directory to a new parent. Since, we are both deleting 3598 * and creating a new directory entry, the link count on the new 3599 * directory should not change. Thus we skip the followup dirrem. 3600 */ 3601 if (dirrem->dm_state & DIRCHG) { 3602 num_dirrem -= 1; 3603 WORKITEM_FREE(dirrem, D_DIRREM); 3604 FREE_LOCK(&lk); 3605 vput(vp); 3606 return; 3607 } 3608 /* 3609 * If the inodedep does not exist, then the zero'ed inode has 3610 * been written to disk. If the allocated inode has never been 3611 * written to disk, then the on-disk inode is zero'ed. In either 3612 * case we can remove the file immediately. 3613 */ 3614 dirrem->dm_state = 0; 3615 oldinum = dirrem->dm_oldinum; 3616 dirrem->dm_oldinum = dirrem->dm_dirinum; 3617 if (inodedep_lookup(dirrem->dm_list.wk_mp, oldinum, 3618 0, &inodedep) == 0 || check_inode_unwritten(inodedep)) { 3619 FREE_LOCK(&lk); 3620 vput(vp); 3621 handle_workitem_remove(dirrem, NULL); 3622 return; 3623 } 3624 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 3625 FREE_LOCK(&lk); 3626 ip->i_flag |= IN_CHANGE; 3627 ffs_update(vp, 0); 3628 vput(vp); 3629 } 3630 3631 /* 3632 * Inode de-allocation dependencies. 3633 * 3634 * When an inode's link count is reduced to zero, it can be de-allocated. We 3635 * found it convenient to postpone de-allocation until after the inode is 3636 * written to disk with its new link count (zero). At this point, all of the 3637 * on-disk inode's block pointers are nullified and, with careful dependency 3638 * list ordering, all dependencies related to the inode will be satisfied and 3639 * the corresponding dependency structures de-allocated. So, if/when the 3640 * inode is reused, there will be no mixing of old dependencies with new 3641 * ones. This artificial dependency is set up by the block de-allocation 3642 * procedure above (softdep_setup_freeblocks) and completed by the 3643 * following procedure. 3644 */ 3645 static void 3646 handle_workitem_freefile(freefile) 3647 struct freefile *freefile; 3648 { 3649 struct fs *fs; 3650 struct inodedep *idp; 3651 struct ufsmount *ump; 3652 int error; 3653 3654 ump = VFSTOUFS(freefile->fx_list.wk_mp); 3655 fs = ump->um_fs; 3656 #ifdef DEBUG 3657 ACQUIRE_LOCK(&lk); 3658 error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp); 3659 FREE_LOCK(&lk); 3660 if (error) 3661 panic("handle_workitem_freefile: inodedep survived"); 3662 #endif 3663 UFS_LOCK(ump); 3664 fs->fs_pendinginodes -= 1; 3665 UFS_UNLOCK(ump); 3666 if ((error = ffs_freefile(ump, fs, freefile->fx_devvp, 3667 freefile->fx_oldinum, freefile->fx_mode)) != 0) 3668 softdep_error("handle_workitem_freefile", error); 3669 ACQUIRE_LOCK(&lk); 3670 WORKITEM_FREE(freefile, D_FREEFILE); 3671 FREE_LOCK(&lk); 3672 } 3673 3674 3675 /* 3676 * Helper function which unlinks marker element from work list and returns 3677 * the next element on the list. 3678 */ 3679 static __inline struct worklist * 3680 markernext(struct worklist *marker) 3681 { 3682 struct worklist *next; 3683 3684 next = LIST_NEXT(marker, wk_list); 3685 LIST_REMOVE(marker, wk_list); 3686 return next; 3687 } 3688 3689 /* 3690 * Disk writes. 3691 * 3692 * The dependency structures constructed above are most actively used when file 3693 * system blocks are written to disk. No constraints are placed on when a 3694 * block can be written, but unsatisfied update dependencies are made safe by 3695 * modifying (or replacing) the source memory for the duration of the disk 3696 * write. When the disk write completes, the memory block is again brought 3697 * up-to-date. 3698 * 3699 * In-core inode structure reclamation. 3700 * 3701 * Because there are a finite number of "in-core" inode structures, they are 3702 * reused regularly. By transferring all inode-related dependencies to the 3703 * in-memory inode block and indexing them separately (via "inodedep"s), we 3704 * can allow "in-core" inode structures to be reused at any time and avoid 3705 * any increase in contention. 3706 * 3707 * Called just before entering the device driver to initiate a new disk I/O. 3708 * The buffer must be locked, thus, no I/O completion operations can occur 3709 * while we are manipulating its associated dependencies. 3710 */ 3711 static void 3712 softdep_disk_io_initiation(bp) 3713 struct buf *bp; /* structure describing disk write to occur */ 3714 { 3715 struct worklist *wk; 3716 struct worklist marker; 3717 struct indirdep *indirdep; 3718 struct inodedep *inodedep; 3719 3720 /* 3721 * We only care about write operations. There should never 3722 * be dependencies for reads. 3723 */ 3724 if (bp->b_iocmd != BIO_WRITE) 3725 panic("softdep_disk_io_initiation: not write"); 3726 3727 marker.wk_type = D_LAST + 1; /* Not a normal workitem */ 3728 PHOLD(curproc); /* Don't swap out kernel stack */ 3729 3730 ACQUIRE_LOCK(&lk); 3731 /* 3732 * Do any necessary pre-I/O processing. 3733 */ 3734 for (wk = LIST_FIRST(&bp->b_dep); wk != NULL; 3735 wk = markernext(&marker)) { 3736 LIST_INSERT_AFTER(wk, &marker, wk_list); 3737 switch (wk->wk_type) { 3738 3739 case D_PAGEDEP: 3740 initiate_write_filepage(WK_PAGEDEP(wk), bp); 3741 continue; 3742 3743 case D_INODEDEP: 3744 inodedep = WK_INODEDEP(wk); 3745 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) 3746 initiate_write_inodeblock_ufs1(inodedep, bp); 3747 else 3748 initiate_write_inodeblock_ufs2(inodedep, bp); 3749 continue; 3750 3751 case D_INDIRDEP: 3752 indirdep = WK_INDIRDEP(wk); 3753 if (indirdep->ir_state & GOINGAWAY) 3754 panic("disk_io_initiation: indirdep gone"); 3755 /* 3756 * If there are no remaining dependencies, this 3757 * will be writing the real pointers, so the 3758 * dependency can be freed. 3759 */ 3760 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) { 3761 struct buf *bp; 3762 3763 bp = indirdep->ir_savebp; 3764 bp->b_flags |= B_INVAL | B_NOCACHE; 3765 /* inline expand WORKLIST_REMOVE(wk); */ 3766 wk->wk_state &= ~ONWORKLIST; 3767 LIST_REMOVE(wk, wk_list); 3768 WORKITEM_FREE(indirdep, D_INDIRDEP); 3769 FREE_LOCK(&lk); 3770 brelse(bp); 3771 ACQUIRE_LOCK(&lk); 3772 continue; 3773 } 3774 /* 3775 * Replace up-to-date version with safe version. 3776 */ 3777 FREE_LOCK(&lk); 3778 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount, 3779 M_INDIRDEP, M_SOFTDEP_FLAGS); 3780 ACQUIRE_LOCK(&lk); 3781 indirdep->ir_state &= ~ATTACHED; 3782 indirdep->ir_state |= UNDONE; 3783 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 3784 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 3785 bp->b_bcount); 3786 continue; 3787 3788 case D_MKDIR: 3789 case D_BMSAFEMAP: 3790 case D_ALLOCDIRECT: 3791 case D_ALLOCINDIR: 3792 continue; 3793 3794 default: 3795 panic("handle_disk_io_initiation: Unexpected type %s", 3796 TYPENAME(wk->wk_type)); 3797 /* NOTREACHED */ 3798 } 3799 } 3800 FREE_LOCK(&lk); 3801 PRELE(curproc); /* Allow swapout of kernel stack */ 3802 } 3803 3804 /* 3805 * Called from within the procedure above to deal with unsatisfied 3806 * allocation dependencies in a directory. The buffer must be locked, 3807 * thus, no I/O completion operations can occur while we are 3808 * manipulating its associated dependencies. 3809 */ 3810 static void 3811 initiate_write_filepage(pagedep, bp) 3812 struct pagedep *pagedep; 3813 struct buf *bp; 3814 { 3815 struct diradd *dap; 3816 struct direct *ep; 3817 int i; 3818 3819 if (pagedep->pd_state & IOSTARTED) { 3820 /* 3821 * This can only happen if there is a driver that does not 3822 * understand chaining. Here biodone will reissue the call 3823 * to strategy for the incomplete buffers. 3824 */ 3825 printf("initiate_write_filepage: already started\n"); 3826 return; 3827 } 3828 pagedep->pd_state |= IOSTARTED; 3829 for (i = 0; i < DAHASHSZ; i++) { 3830 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 3831 ep = (struct direct *) 3832 ((char *)bp->b_data + dap->da_offset); 3833 if (ep->d_ino != dap->da_newinum) 3834 panic("%s: dir inum %d != new %d", 3835 "initiate_write_filepage", 3836 ep->d_ino, dap->da_newinum); 3837 if (dap->da_state & DIRCHG) 3838 ep->d_ino = dap->da_previous->dm_oldinum; 3839 else 3840 ep->d_ino = 0; 3841 dap->da_state &= ~ATTACHED; 3842 dap->da_state |= UNDONE; 3843 } 3844 } 3845 } 3846 3847 /* 3848 * Version of initiate_write_inodeblock that handles UFS1 dinodes. 3849 * Note that any bug fixes made to this routine must be done in the 3850 * version found below. 3851 * 3852 * Called from within the procedure above to deal with unsatisfied 3853 * allocation dependencies in an inodeblock. The buffer must be 3854 * locked, thus, no I/O completion operations can occur while we 3855 * are manipulating its associated dependencies. 3856 */ 3857 static void 3858 initiate_write_inodeblock_ufs1(inodedep, bp) 3859 struct inodedep *inodedep; 3860 struct buf *bp; /* The inode block */ 3861 { 3862 struct allocdirect *adp, *lastadp; 3863 struct ufs1_dinode *dp; 3864 struct ufs1_dinode *sip; 3865 struct fs *fs; 3866 ufs_lbn_t i, prevlbn = 0; 3867 int deplist; 3868 3869 if (inodedep->id_state & IOSTARTED) 3870 panic("initiate_write_inodeblock_ufs1: already started"); 3871 inodedep->id_state |= IOSTARTED; 3872 fs = inodedep->id_fs; 3873 dp = (struct ufs1_dinode *)bp->b_data + 3874 ino_to_fsbo(fs, inodedep->id_ino); 3875 /* 3876 * If the bitmap is not yet written, then the allocated 3877 * inode cannot be written to disk. 3878 */ 3879 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 3880 if (inodedep->id_savedino1 != NULL) 3881 panic("initiate_write_inodeblock_ufs1: I/O underway"); 3882 FREE_LOCK(&lk); 3883 MALLOC(sip, struct ufs1_dinode *, 3884 sizeof(struct ufs1_dinode), M_SAVEDINO, M_SOFTDEP_FLAGS); 3885 ACQUIRE_LOCK(&lk); 3886 inodedep->id_savedino1 = sip; 3887 *inodedep->id_savedino1 = *dp; 3888 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 3889 dp->di_gen = inodedep->id_savedino1->di_gen; 3890 return; 3891 } 3892 /* 3893 * If no dependencies, then there is nothing to roll back. 3894 */ 3895 inodedep->id_savedsize = dp->di_size; 3896 inodedep->id_savedextsize = 0; 3897 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 3898 return; 3899 /* 3900 * Set the dependencies to busy. 3901 */ 3902 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3903 adp = TAILQ_NEXT(adp, ad_next)) { 3904 #ifdef DIAGNOSTIC 3905 if (deplist != 0 && prevlbn >= adp->ad_lbn) 3906 panic("softdep_write_inodeblock: lbn order"); 3907 prevlbn = adp->ad_lbn; 3908 if (adp->ad_lbn < NDADDR && 3909 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) 3910 panic("%s: direct pointer #%jd mismatch %d != %jd", 3911 "softdep_write_inodeblock", 3912 (intmax_t)adp->ad_lbn, 3913 dp->di_db[adp->ad_lbn], 3914 (intmax_t)adp->ad_newblkno); 3915 if (adp->ad_lbn >= NDADDR && 3916 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) 3917 panic("%s: indirect pointer #%jd mismatch %d != %jd", 3918 "softdep_write_inodeblock", 3919 (intmax_t)adp->ad_lbn - NDADDR, 3920 dp->di_ib[adp->ad_lbn - NDADDR], 3921 (intmax_t)adp->ad_newblkno); 3922 deplist |= 1 << adp->ad_lbn; 3923 if ((adp->ad_state & ATTACHED) == 0) 3924 panic("softdep_write_inodeblock: Unknown state 0x%x", 3925 adp->ad_state); 3926 #endif /* DIAGNOSTIC */ 3927 adp->ad_state &= ~ATTACHED; 3928 adp->ad_state |= UNDONE; 3929 } 3930 /* 3931 * The on-disk inode cannot claim to be any larger than the last 3932 * fragment that has been written. Otherwise, the on-disk inode 3933 * might have fragments that were not the last block in the file 3934 * which would corrupt the filesystem. 3935 */ 3936 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3937 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 3938 if (adp->ad_lbn >= NDADDR) 3939 break; 3940 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 3941 /* keep going until hitting a rollback to a frag */ 3942 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 3943 continue; 3944 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 3945 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 3946 #ifdef DIAGNOSTIC 3947 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 3948 panic("softdep_write_inodeblock: lost dep1"); 3949 #endif /* DIAGNOSTIC */ 3950 dp->di_db[i] = 0; 3951 } 3952 for (i = 0; i < NIADDR; i++) { 3953 #ifdef DIAGNOSTIC 3954 if (dp->di_ib[i] != 0 && 3955 (deplist & ((1 << NDADDR) << i)) == 0) 3956 panic("softdep_write_inodeblock: lost dep2"); 3957 #endif /* DIAGNOSTIC */ 3958 dp->di_ib[i] = 0; 3959 } 3960 return; 3961 } 3962 /* 3963 * If we have zero'ed out the last allocated block of the file, 3964 * roll back the size to the last currently allocated block. 3965 * We know that this last allocated block is a full-sized as 3966 * we already checked for fragments in the loop above. 3967 */ 3968 if (lastadp != NULL && 3969 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 3970 for (i = lastadp->ad_lbn; i >= 0; i--) 3971 if (dp->di_db[i] != 0) 3972 break; 3973 dp->di_size = (i + 1) * fs->fs_bsize; 3974 } 3975 /* 3976 * The only dependencies are for indirect blocks. 3977 * 3978 * The file size for indirect block additions is not guaranteed. 3979 * Such a guarantee would be non-trivial to achieve. The conventional 3980 * synchronous write implementation also does not make this guarantee. 3981 * Fsck should catch and fix discrepancies. Arguably, the file size 3982 * can be over-estimated without destroying integrity when the file 3983 * moves into the indirect blocks (i.e., is large). If we want to 3984 * postpone fsck, we are stuck with this argument. 3985 */ 3986 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 3987 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 3988 } 3989 3990 /* 3991 * Version of initiate_write_inodeblock that handles UFS2 dinodes. 3992 * Note that any bug fixes made to this routine must be done in the 3993 * version found above. 3994 * 3995 * Called from within the procedure above to deal with unsatisfied 3996 * allocation dependencies in an inodeblock. The buffer must be 3997 * locked, thus, no I/O completion operations can occur while we 3998 * are manipulating its associated dependencies. 3999 */ 4000 static void 4001 initiate_write_inodeblock_ufs2(inodedep, bp) 4002 struct inodedep *inodedep; 4003 struct buf *bp; /* The inode block */ 4004 { 4005 struct allocdirect *adp, *lastadp; 4006 struct ufs2_dinode *dp; 4007 struct ufs2_dinode *sip; 4008 struct fs *fs; 4009 ufs_lbn_t i, prevlbn = 0; 4010 int deplist; 4011 4012 if (inodedep->id_state & IOSTARTED) 4013 panic("initiate_write_inodeblock_ufs2: already started"); 4014 inodedep->id_state |= IOSTARTED; 4015 fs = inodedep->id_fs; 4016 dp = (struct ufs2_dinode *)bp->b_data + 4017 ino_to_fsbo(fs, inodedep->id_ino); 4018 /* 4019 * If the bitmap is not yet written, then the allocated 4020 * inode cannot be written to disk. 4021 */ 4022 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4023 if (inodedep->id_savedino2 != NULL) 4024 panic("initiate_write_inodeblock_ufs2: I/O underway"); 4025 FREE_LOCK(&lk); 4026 MALLOC(sip, struct ufs2_dinode *, 4027 sizeof(struct ufs2_dinode), M_SAVEDINO, M_SOFTDEP_FLAGS); 4028 ACQUIRE_LOCK(&lk); 4029 inodedep->id_savedino2 = sip; 4030 *inodedep->id_savedino2 = *dp; 4031 bzero((caddr_t)dp, sizeof(struct ufs2_dinode)); 4032 dp->di_gen = inodedep->id_savedino2->di_gen; 4033 return; 4034 } 4035 /* 4036 * If no dependencies, then there is nothing to roll back. 4037 */ 4038 inodedep->id_savedsize = dp->di_size; 4039 inodedep->id_savedextsize = dp->di_extsize; 4040 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL && 4041 TAILQ_FIRST(&inodedep->id_extupdt) == NULL) 4042 return; 4043 /* 4044 * Set the ext data dependencies to busy. 4045 */ 4046 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 4047 adp = TAILQ_NEXT(adp, ad_next)) { 4048 #ifdef DIAGNOSTIC 4049 if (deplist != 0 && prevlbn >= adp->ad_lbn) 4050 panic("softdep_write_inodeblock: lbn order"); 4051 prevlbn = adp->ad_lbn; 4052 if (dp->di_extb[adp->ad_lbn] != adp->ad_newblkno) 4053 panic("%s: direct pointer #%jd mismatch %jd != %jd", 4054 "softdep_write_inodeblock", 4055 (intmax_t)adp->ad_lbn, 4056 (intmax_t)dp->di_extb[adp->ad_lbn], 4057 (intmax_t)adp->ad_newblkno); 4058 deplist |= 1 << adp->ad_lbn; 4059 if ((adp->ad_state & ATTACHED) == 0) 4060 panic("softdep_write_inodeblock: Unknown state 0x%x", 4061 adp->ad_state); 4062 #endif /* DIAGNOSTIC */ 4063 adp->ad_state &= ~ATTACHED; 4064 adp->ad_state |= UNDONE; 4065 } 4066 /* 4067 * The on-disk inode cannot claim to be any larger than the last 4068 * fragment that has been written. Otherwise, the on-disk inode 4069 * might have fragments that were not the last block in the ext 4070 * data which would corrupt the filesystem. 4071 */ 4072 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 4073 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 4074 dp->di_extb[adp->ad_lbn] = adp->ad_oldblkno; 4075 /* keep going until hitting a rollback to a frag */ 4076 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 4077 continue; 4078 dp->di_extsize = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 4079 for (i = adp->ad_lbn + 1; i < NXADDR; i++) { 4080 #ifdef DIAGNOSTIC 4081 if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0) 4082 panic("softdep_write_inodeblock: lost dep1"); 4083 #endif /* DIAGNOSTIC */ 4084 dp->di_extb[i] = 0; 4085 } 4086 lastadp = NULL; 4087 break; 4088 } 4089 /* 4090 * If we have zero'ed out the last allocated block of the ext 4091 * data, roll back the size to the last currently allocated block. 4092 * We know that this last allocated block is a full-sized as 4093 * we already checked for fragments in the loop above. 4094 */ 4095 if (lastadp != NULL && 4096 dp->di_extsize <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 4097 for (i = lastadp->ad_lbn; i >= 0; i--) 4098 if (dp->di_extb[i] != 0) 4099 break; 4100 dp->di_extsize = (i + 1) * fs->fs_bsize; 4101 } 4102 /* 4103 * Set the file data dependencies to busy. 4104 */ 4105 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 4106 adp = TAILQ_NEXT(adp, ad_next)) { 4107 #ifdef DIAGNOSTIC 4108 if (deplist != 0 && prevlbn >= adp->ad_lbn) 4109 panic("softdep_write_inodeblock: lbn order"); 4110 prevlbn = adp->ad_lbn; 4111 if (adp->ad_lbn < NDADDR && 4112 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) 4113 panic("%s: direct pointer #%jd mismatch %jd != %jd", 4114 "softdep_write_inodeblock", 4115 (intmax_t)adp->ad_lbn, 4116 (intmax_t)dp->di_db[adp->ad_lbn], 4117 (intmax_t)adp->ad_newblkno); 4118 if (adp->ad_lbn >= NDADDR && 4119 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) 4120 panic("%s indirect pointer #%jd mismatch %jd != %jd", 4121 "softdep_write_inodeblock:", 4122 (intmax_t)adp->ad_lbn - NDADDR, 4123 (intmax_t)dp->di_ib[adp->ad_lbn - NDADDR], 4124 (intmax_t)adp->ad_newblkno); 4125 deplist |= 1 << adp->ad_lbn; 4126 if ((adp->ad_state & ATTACHED) == 0) 4127 panic("softdep_write_inodeblock: Unknown state 0x%x", 4128 adp->ad_state); 4129 #endif /* DIAGNOSTIC */ 4130 adp->ad_state &= ~ATTACHED; 4131 adp->ad_state |= UNDONE; 4132 } 4133 /* 4134 * The on-disk inode cannot claim to be any larger than the last 4135 * fragment that has been written. Otherwise, the on-disk inode 4136 * might have fragments that were not the last block in the file 4137 * which would corrupt the filesystem. 4138 */ 4139 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 4140 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 4141 if (adp->ad_lbn >= NDADDR) 4142 break; 4143 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 4144 /* keep going until hitting a rollback to a frag */ 4145 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 4146 continue; 4147 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 4148 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 4149 #ifdef DIAGNOSTIC 4150 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 4151 panic("softdep_write_inodeblock: lost dep2"); 4152 #endif /* DIAGNOSTIC */ 4153 dp->di_db[i] = 0; 4154 } 4155 for (i = 0; i < NIADDR; i++) { 4156 #ifdef DIAGNOSTIC 4157 if (dp->di_ib[i] != 0 && 4158 (deplist & ((1 << NDADDR) << i)) == 0) 4159 panic("softdep_write_inodeblock: lost dep3"); 4160 #endif /* DIAGNOSTIC */ 4161 dp->di_ib[i] = 0; 4162 } 4163 return; 4164 } 4165 /* 4166 * If we have zero'ed out the last allocated block of the file, 4167 * roll back the size to the last currently allocated block. 4168 * We know that this last allocated block is a full-sized as 4169 * we already checked for fragments in the loop above. 4170 */ 4171 if (lastadp != NULL && 4172 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 4173 for (i = lastadp->ad_lbn; i >= 0; i--) 4174 if (dp->di_db[i] != 0) 4175 break; 4176 dp->di_size = (i + 1) * fs->fs_bsize; 4177 } 4178 /* 4179 * The only dependencies are for indirect blocks. 4180 * 4181 * The file size for indirect block additions is not guaranteed. 4182 * Such a guarantee would be non-trivial to achieve. The conventional 4183 * synchronous write implementation also does not make this guarantee. 4184 * Fsck should catch and fix discrepancies. Arguably, the file size 4185 * can be over-estimated without destroying integrity when the file 4186 * moves into the indirect blocks (i.e., is large). If we want to 4187 * postpone fsck, we are stuck with this argument. 4188 */ 4189 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 4190 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 4191 } 4192 4193 /* 4194 * This routine is called during the completion interrupt 4195 * service routine for a disk write (from the procedure called 4196 * by the device driver to inform the filesystem caches of 4197 * a request completion). It should be called early in this 4198 * procedure, before the block is made available to other 4199 * processes or other routines are called. 4200 */ 4201 static void 4202 softdep_disk_write_complete(bp) 4203 struct buf *bp; /* describes the completed disk write */ 4204 { 4205 struct worklist *wk; 4206 struct worklist *owk; 4207 struct workhead reattach; 4208 struct newblk *newblk; 4209 struct allocindir *aip; 4210 struct allocdirect *adp; 4211 struct indirdep *indirdep; 4212 struct inodedep *inodedep; 4213 struct bmsafemap *bmsafemap; 4214 4215 /* 4216 * If an error occurred while doing the write, then the data 4217 * has not hit the disk and the dependencies cannot be unrolled. 4218 */ 4219 if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0) 4220 return; 4221 LIST_INIT(&reattach); 4222 /* 4223 * This lock must not be released anywhere in this code segment. 4224 */ 4225 ACQUIRE_LOCK(&lk); 4226 owk = NULL; 4227 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 4228 WORKLIST_REMOVE(wk); 4229 if (wk == owk) 4230 panic("duplicate worklist: %p\n", wk); 4231 owk = wk; 4232 switch (wk->wk_type) { 4233 4234 case D_PAGEDEP: 4235 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 4236 WORKLIST_INSERT(&reattach, wk); 4237 continue; 4238 4239 case D_INODEDEP: 4240 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 4241 WORKLIST_INSERT(&reattach, wk); 4242 continue; 4243 4244 case D_BMSAFEMAP: 4245 bmsafemap = WK_BMSAFEMAP(wk); 4246 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) { 4247 newblk->nb_state |= DEPCOMPLETE; 4248 newblk->nb_bmsafemap = NULL; 4249 LIST_REMOVE(newblk, nb_deps); 4250 } 4251 while ((adp = 4252 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) { 4253 adp->ad_state |= DEPCOMPLETE; 4254 adp->ad_buf = NULL; 4255 LIST_REMOVE(adp, ad_deps); 4256 handle_allocdirect_partdone(adp); 4257 } 4258 while ((aip = 4259 LIST_FIRST(&bmsafemap->sm_allocindirhd))) { 4260 aip->ai_state |= DEPCOMPLETE; 4261 aip->ai_buf = NULL; 4262 LIST_REMOVE(aip, ai_deps); 4263 handle_allocindir_partdone(aip); 4264 } 4265 while ((inodedep = 4266 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { 4267 inodedep->id_state |= DEPCOMPLETE; 4268 LIST_REMOVE(inodedep, id_deps); 4269 inodedep->id_buf = NULL; 4270 } 4271 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 4272 continue; 4273 4274 case D_MKDIR: 4275 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 4276 continue; 4277 4278 case D_ALLOCDIRECT: 4279 adp = WK_ALLOCDIRECT(wk); 4280 adp->ad_state |= COMPLETE; 4281 handle_allocdirect_partdone(adp); 4282 continue; 4283 4284 case D_ALLOCINDIR: 4285 aip = WK_ALLOCINDIR(wk); 4286 aip->ai_state |= COMPLETE; 4287 handle_allocindir_partdone(aip); 4288 continue; 4289 4290 case D_INDIRDEP: 4291 indirdep = WK_INDIRDEP(wk); 4292 if (indirdep->ir_state & GOINGAWAY) 4293 panic("disk_write_complete: indirdep gone"); 4294 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 4295 FREE(indirdep->ir_saveddata, M_INDIRDEP); 4296 indirdep->ir_saveddata = 0; 4297 indirdep->ir_state &= ~UNDONE; 4298 indirdep->ir_state |= ATTACHED; 4299 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 4300 handle_allocindir_partdone(aip); 4301 if (aip == LIST_FIRST(&indirdep->ir_donehd)) 4302 panic("disk_write_complete: not gone"); 4303 } 4304 WORKLIST_INSERT(&reattach, wk); 4305 if ((bp->b_flags & B_DELWRI) == 0) 4306 stat_indir_blk_ptrs++; 4307 bdirty(bp); 4308 continue; 4309 4310 default: 4311 panic("handle_disk_write_complete: Unknown type %s", 4312 TYPENAME(wk->wk_type)); 4313 /* NOTREACHED */ 4314 } 4315 } 4316 /* 4317 * Reattach any requests that must be redone. 4318 */ 4319 while ((wk = LIST_FIRST(&reattach)) != NULL) { 4320 WORKLIST_REMOVE(wk); 4321 WORKLIST_INSERT(&bp->b_dep, wk); 4322 } 4323 FREE_LOCK(&lk); 4324 } 4325 4326 /* 4327 * Called from within softdep_disk_write_complete above. Note that 4328 * this routine is always called from interrupt level with further 4329 * splbio interrupts blocked. 4330 */ 4331 static void 4332 handle_allocdirect_partdone(adp) 4333 struct allocdirect *adp; /* the completed allocdirect */ 4334 { 4335 struct allocdirectlst *listhead; 4336 struct allocdirect *listadp; 4337 struct inodedep *inodedep; 4338 long bsize, delay; 4339 4340 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 4341 return; 4342 if (adp->ad_buf != NULL) 4343 panic("handle_allocdirect_partdone: dangling dep"); 4344 /* 4345 * The on-disk inode cannot claim to be any larger than the last 4346 * fragment that has been written. Otherwise, the on-disk inode 4347 * might have fragments that were not the last block in the file 4348 * which would corrupt the filesystem. Thus, we cannot free any 4349 * allocdirects after one whose ad_oldblkno claims a fragment as 4350 * these blocks must be rolled back to zero before writing the inode. 4351 * We check the currently active set of allocdirects in id_inoupdt 4352 * or id_extupdt as appropriate. 4353 */ 4354 inodedep = adp->ad_inodedep; 4355 bsize = inodedep->id_fs->fs_bsize; 4356 if (adp->ad_state & EXTDATA) 4357 listhead = &inodedep->id_extupdt; 4358 else 4359 listhead = &inodedep->id_inoupdt; 4360 TAILQ_FOREACH(listadp, listhead, ad_next) { 4361 /* found our block */ 4362 if (listadp == adp) 4363 break; 4364 /* continue if ad_oldlbn is not a fragment */ 4365 if (listadp->ad_oldsize == 0 || 4366 listadp->ad_oldsize == bsize) 4367 continue; 4368 /* hit a fragment */ 4369 return; 4370 } 4371 /* 4372 * If we have reached the end of the current list without 4373 * finding the just finished dependency, then it must be 4374 * on the future dependency list. Future dependencies cannot 4375 * be freed until they are moved to the current list. 4376 */ 4377 if (listadp == NULL) { 4378 #ifdef DEBUG 4379 if (adp->ad_state & EXTDATA) 4380 listhead = &inodedep->id_newextupdt; 4381 else 4382 listhead = &inodedep->id_newinoupdt; 4383 TAILQ_FOREACH(listadp, listhead, ad_next) 4384 /* found our block */ 4385 if (listadp == adp) 4386 break; 4387 if (listadp == NULL) 4388 panic("handle_allocdirect_partdone: lost dep"); 4389 #endif /* DEBUG */ 4390 return; 4391 } 4392 /* 4393 * If we have found the just finished dependency, then free 4394 * it along with anything that follows it that is complete. 4395 * If the inode still has a bitmap dependency, then it has 4396 * never been written to disk, hence the on-disk inode cannot 4397 * reference the old fragment so we can free it without delay. 4398 */ 4399 delay = (inodedep->id_state & DEPCOMPLETE); 4400 for (; adp; adp = listadp) { 4401 listadp = TAILQ_NEXT(adp, ad_next); 4402 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 4403 return; 4404 free_allocdirect(listhead, adp, delay); 4405 } 4406 } 4407 4408 /* 4409 * Called from within softdep_disk_write_complete above. Note that 4410 * this routine is always called from interrupt level with further 4411 * splbio interrupts blocked. 4412 */ 4413 static void 4414 handle_allocindir_partdone(aip) 4415 struct allocindir *aip; /* the completed allocindir */ 4416 { 4417 struct indirdep *indirdep; 4418 4419 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 4420 return; 4421 if (aip->ai_buf != NULL) 4422 panic("handle_allocindir_partdone: dangling dependency"); 4423 indirdep = aip->ai_indirdep; 4424 if (indirdep->ir_state & UNDONE) { 4425 LIST_REMOVE(aip, ai_next); 4426 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 4427 return; 4428 } 4429 if (indirdep->ir_state & UFS1FMT) 4430 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 4431 aip->ai_newblkno; 4432 else 4433 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 4434 aip->ai_newblkno; 4435 LIST_REMOVE(aip, ai_next); 4436 if (aip->ai_freefrag != NULL) 4437 add_to_worklist(&aip->ai_freefrag->ff_list); 4438 WORKITEM_FREE(aip, D_ALLOCINDIR); 4439 } 4440 4441 /* 4442 * Called from within softdep_disk_write_complete above to restore 4443 * in-memory inode block contents to their most up-to-date state. Note 4444 * that this routine is always called from interrupt level with further 4445 * splbio interrupts blocked. 4446 */ 4447 static int 4448 handle_written_inodeblock(inodedep, bp) 4449 struct inodedep *inodedep; 4450 struct buf *bp; /* buffer containing the inode block */ 4451 { 4452 struct worklist *wk, *filefree; 4453 struct allocdirect *adp, *nextadp; 4454 struct ufs1_dinode *dp1 = NULL; 4455 struct ufs2_dinode *dp2 = NULL; 4456 int hadchanges, fstype; 4457 4458 if ((inodedep->id_state & IOSTARTED) == 0) 4459 panic("handle_written_inodeblock: not started"); 4460 inodedep->id_state &= ~IOSTARTED; 4461 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) { 4462 fstype = UFS1; 4463 dp1 = (struct ufs1_dinode *)bp->b_data + 4464 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 4465 } else { 4466 fstype = UFS2; 4467 dp2 = (struct ufs2_dinode *)bp->b_data + 4468 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 4469 } 4470 /* 4471 * If we had to rollback the inode allocation because of 4472 * bitmaps being incomplete, then simply restore it. 4473 * Keep the block dirty so that it will not be reclaimed until 4474 * all associated dependencies have been cleared and the 4475 * corresponding updates written to disk. 4476 */ 4477 if (inodedep->id_savedino1 != NULL) { 4478 if (fstype == UFS1) 4479 *dp1 = *inodedep->id_savedino1; 4480 else 4481 *dp2 = *inodedep->id_savedino2; 4482 FREE(inodedep->id_savedino1, M_SAVEDINO); 4483 inodedep->id_savedino1 = NULL; 4484 if ((bp->b_flags & B_DELWRI) == 0) 4485 stat_inode_bitmap++; 4486 bdirty(bp); 4487 return (1); 4488 } 4489 inodedep->id_state |= COMPLETE; 4490 /* 4491 * Roll forward anything that had to be rolled back before 4492 * the inode could be updated. 4493 */ 4494 hadchanges = 0; 4495 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 4496 nextadp = TAILQ_NEXT(adp, ad_next); 4497 if (adp->ad_state & ATTACHED) 4498 panic("handle_written_inodeblock: new entry"); 4499 if (fstype == UFS1) { 4500 if (adp->ad_lbn < NDADDR) { 4501 if (dp1->di_db[adp->ad_lbn]!=adp->ad_oldblkno) 4502 panic("%s %s #%jd mismatch %d != %jd", 4503 "handle_written_inodeblock:", 4504 "direct pointer", 4505 (intmax_t)adp->ad_lbn, 4506 dp1->di_db[adp->ad_lbn], 4507 (intmax_t)adp->ad_oldblkno); 4508 dp1->di_db[adp->ad_lbn] = adp->ad_newblkno; 4509 } else { 4510 if (dp1->di_ib[adp->ad_lbn - NDADDR] != 0) 4511 panic("%s: %s #%jd allocated as %d", 4512 "handle_written_inodeblock", 4513 "indirect pointer", 4514 (intmax_t)adp->ad_lbn - NDADDR, 4515 dp1->di_ib[adp->ad_lbn - NDADDR]); 4516 dp1->di_ib[adp->ad_lbn - NDADDR] = 4517 adp->ad_newblkno; 4518 } 4519 } else { 4520 if (adp->ad_lbn < NDADDR) { 4521 if (dp2->di_db[adp->ad_lbn]!=adp->ad_oldblkno) 4522 panic("%s: %s #%jd %s %jd != %jd", 4523 "handle_written_inodeblock", 4524 "direct pointer", 4525 (intmax_t)adp->ad_lbn, "mismatch", 4526 (intmax_t)dp2->di_db[adp->ad_lbn], 4527 (intmax_t)adp->ad_oldblkno); 4528 dp2->di_db[adp->ad_lbn] = adp->ad_newblkno; 4529 } else { 4530 if (dp2->di_ib[adp->ad_lbn - NDADDR] != 0) 4531 panic("%s: %s #%jd allocated as %jd", 4532 "handle_written_inodeblock", 4533 "indirect pointer", 4534 (intmax_t)adp->ad_lbn - NDADDR, 4535 (intmax_t) 4536 dp2->di_ib[adp->ad_lbn - NDADDR]); 4537 dp2->di_ib[adp->ad_lbn - NDADDR] = 4538 adp->ad_newblkno; 4539 } 4540 } 4541 adp->ad_state &= ~UNDONE; 4542 adp->ad_state |= ATTACHED; 4543 hadchanges = 1; 4544 } 4545 for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) { 4546 nextadp = TAILQ_NEXT(adp, ad_next); 4547 if (adp->ad_state & ATTACHED) 4548 panic("handle_written_inodeblock: new entry"); 4549 if (dp2->di_extb[adp->ad_lbn] != adp->ad_oldblkno) 4550 panic("%s: direct pointers #%jd %s %jd != %jd", 4551 "handle_written_inodeblock", 4552 (intmax_t)adp->ad_lbn, "mismatch", 4553 (intmax_t)dp2->di_extb[adp->ad_lbn], 4554 (intmax_t)adp->ad_oldblkno); 4555 dp2->di_extb[adp->ad_lbn] = adp->ad_newblkno; 4556 adp->ad_state &= ~UNDONE; 4557 adp->ad_state |= ATTACHED; 4558 hadchanges = 1; 4559 } 4560 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 4561 stat_direct_blk_ptrs++; 4562 /* 4563 * Reset the file size to its most up-to-date value. 4564 */ 4565 if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1) 4566 panic("handle_written_inodeblock: bad size"); 4567 if (fstype == UFS1) { 4568 if (dp1->di_size != inodedep->id_savedsize) { 4569 dp1->di_size = inodedep->id_savedsize; 4570 hadchanges = 1; 4571 } 4572 } else { 4573 if (dp2->di_size != inodedep->id_savedsize) { 4574 dp2->di_size = inodedep->id_savedsize; 4575 hadchanges = 1; 4576 } 4577 if (dp2->di_extsize != inodedep->id_savedextsize) { 4578 dp2->di_extsize = inodedep->id_savedextsize; 4579 hadchanges = 1; 4580 } 4581 } 4582 inodedep->id_savedsize = -1; 4583 inodedep->id_savedextsize = -1; 4584 /* 4585 * If there were any rollbacks in the inode block, then it must be 4586 * marked dirty so that its will eventually get written back in 4587 * its correct form. 4588 */ 4589 if (hadchanges) 4590 bdirty(bp); 4591 /* 4592 * Process any allocdirects that completed during the update. 4593 */ 4594 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 4595 handle_allocdirect_partdone(adp); 4596 if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 4597 handle_allocdirect_partdone(adp); 4598 /* 4599 * Process deallocations that were held pending until the 4600 * inode had been written to disk. Freeing of the inode 4601 * is delayed until after all blocks have been freed to 4602 * avoid creation of new <vfsid, inum, lbn> triples 4603 * before the old ones have been deleted. 4604 */ 4605 filefree = NULL; 4606 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 4607 WORKLIST_REMOVE(wk); 4608 switch (wk->wk_type) { 4609 4610 case D_FREEFILE: 4611 /* 4612 * We defer adding filefree to the worklist until 4613 * all other additions have been made to ensure 4614 * that it will be done after all the old blocks 4615 * have been freed. 4616 */ 4617 if (filefree != NULL) 4618 panic("handle_written_inodeblock: filefree"); 4619 filefree = wk; 4620 continue; 4621 4622 case D_MKDIR: 4623 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 4624 continue; 4625 4626 case D_DIRADD: 4627 diradd_inode_written(WK_DIRADD(wk), inodedep); 4628 continue; 4629 4630 case D_FREEBLKS: 4631 wk->wk_state |= COMPLETE; 4632 if ((wk->wk_state & ALLCOMPLETE) != ALLCOMPLETE) 4633 continue; 4634 /* -- fall through -- */ 4635 case D_FREEFRAG: 4636 case D_DIRREM: 4637 add_to_worklist(wk); 4638 continue; 4639 4640 case D_NEWDIRBLK: 4641 free_newdirblk(WK_NEWDIRBLK(wk)); 4642 continue; 4643 4644 default: 4645 panic("handle_written_inodeblock: Unknown type %s", 4646 TYPENAME(wk->wk_type)); 4647 /* NOTREACHED */ 4648 } 4649 } 4650 if (filefree != NULL) { 4651 if (free_inodedep(inodedep) == 0) 4652 panic("handle_written_inodeblock: live inodedep"); 4653 add_to_worklist(filefree); 4654 return (0); 4655 } 4656 4657 /* 4658 * If no outstanding dependencies, free it. 4659 */ 4660 if (free_inodedep(inodedep) || 4661 (TAILQ_FIRST(&inodedep->id_inoupdt) == 0 && 4662 TAILQ_FIRST(&inodedep->id_extupdt) == 0)) 4663 return (0); 4664 return (hadchanges); 4665 } 4666 4667 /* 4668 * Process a diradd entry after its dependent inode has been written. 4669 * This routine must be called with splbio interrupts blocked. 4670 */ 4671 static void 4672 diradd_inode_written(dap, inodedep) 4673 struct diradd *dap; 4674 struct inodedep *inodedep; 4675 { 4676 struct pagedep *pagedep; 4677 4678 dap->da_state |= COMPLETE; 4679 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 4680 if (dap->da_state & DIRCHG) 4681 pagedep = dap->da_previous->dm_pagedep; 4682 else 4683 pagedep = dap->da_pagedep; 4684 LIST_REMOVE(dap, da_pdlist); 4685 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 4686 } 4687 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 4688 } 4689 4690 /* 4691 * Handle the completion of a mkdir dependency. 4692 */ 4693 static void 4694 handle_written_mkdir(mkdir, type) 4695 struct mkdir *mkdir; 4696 int type; 4697 { 4698 struct diradd *dap; 4699 struct pagedep *pagedep; 4700 4701 if (mkdir->md_state != type) 4702 panic("handle_written_mkdir: bad type"); 4703 dap = mkdir->md_diradd; 4704 dap->da_state &= ~type; 4705 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 4706 dap->da_state |= DEPCOMPLETE; 4707 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 4708 if (dap->da_state & DIRCHG) 4709 pagedep = dap->da_previous->dm_pagedep; 4710 else 4711 pagedep = dap->da_pagedep; 4712 LIST_REMOVE(dap, da_pdlist); 4713 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 4714 } 4715 LIST_REMOVE(mkdir, md_mkdirs); 4716 WORKITEM_FREE(mkdir, D_MKDIR); 4717 } 4718 4719 /* 4720 * Called from within softdep_disk_write_complete above. 4721 * A write operation was just completed. Removed inodes can 4722 * now be freed and associated block pointers may be committed. 4723 * Note that this routine is always called from interrupt level 4724 * with further splbio interrupts blocked. 4725 */ 4726 static int 4727 handle_written_filepage(pagedep, bp) 4728 struct pagedep *pagedep; 4729 struct buf *bp; /* buffer containing the written page */ 4730 { 4731 struct dirrem *dirrem; 4732 struct diradd *dap, *nextdap; 4733 struct direct *ep; 4734 int i, chgs; 4735 4736 if ((pagedep->pd_state & IOSTARTED) == 0) 4737 panic("handle_written_filepage: not started"); 4738 pagedep->pd_state &= ~IOSTARTED; 4739 /* 4740 * Process any directory removals that have been committed. 4741 */ 4742 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 4743 LIST_REMOVE(dirrem, dm_next); 4744 dirrem->dm_dirinum = pagedep->pd_ino; 4745 add_to_worklist(&dirrem->dm_list); 4746 } 4747 /* 4748 * Free any directory additions that have been committed. 4749 * If it is a newly allocated block, we have to wait until 4750 * the on-disk directory inode claims the new block. 4751 */ 4752 if ((pagedep->pd_state & NEWBLOCK) == 0) 4753 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 4754 free_diradd(dap); 4755 /* 4756 * Uncommitted directory entries must be restored. 4757 */ 4758 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 4759 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 4760 dap = nextdap) { 4761 nextdap = LIST_NEXT(dap, da_pdlist); 4762 if (dap->da_state & ATTACHED) 4763 panic("handle_written_filepage: attached"); 4764 ep = (struct direct *) 4765 ((char *)bp->b_data + dap->da_offset); 4766 ep->d_ino = dap->da_newinum; 4767 dap->da_state &= ~UNDONE; 4768 dap->da_state |= ATTACHED; 4769 chgs = 1; 4770 /* 4771 * If the inode referenced by the directory has 4772 * been written out, then the dependency can be 4773 * moved to the pending list. 4774 */ 4775 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 4776 LIST_REMOVE(dap, da_pdlist); 4777 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 4778 da_pdlist); 4779 } 4780 } 4781 } 4782 /* 4783 * If there were any rollbacks in the directory, then it must be 4784 * marked dirty so that its will eventually get written back in 4785 * its correct form. 4786 */ 4787 if (chgs) { 4788 if ((bp->b_flags & B_DELWRI) == 0) 4789 stat_dir_entry++; 4790 bdirty(bp); 4791 return (1); 4792 } 4793 /* 4794 * If we are not waiting for a new directory block to be 4795 * claimed by its inode, then the pagedep will be freed. 4796 * Otherwise it will remain to track any new entries on 4797 * the page in case they are fsync'ed. 4798 */ 4799 if ((pagedep->pd_state & NEWBLOCK) == 0) { 4800 LIST_REMOVE(pagedep, pd_hash); 4801 WORKITEM_FREE(pagedep, D_PAGEDEP); 4802 } 4803 return (0); 4804 } 4805 4806 /* 4807 * Writing back in-core inode structures. 4808 * 4809 * The filesystem only accesses an inode's contents when it occupies an 4810 * "in-core" inode structure. These "in-core" structures are separate from 4811 * the page frames used to cache inode blocks. Only the latter are 4812 * transferred to/from the disk. So, when the updated contents of the 4813 * "in-core" inode structure are copied to the corresponding in-memory inode 4814 * block, the dependencies are also transferred. The following procedure is 4815 * called when copying a dirty "in-core" inode to a cached inode block. 4816 */ 4817 4818 /* 4819 * Called when an inode is loaded from disk. If the effective link count 4820 * differed from the actual link count when it was last flushed, then we 4821 * need to ensure that the correct effective link count is put back. 4822 */ 4823 void 4824 softdep_load_inodeblock(ip) 4825 struct inode *ip; /* the "in_core" copy of the inode */ 4826 { 4827 struct inodedep *inodedep; 4828 4829 /* 4830 * Check for alternate nlink count. 4831 */ 4832 ip->i_effnlink = ip->i_nlink; 4833 ACQUIRE_LOCK(&lk); 4834 if (inodedep_lookup(UFSTOVFS(ip->i_ump), 4835 ip->i_number, 0, &inodedep) == 0) { 4836 FREE_LOCK(&lk); 4837 return; 4838 } 4839 ip->i_effnlink -= inodedep->id_nlinkdelta; 4840 if (inodedep->id_state & SPACECOUNTED) 4841 ip->i_flag |= IN_SPACECOUNTED; 4842 FREE_LOCK(&lk); 4843 } 4844 4845 /* 4846 * This routine is called just before the "in-core" inode 4847 * information is to be copied to the in-memory inode block. 4848 * Recall that an inode block contains several inodes. If 4849 * the force flag is set, then the dependencies will be 4850 * cleared so that the update can always be made. Note that 4851 * the buffer is locked when this routine is called, so we 4852 * will never be in the middle of writing the inode block 4853 * to disk. 4854 */ 4855 void 4856 softdep_update_inodeblock(ip, bp, waitfor) 4857 struct inode *ip; /* the "in_core" copy of the inode */ 4858 struct buf *bp; /* the buffer containing the inode block */ 4859 int waitfor; /* nonzero => update must be allowed */ 4860 { 4861 struct inodedep *inodedep; 4862 struct worklist *wk; 4863 struct mount *mp; 4864 struct buf *ibp; 4865 int error; 4866 4867 /* 4868 * If the effective link count is not equal to the actual link 4869 * count, then we must track the difference in an inodedep while 4870 * the inode is (potentially) tossed out of the cache. Otherwise, 4871 * if there is no existing inodedep, then there are no dependencies 4872 * to track. 4873 */ 4874 mp = UFSTOVFS(ip->i_ump); 4875 ACQUIRE_LOCK(&lk); 4876 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 4877 FREE_LOCK(&lk); 4878 if (ip->i_effnlink != ip->i_nlink) 4879 panic("softdep_update_inodeblock: bad link count"); 4880 return; 4881 } 4882 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) 4883 panic("softdep_update_inodeblock: bad delta"); 4884 /* 4885 * Changes have been initiated. Anything depending on these 4886 * changes cannot occur until this inode has been written. 4887 */ 4888 inodedep->id_state &= ~COMPLETE; 4889 if ((inodedep->id_state & ONWORKLIST) == 0) 4890 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 4891 /* 4892 * Any new dependencies associated with the incore inode must 4893 * now be moved to the list associated with the buffer holding 4894 * the in-memory copy of the inode. Once merged process any 4895 * allocdirects that are completed by the merger. 4896 */ 4897 merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt); 4898 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL) 4899 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); 4900 merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt); 4901 if (TAILQ_FIRST(&inodedep->id_extupdt) != NULL) 4902 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt)); 4903 /* 4904 * Now that the inode has been pushed into the buffer, the 4905 * operations dependent on the inode being written to disk 4906 * can be moved to the id_bufwait so that they will be 4907 * processed when the buffer I/O completes. 4908 */ 4909 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 4910 WORKLIST_REMOVE(wk); 4911 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 4912 } 4913 /* 4914 * Newly allocated inodes cannot be written until the bitmap 4915 * that allocates them have been written (indicated by 4916 * DEPCOMPLETE being set in id_state). If we are doing a 4917 * forced sync (e.g., an fsync on a file), we force the bitmap 4918 * to be written so that the update can be done. 4919 */ 4920 if (waitfor == 0) { 4921 FREE_LOCK(&lk); 4922 return; 4923 } 4924 retry: 4925 if ((inodedep->id_state & DEPCOMPLETE) != 0) { 4926 FREE_LOCK(&lk); 4927 return; 4928 } 4929 ibp = inodedep->id_buf; 4930 ibp = getdirtybuf(ibp, &lk, MNT_WAIT); 4931 if (ibp == NULL) { 4932 /* 4933 * If ibp came back as NULL, the dependency could have been 4934 * freed while we slept. Look it up again, and check to see 4935 * that it has completed. 4936 */ 4937 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 4938 goto retry; 4939 FREE_LOCK(&lk); 4940 return; 4941 } 4942 FREE_LOCK(&lk); 4943 if ((error = bwrite(ibp)) != 0) 4944 softdep_error("softdep_update_inodeblock: bwrite", error); 4945 } 4946 4947 /* 4948 * Merge the a new inode dependency list (such as id_newinoupdt) into an 4949 * old inode dependency list (such as id_inoupdt). This routine must be 4950 * called with splbio interrupts blocked. 4951 */ 4952 static void 4953 merge_inode_lists(newlisthead, oldlisthead) 4954 struct allocdirectlst *newlisthead; 4955 struct allocdirectlst *oldlisthead; 4956 { 4957 struct allocdirect *listadp, *newadp; 4958 4959 newadp = TAILQ_FIRST(newlisthead); 4960 for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) { 4961 if (listadp->ad_lbn < newadp->ad_lbn) { 4962 listadp = TAILQ_NEXT(listadp, ad_next); 4963 continue; 4964 } 4965 TAILQ_REMOVE(newlisthead, newadp, ad_next); 4966 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 4967 if (listadp->ad_lbn == newadp->ad_lbn) { 4968 allocdirect_merge(oldlisthead, newadp, 4969 listadp); 4970 listadp = newadp; 4971 } 4972 newadp = TAILQ_FIRST(newlisthead); 4973 } 4974 while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) { 4975 TAILQ_REMOVE(newlisthead, newadp, ad_next); 4976 TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next); 4977 } 4978 } 4979 4980 /* 4981 * If we are doing an fsync, then we must ensure that any directory 4982 * entries for the inode have been written after the inode gets to disk. 4983 */ 4984 int 4985 softdep_fsync(vp) 4986 struct vnode *vp; /* the "in_core" copy of the inode */ 4987 { 4988 struct inodedep *inodedep; 4989 struct pagedep *pagedep; 4990 struct worklist *wk; 4991 struct diradd *dap; 4992 struct mount *mp; 4993 struct vnode *pvp; 4994 struct inode *ip; 4995 struct buf *bp; 4996 struct fs *fs; 4997 struct thread *td = curthread; 4998 int error, flushparent; 4999 ino_t parentino; 5000 ufs_lbn_t lbn; 5001 5002 ip = VTOI(vp); 5003 fs = ip->i_fs; 5004 mp = vp->v_mount; 5005 ACQUIRE_LOCK(&lk); 5006 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 5007 FREE_LOCK(&lk); 5008 return (0); 5009 } 5010 if (LIST_FIRST(&inodedep->id_inowait) != NULL || 5011 LIST_FIRST(&inodedep->id_bufwait) != NULL || 5012 TAILQ_FIRST(&inodedep->id_extupdt) != NULL || 5013 TAILQ_FIRST(&inodedep->id_newextupdt) != NULL || 5014 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 5015 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) 5016 panic("softdep_fsync: pending ops"); 5017 for (error = 0, flushparent = 0; ; ) { 5018 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 5019 break; 5020 if (wk->wk_type != D_DIRADD) 5021 panic("softdep_fsync: Unexpected type %s", 5022 TYPENAME(wk->wk_type)); 5023 dap = WK_DIRADD(wk); 5024 /* 5025 * Flush our parent if this directory entry has a MKDIR_PARENT 5026 * dependency or is contained in a newly allocated block. 5027 */ 5028 if (dap->da_state & DIRCHG) 5029 pagedep = dap->da_previous->dm_pagedep; 5030 else 5031 pagedep = dap->da_pagedep; 5032 parentino = pagedep->pd_ino; 5033 lbn = pagedep->pd_lbn; 5034 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) 5035 panic("softdep_fsync: dirty"); 5036 if ((dap->da_state & MKDIR_PARENT) || 5037 (pagedep->pd_state & NEWBLOCK)) 5038 flushparent = 1; 5039 else 5040 flushparent = 0; 5041 /* 5042 * If we are being fsync'ed as part of vgone'ing this vnode, 5043 * then we will not be able to release and recover the 5044 * vnode below, so we just have to give up on writing its 5045 * directory entry out. It will eventually be written, just 5046 * not now, but then the user was not asking to have it 5047 * written, so we are not breaking any promises. 5048 */ 5049 if (vp->v_iflag & VI_DOOMED) 5050 break; 5051 /* 5052 * We prevent deadlock by always fetching inodes from the 5053 * root, moving down the directory tree. Thus, when fetching 5054 * our parent directory, we first try to get the lock. If 5055 * that fails, we must unlock ourselves before requesting 5056 * the lock on our parent. See the comment in ufs_lookup 5057 * for details on possible races. 5058 */ 5059 FREE_LOCK(&lk); 5060 if (ffs_vget(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp)) { 5061 VOP_UNLOCK(vp, 0, td); 5062 error = ffs_vget(mp, parentino, LK_EXCLUSIVE, &pvp); 5063 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 5064 if (error != 0) 5065 return (error); 5066 } 5067 /* 5068 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps 5069 * that are contained in direct blocks will be resolved by 5070 * doing a ffs_update. Pagedeps contained in indirect blocks 5071 * may require a complete sync'ing of the directory. So, we 5072 * try the cheap and fast ffs_update first, and if that fails, 5073 * then we do the slower ffs_syncvnode of the directory. 5074 */ 5075 if (flushparent) { 5076 if ((error = ffs_update(pvp, 1)) != 0) { 5077 vput(pvp); 5078 return (error); 5079 } 5080 if ((pagedep->pd_state & NEWBLOCK) && 5081 (error = ffs_syncvnode(pvp, MNT_WAIT))) { 5082 vput(pvp); 5083 return (error); 5084 } 5085 } 5086 /* 5087 * Flush directory page containing the inode's name. 5088 */ 5089 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred, 5090 &bp); 5091 if (error == 0) 5092 error = bwrite(bp); 5093 else 5094 brelse(bp); 5095 vput(pvp); 5096 if (error != 0) 5097 return (error); 5098 ACQUIRE_LOCK(&lk); 5099 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 5100 break; 5101 } 5102 FREE_LOCK(&lk); 5103 return (0); 5104 } 5105 5106 /* 5107 * Flush all the dirty bitmaps associated with the block device 5108 * before flushing the rest of the dirty blocks so as to reduce 5109 * the number of dependencies that will have to be rolled back. 5110 */ 5111 void 5112 softdep_fsync_mountdev(vp) 5113 struct vnode *vp; 5114 { 5115 struct buf *bp, *nbp; 5116 struct worklist *wk; 5117 5118 if (!vn_isdisk(vp, NULL)) 5119 panic("softdep_fsync_mountdev: vnode not a disk"); 5120 restart: 5121 ACQUIRE_LOCK(&lk); 5122 VI_LOCK(vp); 5123 TAILQ_FOREACH_SAFE(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs, nbp) { 5124 /* 5125 * If it is already scheduled, skip to the next buffer. 5126 */ 5127 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 5128 continue; 5129 5130 if ((bp->b_flags & B_DELWRI) == 0) 5131 panic("softdep_fsync_mountdev: not dirty"); 5132 /* 5133 * We are only interested in bitmaps with outstanding 5134 * dependencies. 5135 */ 5136 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 5137 wk->wk_type != D_BMSAFEMAP || 5138 (bp->b_vflags & BV_BKGRDINPROG)) { 5139 BUF_UNLOCK(bp); 5140 continue; 5141 } 5142 VI_UNLOCK(vp); 5143 FREE_LOCK(&lk); 5144 bremfree(bp); 5145 (void) bawrite(bp); 5146 goto restart; 5147 } 5148 FREE_LOCK(&lk); 5149 drain_output(vp); 5150 VI_UNLOCK(vp); 5151 } 5152 5153 /* 5154 * This routine is called when we are trying to synchronously flush a 5155 * file. This routine must eliminate any filesystem metadata dependencies 5156 * so that the syncing routine can succeed by pushing the dirty blocks 5157 * associated with the file. If any I/O errors occur, they are returned. 5158 */ 5159 int 5160 softdep_sync_metadata(struct vnode *vp) 5161 { 5162 struct pagedep *pagedep; 5163 struct allocdirect *adp; 5164 struct allocindir *aip; 5165 struct buf *bp, *nbp; 5166 struct worklist *wk; 5167 int i, error, waitfor; 5168 5169 if (!DOINGSOFTDEP(vp)) 5170 return (0); 5171 /* 5172 * Ensure that any direct block dependencies have been cleared. 5173 */ 5174 ACQUIRE_LOCK(&lk); 5175 if ((error = flush_inodedep_deps(vp->v_mount, VTOI(vp)->i_number))) { 5176 FREE_LOCK(&lk); 5177 return (error); 5178 } 5179 FREE_LOCK(&lk); 5180 /* 5181 * For most files, the only metadata dependencies are the 5182 * cylinder group maps that allocate their inode or blocks. 5183 * The block allocation dependencies can be found by traversing 5184 * the dependency lists for any buffers that remain on their 5185 * dirty buffer list. The inode allocation dependency will 5186 * be resolved when the inode is updated with MNT_WAIT. 5187 * This work is done in two passes. The first pass grabs most 5188 * of the buffers and begins asynchronously writing them. The 5189 * only way to wait for these asynchronous writes is to sleep 5190 * on the filesystem vnode which may stay busy for a long time 5191 * if the filesystem is active. So, instead, we make a second 5192 * pass over the dependencies blocking on each write. In the 5193 * usual case we will be blocking against a write that we 5194 * initiated, so when it is done the dependency will have been 5195 * resolved. Thus the second pass is expected to end quickly. 5196 */ 5197 waitfor = MNT_NOWAIT; 5198 5199 top: 5200 /* 5201 * We must wait for any I/O in progress to finish so that 5202 * all potential buffers on the dirty list will be visible. 5203 */ 5204 VI_LOCK(vp); 5205 drain_output(vp); 5206 while ((bp = TAILQ_FIRST(&vp->v_bufobj.bo_dirty.bv_hd)) != NULL) { 5207 bp = getdirtybuf(bp, VI_MTX(vp), MNT_WAIT); 5208 if (bp) 5209 break; 5210 } 5211 VI_UNLOCK(vp); 5212 if (bp == NULL) 5213 return (0); 5214 loop: 5215 /* While syncing snapshots, we must allow recursive lookups */ 5216 bp->b_lock.lk_flags |= LK_CANRECURSE; 5217 ACQUIRE_LOCK(&lk); 5218 /* 5219 * As we hold the buffer locked, none of its dependencies 5220 * will disappear. 5221 */ 5222 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5223 switch (wk->wk_type) { 5224 5225 case D_ALLOCDIRECT: 5226 adp = WK_ALLOCDIRECT(wk); 5227 if (adp->ad_state & DEPCOMPLETE) 5228 continue; 5229 nbp = adp->ad_buf; 5230 nbp = getdirtybuf(nbp, &lk, waitfor); 5231 if (nbp == NULL) 5232 continue; 5233 FREE_LOCK(&lk); 5234 if (waitfor == MNT_NOWAIT) { 5235 bawrite(nbp); 5236 } else if ((error = bwrite(nbp)) != 0) { 5237 break; 5238 } 5239 ACQUIRE_LOCK(&lk); 5240 continue; 5241 5242 case D_ALLOCINDIR: 5243 aip = WK_ALLOCINDIR(wk); 5244 if (aip->ai_state & DEPCOMPLETE) 5245 continue; 5246 nbp = aip->ai_buf; 5247 nbp = getdirtybuf(nbp, &lk, waitfor); 5248 if (nbp == NULL) 5249 continue; 5250 FREE_LOCK(&lk); 5251 if (waitfor == MNT_NOWAIT) { 5252 bawrite(nbp); 5253 } else if ((error = bwrite(nbp)) != 0) { 5254 break; 5255 } 5256 ACQUIRE_LOCK(&lk); 5257 continue; 5258 5259 case D_INDIRDEP: 5260 restart: 5261 5262 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) { 5263 if (aip->ai_state & DEPCOMPLETE) 5264 continue; 5265 nbp = aip->ai_buf; 5266 nbp = getdirtybuf(nbp, &lk, MNT_WAIT); 5267 if (nbp == NULL) 5268 goto restart; 5269 FREE_LOCK(&lk); 5270 if ((error = bwrite(nbp)) != 0) { 5271 break; 5272 } 5273 ACQUIRE_LOCK(&lk); 5274 goto restart; 5275 } 5276 continue; 5277 5278 case D_INODEDEP: 5279 if ((error = flush_inodedep_deps(wk->wk_mp, 5280 WK_INODEDEP(wk)->id_ino)) != 0) { 5281 FREE_LOCK(&lk); 5282 break; 5283 } 5284 continue; 5285 5286 case D_PAGEDEP: 5287 /* 5288 * We are trying to sync a directory that may 5289 * have dependencies on both its own metadata 5290 * and/or dependencies on the inodes of any 5291 * recently allocated files. We walk its diradd 5292 * lists pushing out the associated inode. 5293 */ 5294 pagedep = WK_PAGEDEP(wk); 5295 for (i = 0; i < DAHASHSZ; i++) { 5296 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 5297 continue; 5298 if ((error = 5299 flush_pagedep_deps(vp, wk->wk_mp, 5300 &pagedep->pd_diraddhd[i]))) { 5301 FREE_LOCK(&lk); 5302 break; 5303 } 5304 } 5305 continue; 5306 5307 case D_MKDIR: 5308 /* 5309 * This case should never happen if the vnode has 5310 * been properly sync'ed. However, if this function 5311 * is used at a place where the vnode has not yet 5312 * been sync'ed, this dependency can show up. So, 5313 * rather than panic, just flush it. 5314 */ 5315 nbp = WK_MKDIR(wk)->md_buf; 5316 nbp = getdirtybuf(nbp, &lk, waitfor); 5317 if (nbp == NULL) 5318 continue; 5319 FREE_LOCK(&lk); 5320 if (waitfor == MNT_NOWAIT) { 5321 bawrite(nbp); 5322 } else if ((error = bwrite(nbp)) != 0) { 5323 break; 5324 } 5325 ACQUIRE_LOCK(&lk); 5326 continue; 5327 5328 case D_BMSAFEMAP: 5329 /* 5330 * This case should never happen if the vnode has 5331 * been properly sync'ed. However, if this function 5332 * is used at a place where the vnode has not yet 5333 * been sync'ed, this dependency can show up. So, 5334 * rather than panic, just flush it. 5335 */ 5336 nbp = WK_BMSAFEMAP(wk)->sm_buf; 5337 nbp = getdirtybuf(nbp, &lk, waitfor); 5338 if (nbp == NULL) 5339 continue; 5340 FREE_LOCK(&lk); 5341 if (waitfor == MNT_NOWAIT) { 5342 bawrite(nbp); 5343 } else if ((error = bwrite(nbp)) != 0) { 5344 break; 5345 } 5346 ACQUIRE_LOCK(&lk); 5347 continue; 5348 5349 default: 5350 panic("softdep_sync_metadata: Unknown type %s", 5351 TYPENAME(wk->wk_type)); 5352 /* NOTREACHED */ 5353 } 5354 /* We reach here only in error and unlocked */ 5355 if (error == 0) 5356 panic("softdep_sync_metadata: zero error"); 5357 bp->b_lock.lk_flags &= ~LK_CANRECURSE; 5358 bawrite(bp); 5359 return (error); 5360 } 5361 FREE_LOCK(&lk); 5362 VI_LOCK(vp); 5363 while ((nbp = TAILQ_NEXT(bp, b_bobufs)) != NULL) { 5364 nbp = getdirtybuf(nbp, VI_MTX(vp), MNT_WAIT); 5365 if (nbp) 5366 break; 5367 } 5368 VI_UNLOCK(vp); 5369 bp->b_lock.lk_flags &= ~LK_CANRECURSE; 5370 bawrite(bp); 5371 if (nbp != NULL) { 5372 bp = nbp; 5373 goto loop; 5374 } 5375 /* 5376 * The brief unlock is to allow any pent up dependency 5377 * processing to be done. Then proceed with the second pass. 5378 */ 5379 if (waitfor == MNT_NOWAIT) { 5380 waitfor = MNT_WAIT; 5381 goto top; 5382 } 5383 5384 /* 5385 * If we have managed to get rid of all the dirty buffers, 5386 * then we are done. For certain directories and block 5387 * devices, we may need to do further work. 5388 * 5389 * We must wait for any I/O in progress to finish so that 5390 * all potential buffers on the dirty list will be visible. 5391 */ 5392 VI_LOCK(vp); 5393 drain_output(vp); 5394 VI_UNLOCK(vp); 5395 return (0); 5396 } 5397 5398 /* 5399 * Flush the dependencies associated with an inodedep. 5400 * Called with splbio blocked. 5401 */ 5402 static int 5403 flush_inodedep_deps(mp, ino) 5404 struct mount *mp; 5405 ino_t ino; 5406 { 5407 struct inodedep *inodedep; 5408 int error, waitfor; 5409 5410 /* 5411 * This work is done in two passes. The first pass grabs most 5412 * of the buffers and begins asynchronously writing them. The 5413 * only way to wait for these asynchronous writes is to sleep 5414 * on the filesystem vnode which may stay busy for a long time 5415 * if the filesystem is active. So, instead, we make a second 5416 * pass over the dependencies blocking on each write. In the 5417 * usual case we will be blocking against a write that we 5418 * initiated, so when it is done the dependency will have been 5419 * resolved. Thus the second pass is expected to end quickly. 5420 * We give a brief window at the top of the loop to allow 5421 * any pending I/O to complete. 5422 */ 5423 for (error = 0, waitfor = MNT_NOWAIT; ; ) { 5424 if (error) 5425 return (error); 5426 FREE_LOCK(&lk); 5427 ACQUIRE_LOCK(&lk); 5428 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 5429 return (0); 5430 if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) || 5431 flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) || 5432 flush_deplist(&inodedep->id_extupdt, waitfor, &error) || 5433 flush_deplist(&inodedep->id_newextupdt, waitfor, &error)) 5434 continue; 5435 /* 5436 * If pass2, we are done, otherwise do pass 2. 5437 */ 5438 if (waitfor == MNT_WAIT) 5439 break; 5440 waitfor = MNT_WAIT; 5441 } 5442 /* 5443 * Try freeing inodedep in case all dependencies have been removed. 5444 */ 5445 if (inodedep_lookup(mp, ino, 0, &inodedep) != 0) 5446 (void) free_inodedep(inodedep); 5447 return (0); 5448 } 5449 5450 /* 5451 * Flush an inode dependency list. 5452 * Called with splbio blocked. 5453 */ 5454 static int 5455 flush_deplist(listhead, waitfor, errorp) 5456 struct allocdirectlst *listhead; 5457 int waitfor; 5458 int *errorp; 5459 { 5460 struct allocdirect *adp; 5461 struct buf *bp; 5462 5463 mtx_assert(&lk, MA_OWNED); 5464 TAILQ_FOREACH(adp, listhead, ad_next) { 5465 if (adp->ad_state & DEPCOMPLETE) 5466 continue; 5467 bp = adp->ad_buf; 5468 bp = getdirtybuf(bp, &lk, waitfor); 5469 if (bp == NULL) { 5470 if (waitfor == MNT_NOWAIT) 5471 continue; 5472 return (1); 5473 } 5474 FREE_LOCK(&lk); 5475 if (waitfor == MNT_NOWAIT) { 5476 bawrite(bp); 5477 } else if ((*errorp = bwrite(bp)) != 0) { 5478 ACQUIRE_LOCK(&lk); 5479 return (1); 5480 } 5481 ACQUIRE_LOCK(&lk); 5482 return (1); 5483 } 5484 return (0); 5485 } 5486 5487 /* 5488 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 5489 * Called with splbio blocked. 5490 */ 5491 static int 5492 flush_pagedep_deps(pvp, mp, diraddhdp) 5493 struct vnode *pvp; 5494 struct mount *mp; 5495 struct diraddhd *diraddhdp; 5496 { 5497 struct inodedep *inodedep; 5498 struct ufsmount *ump; 5499 struct diradd *dap; 5500 struct vnode *vp; 5501 int error = 0; 5502 struct buf *bp; 5503 ino_t inum; 5504 5505 ump = VFSTOUFS(mp); 5506 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 5507 /* 5508 * Flush ourselves if this directory entry 5509 * has a MKDIR_PARENT dependency. 5510 */ 5511 if (dap->da_state & MKDIR_PARENT) { 5512 FREE_LOCK(&lk); 5513 if ((error = ffs_update(pvp, 1)) != 0) 5514 break; 5515 ACQUIRE_LOCK(&lk); 5516 /* 5517 * If that cleared dependencies, go on to next. 5518 */ 5519 if (dap != LIST_FIRST(diraddhdp)) 5520 continue; 5521 if (dap->da_state & MKDIR_PARENT) 5522 panic("flush_pagedep_deps: MKDIR_PARENT"); 5523 } 5524 /* 5525 * A newly allocated directory must have its "." and 5526 * ".." entries written out before its name can be 5527 * committed in its parent. We do not want or need 5528 * the full semantics of a synchronous ffs_syncvnode as 5529 * that may end up here again, once for each directory 5530 * level in the filesystem. Instead, we push the blocks 5531 * and wait for them to clear. We have to fsync twice 5532 * because the first call may choose to defer blocks 5533 * that still have dependencies, but deferral will 5534 * happen at most once. 5535 */ 5536 inum = dap->da_newinum; 5537 if (dap->da_state & MKDIR_BODY) { 5538 FREE_LOCK(&lk); 5539 if ((error = ffs_vget(mp, inum, LK_EXCLUSIVE, &vp))) 5540 break; 5541 if ((error=ffs_syncvnode(vp, MNT_NOWAIT)) || 5542 (error=ffs_syncvnode(vp, MNT_NOWAIT))) { 5543 vput(vp); 5544 break; 5545 } 5546 VI_LOCK(vp); 5547 drain_output(vp); 5548 VI_UNLOCK(vp); 5549 vput(vp); 5550 ACQUIRE_LOCK(&lk); 5551 /* 5552 * If that cleared dependencies, go on to next. 5553 */ 5554 if (dap != LIST_FIRST(diraddhdp)) 5555 continue; 5556 if (dap->da_state & MKDIR_BODY) 5557 panic("flush_pagedep_deps: MKDIR_BODY"); 5558 } 5559 /* 5560 * Flush the inode on which the directory entry depends. 5561 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 5562 * the only remaining dependency is that the updated inode 5563 * count must get pushed to disk. The inode has already 5564 * been pushed into its inode buffer (via VOP_UPDATE) at 5565 * the time of the reference count change. So we need only 5566 * locate that buffer, ensure that there will be no rollback 5567 * caused by a bitmap dependency, then write the inode buffer. 5568 */ 5569 retry: 5570 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 5571 panic("flush_pagedep_deps: lost inode"); 5572 /* 5573 * If the inode still has bitmap dependencies, 5574 * push them to disk. 5575 */ 5576 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 5577 bp = inodedep->id_buf; 5578 bp = getdirtybuf(bp, &lk, MNT_WAIT); 5579 if (bp == NULL) 5580 goto retry; 5581 FREE_LOCK(&lk); 5582 if ((error = bwrite(bp)) != 0) 5583 break; 5584 ACQUIRE_LOCK(&lk); 5585 if (dap != LIST_FIRST(diraddhdp)) 5586 continue; 5587 } 5588 /* 5589 * If the inode is still sitting in a buffer waiting 5590 * to be written, push it to disk. 5591 */ 5592 FREE_LOCK(&lk); 5593 if ((error = bread(ump->um_devvp, 5594 fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), 5595 (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0) { 5596 brelse(bp); 5597 break; 5598 } 5599 if ((error = bwrite(bp)) != 0) 5600 break; 5601 ACQUIRE_LOCK(&lk); 5602 /* 5603 * If we have failed to get rid of all the dependencies 5604 * then something is seriously wrong. 5605 */ 5606 if (dap == LIST_FIRST(diraddhdp)) 5607 panic("flush_pagedep_deps: flush failed"); 5608 } 5609 if (error) 5610 ACQUIRE_LOCK(&lk); 5611 return (error); 5612 } 5613 5614 /* 5615 * A large burst of file addition or deletion activity can drive the 5616 * memory load excessively high. First attempt to slow things down 5617 * using the techniques below. If that fails, this routine requests 5618 * the offending operations to fall back to running synchronously 5619 * until the memory load returns to a reasonable level. 5620 */ 5621 int 5622 softdep_slowdown(vp) 5623 struct vnode *vp; 5624 { 5625 int max_softdeps_hard; 5626 5627 ACQUIRE_LOCK(&lk); 5628 max_softdeps_hard = max_softdeps * 11 / 10; 5629 if (num_dirrem < max_softdeps_hard / 2 && 5630 num_inodedep < max_softdeps_hard && 5631 VFSTOUFS(vp->v_mount)->um_numindirdeps < maxindirdeps) { 5632 FREE_LOCK(&lk); 5633 return (0); 5634 } 5635 if (VFSTOUFS(vp->v_mount)->um_numindirdeps >= maxindirdeps) 5636 softdep_speedup(); 5637 stat_sync_limit_hit += 1; 5638 FREE_LOCK(&lk); 5639 return (1); 5640 } 5641 5642 /* 5643 * Called by the allocation routines when they are about to fail 5644 * in the hope that we can free up some disk space. 5645 * 5646 * First check to see if the work list has anything on it. If it has, 5647 * clean up entries until we successfully free some space. Because this 5648 * process holds inodes locked, we cannot handle any remove requests 5649 * that might block on a locked inode as that could lead to deadlock. 5650 * If the worklist yields no free space, encourage the syncer daemon 5651 * to help us. In no event will we try for longer than tickdelay seconds. 5652 */ 5653 int 5654 softdep_request_cleanup(fs, vp) 5655 struct fs *fs; 5656 struct vnode *vp; 5657 { 5658 struct ufsmount *ump; 5659 long starttime; 5660 ufs2_daddr_t needed; 5661 int error; 5662 5663 ump = VTOI(vp)->i_ump; 5664 mtx_assert(UFS_MTX(ump), MA_OWNED); 5665 needed = fs->fs_cstotal.cs_nbfree + fs->fs_contigsumsize; 5666 starttime = time_second + tickdelay; 5667 /* 5668 * If we are being called because of a process doing a 5669 * copy-on-write, then it is not safe to update the vnode 5670 * as we may recurse into the copy-on-write routine. 5671 */ 5672 if (!(curthread->td_pflags & TDP_COWINPROGRESS)) { 5673 UFS_UNLOCK(ump); 5674 error = ffs_update(vp, 1); 5675 UFS_LOCK(ump); 5676 if (error != 0) 5677 return (0); 5678 } 5679 while (fs->fs_pendingblocks > 0 && fs->fs_cstotal.cs_nbfree <= needed) { 5680 if (time_second > starttime) 5681 return (0); 5682 UFS_UNLOCK(ump); 5683 ACQUIRE_LOCK(&lk); 5684 if (ump->softdep_on_worklist > 0 && 5685 process_worklist_item(UFSTOVFS(ump), LK_NOWAIT) != -1) { 5686 stat_worklist_push += 1; 5687 FREE_LOCK(&lk); 5688 UFS_LOCK(ump); 5689 continue; 5690 } 5691 request_cleanup(UFSTOVFS(ump), FLUSH_REMOVE_WAIT); 5692 FREE_LOCK(&lk); 5693 UFS_LOCK(ump); 5694 } 5695 return (1); 5696 } 5697 5698 /* 5699 * If memory utilization has gotten too high, deliberately slow things 5700 * down and speed up the I/O processing. 5701 */ 5702 extern struct thread *syncertd; 5703 static int 5704 request_cleanup(mp, resource) 5705 struct mount *mp; 5706 int resource; 5707 { 5708 struct thread *td = curthread; 5709 struct ufsmount *ump; 5710 5711 mtx_assert(&lk, MA_OWNED); 5712 /* 5713 * We never hold up the filesystem syncer or buf daemon. 5714 */ 5715 if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF)) 5716 return (0); 5717 ump = VFSTOUFS(mp); 5718 /* 5719 * First check to see if the work list has gotten backlogged. 5720 * If it has, co-opt this process to help clean up two entries. 5721 * Because this process may hold inodes locked, we cannot 5722 * handle any remove requests that might block on a locked 5723 * inode as that could lead to deadlock. We set TDP_SOFTDEP 5724 * to avoid recursively processing the worklist. 5725 */ 5726 if (ump->softdep_on_worklist > max_softdeps / 10) { 5727 td->td_pflags |= TDP_SOFTDEP; 5728 process_worklist_item(mp, LK_NOWAIT); 5729 process_worklist_item(mp, LK_NOWAIT); 5730 td->td_pflags &= ~TDP_SOFTDEP; 5731 stat_worklist_push += 2; 5732 return(1); 5733 } 5734 /* 5735 * Next, we attempt to speed up the syncer process. If that 5736 * is successful, then we allow the process to continue. 5737 */ 5738 if (softdep_speedup() && resource != FLUSH_REMOVE_WAIT) 5739 return(0); 5740 /* 5741 * If we are resource constrained on inode dependencies, try 5742 * flushing some dirty inodes. Otherwise, we are constrained 5743 * by file deletions, so try accelerating flushes of directories 5744 * with removal dependencies. We would like to do the cleanup 5745 * here, but we probably hold an inode locked at this point and 5746 * that might deadlock against one that we try to clean. So, 5747 * the best that we can do is request the syncer daemon to do 5748 * the cleanup for us. 5749 */ 5750 switch (resource) { 5751 5752 case FLUSH_INODES: 5753 stat_ino_limit_push += 1; 5754 req_clear_inodedeps += 1; 5755 stat_countp = &stat_ino_limit_hit; 5756 break; 5757 5758 case FLUSH_REMOVE: 5759 case FLUSH_REMOVE_WAIT: 5760 stat_blk_limit_push += 1; 5761 req_clear_remove += 1; 5762 stat_countp = &stat_blk_limit_hit; 5763 break; 5764 5765 default: 5766 panic("request_cleanup: unknown type"); 5767 } 5768 /* 5769 * Hopefully the syncer daemon will catch up and awaken us. 5770 * We wait at most tickdelay before proceeding in any case. 5771 */ 5772 proc_waiting += 1; 5773 if (handle.callout == NULL) 5774 handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2); 5775 msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0); 5776 proc_waiting -= 1; 5777 return (1); 5778 } 5779 5780 /* 5781 * Awaken processes pausing in request_cleanup and clear proc_waiting 5782 * to indicate that there is no longer a timer running. 5783 */ 5784 static void 5785 pause_timer(arg) 5786 void *arg; 5787 { 5788 5789 ACQUIRE_LOCK(&lk); 5790 *stat_countp += 1; 5791 wakeup_one(&proc_waiting); 5792 if (proc_waiting > 0) 5793 handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2); 5794 else 5795 handle.callout = NULL; 5796 FREE_LOCK(&lk); 5797 } 5798 5799 /* 5800 * Flush out a directory with at least one removal dependency in an effort to 5801 * reduce the number of dirrem, freefile, and freeblks dependency structures. 5802 */ 5803 static void 5804 clear_remove(td) 5805 struct thread *td; 5806 { 5807 struct pagedep_hashhead *pagedephd; 5808 struct pagedep *pagedep; 5809 static int next = 0; 5810 struct mount *mp; 5811 struct vnode *vp; 5812 int error, cnt; 5813 ino_t ino; 5814 5815 mtx_assert(&lk, MA_OWNED); 5816 5817 for (cnt = 0; cnt < pagedep_hash; cnt++) { 5818 pagedephd = &pagedep_hashtbl[next++]; 5819 if (next >= pagedep_hash) 5820 next = 0; 5821 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 5822 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL) 5823 continue; 5824 mp = pagedep->pd_list.wk_mp; 5825 ino = pagedep->pd_ino; 5826 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 5827 continue; 5828 FREE_LOCK(&lk); 5829 if ((error = ffs_vget(mp, ino, LK_EXCLUSIVE, &vp))) { 5830 softdep_error("clear_remove: vget", error); 5831 vn_finished_write(mp); 5832 ACQUIRE_LOCK(&lk); 5833 return; 5834 } 5835 if ((error = ffs_syncvnode(vp, MNT_NOWAIT))) 5836 softdep_error("clear_remove: fsync", error); 5837 VI_LOCK(vp); 5838 drain_output(vp); 5839 VI_UNLOCK(vp); 5840 vput(vp); 5841 vn_finished_write(mp); 5842 ACQUIRE_LOCK(&lk); 5843 return; 5844 } 5845 } 5846 } 5847 5848 /* 5849 * Clear out a block of dirty inodes in an effort to reduce 5850 * the number of inodedep dependency structures. 5851 */ 5852 static void 5853 clear_inodedeps(td) 5854 struct thread *td; 5855 { 5856 struct inodedep_hashhead *inodedephd; 5857 struct inodedep *inodedep; 5858 static int next = 0; 5859 struct mount *mp; 5860 struct vnode *vp; 5861 struct fs *fs; 5862 int error, cnt; 5863 ino_t firstino, lastino, ino; 5864 5865 mtx_assert(&lk, MA_OWNED); 5866 /* 5867 * Pick a random inode dependency to be cleared. 5868 * We will then gather up all the inodes in its block 5869 * that have dependencies and flush them out. 5870 */ 5871 for (cnt = 0; cnt < inodedep_hash; cnt++) { 5872 inodedephd = &inodedep_hashtbl[next++]; 5873 if (next >= inodedep_hash) 5874 next = 0; 5875 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 5876 break; 5877 } 5878 if (inodedep == NULL) 5879 return; 5880 fs = inodedep->id_fs; 5881 mp = inodedep->id_list.wk_mp; 5882 /* 5883 * Find the last inode in the block with dependencies. 5884 */ 5885 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 5886 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 5887 if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0) 5888 break; 5889 /* 5890 * Asynchronously push all but the last inode with dependencies. 5891 * Synchronously push the last inode with dependencies to ensure 5892 * that the inode block gets written to free up the inodedeps. 5893 */ 5894 for (ino = firstino; ino <= lastino; ino++) { 5895 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 5896 continue; 5897 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 5898 continue; 5899 FREE_LOCK(&lk); 5900 if ((error = ffs_vget(mp, ino, LK_EXCLUSIVE, &vp)) != 0) { 5901 softdep_error("clear_inodedeps: vget", error); 5902 vn_finished_write(mp); 5903 ACQUIRE_LOCK(&lk); 5904 return; 5905 } 5906 if (ino == lastino) { 5907 if ((error = ffs_syncvnode(vp, MNT_WAIT))) 5908 softdep_error("clear_inodedeps: fsync1", error); 5909 } else { 5910 if ((error = ffs_syncvnode(vp, MNT_NOWAIT))) 5911 softdep_error("clear_inodedeps: fsync2", error); 5912 VI_LOCK(vp); 5913 drain_output(vp); 5914 VI_UNLOCK(vp); 5915 } 5916 vput(vp); 5917 vn_finished_write(mp); 5918 ACQUIRE_LOCK(&lk); 5919 } 5920 } 5921 5922 /* 5923 * Function to determine if the buffer has outstanding dependencies 5924 * that will cause a roll-back if the buffer is written. If wantcount 5925 * is set, return number of dependencies, otherwise just yes or no. 5926 */ 5927 static int 5928 softdep_count_dependencies(bp, wantcount) 5929 struct buf *bp; 5930 int wantcount; 5931 { 5932 struct worklist *wk; 5933 struct inodedep *inodedep; 5934 struct indirdep *indirdep; 5935 struct allocindir *aip; 5936 struct pagedep *pagedep; 5937 struct diradd *dap; 5938 int i, retval; 5939 5940 retval = 0; 5941 ACQUIRE_LOCK(&lk); 5942 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5943 switch (wk->wk_type) { 5944 5945 case D_INODEDEP: 5946 inodedep = WK_INODEDEP(wk); 5947 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 5948 /* bitmap allocation dependency */ 5949 retval += 1; 5950 if (!wantcount) 5951 goto out; 5952 } 5953 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 5954 /* direct block pointer dependency */ 5955 retval += 1; 5956 if (!wantcount) 5957 goto out; 5958 } 5959 if (TAILQ_FIRST(&inodedep->id_extupdt)) { 5960 /* direct block pointer dependency */ 5961 retval += 1; 5962 if (!wantcount) 5963 goto out; 5964 } 5965 continue; 5966 5967 case D_INDIRDEP: 5968 indirdep = WK_INDIRDEP(wk); 5969 5970 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 5971 /* indirect block pointer dependency */ 5972 retval += 1; 5973 if (!wantcount) 5974 goto out; 5975 } 5976 continue; 5977 5978 case D_PAGEDEP: 5979 pagedep = WK_PAGEDEP(wk); 5980 for (i = 0; i < DAHASHSZ; i++) { 5981 5982 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 5983 /* directory entry dependency */ 5984 retval += 1; 5985 if (!wantcount) 5986 goto out; 5987 } 5988 } 5989 continue; 5990 5991 case D_BMSAFEMAP: 5992 case D_ALLOCDIRECT: 5993 case D_ALLOCINDIR: 5994 case D_MKDIR: 5995 /* never a dependency on these blocks */ 5996 continue; 5997 5998 default: 5999 panic("softdep_check_for_rollback: Unexpected type %s", 6000 TYPENAME(wk->wk_type)); 6001 /* NOTREACHED */ 6002 } 6003 } 6004 out: 6005 FREE_LOCK(&lk); 6006 return retval; 6007 } 6008 6009 /* 6010 * Acquire exclusive access to a buffer. 6011 * Must be called with a locked mtx parameter. 6012 * Return acquired buffer or NULL on failure. 6013 */ 6014 static struct buf * 6015 getdirtybuf(bp, mtx, waitfor) 6016 struct buf *bp; 6017 struct mtx *mtx; 6018 int waitfor; 6019 { 6020 int error; 6021 6022 mtx_assert(mtx, MA_OWNED); 6023 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) { 6024 if (waitfor != MNT_WAIT) 6025 return (NULL); 6026 error = BUF_LOCK(bp, 6027 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, mtx); 6028 /* 6029 * Even if we sucessfully acquire bp here, we have dropped 6030 * mtx, which may violates our guarantee. 6031 */ 6032 if (error == 0) 6033 BUF_UNLOCK(bp); 6034 else if (error != ENOLCK) 6035 panic("getdirtybuf: inconsistent lock: %d", error); 6036 mtx_lock(mtx); 6037 return (NULL); 6038 } 6039 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 6040 if (mtx == &lk && waitfor == MNT_WAIT) { 6041 mtx_unlock(mtx); 6042 BO_LOCK(bp->b_bufobj); 6043 BUF_UNLOCK(bp); 6044 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 6045 bp->b_vflags |= BV_BKGRDWAIT; 6046 msleep(&bp->b_xflags, BO_MTX(bp->b_bufobj), 6047 PRIBIO | PDROP, "getbuf", 0); 6048 } else 6049 BO_UNLOCK(bp->b_bufobj); 6050 mtx_lock(mtx); 6051 return (NULL); 6052 } 6053 BUF_UNLOCK(bp); 6054 if (waitfor != MNT_WAIT) 6055 return (NULL); 6056 /* 6057 * The mtx argument must be bp->b_vp's mutex in 6058 * this case. 6059 */ 6060 #ifdef DEBUG_VFS_LOCKS 6061 if (bp->b_vp->v_type != VCHR) 6062 ASSERT_VI_LOCKED(bp->b_vp, "getdirtybuf"); 6063 #endif 6064 bp->b_vflags |= BV_BKGRDWAIT; 6065 msleep(&bp->b_xflags, mtx, PRIBIO, "getbuf", 0); 6066 return (NULL); 6067 } 6068 if ((bp->b_flags & B_DELWRI) == 0) { 6069 BUF_UNLOCK(bp); 6070 return (NULL); 6071 } 6072 bremfree(bp); 6073 return (bp); 6074 } 6075 6076 6077 /* 6078 * Check if it is safe to suspend the file system now. On entry, 6079 * the vnode interlock for devvp should be held. Return 0 with 6080 * the mount interlock held if the file system can be suspended now, 6081 * otherwise return EAGAIN with the mount interlock held. 6082 */ 6083 int 6084 softdep_check_suspend(struct mount *mp, 6085 struct vnode *devvp, 6086 int softdep_deps, 6087 int softdep_accdeps, 6088 int secondary_writes, 6089 int secondary_accwrites) 6090 { 6091 struct bufobj *bo; 6092 struct ufsmount *ump; 6093 int error; 6094 6095 ASSERT_VI_LOCKED(devvp, "softdep_check_suspend"); 6096 ump = VFSTOUFS(mp); 6097 bo = &devvp->v_bufobj; 6098 6099 for (;;) { 6100 if (!TRY_ACQUIRE_LOCK(&lk)) { 6101 VI_UNLOCK(devvp); 6102 ACQUIRE_LOCK(&lk); 6103 FREE_LOCK(&lk); 6104 VI_LOCK(devvp); 6105 continue; 6106 } 6107 if (!MNT_ITRYLOCK(mp)) { 6108 FREE_LOCK(&lk); 6109 VI_UNLOCK(devvp); 6110 MNT_ILOCK(mp); 6111 MNT_IUNLOCK(mp); 6112 VI_LOCK(devvp); 6113 continue; 6114 } 6115 if (mp->mnt_secondary_writes != 0) { 6116 FREE_LOCK(&lk); 6117 VI_UNLOCK(devvp); 6118 msleep(&mp->mnt_secondary_writes, 6119 MNT_MTX(mp), 6120 (PUSER - 1) | PDROP, "secwr", 0); 6121 VI_LOCK(devvp); 6122 continue; 6123 } 6124 break; 6125 } 6126 6127 /* 6128 * Reasons for needing more work before suspend: 6129 * - Dirty buffers on devvp. 6130 * - Softdep activity occurred after start of vnode sync loop 6131 * - Secondary writes occurred after start of vnode sync loop 6132 */ 6133 error = 0; 6134 if (bo->bo_numoutput > 0 || 6135 bo->bo_dirty.bv_cnt > 0 || 6136 softdep_deps != 0 || 6137 ump->softdep_deps != 0 || 6138 softdep_accdeps != ump->softdep_accdeps || 6139 secondary_writes != 0 || 6140 mp->mnt_secondary_writes != 0 || 6141 secondary_accwrites != mp->mnt_secondary_accwrites) 6142 error = EAGAIN; 6143 FREE_LOCK(&lk); 6144 VI_UNLOCK(devvp); 6145 return (error); 6146 } 6147 6148 6149 /* 6150 * Get the number of dependency structures for the file system, both 6151 * the current number and the total number allocated. These will 6152 * later be used to detect that softdep processing has occurred. 6153 */ 6154 void 6155 softdep_get_depcounts(struct mount *mp, 6156 int *softdep_depsp, 6157 int *softdep_accdepsp) 6158 { 6159 struct ufsmount *ump; 6160 6161 ump = VFSTOUFS(mp); 6162 ACQUIRE_LOCK(&lk); 6163 *softdep_depsp = ump->softdep_deps; 6164 *softdep_accdepsp = ump->softdep_accdeps; 6165 FREE_LOCK(&lk); 6166 } 6167 6168 /* 6169 * Wait for pending output on a vnode to complete. 6170 * Must be called with vnode lock and interlock locked. 6171 * 6172 * XXX: Should just be a call to bufobj_wwait(). 6173 */ 6174 static void 6175 drain_output(vp) 6176 struct vnode *vp; 6177 { 6178 ASSERT_VOP_LOCKED(vp, "drain_output"); 6179 ASSERT_VI_LOCKED(vp, "drain_output"); 6180 6181 while (vp->v_bufobj.bo_numoutput) { 6182 vp->v_bufobj.bo_flag |= BO_WWAIT; 6183 msleep((caddr_t)&vp->v_bufobj.bo_numoutput, 6184 VI_MTX(vp), PRIBIO + 1, "drainvp", 0); 6185 } 6186 } 6187 6188 /* 6189 * Called whenever a buffer that is being invalidated or reallocated 6190 * contains dependencies. This should only happen if an I/O error has 6191 * occurred. The routine is called with the buffer locked. 6192 */ 6193 static void 6194 softdep_deallocate_dependencies(bp) 6195 struct buf *bp; 6196 { 6197 6198 if ((bp->b_ioflags & BIO_ERROR) == 0) 6199 panic("softdep_deallocate_dependencies: dangling deps"); 6200 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); 6201 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 6202 } 6203 6204 /* 6205 * Function to handle asynchronous write errors in the filesystem. 6206 */ 6207 static void 6208 softdep_error(func, error) 6209 char *func; 6210 int error; 6211 { 6212 6213 /* XXX should do something better! */ 6214 printf("%s: got error %d while accessing filesystem\n", func, error); 6215 } 6216 6217 #endif /* SOFTUPDATES */ 6218