ffs_softdep.c revision 156203
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: head/sys/ufs/ffs/ffs_softdep.c 156203 2006-03-02 05:50:23Z jeff $"); 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 87int 88softdep_flushfiles(oldmnt, flags, td) 89 struct mount *oldmnt; 90 int flags; 91 struct thread *td; 92{ 93 94 panic("softdep_flushfiles called"); 95} 96 97int 98softdep_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 108void 109softdep_initialize() 110{ 111 112 return; 113} 114 115void 116softdep_uninitialize() 117{ 118 119 return; 120} 121 122void 123softdep_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 132void 133softdep_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 142void 143softdep_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 156void 157softdep_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 170void 171softdep_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 184void 185softdep_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 196void 197softdep_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 206void 207softdep_freefile(pvp, ino, mode) 208 struct vnode *pvp; 209 ino_t ino; 210 int mode; 211{ 212 213 panic("softdep_freefile called"); 214} 215 216int 217softdep_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 229void 230softdep_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 241void 242softdep_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 252void 253softdep_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 264void 265softdep_change_linkcnt(ip) 266 struct inode *ip; 267{ 268 269 panic("softdep_change_linkcnt called"); 270} 271 272void 273softdep_load_inodeblock(ip) 274 struct inode *ip; 275{ 276 277 panic("softdep_load_inodeblock called"); 278} 279 280void 281softdep_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 290int 291softdep_fsync(vp) 292 struct vnode *vp; /* the "in_core" copy of the inode */ 293{ 294 295 return (0); 296} 297 298void 299softdep_fsync_mountdev(vp) 300 struct vnode *vp; 301{ 302 303 return; 304} 305 306int 307softdep_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 317int 318softdep_sync_metadata(struct vnode *vp) 319{ 320 321 return (0); 322} 323 324int 325softdep_slowdown(vp) 326 struct vnode *vp; 327{ 328 329 panic("softdep_slowdown called"); 330} 331 332void 333softdep_releasefile(ip) 334 struct inode *ip; /* inode with the zero effective link count */ 335{ 336 337 panic("softdep_releasefile called"); 338} 339 340int 341softdep_request_cleanup(fs, vp) 342 struct fs *fs; 343 struct vnode *vp; 344{ 345 346 return (0); 347} 348 349#else 350/* 351 * These definitions need to be adapted to the system to which 352 * this file is being ported. 353 */ 354/* 355 * malloc types defined for the softdep system. 356 */ 357static MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies"); 358static MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies"); 359static MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation"); 360static MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map"); 361static MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode"); 362static MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies"); 363static MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block"); 364static MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode"); 365static MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode"); 366static MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated"); 367static MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry"); 368static MALLOC_DEFINE(M_MKDIR, "mkdir","New directory"); 369static MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted"); 370static MALLOC_DEFINE(M_NEWDIRBLK, "newdirblk","Unclaimed new directory block"); 371static MALLOC_DEFINE(M_SAVEDINO, "savedino","Saved inodes"); 372 373#define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE) 374 375#define D_PAGEDEP 0 376#define D_INODEDEP 1 377#define D_NEWBLK 2 378#define D_BMSAFEMAP 3 379#define D_ALLOCDIRECT 4 380#define D_INDIRDEP 5 381#define D_ALLOCINDIR 6 382#define D_FREEFRAG 7 383#define D_FREEBLKS 8 384#define D_FREEFILE 9 385#define D_DIRADD 10 386#define D_MKDIR 11 387#define D_DIRREM 12 388#define D_NEWDIRBLK 13 389#define D_LAST D_NEWDIRBLK 390 391/* 392 * translate from workitem type to memory type 393 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 394 */ 395static struct malloc_type *memtype[] = { 396 M_PAGEDEP, 397 M_INODEDEP, 398 M_NEWBLK, 399 M_BMSAFEMAP, 400 M_ALLOCDIRECT, 401 M_INDIRDEP, 402 M_ALLOCINDIR, 403 M_FREEFRAG, 404 M_FREEBLKS, 405 M_FREEFILE, 406 M_DIRADD, 407 M_MKDIR, 408 M_DIRREM, 409 M_NEWDIRBLK 410}; 411 412#define DtoM(type) (memtype[type]) 413 414/* 415 * Names of malloc types. 416 */ 417#define TYPENAME(type) \ 418 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???") 419/* 420 * End system adaptation definitions. 421 */ 422 423/* 424 * Forward declarations. 425 */ 426struct inodedep_hashhead; 427struct newblk_hashhead; 428struct pagedep_hashhead; 429 430/* 431 * Internal function prototypes. 432 */ 433static void softdep_error(char *, int); 434static void drain_output(struct vnode *); 435static struct buf *getdirtybuf(struct buf *, struct mtx *, int); 436static void clear_remove(struct thread *); 437static void clear_inodedeps(struct thread *); 438static int flush_pagedep_deps(struct vnode *, struct mount *, 439 struct diraddhd *); 440static int flush_inodedep_deps(struct mount *, ino_t); 441static int flush_deplist(struct allocdirectlst *, int, int *); 442static int handle_written_filepage(struct pagedep *, struct buf *); 443static void diradd_inode_written(struct diradd *, struct inodedep *); 444static int handle_written_inodeblock(struct inodedep *, struct buf *); 445static void handle_allocdirect_partdone(struct allocdirect *); 446static void handle_allocindir_partdone(struct allocindir *); 447static void initiate_write_filepage(struct pagedep *, struct buf *); 448static void handle_written_mkdir(struct mkdir *, int); 449static void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *); 450static void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *); 451static void handle_workitem_freefile(struct freefile *); 452static void handle_workitem_remove(struct dirrem *, struct vnode *); 453static struct dirrem *newdirrem(struct buf *, struct inode *, 454 struct inode *, int, struct dirrem **); 455static void free_diradd(struct diradd *); 456static void free_allocindir(struct allocindir *, struct inodedep *); 457static void free_newdirblk(struct newdirblk *); 458static int indir_trunc(struct freeblks *, ufs2_daddr_t, int, ufs_lbn_t, 459 ufs2_daddr_t *); 460static void deallocate_dependencies(struct buf *, struct inodedep *); 461static void free_allocdirect(struct allocdirectlst *, 462 struct allocdirect *, int); 463static int check_inode_unwritten(struct inodedep *); 464static int free_inodedep(struct inodedep *); 465static void handle_workitem_freeblocks(struct freeblks *, int); 466static void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *); 467static void setup_allocindir_phase2(struct buf *, struct inode *, 468 struct allocindir *); 469static struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t, 470 ufs2_daddr_t); 471static void handle_workitem_freefrag(struct freefrag *); 472static struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long); 473static void allocdirect_merge(struct allocdirectlst *, 474 struct allocdirect *, struct allocdirect *); 475static struct bmsafemap *bmsafemap_lookup(struct mount *, struct buf *); 476static int newblk_find(struct newblk_hashhead *, struct fs *, ufs2_daddr_t, 477 struct newblk **); 478static int newblk_lookup(struct fs *, ufs2_daddr_t, int, struct newblk **); 479static int inodedep_find(struct inodedep_hashhead *, struct fs *, ino_t, 480 struct inodedep **); 481static int inodedep_lookup(struct mount *, ino_t, int, struct inodedep **); 482static int pagedep_lookup(struct inode *, ufs_lbn_t, int, struct pagedep **); 483static int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t, 484 struct mount *mp, int, struct pagedep **); 485static void pause_timer(void *); 486static int request_cleanup(struct mount *, int); 487static int process_worklist_item(struct mount *, int); 488static void add_to_worklist(struct worklist *); 489static void softdep_flush(void); 490static int softdep_speedup(void); 491 492/* 493 * Exported softdep operations. 494 */ 495static void softdep_disk_io_initiation(struct buf *); 496static void softdep_disk_write_complete(struct buf *); 497static void softdep_deallocate_dependencies(struct buf *); 498static int softdep_count_dependencies(struct buf *bp, int); 499 500static struct mtx lk; 501MTX_SYSINIT(softdep_lock, &lk, "Softdep Lock", MTX_DEF); 502 503#define ACQUIRE_LOCK(lk) mtx_lock(lk) 504#define FREE_LOCK(lk) mtx_unlock(lk) 505 506/* 507 * Worklist queue management. 508 * These routines require that the lock be held. 509 */ 510#ifndef /* NOT */ DEBUG 511#define WORKLIST_INSERT(head, item) do { \ 512 (item)->wk_state |= ONWORKLIST; \ 513 LIST_INSERT_HEAD(head, item, wk_list); \ 514} while (0) 515#define WORKLIST_REMOVE(item) do { \ 516 (item)->wk_state &= ~ONWORKLIST; \ 517 LIST_REMOVE(item, wk_list); \ 518} while (0) 519#else /* DEBUG */ 520static void worklist_insert(struct workhead *, struct worklist *); 521static void worklist_remove(struct worklist *); 522 523#define WORKLIST_INSERT(head, item) worklist_insert(head, item) 524#define WORKLIST_REMOVE(item) worklist_remove(item) 525 526static void 527worklist_insert(head, item) 528 struct workhead *head; 529 struct worklist *item; 530{ 531 532 mtx_assert(&lk, MA_OWNED); 533 if (item->wk_state & ONWORKLIST) 534 panic("worklist_insert: already on list"); 535 item->wk_state |= ONWORKLIST; 536 LIST_INSERT_HEAD(head, item, wk_list); 537} 538 539static void 540worklist_remove(item) 541 struct worklist *item; 542{ 543 544 mtx_assert(&lk, MA_OWNED); 545 if ((item->wk_state & ONWORKLIST) == 0) 546 panic("worklist_remove: not on list"); 547 item->wk_state &= ~ONWORKLIST; 548 LIST_REMOVE(item, wk_list); 549} 550#endif /* DEBUG */ 551 552/* 553 * Routines for tracking and managing workitems. 554 */ 555static void workitem_free(struct worklist *, int); 556static void workitem_alloc(struct worklist *, int, struct mount *); 557 558#define WORKITEM_FREE(item, type) workitem_free((struct worklist *)(item), (type)) 559 560static void 561workitem_free(item, type) 562 struct worklist *item; 563 int type; 564{ 565 struct ufsmount *ump; 566 mtx_assert(&lk, MA_OWNED); 567 568#ifdef DEBUG 569 if (item->wk_state & ONWORKLIST) 570 panic("workitem_free: still on list"); 571 if (item->wk_type != type) 572 panic("workitem_free: type mismatch"); 573#endif 574 ump = VFSTOUFS(item->wk_mp); 575 if (--ump->softdep_deps == 0 && ump->softdep_req) 576 wakeup(&ump->softdep_deps); 577 FREE(item, DtoM(type)); 578} 579 580static void 581workitem_alloc(item, type, mp) 582 struct worklist *item; 583 int type; 584 struct mount *mp; 585{ 586 item->wk_type = type; 587 item->wk_mp = mp; 588 item->wk_state = 0; 589 ACQUIRE_LOCK(&lk); 590 VFSTOUFS(mp)->softdep_deps++; 591 FREE_LOCK(&lk); 592} 593 594/* 595 * Workitem queue management 596 */ 597static int max_softdeps; /* maximum number of structs before slowdown */ 598static int maxindirdeps = 50; /* max number of indirdeps before slowdown */ 599static int tickdelay = 2; /* number of ticks to pause during slowdown */ 600static int proc_waiting; /* tracks whether we have a timeout posted */ 601static int *stat_countp; /* statistic to count in proc_waiting timeout */ 602static struct callout_handle handle; /* handle on posted proc_waiting timeout */ 603static int req_pending; 604static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 605#define FLUSH_INODES 1 606static int req_clear_remove; /* syncer process flush some freeblks */ 607#define FLUSH_REMOVE 2 608#define FLUSH_REMOVE_WAIT 3 609/* 610 * runtime statistics 611 */ 612static int stat_worklist_push; /* number of worklist cleanups */ 613static int stat_blk_limit_push; /* number of times block limit neared */ 614static int stat_ino_limit_push; /* number of times inode limit neared */ 615static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 616static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 617static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 618static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 619static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 620static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 621static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 622 623SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, ""); 624SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, ""); 625SYSCTL_INT(_debug, OID_AUTO, maxindirdeps, CTLFLAG_RW, &maxindirdeps, 0, ""); 626SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,""); 627SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,""); 628SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,""); 629SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, ""); 630SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, ""); 631SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, ""); 632SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, ""); 633SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, ""); 634SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, ""); 635SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, ""); 636/* SYSCTL_INT(_debug, OID_AUTO, worklist_num, CTLFLAG_RD, &softdep_on_worklist, 0, ""); */ 637 638SYSCTL_DECL(_vfs_ffs); 639 640static int compute_summary_at_mount = 0; /* Whether to recompute the summary at mount time */ 641SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW, 642 &compute_summary_at_mount, 0, "Recompute summary at mount"); 643 644static struct proc *softdepproc; 645static struct kproc_desc softdep_kp = { 646 "softdepflush", 647 softdep_flush, 648 &softdepproc 649}; 650SYSINIT(sdproc, SI_SUB_KTHREAD_UPDATE, SI_ORDER_ANY, kproc_start, &softdep_kp) 651 652static void 653softdep_flush(void) 654{ 655 struct mount *nmp; 656 struct mount *mp; 657 struct thread *td; 658 int remaining; 659 660 td = curthread; 661 td->td_pflags |= TDP_NORUNNINGBUF; 662 663 for (;;) { 664 kthread_suspend_check(softdepproc); 665 ACQUIRE_LOCK(&lk); 666 /* 667 * If requested, try removing inode or removal dependencies. 668 */ 669 if (req_clear_inodedeps) { 670 clear_inodedeps(td); 671 req_clear_inodedeps -= 1; 672 wakeup_one(&proc_waiting); 673 } 674 if (req_clear_remove) { 675 clear_remove(td); 676 req_clear_remove -= 1; 677 wakeup_one(&proc_waiting); 678 } 679 FREE_LOCK(&lk); 680 remaining = 0; 681 mtx_lock(&mountlist_mtx); 682 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 683 nmp = TAILQ_NEXT(mp, mnt_list); 684 if ((mp->mnt_flag & MNT_SOFTDEP) == 0) 685 continue; 686 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) 687 continue; 688 softdep_process_worklist(mp, 0); 689 remaining += VFSTOUFS(mp)->softdep_on_worklist; 690 mtx_lock(&mountlist_mtx); 691 nmp = TAILQ_NEXT(mp, mnt_list); 692 vfs_unbusy(mp, td); 693 } 694 mtx_unlock(&mountlist_mtx); 695 if (remaining) 696 continue; 697 ACQUIRE_LOCK(&lk); 698 if (!req_pending) 699 msleep(&req_pending, &lk, PVM, "sdflush", hz); 700 req_pending = 0; 701 FREE_LOCK(&lk); 702 } 703} 704 705static int 706softdep_speedup(void) 707{ 708 ACQUIRE_LOCK(&lk); 709 if (req_pending == 0) { 710 req_pending = 1; 711 wakeup(&req_pending); 712 } 713 FREE_LOCK(&lk); 714 715 return speedup_syncer(); 716} 717 718/* 719 * Add an item to the end of the work queue. 720 * This routine requires that the lock be held. 721 * This is the only routine that adds items to the list. 722 * The following routine is the only one that removes items 723 * and does so in order from first to last. 724 */ 725static void 726add_to_worklist(wk) 727 struct worklist *wk; 728{ 729 struct ufsmount *ump; 730 731 mtx_assert(&lk, MA_OWNED); 732 ump = VFSTOUFS(wk->wk_mp); 733 if (wk->wk_state & ONWORKLIST) 734 panic("add_to_worklist: already on list"); 735 wk->wk_state |= ONWORKLIST; 736 if (LIST_FIRST(&ump->softdep_workitem_pending) == NULL) 737 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); 738 else 739 LIST_INSERT_AFTER(ump->softdep_worklist_tail, wk, wk_list); 740 ump->softdep_worklist_tail = wk; 741 ump->softdep_on_worklist += 1; 742} 743 744/* 745 * Process that runs once per second to handle items in the background queue. 746 * 747 * Note that we ensure that everything is done in the order in which they 748 * appear in the queue. The code below depends on this property to ensure 749 * that blocks of a file are freed before the inode itself is freed. This 750 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 751 * until all the old ones have been purged from the dependency lists. 752 */ 753int 754softdep_process_worklist(mp, full) 755 struct mount *mp; 756 int full; 757{ 758 struct thread *td = curthread; 759 int cnt, matchcnt, loopcount; 760 struct ufsmount *ump; 761 long starttime; 762 763 KASSERT(mp != NULL, ("softdep_process_worklist: NULL mp")); 764 /* 765 * Record the process identifier of our caller so that we can give 766 * this process preferential treatment in request_cleanup below. 767 */ 768 matchcnt = 0; 769 ump = VFSTOUFS(mp); 770 ACQUIRE_LOCK(&lk); 771 loopcount = 1; 772 starttime = time_second; 773 while (ump->softdep_on_worklist > 0) { 774 if ((cnt = process_worklist_item(mp, 0)) == -1) 775 break; 776 else 777 matchcnt += cnt; 778 /* 779 * If requested, try removing inode or removal dependencies. 780 */ 781 if (req_clear_inodedeps) { 782 clear_inodedeps(td); 783 req_clear_inodedeps -= 1; 784 wakeup_one(&proc_waiting); 785 } 786 if (req_clear_remove) { 787 clear_remove(td); 788 req_clear_remove -= 1; 789 wakeup_one(&proc_waiting); 790 } 791 /* 792 * We do not generally want to stop for buffer space, but if 793 * we are really being a buffer hog, we will stop and wait. 794 */ 795 if (loopcount++ % 128 == 0) { 796 FREE_LOCK(&lk); 797 bwillwrite(); 798 ACQUIRE_LOCK(&lk); 799 } 800 /* 801 * Never allow processing to run for more than one 802 * second. Otherwise the other mountpoints may get 803 * excessively backlogged. 804 */ 805 if (!full && starttime != time_second) { 806 matchcnt = -1; 807 break; 808 } 809 } 810 FREE_LOCK(&lk); 811 return (matchcnt); 812} 813 814/* 815 * Process one item on the worklist. 816 */ 817static int 818process_worklist_item(mp, flags) 819 struct mount *mp; 820 int flags; 821{ 822 struct worklist *wk, *wkend; 823 struct ufsmount *ump; 824 struct vnode *vp; 825 int matchcnt = 0; 826 827 mtx_assert(&lk, MA_OWNED); 828 KASSERT(mp != NULL, ("process_worklist_item: NULL mp")); 829 /* 830 * If we are being called because of a process doing a 831 * copy-on-write, then it is not safe to write as we may 832 * recurse into the copy-on-write routine. 833 */ 834 if (curthread->td_pflags & TDP_COWINPROGRESS) 835 return (-1); 836 /* 837 * Normally we just process each item on the worklist in order. 838 * However, if we are in a situation where we cannot lock any 839 * inodes, we have to skip over any dirrem requests whose 840 * vnodes are resident and locked. 841 */ 842 ump = VFSTOUFS(mp); 843 vp = NULL; 844 LIST_FOREACH(wk, &ump->softdep_workitem_pending, wk_list) { 845 if (wk->wk_state & INPROGRESS) 846 continue; 847 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM) 848 break; 849 wk->wk_state |= INPROGRESS; 850 FREE_LOCK(&lk); 851 ffs_vget(mp, WK_DIRREM(wk)->dm_oldinum, 852 LK_NOWAIT | LK_EXCLUSIVE, &vp); 853 ACQUIRE_LOCK(&lk); 854 wk->wk_state &= ~INPROGRESS; 855 if (vp != NULL) 856 break; 857 } 858 if (wk == 0) 859 return (-1); 860 /* 861 * Remove the item to be processed. If we are removing the last 862 * item on the list, we need to recalculate the tail pointer. 863 * As this happens rarely and usually when the list is short, 864 * we just run down the list to find it rather than tracking it 865 * in the above loop. 866 */ 867 WORKLIST_REMOVE(wk); 868 if (wk == ump->softdep_worklist_tail) { 869 LIST_FOREACH(wkend, &ump->softdep_workitem_pending, wk_list) 870 if (LIST_NEXT(wkend, wk_list) == NULL) 871 break; 872 ump->softdep_worklist_tail = wkend; 873 } 874 ump->softdep_on_worklist -= 1; 875 FREE_LOCK(&lk); 876 if (vn_write_suspend_wait(NULL, mp, V_NOWAIT)) 877 panic("process_worklist_item: suspended filesystem"); 878 matchcnt++; 879 switch (wk->wk_type) { 880 881 case D_DIRREM: 882 /* removal of a directory entry */ 883 handle_workitem_remove(WK_DIRREM(wk), vp); 884 break; 885 886 case D_FREEBLKS: 887 /* releasing blocks and/or fragments from a file */ 888 handle_workitem_freeblocks(WK_FREEBLKS(wk), flags & LK_NOWAIT); 889 break; 890 891 case D_FREEFRAG: 892 /* releasing a fragment when replaced as a file grows */ 893 handle_workitem_freefrag(WK_FREEFRAG(wk)); 894 break; 895 896 case D_FREEFILE: 897 /* releasing an inode when its link count drops to 0 */ 898 handle_workitem_freefile(WK_FREEFILE(wk)); 899 break; 900 901 default: 902 panic("%s_process_worklist: Unknown type %s", 903 "softdep", TYPENAME(wk->wk_type)); 904 /* NOTREACHED */ 905 } 906 ACQUIRE_LOCK(&lk); 907 return (matchcnt); 908} 909 910/* 911 * Move dependencies from one buffer to another. 912 */ 913void 914softdep_move_dependencies(oldbp, newbp) 915 struct buf *oldbp; 916 struct buf *newbp; 917{ 918 struct worklist *wk, *wktail; 919 920 if (LIST_FIRST(&newbp->b_dep) != NULL) 921 panic("softdep_move_dependencies: need merge code"); 922 wktail = 0; 923 ACQUIRE_LOCK(&lk); 924 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 925 LIST_REMOVE(wk, wk_list); 926 if (wktail == 0) 927 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 928 else 929 LIST_INSERT_AFTER(wktail, wk, wk_list); 930 wktail = wk; 931 } 932 FREE_LOCK(&lk); 933} 934 935/* 936 * Purge the work list of all items associated with a particular mount point. 937 */ 938int 939softdep_flushworklist(oldmnt, countp, td) 940 struct mount *oldmnt; 941 int *countp; 942 struct thread *td; 943{ 944 struct vnode *devvp; 945 int count, error = 0; 946 struct ufsmount *ump; 947 948 /* 949 * Alternately flush the block device associated with the mount 950 * point and process any dependencies that the flushing 951 * creates. We continue until no more worklist dependencies 952 * are found. 953 */ 954 *countp = 0; 955 ump = VFSTOUFS(oldmnt); 956 devvp = ump->um_devvp; 957 while ((count = softdep_process_worklist(oldmnt, 1)) > 0) { 958 *countp += count; 959 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, td); 960 error = VOP_FSYNC(devvp, MNT_WAIT, td); 961 VOP_UNLOCK(devvp, 0, td); 962 if (error) 963 break; 964 } 965 return (error); 966} 967 968int 969softdep_waitidle(struct mount *mp) 970{ 971 struct ufsmount *ump; 972 int error; 973 int i; 974 975 ump = VFSTOUFS(mp); 976 ACQUIRE_LOCK(&lk); 977 for (i = 0; i < 10 && ump->softdep_deps; i++) { 978 ump->softdep_req = 1; 979 if (ump->softdep_on_worklist) 980 panic("softdep_waitidle: work added after flush."); 981 msleep(&ump->softdep_deps, &lk, PVM, "softdeps", 1); 982 } 983 ump->softdep_req = 0; 984 FREE_LOCK(&lk); 985 error = 0; 986 if (i == 10) { 987 error = EBUSY; 988 printf("softdep_waitidle: Failed to flush worklist for %p", 989 mp); 990 } 991 992 return (error); 993} 994 995/* 996 * Flush all vnodes and worklist items associated with a specified mount point. 997 */ 998int 999softdep_flushfiles(oldmnt, flags, td) 1000 struct mount *oldmnt; 1001 int flags; 1002 struct thread *td; 1003{ 1004 int error, count, loopcnt; 1005 1006 error = 0; 1007 1008 /* 1009 * Alternately flush the vnodes associated with the mount 1010 * point and process any dependencies that the flushing 1011 * creates. In theory, this loop can happen at most twice, 1012 * but we give it a few extra just to be sure. 1013 */ 1014 for (loopcnt = 10; loopcnt > 0; loopcnt--) { 1015 /* 1016 * Do another flush in case any vnodes were brought in 1017 * as part of the cleanup operations. 1018 */ 1019 if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0) 1020 break; 1021 if ((error = softdep_flushworklist(oldmnt, &count, td)) != 0 || 1022 count == 0) 1023 break; 1024 } 1025 /* 1026 * If we are unmounting then it is an error to fail. If we 1027 * are simply trying to downgrade to read-only, then filesystem 1028 * activity can keep us busy forever, so we just fail with EBUSY. 1029 */ 1030 if (loopcnt == 0) { 1031 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 1032 panic("softdep_flushfiles: looping"); 1033 error = EBUSY; 1034 } 1035 if (!error) 1036 error = softdep_waitidle(oldmnt); 1037 return (error); 1038} 1039 1040/* 1041 * Structure hashing. 1042 * 1043 * There are three types of structures that can be looked up: 1044 * 1) pagedep structures identified by mount point, inode number, 1045 * and logical block. 1046 * 2) inodedep structures identified by mount point and inode number. 1047 * 3) newblk structures identified by mount point and 1048 * physical block number. 1049 * 1050 * The "pagedep" and "inodedep" dependency structures are hashed 1051 * separately from the file blocks and inodes to which they correspond. 1052 * This separation helps when the in-memory copy of an inode or 1053 * file block must be replaced. It also obviates the need to access 1054 * an inode or file page when simply updating (or de-allocating) 1055 * dependency structures. Lookup of newblk structures is needed to 1056 * find newly allocated blocks when trying to associate them with 1057 * their allocdirect or allocindir structure. 1058 * 1059 * The lookup routines optionally create and hash a new instance when 1060 * an existing entry is not found. 1061 */ 1062#define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 1063#define NODELAY 0x0002 /* cannot do background work */ 1064 1065/* 1066 * Structures and routines associated with pagedep caching. 1067 */ 1068LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl; 1069u_long pagedep_hash; /* size of hash table - 1 */ 1070#define PAGEDEP_HASH(mp, inum, lbn) \ 1071 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \ 1072 pagedep_hash]) 1073 1074static int 1075pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp) 1076 struct pagedep_hashhead *pagedephd; 1077 ino_t ino; 1078 ufs_lbn_t lbn; 1079 struct mount *mp; 1080 int flags; 1081 struct pagedep **pagedeppp; 1082{ 1083 struct pagedep *pagedep; 1084 1085 LIST_FOREACH(pagedep, pagedephd, pd_hash) 1086 if (ino == pagedep->pd_ino && 1087 lbn == pagedep->pd_lbn && 1088 mp == pagedep->pd_list.wk_mp) 1089 break; 1090 if (pagedep) { 1091 *pagedeppp = pagedep; 1092 if ((flags & DEPALLOC) != 0 && 1093 (pagedep->pd_state & ONWORKLIST) == 0) 1094 return (0); 1095 return (1); 1096 } 1097 *pagedeppp = NULL; 1098 return (0); 1099} 1100/* 1101 * Look up a pagedep. Return 1 if found, 0 if not found or found 1102 * when asked to allocate but not associated with any buffer. 1103 * If not found, allocate if DEPALLOC flag is passed. 1104 * Found or allocated entry is returned in pagedeppp. 1105 * This routine must be called with splbio interrupts blocked. 1106 */ 1107static int 1108pagedep_lookup(ip, lbn, flags, pagedeppp) 1109 struct inode *ip; 1110 ufs_lbn_t lbn; 1111 int flags; 1112 struct pagedep **pagedeppp; 1113{ 1114 struct pagedep *pagedep; 1115 struct pagedep_hashhead *pagedephd; 1116 struct mount *mp; 1117 int ret; 1118 int i; 1119 1120 mtx_assert(&lk, MA_OWNED); 1121 mp = ITOV(ip)->v_mount; 1122 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn); 1123 1124 ret = pagedep_find(pagedephd, ip->i_number, lbn, mp, flags, pagedeppp); 1125 if (*pagedeppp || (flags & DEPALLOC) == 0) 1126 return (ret); 1127 FREE_LOCK(&lk); 1128 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), 1129 M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO); 1130 workitem_alloc(&pagedep->pd_list, D_PAGEDEP, mp); 1131 ACQUIRE_LOCK(&lk); 1132 ret = pagedep_find(pagedephd, ip->i_number, lbn, mp, flags, pagedeppp); 1133 if (*pagedeppp) { 1134 WORKITEM_FREE(pagedep, D_PAGEDEP); 1135 return (ret); 1136 } 1137 pagedep->pd_ino = ip->i_number; 1138 pagedep->pd_lbn = lbn; 1139 LIST_INIT(&pagedep->pd_dirremhd); 1140 LIST_INIT(&pagedep->pd_pendinghd); 1141 for (i = 0; i < DAHASHSZ; i++) 1142 LIST_INIT(&pagedep->pd_diraddhd[i]); 1143 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 1144 *pagedeppp = pagedep; 1145 return (0); 1146} 1147 1148/* 1149 * Structures and routines associated with inodedep caching. 1150 */ 1151LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl; 1152static u_long inodedep_hash; /* size of hash table - 1 */ 1153static long num_inodedep; /* number of inodedep allocated */ 1154#define INODEDEP_HASH(fs, inum) \ 1155 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash]) 1156 1157static int 1158inodedep_find(inodedephd, fs, inum, inodedeppp) 1159 struct inodedep_hashhead *inodedephd; 1160 struct fs *fs; 1161 ino_t inum; 1162 struct inodedep **inodedeppp; 1163{ 1164 struct inodedep *inodedep; 1165 1166 LIST_FOREACH(inodedep, inodedephd, id_hash) 1167 if (inum == inodedep->id_ino && fs == inodedep->id_fs) 1168 break; 1169 if (inodedep) { 1170 *inodedeppp = inodedep; 1171 return (1); 1172 } 1173 *inodedeppp = NULL; 1174 1175 return (0); 1176} 1177/* 1178 * Look up an inodedep. Return 1 if found, 0 if not found. 1179 * If not found, allocate if DEPALLOC flag is passed. 1180 * Found or allocated entry is returned in inodedeppp. 1181 * This routine must be called with splbio interrupts blocked. 1182 */ 1183static int 1184inodedep_lookup(mp, inum, flags, inodedeppp) 1185 struct mount *mp; 1186 ino_t inum; 1187 int flags; 1188 struct inodedep **inodedeppp; 1189{ 1190 struct inodedep *inodedep; 1191 struct inodedep_hashhead *inodedephd; 1192 struct fs *fs; 1193 1194 mtx_assert(&lk, MA_OWNED); 1195 fs = VFSTOUFS(mp)->um_fs; 1196 inodedephd = INODEDEP_HASH(fs, inum); 1197 1198 if (inodedep_find(inodedephd, fs, inum, inodedeppp)) 1199 return (1); 1200 if ((flags & DEPALLOC) == 0) 1201 return (0); 1202 /* 1203 * If we are over our limit, try to improve the situation. 1204 */ 1205 if (num_inodedep > max_softdeps && (flags & NODELAY) == 0) 1206 request_cleanup(mp, FLUSH_INODES); 1207 FREE_LOCK(&lk); 1208 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep), 1209 M_INODEDEP, M_SOFTDEP_FLAGS); 1210 workitem_alloc(&inodedep->id_list, D_INODEDEP, mp); 1211 ACQUIRE_LOCK(&lk); 1212 if (inodedep_find(inodedephd, fs, inum, inodedeppp)) { 1213 WORKITEM_FREE(inodedep, D_INODEDEP); 1214 return (1); 1215 } 1216 num_inodedep += 1; 1217 inodedep->id_fs = fs; 1218 inodedep->id_ino = inum; 1219 inodedep->id_state = ALLCOMPLETE; 1220 inodedep->id_nlinkdelta = 0; 1221 inodedep->id_savedino1 = NULL; 1222 inodedep->id_savedsize = -1; 1223 inodedep->id_savedextsize = -1; 1224 inodedep->id_buf = NULL; 1225 LIST_INIT(&inodedep->id_pendinghd); 1226 LIST_INIT(&inodedep->id_inowait); 1227 LIST_INIT(&inodedep->id_bufwait); 1228 TAILQ_INIT(&inodedep->id_inoupdt); 1229 TAILQ_INIT(&inodedep->id_newinoupdt); 1230 TAILQ_INIT(&inodedep->id_extupdt); 1231 TAILQ_INIT(&inodedep->id_newextupdt); 1232 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 1233 *inodedeppp = inodedep; 1234 return (0); 1235} 1236 1237/* 1238 * Structures and routines associated with newblk caching. 1239 */ 1240LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl; 1241u_long newblk_hash; /* size of hash table - 1 */ 1242#define NEWBLK_HASH(fs, inum) \ 1243 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash]) 1244 1245static int 1246newblk_find(newblkhd, fs, newblkno, newblkpp) 1247 struct newblk_hashhead *newblkhd; 1248 struct fs *fs; 1249 ufs2_daddr_t newblkno; 1250 struct newblk **newblkpp; 1251{ 1252 struct newblk *newblk; 1253 1254 LIST_FOREACH(newblk, newblkhd, nb_hash) 1255 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs) 1256 break; 1257 if (newblk) { 1258 *newblkpp = newblk; 1259 return (1); 1260 } 1261 *newblkpp = NULL; 1262 return (0); 1263} 1264 1265/* 1266 * Look up a newblk. Return 1 if found, 0 if not found. 1267 * If not found, allocate if DEPALLOC flag is passed. 1268 * Found or allocated entry is returned in newblkpp. 1269 */ 1270static int 1271newblk_lookup(fs, newblkno, flags, newblkpp) 1272 struct fs *fs; 1273 ufs2_daddr_t newblkno; 1274 int flags; 1275 struct newblk **newblkpp; 1276{ 1277 struct newblk *newblk; 1278 struct newblk_hashhead *newblkhd; 1279 1280 newblkhd = NEWBLK_HASH(fs, newblkno); 1281 if (newblk_find(newblkhd, fs, newblkno, newblkpp)) 1282 return (1); 1283 if ((flags & DEPALLOC) == 0) 1284 return (0); 1285 FREE_LOCK(&lk); 1286 MALLOC(newblk, struct newblk *, sizeof(struct newblk), 1287 M_NEWBLK, M_SOFTDEP_FLAGS); 1288 ACQUIRE_LOCK(&lk); 1289 if (newblk_find(newblkhd, fs, newblkno, newblkpp)) { 1290 FREE(newblk, M_NEWBLK); 1291 return (1); 1292 } 1293 newblk->nb_state = 0; 1294 newblk->nb_fs = fs; 1295 newblk->nb_newblkno = newblkno; 1296 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 1297 *newblkpp = newblk; 1298 return (0); 1299} 1300 1301/* 1302 * Executed during filesystem system initialization before 1303 * mounting any filesystems. 1304 */ 1305void 1306softdep_initialize() 1307{ 1308 1309 LIST_INIT(&mkdirlisthd); 1310 max_softdeps = desiredvnodes * 4; 1311 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 1312 &pagedep_hash); 1313 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash); 1314 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash); 1315 1316 /* initialise bioops hack */ 1317 bioops.io_start = softdep_disk_io_initiation; 1318 bioops.io_complete = softdep_disk_write_complete; 1319 bioops.io_deallocate = softdep_deallocate_dependencies; 1320 bioops.io_countdeps = softdep_count_dependencies; 1321} 1322 1323/* 1324 * Executed after all filesystems have been unmounted during 1325 * filesystem module unload. 1326 */ 1327void 1328softdep_uninitialize() 1329{ 1330 1331 hashdestroy(pagedep_hashtbl, M_PAGEDEP, pagedep_hash); 1332 hashdestroy(inodedep_hashtbl, M_INODEDEP, inodedep_hash); 1333 hashdestroy(newblk_hashtbl, M_NEWBLK, newblk_hash); 1334} 1335 1336/* 1337 * Called at mount time to notify the dependency code that a 1338 * filesystem wishes to use it. 1339 */ 1340int 1341softdep_mount(devvp, mp, fs, cred) 1342 struct vnode *devvp; 1343 struct mount *mp; 1344 struct fs *fs; 1345 struct ucred *cred; 1346{ 1347 struct csum_total cstotal; 1348 struct ufsmount *ump; 1349 struct cg *cgp; 1350 struct buf *bp; 1351 int error, cyl; 1352 1353 mp->mnt_flag &= ~MNT_ASYNC; 1354 mp->mnt_flag |= MNT_SOFTDEP; 1355 ump = VFSTOUFS(mp); 1356 LIST_INIT(&ump->softdep_workitem_pending); 1357 ump->softdep_worklist_tail = NULL; 1358 ump->softdep_on_worklist = 0; 1359 ump->softdep_deps = 0; 1360 /* 1361 * When doing soft updates, the counters in the 1362 * superblock may have gotten out of sync. Recomputation 1363 * can take a long time and can be deferred for background 1364 * fsck. However, the old behavior of scanning the cylinder 1365 * groups and recalculating them at mount time is available 1366 * by setting vfs.ffs.compute_summary_at_mount to one. 1367 */ 1368 if (compute_summary_at_mount == 0 || fs->fs_clean != 0) 1369 return (0); 1370 bzero(&cstotal, sizeof cstotal); 1371 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 1372 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 1373 fs->fs_cgsize, cred, &bp)) != 0) { 1374 brelse(bp); 1375 return (error); 1376 } 1377 cgp = (struct cg *)bp->b_data; 1378 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 1379 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 1380 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 1381 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 1382 fs->fs_cs(fs, cyl) = cgp->cg_cs; 1383 brelse(bp); 1384 } 1385#ifdef DEBUG 1386 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 1387 printf("%s: superblock summary recomputed\n", fs->fs_fsmnt); 1388#endif 1389 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 1390 return (0); 1391} 1392 1393/* 1394 * Protecting the freemaps (or bitmaps). 1395 * 1396 * To eliminate the need to execute fsck before mounting a filesystem 1397 * after a power failure, one must (conservatively) guarantee that the 1398 * on-disk copy of the bitmaps never indicate that a live inode or block is 1399 * free. So, when a block or inode is allocated, the bitmap should be 1400 * updated (on disk) before any new pointers. When a block or inode is 1401 * freed, the bitmap should not be updated until all pointers have been 1402 * reset. The latter dependency is handled by the delayed de-allocation 1403 * approach described below for block and inode de-allocation. The former 1404 * dependency is handled by calling the following procedure when a block or 1405 * inode is allocated. When an inode is allocated an "inodedep" is created 1406 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 1407 * Each "inodedep" is also inserted into the hash indexing structure so 1408 * that any additional link additions can be made dependent on the inode 1409 * allocation. 1410 * 1411 * The ufs filesystem maintains a number of free block counts (e.g., per 1412 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 1413 * in addition to the bitmaps. These counts are used to improve efficiency 1414 * during allocation and therefore must be consistent with the bitmaps. 1415 * There is no convenient way to guarantee post-crash consistency of these 1416 * counts with simple update ordering, for two main reasons: (1) The counts 1417 * and bitmaps for a single cylinder group block are not in the same disk 1418 * sector. If a disk write is interrupted (e.g., by power failure), one may 1419 * be written and the other not. (2) Some of the counts are located in the 1420 * superblock rather than the cylinder group block. So, we focus our soft 1421 * updates implementation on protecting the bitmaps. When mounting a 1422 * filesystem, we recompute the auxiliary counts from the bitmaps. 1423 */ 1424 1425/* 1426 * Called just after updating the cylinder group block to allocate an inode. 1427 */ 1428void 1429softdep_setup_inomapdep(bp, ip, newinum) 1430 struct buf *bp; /* buffer for cylgroup block with inode map */ 1431 struct inode *ip; /* inode related to allocation */ 1432 ino_t newinum; /* new inode number being allocated */ 1433{ 1434 struct inodedep *inodedep; 1435 struct bmsafemap *bmsafemap; 1436 1437 /* 1438 * Create a dependency for the newly allocated inode. 1439 * Panic if it already exists as something is seriously wrong. 1440 * Otherwise add it to the dependency list for the buffer holding 1441 * the cylinder group map from which it was allocated. 1442 */ 1443 ACQUIRE_LOCK(&lk); 1444 if ((inodedep_lookup(UFSTOVFS(ip->i_ump), newinum, DEPALLOC|NODELAY, 1445 &inodedep))) 1446 panic("softdep_setup_inomapdep: found inode"); 1447 inodedep->id_buf = bp; 1448 inodedep->id_state &= ~DEPCOMPLETE; 1449 bmsafemap = bmsafemap_lookup(inodedep->id_list.wk_mp, bp); 1450 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 1451 FREE_LOCK(&lk); 1452} 1453 1454/* 1455 * Called just after updating the cylinder group block to 1456 * allocate block or fragment. 1457 */ 1458void 1459softdep_setup_blkmapdep(bp, mp, newblkno) 1460 struct buf *bp; /* buffer for cylgroup block with block map */ 1461 struct mount *mp; /* filesystem doing allocation */ 1462 ufs2_daddr_t newblkno; /* number of newly allocated block */ 1463{ 1464 struct newblk *newblk; 1465 struct bmsafemap *bmsafemap; 1466 struct fs *fs; 1467 1468 fs = VFSTOUFS(mp)->um_fs; 1469 /* 1470 * Create a dependency for the newly allocated block. 1471 * Add it to the dependency list for the buffer holding 1472 * the cylinder group map from which it was allocated. 1473 */ 1474 ACQUIRE_LOCK(&lk); 1475 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0) 1476 panic("softdep_setup_blkmapdep: found block"); 1477 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp); 1478 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 1479 FREE_LOCK(&lk); 1480} 1481 1482/* 1483 * Find the bmsafemap associated with a cylinder group buffer. 1484 * If none exists, create one. The buffer must be locked when 1485 * this routine is called and this routine must be called with 1486 * splbio interrupts blocked. 1487 */ 1488static struct bmsafemap * 1489bmsafemap_lookup(mp, bp) 1490 struct mount *mp; 1491 struct buf *bp; 1492{ 1493 struct bmsafemap *bmsafemap; 1494 struct worklist *wk; 1495 1496 mtx_assert(&lk, MA_OWNED); 1497 LIST_FOREACH(wk, &bp->b_dep, wk_list) 1498 if (wk->wk_type == D_BMSAFEMAP) 1499 return (WK_BMSAFEMAP(wk)); 1500 FREE_LOCK(&lk); 1501 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap), 1502 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 1503 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); 1504 bmsafemap->sm_buf = bp; 1505 LIST_INIT(&bmsafemap->sm_allocdirecthd); 1506 LIST_INIT(&bmsafemap->sm_allocindirhd); 1507 LIST_INIT(&bmsafemap->sm_inodedephd); 1508 LIST_INIT(&bmsafemap->sm_newblkhd); 1509 ACQUIRE_LOCK(&lk); 1510 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 1511 return (bmsafemap); 1512} 1513 1514/* 1515 * Direct block allocation dependencies. 1516 * 1517 * When a new block is allocated, the corresponding disk locations must be 1518 * initialized (with zeros or new data) before the on-disk inode points to 1519 * them. Also, the freemap from which the block was allocated must be 1520 * updated (on disk) before the inode's pointer. These two dependencies are 1521 * independent of each other and are needed for all file blocks and indirect 1522 * blocks that are pointed to directly by the inode. Just before the 1523 * "in-core" version of the inode is updated with a newly allocated block 1524 * number, a procedure (below) is called to setup allocation dependency 1525 * structures. These structures are removed when the corresponding 1526 * dependencies are satisfied or when the block allocation becomes obsolete 1527 * (i.e., the file is deleted, the block is de-allocated, or the block is a 1528 * fragment that gets upgraded). All of these cases are handled in 1529 * procedures described later. 1530 * 1531 * When a file extension causes a fragment to be upgraded, either to a larger 1532 * fragment or to a full block, the on-disk location may change (if the 1533 * previous fragment could not simply be extended). In this case, the old 1534 * fragment must be de-allocated, but not until after the inode's pointer has 1535 * been updated. In most cases, this is handled by later procedures, which 1536 * will construct a "freefrag" structure to be added to the workitem queue 1537 * when the inode update is complete (or obsolete). The main exception to 1538 * this is when an allocation occurs while a pending allocation dependency 1539 * (for the same block pointer) remains. This case is handled in the main 1540 * allocation dependency setup procedure by immediately freeing the 1541 * unreferenced fragments. 1542 */ 1543void 1544softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 1545 struct inode *ip; /* inode to which block is being added */ 1546 ufs_lbn_t lbn; /* block pointer within inode */ 1547 ufs2_daddr_t newblkno; /* disk block number being added */ 1548 ufs2_daddr_t oldblkno; /* previous block number, 0 unless frag */ 1549 long newsize; /* size of new block */ 1550 long oldsize; /* size of new block */ 1551 struct buf *bp; /* bp for allocated block */ 1552{ 1553 struct allocdirect *adp, *oldadp; 1554 struct allocdirectlst *adphead; 1555 struct bmsafemap *bmsafemap; 1556 struct inodedep *inodedep; 1557 struct pagedep *pagedep; 1558 struct newblk *newblk; 1559 struct mount *mp; 1560 1561 mp = UFSTOVFS(ip->i_ump); 1562 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect), 1563 M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO); 1564 workitem_alloc(&adp->ad_list, D_ALLOCDIRECT, mp); 1565 adp->ad_lbn = lbn; 1566 adp->ad_newblkno = newblkno; 1567 adp->ad_oldblkno = oldblkno; 1568 adp->ad_newsize = newsize; 1569 adp->ad_oldsize = oldsize; 1570 adp->ad_state = ATTACHED; 1571 LIST_INIT(&adp->ad_newdirblk); 1572 if (newblkno == oldblkno) 1573 adp->ad_freefrag = NULL; 1574 else 1575 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1576 1577 ACQUIRE_LOCK(&lk); 1578 if (lbn >= NDADDR) { 1579 /* allocating an indirect block */ 1580 if (oldblkno != 0) 1581 panic("softdep_setup_allocdirect: non-zero indir"); 1582 } else { 1583 /* 1584 * Allocating a direct block. 1585 * 1586 * If we are allocating a directory block, then we must 1587 * allocate an associated pagedep to track additions and 1588 * deletions. 1589 */ 1590 if ((ip->i_mode & IFMT) == IFDIR && 1591 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1592 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 1593 } 1594 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1595 panic("softdep_setup_allocdirect: lost block"); 1596 if (newblk->nb_state == DEPCOMPLETE) { 1597 adp->ad_state |= DEPCOMPLETE; 1598 adp->ad_buf = NULL; 1599 } else { 1600 bmsafemap = newblk->nb_bmsafemap; 1601 adp->ad_buf = bmsafemap->sm_buf; 1602 LIST_REMOVE(newblk, nb_deps); 1603 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1604 } 1605 LIST_REMOVE(newblk, nb_hash); 1606 FREE(newblk, M_NEWBLK); 1607 1608 inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep); 1609 adp->ad_inodedep = inodedep; 1610 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list); 1611 /* 1612 * The list of allocdirects must be kept in sorted and ascending 1613 * order so that the rollback routines can quickly determine the 1614 * first uncommitted block (the size of the file stored on disk 1615 * ends at the end of the lowest committed fragment, or if there 1616 * are no fragments, at the end of the highest committed block). 1617 * Since files generally grow, the typical case is that the new 1618 * block is to be added at the end of the list. We speed this 1619 * special case by checking against the last allocdirect in the 1620 * list before laboriously traversing the list looking for the 1621 * insertion point. 1622 */ 1623 adphead = &inodedep->id_newinoupdt; 1624 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1625 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1626 /* insert at end of list */ 1627 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1628 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1629 allocdirect_merge(adphead, adp, oldadp); 1630 FREE_LOCK(&lk); 1631 return; 1632 } 1633 TAILQ_FOREACH(oldadp, adphead, ad_next) { 1634 if (oldadp->ad_lbn >= lbn) 1635 break; 1636 } 1637 if (oldadp == NULL) 1638 panic("softdep_setup_allocdirect: lost entry"); 1639 /* insert in middle of list */ 1640 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1641 if (oldadp->ad_lbn == lbn) 1642 allocdirect_merge(adphead, adp, oldadp); 1643 FREE_LOCK(&lk); 1644} 1645 1646/* 1647 * Replace an old allocdirect dependency with a newer one. 1648 * This routine must be called with splbio interrupts blocked. 1649 */ 1650static void 1651allocdirect_merge(adphead, newadp, oldadp) 1652 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 1653 struct allocdirect *newadp; /* allocdirect being added */ 1654 struct allocdirect *oldadp; /* existing allocdirect being checked */ 1655{ 1656 struct worklist *wk; 1657 struct freefrag *freefrag; 1658 struct newdirblk *newdirblk; 1659 1660 mtx_assert(&lk, MA_OWNED); 1661 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 1662 newadp->ad_oldsize != oldadp->ad_newsize || 1663 newadp->ad_lbn >= NDADDR) 1664 panic("%s %jd != new %jd || old size %ld != new %ld", 1665 "allocdirect_merge: old blkno", 1666 (intmax_t)newadp->ad_oldblkno, 1667 (intmax_t)oldadp->ad_newblkno, 1668 newadp->ad_oldsize, oldadp->ad_newsize); 1669 newadp->ad_oldblkno = oldadp->ad_oldblkno; 1670 newadp->ad_oldsize = oldadp->ad_oldsize; 1671 /* 1672 * If the old dependency had a fragment to free or had never 1673 * previously had a block allocated, then the new dependency 1674 * can immediately post its freefrag and adopt the old freefrag. 1675 * This action is done by swapping the freefrag dependencies. 1676 * The new dependency gains the old one's freefrag, and the 1677 * old one gets the new one and then immediately puts it on 1678 * the worklist when it is freed by free_allocdirect. It is 1679 * not possible to do this swap when the old dependency had a 1680 * non-zero size but no previous fragment to free. This condition 1681 * arises when the new block is an extension of the old block. 1682 * Here, the first part of the fragment allocated to the new 1683 * dependency is part of the block currently claimed on disk by 1684 * the old dependency, so cannot legitimately be freed until the 1685 * conditions for the new dependency are fulfilled. 1686 */ 1687 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 1688 freefrag = newadp->ad_freefrag; 1689 newadp->ad_freefrag = oldadp->ad_freefrag; 1690 oldadp->ad_freefrag = freefrag; 1691 } 1692 /* 1693 * If we are tracking a new directory-block allocation, 1694 * move it from the old allocdirect to the new allocdirect. 1695 */ 1696 if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) { 1697 newdirblk = WK_NEWDIRBLK(wk); 1698 WORKLIST_REMOVE(&newdirblk->db_list); 1699 if (LIST_FIRST(&oldadp->ad_newdirblk) != NULL) 1700 panic("allocdirect_merge: extra newdirblk"); 1701 WORKLIST_INSERT(&newadp->ad_newdirblk, &newdirblk->db_list); 1702 } 1703 free_allocdirect(adphead, oldadp, 0); 1704} 1705 1706/* 1707 * Allocate a new freefrag structure if needed. 1708 */ 1709static struct freefrag * 1710newfreefrag(ip, blkno, size) 1711 struct inode *ip; 1712 ufs2_daddr_t blkno; 1713 long size; 1714{ 1715 struct freefrag *freefrag; 1716 struct fs *fs; 1717 1718 if (blkno == 0) 1719 return (NULL); 1720 fs = ip->i_fs; 1721 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 1722 panic("newfreefrag: frag size"); 1723 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag), 1724 M_FREEFRAG, M_SOFTDEP_FLAGS); 1725 workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ip->i_ump)); 1726 freefrag->ff_inum = ip->i_number; 1727 freefrag->ff_blkno = blkno; 1728 freefrag->ff_fragsize = size; 1729 return (freefrag); 1730} 1731 1732/* 1733 * This workitem de-allocates fragments that were replaced during 1734 * file block allocation. 1735 */ 1736static void 1737handle_workitem_freefrag(freefrag) 1738 struct freefrag *freefrag; 1739{ 1740 struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp); 1741 1742 ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno, 1743 freefrag->ff_fragsize, freefrag->ff_inum); 1744 ACQUIRE_LOCK(&lk); 1745 WORKITEM_FREE(freefrag, D_FREEFRAG); 1746 FREE_LOCK(&lk); 1747} 1748 1749/* 1750 * Set up a dependency structure for an external attributes data block. 1751 * This routine follows much of the structure of softdep_setup_allocdirect. 1752 * See the description of softdep_setup_allocdirect above for details. 1753 */ 1754void 1755softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 1756 struct inode *ip; 1757 ufs_lbn_t lbn; 1758 ufs2_daddr_t newblkno; 1759 ufs2_daddr_t oldblkno; 1760 long newsize; 1761 long oldsize; 1762 struct buf *bp; 1763{ 1764 struct allocdirect *adp, *oldadp; 1765 struct allocdirectlst *adphead; 1766 struct bmsafemap *bmsafemap; 1767 struct inodedep *inodedep; 1768 struct newblk *newblk; 1769 struct mount *mp; 1770 1771 mp = UFSTOVFS(ip->i_ump); 1772 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect), 1773 M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO); 1774 workitem_alloc(&adp->ad_list, D_ALLOCDIRECT, mp); 1775 adp->ad_lbn = lbn; 1776 adp->ad_newblkno = newblkno; 1777 adp->ad_oldblkno = oldblkno; 1778 adp->ad_newsize = newsize; 1779 adp->ad_oldsize = oldsize; 1780 adp->ad_state = ATTACHED | EXTDATA; 1781 LIST_INIT(&adp->ad_newdirblk); 1782 if (newblkno == oldblkno) 1783 adp->ad_freefrag = NULL; 1784 else 1785 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1786 1787 ACQUIRE_LOCK(&lk); 1788 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1789 panic("softdep_setup_allocext: lost block"); 1790 1791 inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep); 1792 adp->ad_inodedep = inodedep; 1793 1794 if (newblk->nb_state == DEPCOMPLETE) { 1795 adp->ad_state |= DEPCOMPLETE; 1796 adp->ad_buf = NULL; 1797 } else { 1798 bmsafemap = newblk->nb_bmsafemap; 1799 adp->ad_buf = bmsafemap->sm_buf; 1800 LIST_REMOVE(newblk, nb_deps); 1801 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1802 } 1803 LIST_REMOVE(newblk, nb_hash); 1804 FREE(newblk, M_NEWBLK); 1805 1806 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list); 1807 if (lbn >= NXADDR) 1808 panic("softdep_setup_allocext: lbn %lld > NXADDR", 1809 (long long)lbn); 1810 /* 1811 * The list of allocdirects must be kept in sorted and ascending 1812 * order so that the rollback routines can quickly determine the 1813 * first uncommitted block (the size of the file stored on disk 1814 * ends at the end of the lowest committed fragment, or if there 1815 * are no fragments, at the end of the highest committed block). 1816 * Since files generally grow, the typical case is that the new 1817 * block is to be added at the end of the list. We speed this 1818 * special case by checking against the last allocdirect in the 1819 * list before laboriously traversing the list looking for the 1820 * insertion point. 1821 */ 1822 adphead = &inodedep->id_newextupdt; 1823 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1824 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1825 /* insert at end of list */ 1826 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1827 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1828 allocdirect_merge(adphead, adp, oldadp); 1829 FREE_LOCK(&lk); 1830 return; 1831 } 1832 TAILQ_FOREACH(oldadp, adphead, ad_next) { 1833 if (oldadp->ad_lbn >= lbn) 1834 break; 1835 } 1836 if (oldadp == NULL) 1837 panic("softdep_setup_allocext: lost entry"); 1838 /* insert in middle of list */ 1839 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1840 if (oldadp->ad_lbn == lbn) 1841 allocdirect_merge(adphead, adp, oldadp); 1842 FREE_LOCK(&lk); 1843} 1844 1845/* 1846 * Indirect block allocation dependencies. 1847 * 1848 * The same dependencies that exist for a direct block also exist when 1849 * a new block is allocated and pointed to by an entry in a block of 1850 * indirect pointers. The undo/redo states described above are also 1851 * used here. Because an indirect block contains many pointers that 1852 * may have dependencies, a second copy of the entire in-memory indirect 1853 * block is kept. The buffer cache copy is always completely up-to-date. 1854 * The second copy, which is used only as a source for disk writes, 1855 * contains only the safe pointers (i.e., those that have no remaining 1856 * update dependencies). The second copy is freed when all pointers 1857 * are safe. The cache is not allowed to replace indirect blocks with 1858 * pending update dependencies. If a buffer containing an indirect 1859 * block with dependencies is written, these routines will mark it 1860 * dirty again. It can only be successfully written once all the 1861 * dependencies are removed. The ffs_fsync routine in conjunction with 1862 * softdep_sync_metadata work together to get all the dependencies 1863 * removed so that a file can be successfully written to disk. Three 1864 * procedures are used when setting up indirect block pointer 1865 * dependencies. The division is necessary because of the organization 1866 * of the "balloc" routine and because of the distinction between file 1867 * pages and file metadata blocks. 1868 */ 1869 1870/* 1871 * Allocate a new allocindir structure. 1872 */ 1873static struct allocindir * 1874newallocindir(ip, ptrno, newblkno, oldblkno) 1875 struct inode *ip; /* inode for file being extended */ 1876 int ptrno; /* offset of pointer in indirect block */ 1877 ufs2_daddr_t newblkno; /* disk block number being added */ 1878 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 1879{ 1880 struct allocindir *aip; 1881 1882 MALLOC(aip, struct allocindir *, sizeof(struct allocindir), 1883 M_ALLOCINDIR, M_SOFTDEP_FLAGS|M_ZERO); 1884 workitem_alloc(&aip->ai_list, D_ALLOCINDIR, UFSTOVFS(ip->i_ump)); 1885 aip->ai_state = ATTACHED; 1886 aip->ai_offset = ptrno; 1887 aip->ai_newblkno = newblkno; 1888 aip->ai_oldblkno = oldblkno; 1889 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize); 1890 return (aip); 1891} 1892 1893/* 1894 * Called just before setting an indirect block pointer 1895 * to a newly allocated file page. 1896 */ 1897void 1898softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 1899 struct inode *ip; /* inode for file being extended */ 1900 ufs_lbn_t lbn; /* allocated block number within file */ 1901 struct buf *bp; /* buffer with indirect blk referencing page */ 1902 int ptrno; /* offset of pointer in indirect block */ 1903 ufs2_daddr_t newblkno; /* disk block number being added */ 1904 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 1905 struct buf *nbp; /* buffer holding allocated page */ 1906{ 1907 struct allocindir *aip; 1908 struct pagedep *pagedep; 1909 1910 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page"); 1911 aip = newallocindir(ip, ptrno, newblkno, oldblkno); 1912 ACQUIRE_LOCK(&lk); 1913 /* 1914 * If we are allocating a directory page, then we must 1915 * allocate an associated pagedep to track additions and 1916 * deletions. 1917 */ 1918 if ((ip->i_mode & IFMT) == IFDIR && 1919 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1920 WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list); 1921 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1922 setup_allocindir_phase2(bp, ip, aip); 1923 FREE_LOCK(&lk); 1924} 1925 1926/* 1927 * Called just before setting an indirect block pointer to a 1928 * newly allocated indirect block. 1929 */ 1930void 1931softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 1932 struct buf *nbp; /* newly allocated indirect block */ 1933 struct inode *ip; /* inode for file being extended */ 1934 struct buf *bp; /* indirect block referencing allocated block */ 1935 int ptrno; /* offset of pointer in indirect block */ 1936 ufs2_daddr_t newblkno; /* disk block number being added */ 1937{ 1938 struct allocindir *aip; 1939 1940 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta"); 1941 aip = newallocindir(ip, ptrno, newblkno, 0); 1942 ACQUIRE_LOCK(&lk); 1943 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 1944 setup_allocindir_phase2(bp, ip, aip); 1945 FREE_LOCK(&lk); 1946} 1947 1948/* 1949 * Called to finish the allocation of the "aip" allocated 1950 * by one of the two routines above. 1951 */ 1952static void 1953setup_allocindir_phase2(bp, ip, aip) 1954 struct buf *bp; /* in-memory copy of the indirect block */ 1955 struct inode *ip; /* inode for file being extended */ 1956 struct allocindir *aip; /* allocindir allocated by the above routines */ 1957{ 1958 struct worklist *wk; 1959 struct indirdep *indirdep, *newindirdep; 1960 struct bmsafemap *bmsafemap; 1961 struct allocindir *oldaip; 1962 struct freefrag *freefrag; 1963 struct newblk *newblk; 1964 ufs2_daddr_t blkno; 1965 1966 mtx_assert(&lk, MA_OWNED); 1967 if (bp->b_lblkno >= 0) 1968 panic("setup_allocindir_phase2: not indir blk"); 1969 for (indirdep = NULL, newindirdep = NULL; ; ) { 1970 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 1971 if (wk->wk_type != D_INDIRDEP) 1972 continue; 1973 indirdep = WK_INDIRDEP(wk); 1974 break; 1975 } 1976 if (indirdep == NULL && newindirdep) { 1977 indirdep = newindirdep; 1978 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 1979 newindirdep = NULL; 1980 } 1981 if (indirdep) { 1982 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0, 1983 &newblk) == 0) 1984 panic("setup_allocindir: lost block"); 1985 if (newblk->nb_state == DEPCOMPLETE) { 1986 aip->ai_state |= DEPCOMPLETE; 1987 aip->ai_buf = NULL; 1988 } else { 1989 bmsafemap = newblk->nb_bmsafemap; 1990 aip->ai_buf = bmsafemap->sm_buf; 1991 LIST_REMOVE(newblk, nb_deps); 1992 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd, 1993 aip, ai_deps); 1994 } 1995 LIST_REMOVE(newblk, nb_hash); 1996 FREE(newblk, M_NEWBLK); 1997 aip->ai_indirdep = indirdep; 1998 /* 1999 * Check to see if there is an existing dependency 2000 * for this block. If there is, merge the old 2001 * dependency into the new one. 2002 */ 2003 if (aip->ai_oldblkno == 0) 2004 oldaip = NULL; 2005 else 2006 2007 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) 2008 if (oldaip->ai_offset == aip->ai_offset) 2009 break; 2010 freefrag = NULL; 2011 if (oldaip != NULL) { 2012 if (oldaip->ai_newblkno != aip->ai_oldblkno) 2013 panic("setup_allocindir_phase2: blkno"); 2014 aip->ai_oldblkno = oldaip->ai_oldblkno; 2015 freefrag = aip->ai_freefrag; 2016 aip->ai_freefrag = oldaip->ai_freefrag; 2017 oldaip->ai_freefrag = NULL; 2018 free_allocindir(oldaip, NULL); 2019 } 2020 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 2021 if (ip->i_ump->um_fstype == UFS1) 2022 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data) 2023 [aip->ai_offset] = aip->ai_oldblkno; 2024 else 2025 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data) 2026 [aip->ai_offset] = aip->ai_oldblkno; 2027 FREE_LOCK(&lk); 2028 if (freefrag != NULL) 2029 handle_workitem_freefrag(freefrag); 2030 } else 2031 FREE_LOCK(&lk); 2032 if (newindirdep) { 2033 newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 2034 brelse(newindirdep->ir_savebp); 2035 ACQUIRE_LOCK(&lk); 2036 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP); 2037 if (indirdep) 2038 break; 2039 FREE_LOCK(&lk); 2040 } 2041 if (indirdep) { 2042 ACQUIRE_LOCK(&lk); 2043 break; 2044 } 2045 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep), 2046 M_INDIRDEP, M_SOFTDEP_FLAGS); 2047 workitem_alloc(&newindirdep->ir_list, D_INDIRDEP, 2048 UFSTOVFS(ip->i_ump)); 2049 newindirdep->ir_state = ATTACHED; 2050 if (ip->i_ump->um_fstype == UFS1) 2051 newindirdep->ir_state |= UFS1FMT; 2052 LIST_INIT(&newindirdep->ir_deplisthd); 2053 LIST_INIT(&newindirdep->ir_donehd); 2054 if (bp->b_blkno == bp->b_lblkno) { 2055 ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp, 2056 NULL, NULL); 2057 bp->b_blkno = blkno; 2058 } 2059 newindirdep->ir_savebp = 2060 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0); 2061 BUF_KERNPROC(newindirdep->ir_savebp); 2062 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 2063 ACQUIRE_LOCK(&lk); 2064 } 2065} 2066 2067/* 2068 * Block de-allocation dependencies. 2069 * 2070 * When blocks are de-allocated, the on-disk pointers must be nullified before 2071 * the blocks are made available for use by other files. (The true 2072 * requirement is that old pointers must be nullified before new on-disk 2073 * pointers are set. We chose this slightly more stringent requirement to 2074 * reduce complexity.) Our implementation handles this dependency by updating 2075 * the inode (or indirect block) appropriately but delaying the actual block 2076 * de-allocation (i.e., freemap and free space count manipulation) until 2077 * after the updated versions reach stable storage. After the disk is 2078 * updated, the blocks can be safely de-allocated whenever it is convenient. 2079 * This implementation handles only the common case of reducing a file's 2080 * length to zero. Other cases are handled by the conventional synchronous 2081 * write approach. 2082 * 2083 * The ffs implementation with which we worked double-checks 2084 * the state of the block pointers and file size as it reduces 2085 * a file's length. Some of this code is replicated here in our 2086 * soft updates implementation. The freeblks->fb_chkcnt field is 2087 * used to transfer a part of this information to the procedure 2088 * that eventually de-allocates the blocks. 2089 * 2090 * This routine should be called from the routine that shortens 2091 * a file's length, before the inode's size or block pointers 2092 * are modified. It will save the block pointer information for 2093 * later release and zero the inode so that the calling routine 2094 * can release it. 2095 */ 2096void 2097softdep_setup_freeblocks(ip, length, flags) 2098 struct inode *ip; /* The inode whose length is to be reduced */ 2099 off_t length; /* The new length for the file */ 2100 int flags; /* IO_EXT and/or IO_NORMAL */ 2101{ 2102 struct freeblks *freeblks; 2103 struct inodedep *inodedep; 2104 struct allocdirect *adp; 2105 struct vnode *vp; 2106 struct buf *bp; 2107 struct fs *fs; 2108 ufs2_daddr_t extblocks, datablocks; 2109 struct mount *mp; 2110 int i, delay, error; 2111 2112 fs = ip->i_fs; 2113 mp = UFSTOVFS(ip->i_ump); 2114 if (length != 0) 2115 panic("softdep_setup_freeblocks: non-zero length"); 2116 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks), 2117 M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO); 2118 workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp); 2119 freeblks->fb_state = ATTACHED; 2120 freeblks->fb_uid = ip->i_uid; 2121 freeblks->fb_previousinum = ip->i_number; 2122 freeblks->fb_devvp = ip->i_devvp; 2123 extblocks = 0; 2124 if (fs->fs_magic == FS_UFS2_MAGIC) 2125 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 2126 datablocks = DIP(ip, i_blocks) - extblocks; 2127 if ((flags & IO_NORMAL) == 0) { 2128 freeblks->fb_oldsize = 0; 2129 freeblks->fb_chkcnt = 0; 2130 } else { 2131 freeblks->fb_oldsize = ip->i_size; 2132 ip->i_size = 0; 2133 DIP_SET(ip, i_size, 0); 2134 freeblks->fb_chkcnt = datablocks; 2135 for (i = 0; i < NDADDR; i++) { 2136 freeblks->fb_dblks[i] = DIP(ip, i_db[i]); 2137 DIP_SET(ip, i_db[i], 0); 2138 } 2139 for (i = 0; i < NIADDR; i++) { 2140 freeblks->fb_iblks[i] = DIP(ip, i_ib[i]); 2141 DIP_SET(ip, i_ib[i], 0); 2142 } 2143 /* 2144 * If the file was removed, then the space being freed was 2145 * accounted for then (see softdep_filereleased()). If the 2146 * file is merely being truncated, then we account for it now. 2147 */ 2148 if ((ip->i_flag & IN_SPACECOUNTED) == 0) { 2149 UFS_LOCK(ip->i_ump); 2150 fs->fs_pendingblocks += datablocks; 2151 UFS_UNLOCK(ip->i_ump); 2152 } 2153 } 2154 if ((flags & IO_EXT) == 0) { 2155 freeblks->fb_oldextsize = 0; 2156 } else { 2157 freeblks->fb_oldextsize = ip->i_din2->di_extsize; 2158 ip->i_din2->di_extsize = 0; 2159 freeblks->fb_chkcnt += extblocks; 2160 for (i = 0; i < NXADDR; i++) { 2161 freeblks->fb_eblks[i] = ip->i_din2->di_extb[i]; 2162 ip->i_din2->di_extb[i] = 0; 2163 } 2164 } 2165 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - freeblks->fb_chkcnt); 2166 /* 2167 * Push the zero'ed inode to to its disk buffer so that we are free 2168 * to delete its dependencies below. Once the dependencies are gone 2169 * the buffer can be safely released. 2170 */ 2171 if ((error = bread(ip->i_devvp, 2172 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 2173 (int)fs->fs_bsize, NOCRED, &bp)) != 0) { 2174 brelse(bp); 2175 softdep_error("softdep_setup_freeblocks", error); 2176 } 2177 if (ip->i_ump->um_fstype == UFS1) 2178 *((struct ufs1_dinode *)bp->b_data + 2179 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1; 2180 else 2181 *((struct ufs2_dinode *)bp->b_data + 2182 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2; 2183 /* 2184 * Find and eliminate any inode dependencies. 2185 */ 2186 ACQUIRE_LOCK(&lk); 2187 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 2188 if ((inodedep->id_state & IOSTARTED) != 0) 2189 panic("softdep_setup_freeblocks: inode busy"); 2190 /* 2191 * Add the freeblks structure to the list of operations that 2192 * must await the zero'ed inode being written to disk. If we 2193 * still have a bitmap dependency (delay == 0), then the inode 2194 * has never been written to disk, so we can process the 2195 * freeblks below once we have deleted the dependencies. 2196 */ 2197 delay = (inodedep->id_state & DEPCOMPLETE); 2198 if (delay) 2199 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list); 2200 /* 2201 * Because the file length has been truncated to zero, any 2202 * pending block allocation dependency structures associated 2203 * with this inode are obsolete and can simply be de-allocated. 2204 * We must first merge the two dependency lists to get rid of 2205 * any duplicate freefrag structures, then purge the merged list. 2206 * If we still have a bitmap dependency, then the inode has never 2207 * been written to disk, so we can free any fragments without delay. 2208 */ 2209 if (flags & IO_NORMAL) { 2210 merge_inode_lists(&inodedep->id_newinoupdt, 2211 &inodedep->id_inoupdt); 2212 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 2213 free_allocdirect(&inodedep->id_inoupdt, adp, delay); 2214 } 2215 if (flags & IO_EXT) { 2216 merge_inode_lists(&inodedep->id_newextupdt, 2217 &inodedep->id_extupdt); 2218 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0) 2219 free_allocdirect(&inodedep->id_extupdt, adp, delay); 2220 } 2221 FREE_LOCK(&lk); 2222 bdwrite(bp); 2223 /* 2224 * We must wait for any I/O in progress to finish so that 2225 * all potential buffers on the dirty list will be visible. 2226 * Once they are all there, walk the list and get rid of 2227 * any dependencies. 2228 */ 2229 vp = ITOV(ip); 2230 VI_LOCK(vp); 2231 drain_output(vp); 2232restart: 2233 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 2234 if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) || 2235 ((flags & IO_NORMAL) == 0 && 2236 (bp->b_xflags & BX_ALTDATA) == 0)) 2237 continue; 2238 if ((bp = getdirtybuf(bp, VI_MTX(vp), MNT_WAIT)) == NULL) 2239 goto restart; 2240 VI_UNLOCK(vp); 2241 ACQUIRE_LOCK(&lk); 2242 (void) inodedep_lookup(mp, ip->i_number, 0, &inodedep); 2243 deallocate_dependencies(bp, inodedep); 2244 FREE_LOCK(&lk); 2245 bp->b_flags |= B_INVAL | B_NOCACHE; 2246 brelse(bp); 2247 VI_LOCK(vp); 2248 goto restart; 2249 } 2250 VI_UNLOCK(vp); 2251 ACQUIRE_LOCK(&lk); 2252 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 2253 (void) free_inodedep(inodedep); 2254 2255 if(delay) { 2256 freeblks->fb_state |= DEPCOMPLETE; 2257 /* 2258 * If the inode with zeroed block pointers is now on disk 2259 * we can start freeing blocks. Add freeblks to the worklist 2260 * instead of calling handle_workitem_freeblocks directly as 2261 * it is more likely that additional IO is needed to complete 2262 * the request here than in the !delay case. 2263 */ 2264 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 2265 add_to_worklist(&freeblks->fb_list); 2266 } 2267 2268 FREE_LOCK(&lk); 2269 /* 2270 * If the inode has never been written to disk (delay == 0), 2271 * then we can process the freeblks now that we have deleted 2272 * the dependencies. 2273 */ 2274 if (!delay) 2275 handle_workitem_freeblocks(freeblks, 0); 2276} 2277 2278/* 2279 * Reclaim any dependency structures from a buffer that is about to 2280 * be reallocated to a new vnode. The buffer must be locked, thus, 2281 * no I/O completion operations can occur while we are manipulating 2282 * its associated dependencies. The mutex is held so that other I/O's 2283 * associated with related dependencies do not occur. 2284 */ 2285static void 2286deallocate_dependencies(bp, inodedep) 2287 struct buf *bp; 2288 struct inodedep *inodedep; 2289{ 2290 struct worklist *wk; 2291 struct indirdep *indirdep; 2292 struct allocindir *aip; 2293 struct pagedep *pagedep; 2294 struct dirrem *dirrem; 2295 struct diradd *dap; 2296 int i; 2297 2298 mtx_assert(&lk, MA_OWNED); 2299 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2300 switch (wk->wk_type) { 2301 2302 case D_INDIRDEP: 2303 indirdep = WK_INDIRDEP(wk); 2304 /* 2305 * None of the indirect pointers will ever be visible, 2306 * so they can simply be tossed. GOINGAWAY ensures 2307 * that allocated pointers will be saved in the buffer 2308 * cache until they are freed. Note that they will 2309 * only be able to be found by their physical address 2310 * since the inode mapping the logical address will 2311 * be gone. The save buffer used for the safe copy 2312 * was allocated in setup_allocindir_phase2 using 2313 * the physical address so it could be used for this 2314 * purpose. Hence we swap the safe copy with the real 2315 * copy, allowing the safe copy to be freed and holding 2316 * on to the real copy for later use in indir_trunc. 2317 */ 2318 if (indirdep->ir_state & GOINGAWAY) 2319 panic("deallocate_dependencies: already gone"); 2320 indirdep->ir_state |= GOINGAWAY; 2321 VFSTOUFS(bp->b_vp->v_mount)->um_numindirdeps += 1; 2322 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 2323 free_allocindir(aip, inodedep); 2324 if (bp->b_lblkno >= 0 || 2325 bp->b_blkno != indirdep->ir_savebp->b_lblkno) 2326 panic("deallocate_dependencies: not indir"); 2327 bcopy(bp->b_data, indirdep->ir_savebp->b_data, 2328 bp->b_bcount); 2329 WORKLIST_REMOVE(wk); 2330 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk); 2331 continue; 2332 2333 case D_PAGEDEP: 2334 pagedep = WK_PAGEDEP(wk); 2335 /* 2336 * None of the directory additions will ever be 2337 * visible, so they can simply be tossed. 2338 */ 2339 for (i = 0; i < DAHASHSZ; i++) 2340 while ((dap = 2341 LIST_FIRST(&pagedep->pd_diraddhd[i]))) 2342 free_diradd(dap); 2343 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0) 2344 free_diradd(dap); 2345 /* 2346 * Copy any directory remove dependencies to the list 2347 * to be processed after the zero'ed inode is written. 2348 * If the inode has already been written, then they 2349 * can be dumped directly onto the work list. 2350 */ 2351 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 2352 LIST_REMOVE(dirrem, dm_next); 2353 dirrem->dm_dirinum = pagedep->pd_ino; 2354 if (inodedep == NULL || 2355 (inodedep->id_state & ALLCOMPLETE) == 2356 ALLCOMPLETE) 2357 add_to_worklist(&dirrem->dm_list); 2358 else 2359 WORKLIST_INSERT(&inodedep->id_bufwait, 2360 &dirrem->dm_list); 2361 } 2362 if ((pagedep->pd_state & NEWBLOCK) != 0) { 2363 LIST_FOREACH(wk, &inodedep->id_bufwait, wk_list) 2364 if (wk->wk_type == D_NEWDIRBLK && 2365 WK_NEWDIRBLK(wk)->db_pagedep == 2366 pagedep) 2367 break; 2368 if (wk != NULL) { 2369 WORKLIST_REMOVE(wk); 2370 free_newdirblk(WK_NEWDIRBLK(wk)); 2371 } else 2372 panic("deallocate_dependencies: " 2373 "lost pagedep"); 2374 } 2375 WORKLIST_REMOVE(&pagedep->pd_list); 2376 LIST_REMOVE(pagedep, pd_hash); 2377 WORKITEM_FREE(pagedep, D_PAGEDEP); 2378 continue; 2379 2380 case D_ALLOCINDIR: 2381 free_allocindir(WK_ALLOCINDIR(wk), inodedep); 2382 continue; 2383 2384 case D_ALLOCDIRECT: 2385 case D_INODEDEP: 2386 panic("deallocate_dependencies: Unexpected type %s", 2387 TYPENAME(wk->wk_type)); 2388 /* NOTREACHED */ 2389 2390 default: 2391 panic("deallocate_dependencies: Unknown type %s", 2392 TYPENAME(wk->wk_type)); 2393 /* NOTREACHED */ 2394 } 2395 } 2396} 2397 2398/* 2399 * Free an allocdirect. Generate a new freefrag work request if appropriate. 2400 * This routine must be called with splbio interrupts blocked. 2401 */ 2402static void 2403free_allocdirect(adphead, adp, delay) 2404 struct allocdirectlst *adphead; 2405 struct allocdirect *adp; 2406 int delay; 2407{ 2408 struct newdirblk *newdirblk; 2409 struct worklist *wk; 2410 2411 mtx_assert(&lk, MA_OWNED); 2412 if ((adp->ad_state & DEPCOMPLETE) == 0) 2413 LIST_REMOVE(adp, ad_deps); 2414 TAILQ_REMOVE(adphead, adp, ad_next); 2415 if ((adp->ad_state & COMPLETE) == 0) 2416 WORKLIST_REMOVE(&adp->ad_list); 2417 if (adp->ad_freefrag != NULL) { 2418 if (delay) 2419 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 2420 &adp->ad_freefrag->ff_list); 2421 else 2422 add_to_worklist(&adp->ad_freefrag->ff_list); 2423 } 2424 if ((wk = LIST_FIRST(&adp->ad_newdirblk)) != NULL) { 2425 newdirblk = WK_NEWDIRBLK(wk); 2426 WORKLIST_REMOVE(&newdirblk->db_list); 2427 if (LIST_FIRST(&adp->ad_newdirblk) != NULL) 2428 panic("free_allocdirect: extra newdirblk"); 2429 if (delay) 2430 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 2431 &newdirblk->db_list); 2432 else 2433 free_newdirblk(newdirblk); 2434 } 2435 WORKITEM_FREE(adp, D_ALLOCDIRECT); 2436} 2437 2438/* 2439 * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep. 2440 * This routine must be called with splbio interrupts blocked. 2441 */ 2442static void 2443free_newdirblk(newdirblk) 2444 struct newdirblk *newdirblk; 2445{ 2446 struct pagedep *pagedep; 2447 struct diradd *dap; 2448 int i; 2449 2450 mtx_assert(&lk, MA_OWNED); 2451 /* 2452 * If the pagedep is still linked onto the directory buffer 2453 * dependency chain, then some of the entries on the 2454 * pd_pendinghd list may not be committed to disk yet. In 2455 * this case, we will simply clear the NEWBLOCK flag and 2456 * let the pd_pendinghd list be processed when the pagedep 2457 * is next written. If the pagedep is no longer on the buffer 2458 * dependency chain, then all the entries on the pd_pending 2459 * list are committed to disk and we can free them here. 2460 */ 2461 pagedep = newdirblk->db_pagedep; 2462 pagedep->pd_state &= ~NEWBLOCK; 2463 if ((pagedep->pd_state & ONWORKLIST) == 0) 2464 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 2465 free_diradd(dap); 2466 /* 2467 * If no dependencies remain, the pagedep will be freed. 2468 */ 2469 for (i = 0; i < DAHASHSZ; i++) 2470 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL) 2471 break; 2472 if (i == DAHASHSZ && (pagedep->pd_state & ONWORKLIST) == 0) { 2473 LIST_REMOVE(pagedep, pd_hash); 2474 WORKITEM_FREE(pagedep, D_PAGEDEP); 2475 } 2476 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 2477} 2478 2479/* 2480 * Prepare an inode to be freed. The actual free operation is not 2481 * done until the zero'ed inode has been written to disk. 2482 */ 2483void 2484softdep_freefile(pvp, ino, mode) 2485 struct vnode *pvp; 2486 ino_t ino; 2487 int mode; 2488{ 2489 struct inode *ip = VTOI(pvp); 2490 struct inodedep *inodedep; 2491 struct freefile *freefile; 2492 2493 /* 2494 * This sets up the inode de-allocation dependency. 2495 */ 2496 MALLOC(freefile, struct freefile *, sizeof(struct freefile), 2497 M_FREEFILE, M_SOFTDEP_FLAGS); 2498 workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount); 2499 freefile->fx_mode = mode; 2500 freefile->fx_oldinum = ino; 2501 freefile->fx_devvp = ip->i_devvp; 2502 if ((ip->i_flag & IN_SPACECOUNTED) == 0) { 2503 UFS_LOCK(ip->i_ump); 2504 ip->i_fs->fs_pendinginodes += 1; 2505 UFS_UNLOCK(ip->i_ump); 2506 } 2507 2508 /* 2509 * If the inodedep does not exist, then the zero'ed inode has 2510 * been written to disk. If the allocated inode has never been 2511 * written to disk, then the on-disk inode is zero'ed. In either 2512 * case we can free the file immediately. 2513 */ 2514 ACQUIRE_LOCK(&lk); 2515 if (inodedep_lookup(pvp->v_mount, ino, 0, &inodedep) == 0 || 2516 check_inode_unwritten(inodedep)) { 2517 FREE_LOCK(&lk); 2518 handle_workitem_freefile(freefile); 2519 return; 2520 } 2521 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 2522 FREE_LOCK(&lk); 2523} 2524 2525/* 2526 * Check to see if an inode has never been written to disk. If 2527 * so free the inodedep and return success, otherwise return failure. 2528 * This routine must be called with splbio interrupts blocked. 2529 * 2530 * If we still have a bitmap dependency, then the inode has never 2531 * been written to disk. Drop the dependency as it is no longer 2532 * necessary since the inode is being deallocated. We set the 2533 * ALLCOMPLETE flags since the bitmap now properly shows that the 2534 * inode is not allocated. Even if the inode is actively being 2535 * written, it has been rolled back to its zero'ed state, so we 2536 * are ensured that a zero inode is what is on the disk. For short 2537 * lived files, this change will usually result in removing all the 2538 * dependencies from the inode so that it can be freed immediately. 2539 */ 2540static int 2541check_inode_unwritten(inodedep) 2542 struct inodedep *inodedep; 2543{ 2544 2545 mtx_assert(&lk, MA_OWNED); 2546 if ((inodedep->id_state & DEPCOMPLETE) != 0 || 2547 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2548 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2549 LIST_FIRST(&inodedep->id_inowait) != NULL || 2550 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2551 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2552 TAILQ_FIRST(&inodedep->id_extupdt) != NULL || 2553 TAILQ_FIRST(&inodedep->id_newextupdt) != NULL || 2554 inodedep->id_nlinkdelta != 0) 2555 return (0); 2556 2557 /* 2558 * Another process might be in initiate_write_inodeblock_ufs[12] 2559 * trying to allocate memory without holding "Softdep Lock". 2560 */ 2561 if ((inodedep->id_state & IOSTARTED) != 0 && 2562 inodedep->id_savedino1 == NULL) 2563 return (0); 2564 2565 inodedep->id_state |= ALLCOMPLETE; 2566 LIST_REMOVE(inodedep, id_deps); 2567 inodedep->id_buf = NULL; 2568 if (inodedep->id_state & ONWORKLIST) 2569 WORKLIST_REMOVE(&inodedep->id_list); 2570 if (inodedep->id_savedino1 != NULL) { 2571 FREE(inodedep->id_savedino1, M_SAVEDINO); 2572 inodedep->id_savedino1 = NULL; 2573 } 2574 if (free_inodedep(inodedep) == 0) 2575 panic("check_inode_unwritten: busy inode"); 2576 return (1); 2577} 2578 2579/* 2580 * Try to free an inodedep structure. Return 1 if it could be freed. 2581 */ 2582static int 2583free_inodedep(inodedep) 2584 struct inodedep *inodedep; 2585{ 2586 2587 mtx_assert(&lk, MA_OWNED); 2588 if ((inodedep->id_state & ONWORKLIST) != 0 || 2589 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 2590 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2591 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2592 LIST_FIRST(&inodedep->id_inowait) != NULL || 2593 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2594 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2595 TAILQ_FIRST(&inodedep->id_extupdt) != NULL || 2596 TAILQ_FIRST(&inodedep->id_newextupdt) != NULL || 2597 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino1 != NULL) 2598 return (0); 2599 LIST_REMOVE(inodedep, id_hash); 2600 WORKITEM_FREE(inodedep, D_INODEDEP); 2601 num_inodedep -= 1; 2602 return (1); 2603} 2604 2605/* 2606 * This workitem routine performs the block de-allocation. 2607 * The workitem is added to the pending list after the updated 2608 * inode block has been written to disk. As mentioned above, 2609 * checks regarding the number of blocks de-allocated (compared 2610 * to the number of blocks allocated for the file) are also 2611 * performed in this function. 2612 */ 2613static void 2614handle_workitem_freeblocks(freeblks, flags) 2615 struct freeblks *freeblks; 2616 int flags; 2617{ 2618 struct inode *ip; 2619 struct vnode *vp; 2620 struct fs *fs; 2621 struct ufsmount *ump; 2622 int i, nblocks, level, bsize; 2623 ufs2_daddr_t bn, blocksreleased = 0; 2624 int error, allerror = 0; 2625 ufs_lbn_t baselbns[NIADDR], tmpval; 2626 int fs_pendingblocks; 2627 2628 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 2629 fs = ump->um_fs; 2630 fs_pendingblocks = 0; 2631 tmpval = 1; 2632 baselbns[0] = NDADDR; 2633 for (i = 1; i < NIADDR; i++) { 2634 tmpval *= NINDIR(fs); 2635 baselbns[i] = baselbns[i - 1] + tmpval; 2636 } 2637 nblocks = btodb(fs->fs_bsize); 2638 blocksreleased = 0; 2639 /* 2640 * Release all extended attribute blocks or frags. 2641 */ 2642 if (freeblks->fb_oldextsize > 0) { 2643 for (i = (NXADDR - 1); i >= 0; i--) { 2644 if ((bn = freeblks->fb_eblks[i]) == 0) 2645 continue; 2646 bsize = sblksize(fs, freeblks->fb_oldextsize, i); 2647 ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, bsize, 2648 freeblks->fb_previousinum); 2649 blocksreleased += btodb(bsize); 2650 } 2651 } 2652 /* 2653 * Release all data blocks or frags. 2654 */ 2655 if (freeblks->fb_oldsize > 0) { 2656 /* 2657 * Indirect blocks first. 2658 */ 2659 for (level = (NIADDR - 1); level >= 0; level--) { 2660 if ((bn = freeblks->fb_iblks[level]) == 0) 2661 continue; 2662 if ((error = indir_trunc(freeblks, fsbtodb(fs, bn), 2663 level, baselbns[level], &blocksreleased)) == 0) 2664 allerror = error; 2665 ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, 2666 fs->fs_bsize, freeblks->fb_previousinum); 2667 fs_pendingblocks += nblocks; 2668 blocksreleased += nblocks; 2669 } 2670 /* 2671 * All direct blocks or frags. 2672 */ 2673 for (i = (NDADDR - 1); i >= 0; i--) { 2674 if ((bn = freeblks->fb_dblks[i]) == 0) 2675 continue; 2676 bsize = sblksize(fs, freeblks->fb_oldsize, i); 2677 ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, bsize, 2678 freeblks->fb_previousinum); 2679 fs_pendingblocks += btodb(bsize); 2680 blocksreleased += btodb(bsize); 2681 } 2682 } 2683 UFS_LOCK(ump); 2684 fs->fs_pendingblocks -= fs_pendingblocks; 2685 UFS_UNLOCK(ump); 2686 /* 2687 * If we still have not finished background cleanup, then check 2688 * to see if the block count needs to be adjusted. 2689 */ 2690 if (freeblks->fb_chkcnt != blocksreleased && 2691 (fs->fs_flags & FS_UNCLEAN) != 0 && 2692 ffs_vget(freeblks->fb_list.wk_mp, freeblks->fb_previousinum, 2693 (flags & LK_NOWAIT) | LK_EXCLUSIVE, &vp) == 0) { 2694 ip = VTOI(vp); 2695 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + \ 2696 freeblks->fb_chkcnt - blocksreleased); 2697 ip->i_flag |= IN_CHANGE; 2698 vput(vp); 2699 } 2700 2701#ifdef DIAGNOSTIC 2702 if (freeblks->fb_chkcnt != blocksreleased && 2703 ((fs->fs_flags & FS_UNCLEAN) == 0 || (flags & LK_NOWAIT) != 0)) 2704 printf("handle_workitem_freeblocks: block count\n"); 2705 if (allerror) 2706 softdep_error("handle_workitem_freeblks", allerror); 2707#endif /* DIAGNOSTIC */ 2708 2709 ACQUIRE_LOCK(&lk); 2710 WORKITEM_FREE(freeblks, D_FREEBLKS); 2711 FREE_LOCK(&lk); 2712} 2713 2714/* 2715 * Release blocks associated with the inode ip and stored in the indirect 2716 * block dbn. If level is greater than SINGLE, the block is an indirect block 2717 * and recursive calls to indirtrunc must be used to cleanse other indirect 2718 * blocks. 2719 */ 2720static int 2721indir_trunc(freeblks, dbn, level, lbn, countp) 2722 struct freeblks *freeblks; 2723 ufs2_daddr_t dbn; 2724 int level; 2725 ufs_lbn_t lbn; 2726 ufs2_daddr_t *countp; 2727{ 2728 struct buf *bp; 2729 struct fs *fs; 2730 struct worklist *wk; 2731 struct indirdep *indirdep; 2732 struct ufsmount *ump; 2733 ufs1_daddr_t *bap1 = 0; 2734 ufs2_daddr_t nb, *bap2 = 0; 2735 ufs_lbn_t lbnadd; 2736 int i, nblocks, ufs1fmt; 2737 int error, allerror = 0; 2738 int fs_pendingblocks; 2739 2740 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 2741 fs = ump->um_fs; 2742 fs_pendingblocks = 0; 2743 lbnadd = 1; 2744 for (i = level; i > 0; i--) 2745 lbnadd *= NINDIR(fs); 2746 /* 2747 * Get buffer of block pointers to be freed. This routine is not 2748 * called until the zero'ed inode has been written, so it is safe 2749 * to free blocks as they are encountered. Because the inode has 2750 * been zero'ed, calls to bmap on these blocks will fail. So, we 2751 * have to use the on-disk address and the block device for the 2752 * filesystem to look them up. If the file was deleted before its 2753 * indirect blocks were all written to disk, the routine that set 2754 * us up (deallocate_dependencies) will have arranged to leave 2755 * a complete copy of the indirect block in memory for our use. 2756 * Otherwise we have to read the blocks in from the disk. 2757 */ 2758#ifdef notyet 2759 bp = getblk(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 0, 0, 2760 GB_NOCREAT); 2761#else 2762 bp = incore(&freeblks->fb_devvp->v_bufobj, dbn); 2763#endif 2764 ACQUIRE_LOCK(&lk); 2765 if (bp != NULL && (wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2766 if (wk->wk_type != D_INDIRDEP || 2767 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp || 2768 (indirdep->ir_state & GOINGAWAY) == 0) 2769 panic("indir_trunc: lost indirdep"); 2770 WORKLIST_REMOVE(wk); 2771 WORKITEM_FREE(indirdep, D_INDIRDEP); 2772 if (LIST_FIRST(&bp->b_dep) != NULL) 2773 panic("indir_trunc: dangling dep"); 2774 ump->um_numindirdeps -= 1; 2775 FREE_LOCK(&lk); 2776 } else { 2777#ifdef notyet 2778 if (bp) 2779 brelse(bp); 2780#endif 2781 FREE_LOCK(&lk); 2782 error = bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 2783 NOCRED, &bp); 2784 if (error) { 2785 brelse(bp); 2786 return (error); 2787 } 2788 } 2789 /* 2790 * Recursively free indirect blocks. 2791 */ 2792 if (ump->um_fstype == UFS1) { 2793 ufs1fmt = 1; 2794 bap1 = (ufs1_daddr_t *)bp->b_data; 2795 } else { 2796 ufs1fmt = 0; 2797 bap2 = (ufs2_daddr_t *)bp->b_data; 2798 } 2799 nblocks = btodb(fs->fs_bsize); 2800 for (i = NINDIR(fs) - 1; i >= 0; i--) { 2801 if (ufs1fmt) 2802 nb = bap1[i]; 2803 else 2804 nb = bap2[i]; 2805 if (nb == 0) 2806 continue; 2807 if (level != 0) { 2808 if ((error = indir_trunc(freeblks, fsbtodb(fs, nb), 2809 level - 1, lbn + (i * lbnadd), countp)) != 0) 2810 allerror = error; 2811 } 2812 ffs_blkfree(ump, fs, freeblks->fb_devvp, nb, fs->fs_bsize, 2813 freeblks->fb_previousinum); 2814 fs_pendingblocks += nblocks; 2815 *countp += nblocks; 2816 } 2817 UFS_LOCK(ump); 2818 fs->fs_pendingblocks -= fs_pendingblocks; 2819 UFS_UNLOCK(ump); 2820 bp->b_flags |= B_INVAL | B_NOCACHE; 2821 brelse(bp); 2822 return (allerror); 2823} 2824 2825/* 2826 * Free an allocindir. 2827 * This routine must be called with splbio interrupts blocked. 2828 */ 2829static void 2830free_allocindir(aip, inodedep) 2831 struct allocindir *aip; 2832 struct inodedep *inodedep; 2833{ 2834 struct freefrag *freefrag; 2835 2836 mtx_assert(&lk, MA_OWNED); 2837 if ((aip->ai_state & DEPCOMPLETE) == 0) 2838 LIST_REMOVE(aip, ai_deps); 2839 if (aip->ai_state & ONWORKLIST) 2840 WORKLIST_REMOVE(&aip->ai_list); 2841 LIST_REMOVE(aip, ai_next); 2842 if ((freefrag = aip->ai_freefrag) != NULL) { 2843 if (inodedep == NULL) 2844 add_to_worklist(&freefrag->ff_list); 2845 else 2846 WORKLIST_INSERT(&inodedep->id_bufwait, 2847 &freefrag->ff_list); 2848 } 2849 WORKITEM_FREE(aip, D_ALLOCINDIR); 2850} 2851 2852/* 2853 * Directory entry addition dependencies. 2854 * 2855 * When adding a new directory entry, the inode (with its incremented link 2856 * count) must be written to disk before the directory entry's pointer to it. 2857 * Also, if the inode is newly allocated, the corresponding freemap must be 2858 * updated (on disk) before the directory entry's pointer. These requirements 2859 * are met via undo/redo on the directory entry's pointer, which consists 2860 * simply of the inode number. 2861 * 2862 * As directory entries are added and deleted, the free space within a 2863 * directory block can become fragmented. The ufs filesystem will compact 2864 * a fragmented directory block to make space for a new entry. When this 2865 * occurs, the offsets of previously added entries change. Any "diradd" 2866 * dependency structures corresponding to these entries must be updated with 2867 * the new offsets. 2868 */ 2869 2870/* 2871 * This routine is called after the in-memory inode's link 2872 * count has been incremented, but before the directory entry's 2873 * pointer to the inode has been set. 2874 */ 2875int 2876softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 2877 struct buf *bp; /* buffer containing directory block */ 2878 struct inode *dp; /* inode for directory */ 2879 off_t diroffset; /* offset of new entry in directory */ 2880 ino_t newinum; /* inode referenced by new directory entry */ 2881 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 2882 int isnewblk; /* entry is in a newly allocated block */ 2883{ 2884 int offset; /* offset of new entry within directory block */ 2885 ufs_lbn_t lbn; /* block in directory containing new entry */ 2886 struct fs *fs; 2887 struct diradd *dap; 2888 struct allocdirect *adp; 2889 struct pagedep *pagedep; 2890 struct inodedep *inodedep; 2891 struct newdirblk *newdirblk = 0; 2892 struct mkdir *mkdir1, *mkdir2; 2893 struct mount *mp; 2894 2895 /* 2896 * Whiteouts have no dependencies. 2897 */ 2898 if (newinum == WINO) { 2899 if (newdirbp != NULL) 2900 bdwrite(newdirbp); 2901 return (0); 2902 } 2903 mp = UFSTOVFS(dp->i_ump); 2904 fs = dp->i_fs; 2905 lbn = lblkno(fs, diroffset); 2906 offset = blkoff(fs, diroffset); 2907 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, 2908 M_SOFTDEP_FLAGS|M_ZERO); 2909 workitem_alloc(&dap->da_list, D_DIRADD, mp); 2910 dap->da_offset = offset; 2911 dap->da_newinum = newinum; 2912 dap->da_state = ATTACHED; 2913 if (isnewblk && lbn < NDADDR && fragoff(fs, diroffset) == 0) { 2914 MALLOC(newdirblk, struct newdirblk *, sizeof(struct newdirblk), 2915 M_NEWDIRBLK, M_SOFTDEP_FLAGS); 2916 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 2917 } 2918 if (newdirbp == NULL) { 2919 dap->da_state |= DEPCOMPLETE; 2920 ACQUIRE_LOCK(&lk); 2921 } else { 2922 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 2923 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2924 M_SOFTDEP_FLAGS); 2925 workitem_alloc(&mkdir1->md_list, D_MKDIR, mp); 2926 mkdir1->md_state = MKDIR_BODY; 2927 mkdir1->md_diradd = dap; 2928 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 2929 M_SOFTDEP_FLAGS); 2930 workitem_alloc(&mkdir2->md_list, D_MKDIR, mp); 2931 mkdir2->md_state = MKDIR_PARENT; 2932 mkdir2->md_diradd = dap; 2933 /* 2934 * Dependency on "." and ".." being written to disk. 2935 */ 2936 mkdir1->md_buf = newdirbp; 2937 ACQUIRE_LOCK(&lk); 2938 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 2939 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list); 2940 FREE_LOCK(&lk); 2941 bdwrite(newdirbp); 2942 /* 2943 * Dependency on link count increase for parent directory 2944 */ 2945 ACQUIRE_LOCK(&lk); 2946 if (inodedep_lookup(mp, dp->i_number, 0, &inodedep) == 0 2947 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 2948 dap->da_state &= ~MKDIR_PARENT; 2949 WORKITEM_FREE(mkdir2, D_MKDIR); 2950 } else { 2951 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 2952 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 2953 } 2954 } 2955 /* 2956 * Link into parent directory pagedep to await its being written. 2957 */ 2958 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 2959 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2960 dap->da_pagedep = pagedep; 2961 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 2962 da_pdlist); 2963 /* 2964 * Link into its inodedep. Put it on the id_bufwait list if the inode 2965 * is not yet written. If it is written, do the post-inode write 2966 * processing to put it on the id_pendinghd list. 2967 */ 2968 (void) inodedep_lookup(mp, newinum, DEPALLOC, &inodedep); 2969 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 2970 diradd_inode_written(dap, inodedep); 2971 else 2972 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 2973 if (isnewblk) { 2974 /* 2975 * Directories growing into indirect blocks are rare 2976 * enough and the frequency of new block allocation 2977 * in those cases even more rare, that we choose not 2978 * to bother tracking them. Rather we simply force the 2979 * new directory entry to disk. 2980 */ 2981 if (lbn >= NDADDR) { 2982 FREE_LOCK(&lk); 2983 /* 2984 * We only have a new allocation when at the 2985 * beginning of a new block, not when we are 2986 * expanding into an existing block. 2987 */ 2988 if (blkoff(fs, diroffset) == 0) 2989 return (1); 2990 return (0); 2991 } 2992 /* 2993 * We only have a new allocation when at the beginning 2994 * of a new fragment, not when we are expanding into an 2995 * existing fragment. Also, there is nothing to do if we 2996 * are already tracking this block. 2997 */ 2998 if (fragoff(fs, diroffset) != 0) { 2999 FREE_LOCK(&lk); 3000 return (0); 3001 } 3002 if ((pagedep->pd_state & NEWBLOCK) != 0) { 3003 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 3004 FREE_LOCK(&lk); 3005 return (0); 3006 } 3007 /* 3008 * Find our associated allocdirect and have it track us. 3009 */ 3010 if (inodedep_lookup(mp, dp->i_number, 0, &inodedep) == 0) 3011 panic("softdep_setup_directory_add: lost inodedep"); 3012 adp = TAILQ_LAST(&inodedep->id_newinoupdt, allocdirectlst); 3013 if (adp == NULL || adp->ad_lbn != lbn) 3014 panic("softdep_setup_directory_add: lost entry"); 3015 pagedep->pd_state |= NEWBLOCK; 3016 newdirblk->db_pagedep = pagedep; 3017 WORKLIST_INSERT(&adp->ad_newdirblk, &newdirblk->db_list); 3018 } 3019 FREE_LOCK(&lk); 3020 return (0); 3021} 3022 3023/* 3024 * This procedure is called to change the offset of a directory 3025 * entry when compacting a directory block which must be owned 3026 * exclusively by the caller. Note that the actual entry movement 3027 * must be done in this procedure to ensure that no I/O completions 3028 * occur while the move is in progress. 3029 */ 3030void 3031softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize) 3032 struct inode *dp; /* inode for directory */ 3033 caddr_t base; /* address of dp->i_offset */ 3034 caddr_t oldloc; /* address of old directory location */ 3035 caddr_t newloc; /* address of new directory location */ 3036 int entrysize; /* size of directory entry */ 3037{ 3038 int offset, oldoffset, newoffset; 3039 struct pagedep *pagedep; 3040 struct diradd *dap; 3041 ufs_lbn_t lbn; 3042 3043 ACQUIRE_LOCK(&lk); 3044 lbn = lblkno(dp->i_fs, dp->i_offset); 3045 offset = blkoff(dp->i_fs, dp->i_offset); 3046 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) 3047 goto done; 3048 oldoffset = offset + (oldloc - base); 3049 newoffset = offset + (newloc - base); 3050 3051 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) { 3052 if (dap->da_offset != oldoffset) 3053 continue; 3054 dap->da_offset = newoffset; 3055 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) 3056 break; 3057 LIST_REMOVE(dap, da_pdlist); 3058 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], 3059 dap, da_pdlist); 3060 break; 3061 } 3062 if (dap == NULL) { 3063 3064 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) { 3065 if (dap->da_offset == oldoffset) { 3066 dap->da_offset = newoffset; 3067 break; 3068 } 3069 } 3070 } 3071done: 3072 bcopy(oldloc, newloc, entrysize); 3073 FREE_LOCK(&lk); 3074} 3075 3076/* 3077 * Free a diradd dependency structure. This routine must be called 3078 * with splbio interrupts blocked. 3079 */ 3080static void 3081free_diradd(dap) 3082 struct diradd *dap; 3083{ 3084 struct dirrem *dirrem; 3085 struct pagedep *pagedep; 3086 struct inodedep *inodedep; 3087 struct mkdir *mkdir, *nextmd; 3088 3089 mtx_assert(&lk, MA_OWNED); 3090 WORKLIST_REMOVE(&dap->da_list); 3091 LIST_REMOVE(dap, da_pdlist); 3092 if ((dap->da_state & DIRCHG) == 0) { 3093 pagedep = dap->da_pagedep; 3094 } else { 3095 dirrem = dap->da_previous; 3096 pagedep = dirrem->dm_pagedep; 3097 dirrem->dm_dirinum = pagedep->pd_ino; 3098 add_to_worklist(&dirrem->dm_list); 3099 } 3100 if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum, 3101 0, &inodedep) != 0) 3102 (void) free_inodedep(inodedep); 3103 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 3104 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 3105 nextmd = LIST_NEXT(mkdir, md_mkdirs); 3106 if (mkdir->md_diradd != dap) 3107 continue; 3108 dap->da_state &= ~mkdir->md_state; 3109 WORKLIST_REMOVE(&mkdir->md_list); 3110 LIST_REMOVE(mkdir, md_mkdirs); 3111 WORKITEM_FREE(mkdir, D_MKDIR); 3112 } 3113 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 3114 panic("free_diradd: unfound ref"); 3115 } 3116 WORKITEM_FREE(dap, D_DIRADD); 3117} 3118 3119/* 3120 * Directory entry removal dependencies. 3121 * 3122 * When removing a directory entry, the entry's inode pointer must be 3123 * zero'ed on disk before the corresponding inode's link count is decremented 3124 * (possibly freeing the inode for re-use). This dependency is handled by 3125 * updating the directory entry but delaying the inode count reduction until 3126 * after the directory block has been written to disk. After this point, the 3127 * inode count can be decremented whenever it is convenient. 3128 */ 3129 3130/* 3131 * This routine should be called immediately after removing 3132 * a directory entry. The inode's link count should not be 3133 * decremented by the calling procedure -- the soft updates 3134 * code will do this task when it is safe. 3135 */ 3136void 3137softdep_setup_remove(bp, dp, ip, isrmdir) 3138 struct buf *bp; /* buffer containing directory block */ 3139 struct inode *dp; /* inode for the directory being modified */ 3140 struct inode *ip; /* inode for directory entry being removed */ 3141 int isrmdir; /* indicates if doing RMDIR */ 3142{ 3143 struct dirrem *dirrem, *prevdirrem; 3144 3145 /* 3146 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 3147 */ 3148 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 3149 3150 /* 3151 * If the COMPLETE flag is clear, then there were no active 3152 * entries and we want to roll back to a zeroed entry until 3153 * the new inode is committed to disk. If the COMPLETE flag is 3154 * set then we have deleted an entry that never made it to 3155 * disk. If the entry we deleted resulted from a name change, 3156 * then the old name still resides on disk. We cannot delete 3157 * its inode (returned to us in prevdirrem) until the zeroed 3158 * directory entry gets to disk. The new inode has never been 3159 * referenced on the disk, so can be deleted immediately. 3160 */ 3161 if ((dirrem->dm_state & COMPLETE) == 0) { 3162 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 3163 dm_next); 3164 FREE_LOCK(&lk); 3165 } else { 3166 if (prevdirrem != NULL) 3167 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 3168 prevdirrem, dm_next); 3169 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 3170 FREE_LOCK(&lk); 3171 handle_workitem_remove(dirrem, NULL); 3172 } 3173} 3174 3175/* 3176 * Allocate a new dirrem if appropriate and return it along with 3177 * its associated pagedep. Called without a lock, returns with lock. 3178 */ 3179static long num_dirrem; /* number of dirrem allocated */ 3180static struct dirrem * 3181newdirrem(bp, dp, ip, isrmdir, prevdirremp) 3182 struct buf *bp; /* buffer containing directory block */ 3183 struct inode *dp; /* inode for the directory being modified */ 3184 struct inode *ip; /* inode for directory entry being removed */ 3185 int isrmdir; /* indicates if doing RMDIR */ 3186 struct dirrem **prevdirremp; /* previously referenced inode, if any */ 3187{ 3188 int offset; 3189 ufs_lbn_t lbn; 3190 struct diradd *dap; 3191 struct dirrem *dirrem; 3192 struct pagedep *pagedep; 3193 3194 /* 3195 * Whiteouts have no deletion dependencies. 3196 */ 3197 if (ip == NULL) 3198 panic("newdirrem: whiteout"); 3199 /* 3200 * If we are over our limit, try to improve the situation. 3201 * Limiting the number of dirrem structures will also limit 3202 * the number of freefile and freeblks structures. 3203 */ 3204 ACQUIRE_LOCK(&lk); 3205 if (num_dirrem > max_softdeps / 2) 3206 (void) request_cleanup(ITOV(dp)->v_mount, FLUSH_REMOVE); 3207 num_dirrem += 1; 3208 FREE_LOCK(&lk); 3209 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem), 3210 M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO); 3211 workitem_alloc(&dirrem->dm_list, D_DIRREM, ITOV(dp)->v_mount); 3212 dirrem->dm_state = isrmdir ? RMDIR : 0; 3213 dirrem->dm_oldinum = ip->i_number; 3214 *prevdirremp = NULL; 3215 3216 ACQUIRE_LOCK(&lk); 3217 lbn = lblkno(dp->i_fs, dp->i_offset); 3218 offset = blkoff(dp->i_fs, dp->i_offset); 3219 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 3220 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 3221 dirrem->dm_pagedep = pagedep; 3222 /* 3223 * Check for a diradd dependency for the same directory entry. 3224 * If present, then both dependencies become obsolete and can 3225 * be de-allocated. Check for an entry on both the pd_dirraddhd 3226 * list and the pd_pendinghd list. 3227 */ 3228 3229 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 3230 if (dap->da_offset == offset) 3231 break; 3232 if (dap == NULL) { 3233 3234 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 3235 if (dap->da_offset == offset) 3236 break; 3237 if (dap == NULL) 3238 return (dirrem); 3239 } 3240 /* 3241 * Must be ATTACHED at this point. 3242 */ 3243 if ((dap->da_state & ATTACHED) == 0) 3244 panic("newdirrem: not ATTACHED"); 3245 if (dap->da_newinum != ip->i_number) 3246 panic("newdirrem: inum %d should be %d", 3247 ip->i_number, dap->da_newinum); 3248 /* 3249 * If we are deleting a changed name that never made it to disk, 3250 * then return the dirrem describing the previous inode (which 3251 * represents the inode currently referenced from this entry on disk). 3252 */ 3253 if ((dap->da_state & DIRCHG) != 0) { 3254 *prevdirremp = dap->da_previous; 3255 dap->da_state &= ~DIRCHG; 3256 dap->da_pagedep = pagedep; 3257 } 3258 /* 3259 * We are deleting an entry that never made it to disk. 3260 * Mark it COMPLETE so we can delete its inode immediately. 3261 */ 3262 dirrem->dm_state |= COMPLETE; 3263 free_diradd(dap); 3264 return (dirrem); 3265} 3266 3267/* 3268 * Directory entry change dependencies. 3269 * 3270 * Changing an existing directory entry requires that an add operation 3271 * be completed first followed by a deletion. The semantics for the addition 3272 * are identical to the description of adding a new entry above except 3273 * that the rollback is to the old inode number rather than zero. Once 3274 * the addition dependency is completed, the removal is done as described 3275 * in the removal routine above. 3276 */ 3277 3278/* 3279 * This routine should be called immediately after changing 3280 * a directory entry. The inode's link count should not be 3281 * decremented by the calling procedure -- the soft updates 3282 * code will perform this task when it is safe. 3283 */ 3284void 3285softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 3286 struct buf *bp; /* buffer containing directory block */ 3287 struct inode *dp; /* inode for the directory being modified */ 3288 struct inode *ip; /* inode for directory entry being removed */ 3289 ino_t newinum; /* new inode number for changed entry */ 3290 int isrmdir; /* indicates if doing RMDIR */ 3291{ 3292 int offset; 3293 struct diradd *dap = NULL; 3294 struct dirrem *dirrem, *prevdirrem; 3295 struct pagedep *pagedep; 3296 struct inodedep *inodedep; 3297 struct mount *mp; 3298 3299 offset = blkoff(dp->i_fs, dp->i_offset); 3300 mp = UFSTOVFS(dp->i_ump); 3301 3302 /* 3303 * Whiteouts do not need diradd dependencies. 3304 */ 3305 if (newinum != WINO) { 3306 MALLOC(dap, struct diradd *, sizeof(struct diradd), 3307 M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO); 3308 workitem_alloc(&dap->da_list, D_DIRADD, mp); 3309 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 3310 dap->da_offset = offset; 3311 dap->da_newinum = newinum; 3312 } 3313 3314 /* 3315 * Allocate a new dirrem and ACQUIRE_LOCK. 3316 */ 3317 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 3318 pagedep = dirrem->dm_pagedep; 3319 /* 3320 * The possible values for isrmdir: 3321 * 0 - non-directory file rename 3322 * 1 - directory rename within same directory 3323 * inum - directory rename to new directory of given inode number 3324 * When renaming to a new directory, we are both deleting and 3325 * creating a new directory entry, so the link count on the new 3326 * directory should not change. Thus we do not need the followup 3327 * dirrem which is usually done in handle_workitem_remove. We set 3328 * the DIRCHG flag to tell handle_workitem_remove to skip the 3329 * followup dirrem. 3330 */ 3331 if (isrmdir > 1) 3332 dirrem->dm_state |= DIRCHG; 3333 3334 /* 3335 * Whiteouts have no additional dependencies, 3336 * so just put the dirrem on the correct list. 3337 */ 3338 if (newinum == WINO) { 3339 if ((dirrem->dm_state & COMPLETE) == 0) { 3340 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 3341 dm_next); 3342 } else { 3343 dirrem->dm_dirinum = pagedep->pd_ino; 3344 add_to_worklist(&dirrem->dm_list); 3345 } 3346 FREE_LOCK(&lk); 3347 return; 3348 } 3349 3350 /* 3351 * If the COMPLETE flag is clear, then there were no active 3352 * entries and we want to roll back to the previous inode until 3353 * the new inode is committed to disk. If the COMPLETE flag is 3354 * set, then we have deleted an entry that never made it to disk. 3355 * If the entry we deleted resulted from a name change, then the old 3356 * inode reference still resides on disk. Any rollback that we do 3357 * needs to be to that old inode (returned to us in prevdirrem). If 3358 * the entry we deleted resulted from a create, then there is 3359 * no entry on the disk, so we want to roll back to zero rather 3360 * than the uncommitted inode. In either of the COMPLETE cases we 3361 * want to immediately free the unwritten and unreferenced inode. 3362 */ 3363 if ((dirrem->dm_state & COMPLETE) == 0) { 3364 dap->da_previous = dirrem; 3365 } else { 3366 if (prevdirrem != NULL) { 3367 dap->da_previous = prevdirrem; 3368 } else { 3369 dap->da_state &= ~DIRCHG; 3370 dap->da_pagedep = pagedep; 3371 } 3372 dirrem->dm_dirinum = pagedep->pd_ino; 3373 add_to_worklist(&dirrem->dm_list); 3374 } 3375 /* 3376 * Link into its inodedep. Put it on the id_bufwait list if the inode 3377 * is not yet written. If it is written, do the post-inode write 3378 * processing to put it on the id_pendinghd list. 3379 */ 3380 if (inodedep_lookup(mp, newinum, DEPALLOC, &inodedep) == 0 || 3381 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 3382 dap->da_state |= COMPLETE; 3383 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3384 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 3385 } else { 3386 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 3387 dap, da_pdlist); 3388 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 3389 } 3390 FREE_LOCK(&lk); 3391} 3392 3393/* 3394 * Called whenever the link count on an inode is changed. 3395 * It creates an inode dependency so that the new reference(s) 3396 * to the inode cannot be committed to disk until the updated 3397 * inode has been written. 3398 */ 3399void 3400softdep_change_linkcnt(ip) 3401 struct inode *ip; /* the inode with the increased link count */ 3402{ 3403 struct inodedep *inodedep; 3404 3405 ACQUIRE_LOCK(&lk); 3406 (void) inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 3407 DEPALLOC, &inodedep); 3408 if (ip->i_nlink < ip->i_effnlink) 3409 panic("softdep_change_linkcnt: bad delta"); 3410 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3411 FREE_LOCK(&lk); 3412} 3413 3414/* 3415 * Called when the effective link count and the reference count 3416 * on an inode drops to zero. At this point there are no names 3417 * referencing the file in the filesystem and no active file 3418 * references. The space associated with the file will be freed 3419 * as soon as the necessary soft dependencies are cleared. 3420 */ 3421void 3422softdep_releasefile(ip) 3423 struct inode *ip; /* inode with the zero effective link count */ 3424{ 3425 struct inodedep *inodedep; 3426 struct fs *fs; 3427 int extblocks; 3428 3429 if (ip->i_effnlink > 0) 3430 panic("softdep_filerelease: file still referenced"); 3431 /* 3432 * We may be called several times as the real reference count 3433 * drops to zero. We only want to account for the space once. 3434 */ 3435 if (ip->i_flag & IN_SPACECOUNTED) 3436 return; 3437 /* 3438 * We have to deactivate a snapshot otherwise copyonwrites may 3439 * add blocks and the cleanup may remove blocks after we have 3440 * tried to account for them. 3441 */ 3442 if ((ip->i_flags & SF_SNAPSHOT) != 0) 3443 ffs_snapremove(ITOV(ip)); 3444 /* 3445 * If we are tracking an nlinkdelta, we have to also remember 3446 * whether we accounted for the freed space yet. 3447 */ 3448 ACQUIRE_LOCK(&lk); 3449 if ((inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, &inodedep))) 3450 inodedep->id_state |= SPACECOUNTED; 3451 FREE_LOCK(&lk); 3452 fs = ip->i_fs; 3453 extblocks = 0; 3454 if (fs->fs_magic == FS_UFS2_MAGIC) 3455 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 3456 UFS_LOCK(ip->i_ump); 3457 ip->i_fs->fs_pendingblocks += DIP(ip, i_blocks) - extblocks; 3458 ip->i_fs->fs_pendinginodes += 1; 3459 UFS_UNLOCK(ip->i_ump); 3460 ip->i_flag |= IN_SPACECOUNTED; 3461} 3462 3463/* 3464 * This workitem decrements the inode's link count. 3465 * If the link count reaches zero, the file is removed. 3466 */ 3467static void 3468handle_workitem_remove(dirrem, xp) 3469 struct dirrem *dirrem; 3470 struct vnode *xp; 3471{ 3472 struct thread *td = curthread; 3473 struct inodedep *inodedep; 3474 struct vnode *vp; 3475 struct inode *ip; 3476 ino_t oldinum; 3477 int error; 3478 3479 if ((vp = xp) == NULL && 3480 (error = ffs_vget(dirrem->dm_list.wk_mp, 3481 dirrem->dm_oldinum, LK_EXCLUSIVE, &vp)) != 0) { 3482 softdep_error("handle_workitem_remove: vget", error); 3483 return; 3484 } 3485 ip = VTOI(vp); 3486 ACQUIRE_LOCK(&lk); 3487 if ((inodedep_lookup(dirrem->dm_list.wk_mp, 3488 dirrem->dm_oldinum, 0, &inodedep)) == 0) 3489 panic("handle_workitem_remove: lost inodedep"); 3490 /* 3491 * Normal file deletion. 3492 */ 3493 if ((dirrem->dm_state & RMDIR) == 0) { 3494 ip->i_nlink--; 3495 DIP_SET(ip, i_nlink, ip->i_nlink); 3496 ip->i_flag |= IN_CHANGE; 3497 if (ip->i_nlink < ip->i_effnlink) 3498 panic("handle_workitem_remove: bad file delta"); 3499 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3500 num_dirrem -= 1; 3501 WORKITEM_FREE(dirrem, D_DIRREM); 3502 FREE_LOCK(&lk); 3503 vput(vp); 3504 return; 3505 } 3506 /* 3507 * Directory deletion. Decrement reference count for both the 3508 * just deleted parent directory entry and the reference for ".". 3509 * Next truncate the directory to length zero. When the 3510 * truncation completes, arrange to have the reference count on 3511 * the parent decremented to account for the loss of "..". 3512 */ 3513 ip->i_nlink -= 2; 3514 DIP_SET(ip, i_nlink, ip->i_nlink); 3515 ip->i_flag |= IN_CHANGE; 3516 if (ip->i_nlink < ip->i_effnlink) 3517 panic("handle_workitem_remove: bad dir delta"); 3518 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3519 FREE_LOCK(&lk); 3520 if ((error = ffs_truncate(vp, (off_t)0, 0, td->td_ucred, td)) != 0) 3521 softdep_error("handle_workitem_remove: truncate", error); 3522 ACQUIRE_LOCK(&lk); 3523 /* 3524 * Rename a directory to a new parent. Since, we are both deleting 3525 * and creating a new directory entry, the link count on the new 3526 * directory should not change. Thus we skip the followup dirrem. 3527 */ 3528 if (dirrem->dm_state & DIRCHG) { 3529 num_dirrem -= 1; 3530 WORKITEM_FREE(dirrem, D_DIRREM); 3531 FREE_LOCK(&lk); 3532 vput(vp); 3533 return; 3534 } 3535 /* 3536 * If the inodedep does not exist, then the zero'ed inode has 3537 * been written to disk. If the allocated inode has never been 3538 * written to disk, then the on-disk inode is zero'ed. In either 3539 * case we can remove the file immediately. 3540 */ 3541 dirrem->dm_state = 0; 3542 oldinum = dirrem->dm_oldinum; 3543 dirrem->dm_oldinum = dirrem->dm_dirinum; 3544 if (inodedep_lookup(dirrem->dm_list.wk_mp, oldinum, 3545 0, &inodedep) == 0 || check_inode_unwritten(inodedep)) { 3546 FREE_LOCK(&lk); 3547 vput(vp); 3548 handle_workitem_remove(dirrem, NULL); 3549 return; 3550 } 3551 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 3552 FREE_LOCK(&lk); 3553 ip->i_flag |= IN_CHANGE; 3554 ffs_update(vp, 0); 3555 vput(vp); 3556} 3557 3558/* 3559 * Inode de-allocation dependencies. 3560 * 3561 * When an inode's link count is reduced to zero, it can be de-allocated. We 3562 * found it convenient to postpone de-allocation until after the inode is 3563 * written to disk with its new link count (zero). At this point, all of the 3564 * on-disk inode's block pointers are nullified and, with careful dependency 3565 * list ordering, all dependencies related to the inode will be satisfied and 3566 * the corresponding dependency structures de-allocated. So, if/when the 3567 * inode is reused, there will be no mixing of old dependencies with new 3568 * ones. This artificial dependency is set up by the block de-allocation 3569 * procedure above (softdep_setup_freeblocks) and completed by the 3570 * following procedure. 3571 */ 3572static void 3573handle_workitem_freefile(freefile) 3574 struct freefile *freefile; 3575{ 3576 struct fs *fs; 3577 struct inodedep *idp; 3578 struct ufsmount *ump; 3579 int error; 3580 3581 ump = VFSTOUFS(freefile->fx_list.wk_mp); 3582 fs = ump->um_fs; 3583#ifdef DEBUG 3584 ACQUIRE_LOCK(&lk); 3585 error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp); 3586 FREE_LOCK(&lk); 3587 if (error) 3588 panic("handle_workitem_freefile: inodedep survived"); 3589#endif 3590 UFS_LOCK(ump); 3591 fs->fs_pendinginodes -= 1; 3592 UFS_UNLOCK(ump); 3593 if ((error = ffs_freefile(ump, fs, freefile->fx_devvp, 3594 freefile->fx_oldinum, freefile->fx_mode)) != 0) 3595 softdep_error("handle_workitem_freefile", error); 3596 ACQUIRE_LOCK(&lk); 3597 WORKITEM_FREE(freefile, D_FREEFILE); 3598 FREE_LOCK(&lk); 3599} 3600 3601 3602/* 3603 * Helper function which unlinks marker element from work list and returns 3604 * the next element on the list. 3605 */ 3606static __inline struct worklist * 3607markernext(struct worklist *marker) 3608{ 3609 struct worklist *next; 3610 3611 next = LIST_NEXT(marker, wk_list); 3612 LIST_REMOVE(marker, wk_list); 3613 return next; 3614} 3615 3616/* 3617 * Disk writes. 3618 * 3619 * The dependency structures constructed above are most actively used when file 3620 * system blocks are written to disk. No constraints are placed on when a 3621 * block can be written, but unsatisfied update dependencies are made safe by 3622 * modifying (or replacing) the source memory for the duration of the disk 3623 * write. When the disk write completes, the memory block is again brought 3624 * up-to-date. 3625 * 3626 * In-core inode structure reclamation. 3627 * 3628 * Because there are a finite number of "in-core" inode structures, they are 3629 * reused regularly. By transferring all inode-related dependencies to the 3630 * in-memory inode block and indexing them separately (via "inodedep"s), we 3631 * can allow "in-core" inode structures to be reused at any time and avoid 3632 * any increase in contention. 3633 * 3634 * Called just before entering the device driver to initiate a new disk I/O. 3635 * The buffer must be locked, thus, no I/O completion operations can occur 3636 * while we are manipulating its associated dependencies. 3637 */ 3638static void 3639softdep_disk_io_initiation(bp) 3640 struct buf *bp; /* structure describing disk write to occur */ 3641{ 3642 struct worklist *wk; 3643 struct worklist marker; 3644 struct indirdep *indirdep; 3645 struct inodedep *inodedep; 3646 3647 /* 3648 * We only care about write operations. There should never 3649 * be dependencies for reads. 3650 */ 3651 if (bp->b_iocmd != BIO_WRITE) 3652 panic("softdep_disk_io_initiation: not write"); 3653 3654 marker.wk_type = D_LAST + 1; /* Not a normal workitem */ 3655 PHOLD(curproc); /* Don't swap out kernel stack */ 3656 3657 ACQUIRE_LOCK(&lk); 3658 /* 3659 * Do any necessary pre-I/O processing. 3660 */ 3661 for (wk = LIST_FIRST(&bp->b_dep); wk != NULL; 3662 wk = markernext(&marker)) { 3663 LIST_INSERT_AFTER(wk, &marker, wk_list); 3664 switch (wk->wk_type) { 3665 3666 case D_PAGEDEP: 3667 initiate_write_filepage(WK_PAGEDEP(wk), bp); 3668 continue; 3669 3670 case D_INODEDEP: 3671 inodedep = WK_INODEDEP(wk); 3672 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) 3673 initiate_write_inodeblock_ufs1(inodedep, bp); 3674 else 3675 initiate_write_inodeblock_ufs2(inodedep, bp); 3676 continue; 3677 3678 case D_INDIRDEP: 3679 indirdep = WK_INDIRDEP(wk); 3680 if (indirdep->ir_state & GOINGAWAY) 3681 panic("disk_io_initiation: indirdep gone"); 3682 /* 3683 * If there are no remaining dependencies, this 3684 * will be writing the real pointers, so the 3685 * dependency can be freed. 3686 */ 3687 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) { 3688 struct buf *bp; 3689 3690 bp = indirdep->ir_savebp; 3691 bp->b_flags |= B_INVAL | B_NOCACHE; 3692 /* inline expand WORKLIST_REMOVE(wk); */ 3693 wk->wk_state &= ~ONWORKLIST; 3694 LIST_REMOVE(wk, wk_list); 3695 WORKITEM_FREE(indirdep, D_INDIRDEP); 3696 FREE_LOCK(&lk); 3697 brelse(bp); 3698 ACQUIRE_LOCK(&lk); 3699 continue; 3700 } 3701 /* 3702 * Replace up-to-date version with safe version. 3703 */ 3704 FREE_LOCK(&lk); 3705 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount, 3706 M_INDIRDEP, M_SOFTDEP_FLAGS); 3707 ACQUIRE_LOCK(&lk); 3708 indirdep->ir_state &= ~ATTACHED; 3709 indirdep->ir_state |= UNDONE; 3710 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 3711 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 3712 bp->b_bcount); 3713 continue; 3714 3715 case D_MKDIR: 3716 case D_BMSAFEMAP: 3717 case D_ALLOCDIRECT: 3718 case D_ALLOCINDIR: 3719 continue; 3720 3721 default: 3722 panic("handle_disk_io_initiation: Unexpected type %s", 3723 TYPENAME(wk->wk_type)); 3724 /* NOTREACHED */ 3725 } 3726 } 3727 FREE_LOCK(&lk); 3728 PRELE(curproc); /* Allow swapout of kernel stack */ 3729} 3730 3731/* 3732 * Called from within the procedure above to deal with unsatisfied 3733 * allocation dependencies in a directory. The buffer must be locked, 3734 * thus, no I/O completion operations can occur while we are 3735 * manipulating its associated dependencies. 3736 */ 3737static void 3738initiate_write_filepage(pagedep, bp) 3739 struct pagedep *pagedep; 3740 struct buf *bp; 3741{ 3742 struct diradd *dap; 3743 struct direct *ep; 3744 int i; 3745 3746 if (pagedep->pd_state & IOSTARTED) { 3747 /* 3748 * This can only happen if there is a driver that does not 3749 * understand chaining. Here biodone will reissue the call 3750 * to strategy for the incomplete buffers. 3751 */ 3752 printf("initiate_write_filepage: already started\n"); 3753 return; 3754 } 3755 pagedep->pd_state |= IOSTARTED; 3756 for (i = 0; i < DAHASHSZ; i++) { 3757 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 3758 ep = (struct direct *) 3759 ((char *)bp->b_data + dap->da_offset); 3760 if (ep->d_ino != dap->da_newinum) 3761 panic("%s: dir inum %d != new %d", 3762 "initiate_write_filepage", 3763 ep->d_ino, dap->da_newinum); 3764 if (dap->da_state & DIRCHG) 3765 ep->d_ino = dap->da_previous->dm_oldinum; 3766 else 3767 ep->d_ino = 0; 3768 dap->da_state &= ~ATTACHED; 3769 dap->da_state |= UNDONE; 3770 } 3771 } 3772} 3773 3774/* 3775 * Version of initiate_write_inodeblock that handles UFS1 dinodes. 3776 * Note that any bug fixes made to this routine must be done in the 3777 * version found below. 3778 * 3779 * Called from within the procedure above to deal with unsatisfied 3780 * allocation dependencies in an inodeblock. The buffer must be 3781 * locked, thus, no I/O completion operations can occur while we 3782 * are manipulating its associated dependencies. 3783 */ 3784static void 3785initiate_write_inodeblock_ufs1(inodedep, bp) 3786 struct inodedep *inodedep; 3787 struct buf *bp; /* The inode block */ 3788{ 3789 struct allocdirect *adp, *lastadp; 3790 struct ufs1_dinode *dp; 3791 struct ufs1_dinode *sip; 3792 struct fs *fs; 3793 ufs_lbn_t i, prevlbn = 0; 3794 int deplist; 3795 3796 if (inodedep->id_state & IOSTARTED) 3797 panic("initiate_write_inodeblock_ufs1: already started"); 3798 inodedep->id_state |= IOSTARTED; 3799 fs = inodedep->id_fs; 3800 dp = (struct ufs1_dinode *)bp->b_data + 3801 ino_to_fsbo(fs, inodedep->id_ino); 3802 /* 3803 * If the bitmap is not yet written, then the allocated 3804 * inode cannot be written to disk. 3805 */ 3806 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 3807 if (inodedep->id_savedino1 != NULL) 3808 panic("initiate_write_inodeblock_ufs1: I/O underway"); 3809 FREE_LOCK(&lk); 3810 MALLOC(sip, struct ufs1_dinode *, 3811 sizeof(struct ufs1_dinode), M_SAVEDINO, M_SOFTDEP_FLAGS); 3812 ACQUIRE_LOCK(&lk); 3813 inodedep->id_savedino1 = sip; 3814 *inodedep->id_savedino1 = *dp; 3815 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 3816 dp->di_gen = inodedep->id_savedino1->di_gen; 3817 return; 3818 } 3819 /* 3820 * If no dependencies, then there is nothing to roll back. 3821 */ 3822 inodedep->id_savedsize = dp->di_size; 3823 inodedep->id_savedextsize = 0; 3824 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 3825 return; 3826 /* 3827 * Set the dependencies to busy. 3828 */ 3829 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3830 adp = TAILQ_NEXT(adp, ad_next)) { 3831#ifdef DIAGNOSTIC 3832 if (deplist != 0 && prevlbn >= adp->ad_lbn) 3833 panic("softdep_write_inodeblock: lbn order"); 3834 prevlbn = adp->ad_lbn; 3835 if (adp->ad_lbn < NDADDR && 3836 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) 3837 panic("%s: direct pointer #%jd mismatch %d != %jd", 3838 "softdep_write_inodeblock", 3839 (intmax_t)adp->ad_lbn, 3840 dp->di_db[adp->ad_lbn], 3841 (intmax_t)adp->ad_newblkno); 3842 if (adp->ad_lbn >= NDADDR && 3843 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) 3844 panic("%s: indirect pointer #%jd mismatch %d != %jd", 3845 "softdep_write_inodeblock", 3846 (intmax_t)adp->ad_lbn - NDADDR, 3847 dp->di_ib[adp->ad_lbn - NDADDR], 3848 (intmax_t)adp->ad_newblkno); 3849 deplist |= 1 << adp->ad_lbn; 3850 if ((adp->ad_state & ATTACHED) == 0) 3851 panic("softdep_write_inodeblock: Unknown state 0x%x", 3852 adp->ad_state); 3853#endif /* DIAGNOSTIC */ 3854 adp->ad_state &= ~ATTACHED; 3855 adp->ad_state |= UNDONE; 3856 } 3857 /* 3858 * The on-disk inode cannot claim to be any larger than the last 3859 * fragment that has been written. Otherwise, the on-disk inode 3860 * might have fragments that were not the last block in the file 3861 * which would corrupt the filesystem. 3862 */ 3863 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3864 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 3865 if (adp->ad_lbn >= NDADDR) 3866 break; 3867 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 3868 /* keep going until hitting a rollback to a frag */ 3869 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 3870 continue; 3871 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 3872 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 3873#ifdef DIAGNOSTIC 3874 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 3875 panic("softdep_write_inodeblock: lost dep1"); 3876#endif /* DIAGNOSTIC */ 3877 dp->di_db[i] = 0; 3878 } 3879 for (i = 0; i < NIADDR; i++) { 3880#ifdef DIAGNOSTIC 3881 if (dp->di_ib[i] != 0 && 3882 (deplist & ((1 << NDADDR) << i)) == 0) 3883 panic("softdep_write_inodeblock: lost dep2"); 3884#endif /* DIAGNOSTIC */ 3885 dp->di_ib[i] = 0; 3886 } 3887 return; 3888 } 3889 /* 3890 * If we have zero'ed out the last allocated block of the file, 3891 * roll back the size to the last currently allocated block. 3892 * We know that this last allocated block is a full-sized as 3893 * we already checked for fragments in the loop above. 3894 */ 3895 if (lastadp != NULL && 3896 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 3897 for (i = lastadp->ad_lbn; i >= 0; i--) 3898 if (dp->di_db[i] != 0) 3899 break; 3900 dp->di_size = (i + 1) * fs->fs_bsize; 3901 } 3902 /* 3903 * The only dependencies are for indirect blocks. 3904 * 3905 * The file size for indirect block additions is not guaranteed. 3906 * Such a guarantee would be non-trivial to achieve. The conventional 3907 * synchronous write implementation also does not make this guarantee. 3908 * Fsck should catch and fix discrepancies. Arguably, the file size 3909 * can be over-estimated without destroying integrity when the file 3910 * moves into the indirect blocks (i.e., is large). If we want to 3911 * postpone fsck, we are stuck with this argument. 3912 */ 3913 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 3914 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 3915} 3916 3917/* 3918 * Version of initiate_write_inodeblock that handles UFS2 dinodes. 3919 * Note that any bug fixes made to this routine must be done in the 3920 * version found above. 3921 * 3922 * Called from within the procedure above to deal with unsatisfied 3923 * allocation dependencies in an inodeblock. The buffer must be 3924 * locked, thus, no I/O completion operations can occur while we 3925 * are manipulating its associated dependencies. 3926 */ 3927static void 3928initiate_write_inodeblock_ufs2(inodedep, bp) 3929 struct inodedep *inodedep; 3930 struct buf *bp; /* The inode block */ 3931{ 3932 struct allocdirect *adp, *lastadp; 3933 struct ufs2_dinode *dp; 3934 struct ufs2_dinode *sip; 3935 struct fs *fs; 3936 ufs_lbn_t i, prevlbn = 0; 3937 int deplist; 3938 3939 if (inodedep->id_state & IOSTARTED) 3940 panic("initiate_write_inodeblock_ufs2: already started"); 3941 inodedep->id_state |= IOSTARTED; 3942 fs = inodedep->id_fs; 3943 dp = (struct ufs2_dinode *)bp->b_data + 3944 ino_to_fsbo(fs, inodedep->id_ino); 3945 /* 3946 * If the bitmap is not yet written, then the allocated 3947 * inode cannot be written to disk. 3948 */ 3949 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 3950 if (inodedep->id_savedino2 != NULL) 3951 panic("initiate_write_inodeblock_ufs2: I/O underway"); 3952 FREE_LOCK(&lk); 3953 MALLOC(sip, struct ufs2_dinode *, 3954 sizeof(struct ufs2_dinode), M_SAVEDINO, M_SOFTDEP_FLAGS); 3955 ACQUIRE_LOCK(&lk); 3956 inodedep->id_savedino2 = sip; 3957 *inodedep->id_savedino2 = *dp; 3958 bzero((caddr_t)dp, sizeof(struct ufs2_dinode)); 3959 dp->di_gen = inodedep->id_savedino2->di_gen; 3960 return; 3961 } 3962 /* 3963 * If no dependencies, then there is nothing to roll back. 3964 */ 3965 inodedep->id_savedsize = dp->di_size; 3966 inodedep->id_savedextsize = dp->di_extsize; 3967 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL && 3968 TAILQ_FIRST(&inodedep->id_extupdt) == NULL) 3969 return; 3970 /* 3971 * Set the ext data dependencies to busy. 3972 */ 3973 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 3974 adp = TAILQ_NEXT(adp, ad_next)) { 3975#ifdef DIAGNOSTIC 3976 if (deplist != 0 && prevlbn >= adp->ad_lbn) 3977 panic("softdep_write_inodeblock: lbn order"); 3978 prevlbn = adp->ad_lbn; 3979 if (dp->di_extb[adp->ad_lbn] != adp->ad_newblkno) 3980 panic("%s: direct pointer #%jd mismatch %jd != %jd", 3981 "softdep_write_inodeblock", 3982 (intmax_t)adp->ad_lbn, 3983 (intmax_t)dp->di_extb[adp->ad_lbn], 3984 (intmax_t)adp->ad_newblkno); 3985 deplist |= 1 << adp->ad_lbn; 3986 if ((adp->ad_state & ATTACHED) == 0) 3987 panic("softdep_write_inodeblock: Unknown state 0x%x", 3988 adp->ad_state); 3989#endif /* DIAGNOSTIC */ 3990 adp->ad_state &= ~ATTACHED; 3991 adp->ad_state |= UNDONE; 3992 } 3993 /* 3994 * The on-disk inode cannot claim to be any larger than the last 3995 * fragment that has been written. Otherwise, the on-disk inode 3996 * might have fragments that were not the last block in the ext 3997 * data which would corrupt the filesystem. 3998 */ 3999 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 4000 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 4001 dp->di_extb[adp->ad_lbn] = adp->ad_oldblkno; 4002 /* keep going until hitting a rollback to a frag */ 4003 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 4004 continue; 4005 dp->di_extsize = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 4006 for (i = adp->ad_lbn + 1; i < NXADDR; i++) { 4007#ifdef DIAGNOSTIC 4008 if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0) 4009 panic("softdep_write_inodeblock: lost dep1"); 4010#endif /* DIAGNOSTIC */ 4011 dp->di_extb[i] = 0; 4012 } 4013 lastadp = NULL; 4014 break; 4015 } 4016 /* 4017 * If we have zero'ed out the last allocated block of the ext 4018 * data, roll back the size to the last currently allocated block. 4019 * We know that this last allocated block is a full-sized as 4020 * we already checked for fragments in the loop above. 4021 */ 4022 if (lastadp != NULL && 4023 dp->di_extsize <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 4024 for (i = lastadp->ad_lbn; i >= 0; i--) 4025 if (dp->di_extb[i] != 0) 4026 break; 4027 dp->di_extsize = (i + 1) * fs->fs_bsize; 4028 } 4029 /* 4030 * Set the file data dependencies to busy. 4031 */ 4032 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 4033 adp = TAILQ_NEXT(adp, ad_next)) { 4034#ifdef DIAGNOSTIC 4035 if (deplist != 0 && prevlbn >= adp->ad_lbn) 4036 panic("softdep_write_inodeblock: lbn order"); 4037 prevlbn = adp->ad_lbn; 4038 if (adp->ad_lbn < NDADDR && 4039 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) 4040 panic("%s: direct pointer #%jd mismatch %jd != %jd", 4041 "softdep_write_inodeblock", 4042 (intmax_t)adp->ad_lbn, 4043 (intmax_t)dp->di_db[adp->ad_lbn], 4044 (intmax_t)adp->ad_newblkno); 4045 if (adp->ad_lbn >= NDADDR && 4046 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) 4047 panic("%s indirect pointer #%jd mismatch %jd != %jd", 4048 "softdep_write_inodeblock:", 4049 (intmax_t)adp->ad_lbn - NDADDR, 4050 (intmax_t)dp->di_ib[adp->ad_lbn - NDADDR], 4051 (intmax_t)adp->ad_newblkno); 4052 deplist |= 1 << adp->ad_lbn; 4053 if ((adp->ad_state & ATTACHED) == 0) 4054 panic("softdep_write_inodeblock: Unknown state 0x%x", 4055 adp->ad_state); 4056#endif /* DIAGNOSTIC */ 4057 adp->ad_state &= ~ATTACHED; 4058 adp->ad_state |= UNDONE; 4059 } 4060 /* 4061 * The on-disk inode cannot claim to be any larger than the last 4062 * fragment that has been written. Otherwise, the on-disk inode 4063 * might have fragments that were not the last block in the file 4064 * which would corrupt the filesystem. 4065 */ 4066 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 4067 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 4068 if (adp->ad_lbn >= NDADDR) 4069 break; 4070 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 4071 /* keep going until hitting a rollback to a frag */ 4072 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 4073 continue; 4074 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 4075 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 4076#ifdef DIAGNOSTIC 4077 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 4078 panic("softdep_write_inodeblock: lost dep2"); 4079#endif /* DIAGNOSTIC */ 4080 dp->di_db[i] = 0; 4081 } 4082 for (i = 0; i < NIADDR; i++) { 4083#ifdef DIAGNOSTIC 4084 if (dp->di_ib[i] != 0 && 4085 (deplist & ((1 << NDADDR) << i)) == 0) 4086 panic("softdep_write_inodeblock: lost dep3"); 4087#endif /* DIAGNOSTIC */ 4088 dp->di_ib[i] = 0; 4089 } 4090 return; 4091 } 4092 /* 4093 * If we have zero'ed out the last allocated block of the file, 4094 * roll back the size to the last currently allocated block. 4095 * We know that this last allocated block is a full-sized as 4096 * we already checked for fragments in the loop above. 4097 */ 4098 if (lastadp != NULL && 4099 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 4100 for (i = lastadp->ad_lbn; i >= 0; i--) 4101 if (dp->di_db[i] != 0) 4102 break; 4103 dp->di_size = (i + 1) * fs->fs_bsize; 4104 } 4105 /* 4106 * The only dependencies are for indirect blocks. 4107 * 4108 * The file size for indirect block additions is not guaranteed. 4109 * Such a guarantee would be non-trivial to achieve. The conventional 4110 * synchronous write implementation also does not make this guarantee. 4111 * Fsck should catch and fix discrepancies. Arguably, the file size 4112 * can be over-estimated without destroying integrity when the file 4113 * moves into the indirect blocks (i.e., is large). If we want to 4114 * postpone fsck, we are stuck with this argument. 4115 */ 4116 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 4117 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 4118} 4119 4120/* 4121 * This routine is called during the completion interrupt 4122 * service routine for a disk write (from the procedure called 4123 * by the device driver to inform the filesystem caches of 4124 * a request completion). It should be called early in this 4125 * procedure, before the block is made available to other 4126 * processes or other routines are called. 4127 */ 4128static void 4129softdep_disk_write_complete(bp) 4130 struct buf *bp; /* describes the completed disk write */ 4131{ 4132 struct worklist *wk; 4133 struct worklist *owk; 4134 struct workhead reattach; 4135 struct newblk *newblk; 4136 struct allocindir *aip; 4137 struct allocdirect *adp; 4138 struct indirdep *indirdep; 4139 struct inodedep *inodedep; 4140 struct bmsafemap *bmsafemap; 4141 4142 /* 4143 * If an error occurred while doing the write, then the data 4144 * has not hit the disk and the dependencies cannot be unrolled. 4145 */ 4146 if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0) 4147 return; 4148 LIST_INIT(&reattach); 4149 /* 4150 * This lock must not be released anywhere in this code segment. 4151 */ 4152 ACQUIRE_LOCK(&lk); 4153 owk = NULL; 4154 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 4155 WORKLIST_REMOVE(wk); 4156 if (wk == owk) 4157 panic("duplicate worklist: %p\n", wk); 4158 owk = wk; 4159 switch (wk->wk_type) { 4160 4161 case D_PAGEDEP: 4162 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 4163 WORKLIST_INSERT(&reattach, wk); 4164 continue; 4165 4166 case D_INODEDEP: 4167 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 4168 WORKLIST_INSERT(&reattach, wk); 4169 continue; 4170 4171 case D_BMSAFEMAP: 4172 bmsafemap = WK_BMSAFEMAP(wk); 4173 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) { 4174 newblk->nb_state |= DEPCOMPLETE; 4175 newblk->nb_bmsafemap = NULL; 4176 LIST_REMOVE(newblk, nb_deps); 4177 } 4178 while ((adp = 4179 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) { 4180 adp->ad_state |= DEPCOMPLETE; 4181 adp->ad_buf = NULL; 4182 LIST_REMOVE(adp, ad_deps); 4183 handle_allocdirect_partdone(adp); 4184 } 4185 while ((aip = 4186 LIST_FIRST(&bmsafemap->sm_allocindirhd))) { 4187 aip->ai_state |= DEPCOMPLETE; 4188 aip->ai_buf = NULL; 4189 LIST_REMOVE(aip, ai_deps); 4190 handle_allocindir_partdone(aip); 4191 } 4192 while ((inodedep = 4193 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { 4194 inodedep->id_state |= DEPCOMPLETE; 4195 LIST_REMOVE(inodedep, id_deps); 4196 inodedep->id_buf = NULL; 4197 } 4198 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 4199 continue; 4200 4201 case D_MKDIR: 4202 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 4203 continue; 4204 4205 case D_ALLOCDIRECT: 4206 adp = WK_ALLOCDIRECT(wk); 4207 adp->ad_state |= COMPLETE; 4208 handle_allocdirect_partdone(adp); 4209 continue; 4210 4211 case D_ALLOCINDIR: 4212 aip = WK_ALLOCINDIR(wk); 4213 aip->ai_state |= COMPLETE; 4214 handle_allocindir_partdone(aip); 4215 continue; 4216 4217 case D_INDIRDEP: 4218 indirdep = WK_INDIRDEP(wk); 4219 if (indirdep->ir_state & GOINGAWAY) 4220 panic("disk_write_complete: indirdep gone"); 4221 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 4222 FREE(indirdep->ir_saveddata, M_INDIRDEP); 4223 indirdep->ir_saveddata = 0; 4224 indirdep->ir_state &= ~UNDONE; 4225 indirdep->ir_state |= ATTACHED; 4226 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 4227 handle_allocindir_partdone(aip); 4228 if (aip == LIST_FIRST(&indirdep->ir_donehd)) 4229 panic("disk_write_complete: not gone"); 4230 } 4231 WORKLIST_INSERT(&reattach, wk); 4232 if ((bp->b_flags & B_DELWRI) == 0) 4233 stat_indir_blk_ptrs++; 4234 bdirty(bp); 4235 continue; 4236 4237 default: 4238 panic("handle_disk_write_complete: Unknown type %s", 4239 TYPENAME(wk->wk_type)); 4240 /* NOTREACHED */ 4241 } 4242 } 4243 /* 4244 * Reattach any requests that must be redone. 4245 */ 4246 while ((wk = LIST_FIRST(&reattach)) != NULL) { 4247 WORKLIST_REMOVE(wk); 4248 WORKLIST_INSERT(&bp->b_dep, wk); 4249 } 4250 FREE_LOCK(&lk); 4251} 4252 4253/* 4254 * Called from within softdep_disk_write_complete above. Note that 4255 * this routine is always called from interrupt level with further 4256 * splbio interrupts blocked. 4257 */ 4258static void 4259handle_allocdirect_partdone(adp) 4260 struct allocdirect *adp; /* the completed allocdirect */ 4261{ 4262 struct allocdirectlst *listhead; 4263 struct allocdirect *listadp; 4264 struct inodedep *inodedep; 4265 long bsize, delay; 4266 4267 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 4268 return; 4269 if (adp->ad_buf != NULL) 4270 panic("handle_allocdirect_partdone: dangling dep"); 4271 /* 4272 * The on-disk inode cannot claim to be any larger than the last 4273 * fragment that has been written. Otherwise, the on-disk inode 4274 * might have fragments that were not the last block in the file 4275 * which would corrupt the filesystem. Thus, we cannot free any 4276 * allocdirects after one whose ad_oldblkno claims a fragment as 4277 * these blocks must be rolled back to zero before writing the inode. 4278 * We check the currently active set of allocdirects in id_inoupdt 4279 * or id_extupdt as appropriate. 4280 */ 4281 inodedep = adp->ad_inodedep; 4282 bsize = inodedep->id_fs->fs_bsize; 4283 if (adp->ad_state & EXTDATA) 4284 listhead = &inodedep->id_extupdt; 4285 else 4286 listhead = &inodedep->id_inoupdt; 4287 TAILQ_FOREACH(listadp, listhead, ad_next) { 4288 /* found our block */ 4289 if (listadp == adp) 4290 break; 4291 /* continue if ad_oldlbn is not a fragment */ 4292 if (listadp->ad_oldsize == 0 || 4293 listadp->ad_oldsize == bsize) 4294 continue; 4295 /* hit a fragment */ 4296 return; 4297 } 4298 /* 4299 * If we have reached the end of the current list without 4300 * finding the just finished dependency, then it must be 4301 * on the future dependency list. Future dependencies cannot 4302 * be freed until they are moved to the current list. 4303 */ 4304 if (listadp == NULL) { 4305#ifdef DEBUG 4306 if (adp->ad_state & EXTDATA) 4307 listhead = &inodedep->id_newextupdt; 4308 else 4309 listhead = &inodedep->id_newinoupdt; 4310 TAILQ_FOREACH(listadp, listhead, ad_next) 4311 /* found our block */ 4312 if (listadp == adp) 4313 break; 4314 if (listadp == NULL) 4315 panic("handle_allocdirect_partdone: lost dep"); 4316#endif /* DEBUG */ 4317 return; 4318 } 4319 /* 4320 * If we have found the just finished dependency, then free 4321 * it along with anything that follows it that is complete. 4322 * If the inode still has a bitmap dependency, then it has 4323 * never been written to disk, hence the on-disk inode cannot 4324 * reference the old fragment so we can free it without delay. 4325 */ 4326 delay = (inodedep->id_state & DEPCOMPLETE); 4327 for (; adp; adp = listadp) { 4328 listadp = TAILQ_NEXT(adp, ad_next); 4329 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 4330 return; 4331 free_allocdirect(listhead, adp, delay); 4332 } 4333} 4334 4335/* 4336 * Called from within softdep_disk_write_complete above. Note that 4337 * this routine is always called from interrupt level with further 4338 * splbio interrupts blocked. 4339 */ 4340static void 4341handle_allocindir_partdone(aip) 4342 struct allocindir *aip; /* the completed allocindir */ 4343{ 4344 struct indirdep *indirdep; 4345 4346 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 4347 return; 4348 if (aip->ai_buf != NULL) 4349 panic("handle_allocindir_partdone: dangling dependency"); 4350 indirdep = aip->ai_indirdep; 4351 if (indirdep->ir_state & UNDONE) { 4352 LIST_REMOVE(aip, ai_next); 4353 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 4354 return; 4355 } 4356 if (indirdep->ir_state & UFS1FMT) 4357 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 4358 aip->ai_newblkno; 4359 else 4360 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 4361 aip->ai_newblkno; 4362 LIST_REMOVE(aip, ai_next); 4363 if (aip->ai_freefrag != NULL) 4364 add_to_worklist(&aip->ai_freefrag->ff_list); 4365 WORKITEM_FREE(aip, D_ALLOCINDIR); 4366} 4367 4368/* 4369 * Called from within softdep_disk_write_complete above to restore 4370 * in-memory inode block contents to their most up-to-date state. Note 4371 * that this routine is always called from interrupt level with further 4372 * splbio interrupts blocked. 4373 */ 4374static int 4375handle_written_inodeblock(inodedep, bp) 4376 struct inodedep *inodedep; 4377 struct buf *bp; /* buffer containing the inode block */ 4378{ 4379 struct worklist *wk, *filefree; 4380 struct allocdirect *adp, *nextadp; 4381 struct ufs1_dinode *dp1 = NULL; 4382 struct ufs2_dinode *dp2 = NULL; 4383 int hadchanges, fstype; 4384 4385 if ((inodedep->id_state & IOSTARTED) == 0) 4386 panic("handle_written_inodeblock: not started"); 4387 inodedep->id_state &= ~IOSTARTED; 4388 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) { 4389 fstype = UFS1; 4390 dp1 = (struct ufs1_dinode *)bp->b_data + 4391 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 4392 } else { 4393 fstype = UFS2; 4394 dp2 = (struct ufs2_dinode *)bp->b_data + 4395 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 4396 } 4397 /* 4398 * If we had to rollback the inode allocation because of 4399 * bitmaps being incomplete, then simply restore it. 4400 * Keep the block dirty so that it will not be reclaimed until 4401 * all associated dependencies have been cleared and the 4402 * corresponding updates written to disk. 4403 */ 4404 if (inodedep->id_savedino1 != NULL) { 4405 if (fstype == UFS1) 4406 *dp1 = *inodedep->id_savedino1; 4407 else 4408 *dp2 = *inodedep->id_savedino2; 4409 FREE(inodedep->id_savedino1, M_SAVEDINO); 4410 inodedep->id_savedino1 = NULL; 4411 if ((bp->b_flags & B_DELWRI) == 0) 4412 stat_inode_bitmap++; 4413 bdirty(bp); 4414 return (1); 4415 } 4416 inodedep->id_state |= COMPLETE; 4417 /* 4418 * Roll forward anything that had to be rolled back before 4419 * the inode could be updated. 4420 */ 4421 hadchanges = 0; 4422 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 4423 nextadp = TAILQ_NEXT(adp, ad_next); 4424 if (adp->ad_state & ATTACHED) 4425 panic("handle_written_inodeblock: new entry"); 4426 if (fstype == UFS1) { 4427 if (adp->ad_lbn < NDADDR) { 4428 if (dp1->di_db[adp->ad_lbn]!=adp->ad_oldblkno) 4429 panic("%s %s #%jd mismatch %d != %jd", 4430 "handle_written_inodeblock:", 4431 "direct pointer", 4432 (intmax_t)adp->ad_lbn, 4433 dp1->di_db[adp->ad_lbn], 4434 (intmax_t)adp->ad_oldblkno); 4435 dp1->di_db[adp->ad_lbn] = adp->ad_newblkno; 4436 } else { 4437 if (dp1->di_ib[adp->ad_lbn - NDADDR] != 0) 4438 panic("%s: %s #%jd allocated as %d", 4439 "handle_written_inodeblock", 4440 "indirect pointer", 4441 (intmax_t)adp->ad_lbn - NDADDR, 4442 dp1->di_ib[adp->ad_lbn - NDADDR]); 4443 dp1->di_ib[adp->ad_lbn - NDADDR] = 4444 adp->ad_newblkno; 4445 } 4446 } else { 4447 if (adp->ad_lbn < NDADDR) { 4448 if (dp2->di_db[adp->ad_lbn]!=adp->ad_oldblkno) 4449 panic("%s: %s #%jd %s %jd != %jd", 4450 "handle_written_inodeblock", 4451 "direct pointer", 4452 (intmax_t)adp->ad_lbn, "mismatch", 4453 (intmax_t)dp2->di_db[adp->ad_lbn], 4454 (intmax_t)adp->ad_oldblkno); 4455 dp2->di_db[adp->ad_lbn] = adp->ad_newblkno; 4456 } else { 4457 if (dp2->di_ib[adp->ad_lbn - NDADDR] != 0) 4458 panic("%s: %s #%jd allocated as %jd", 4459 "handle_written_inodeblock", 4460 "indirect pointer", 4461 (intmax_t)adp->ad_lbn - NDADDR, 4462 (intmax_t) 4463 dp2->di_ib[adp->ad_lbn - NDADDR]); 4464 dp2->di_ib[adp->ad_lbn - NDADDR] = 4465 adp->ad_newblkno; 4466 } 4467 } 4468 adp->ad_state &= ~UNDONE; 4469 adp->ad_state |= ATTACHED; 4470 hadchanges = 1; 4471 } 4472 for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) { 4473 nextadp = TAILQ_NEXT(adp, ad_next); 4474 if (adp->ad_state & ATTACHED) 4475 panic("handle_written_inodeblock: new entry"); 4476 if (dp2->di_extb[adp->ad_lbn] != adp->ad_oldblkno) 4477 panic("%s: direct pointers #%jd %s %jd != %jd", 4478 "handle_written_inodeblock", 4479 (intmax_t)adp->ad_lbn, "mismatch", 4480 (intmax_t)dp2->di_extb[adp->ad_lbn], 4481 (intmax_t)adp->ad_oldblkno); 4482 dp2->di_extb[adp->ad_lbn] = adp->ad_newblkno; 4483 adp->ad_state &= ~UNDONE; 4484 adp->ad_state |= ATTACHED; 4485 hadchanges = 1; 4486 } 4487 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 4488 stat_direct_blk_ptrs++; 4489 /* 4490 * Reset the file size to its most up-to-date value. 4491 */ 4492 if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1) 4493 panic("handle_written_inodeblock: bad size"); 4494 if (fstype == UFS1) { 4495 if (dp1->di_size != inodedep->id_savedsize) { 4496 dp1->di_size = inodedep->id_savedsize; 4497 hadchanges = 1; 4498 } 4499 } else { 4500 if (dp2->di_size != inodedep->id_savedsize) { 4501 dp2->di_size = inodedep->id_savedsize; 4502 hadchanges = 1; 4503 } 4504 if (dp2->di_extsize != inodedep->id_savedextsize) { 4505 dp2->di_extsize = inodedep->id_savedextsize; 4506 hadchanges = 1; 4507 } 4508 } 4509 inodedep->id_savedsize = -1; 4510 inodedep->id_savedextsize = -1; 4511 /* 4512 * If there were any rollbacks in the inode block, then it must be 4513 * marked dirty so that its will eventually get written back in 4514 * its correct form. 4515 */ 4516 if (hadchanges) 4517 bdirty(bp); 4518 /* 4519 * Process any allocdirects that completed during the update. 4520 */ 4521 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 4522 handle_allocdirect_partdone(adp); 4523 if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 4524 handle_allocdirect_partdone(adp); 4525 /* 4526 * Process deallocations that were held pending until the 4527 * inode had been written to disk. Freeing of the inode 4528 * is delayed until after all blocks have been freed to 4529 * avoid creation of new <vfsid, inum, lbn> triples 4530 * before the old ones have been deleted. 4531 */ 4532 filefree = NULL; 4533 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 4534 WORKLIST_REMOVE(wk); 4535 switch (wk->wk_type) { 4536 4537 case D_FREEFILE: 4538 /* 4539 * We defer adding filefree to the worklist until 4540 * all other additions have been made to ensure 4541 * that it will be done after all the old blocks 4542 * have been freed. 4543 */ 4544 if (filefree != NULL) 4545 panic("handle_written_inodeblock: filefree"); 4546 filefree = wk; 4547 continue; 4548 4549 case D_MKDIR: 4550 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 4551 continue; 4552 4553 case D_DIRADD: 4554 diradd_inode_written(WK_DIRADD(wk), inodedep); 4555 continue; 4556 4557 case D_FREEBLKS: 4558 wk->wk_state |= COMPLETE; 4559 if ((wk->wk_state & ALLCOMPLETE) != ALLCOMPLETE) 4560 continue; 4561 /* -- fall through -- */ 4562 case D_FREEFRAG: 4563 case D_DIRREM: 4564 add_to_worklist(wk); 4565 continue; 4566 4567 case D_NEWDIRBLK: 4568 free_newdirblk(WK_NEWDIRBLK(wk)); 4569 continue; 4570 4571 default: 4572 panic("handle_written_inodeblock: Unknown type %s", 4573 TYPENAME(wk->wk_type)); 4574 /* NOTREACHED */ 4575 } 4576 } 4577 if (filefree != NULL) { 4578 if (free_inodedep(inodedep) == 0) 4579 panic("handle_written_inodeblock: live inodedep"); 4580 add_to_worklist(filefree); 4581 return (0); 4582 } 4583 4584 /* 4585 * If no outstanding dependencies, free it. 4586 */ 4587 if (free_inodedep(inodedep) || 4588 (TAILQ_FIRST(&inodedep->id_inoupdt) == 0 && 4589 TAILQ_FIRST(&inodedep->id_extupdt) == 0)) 4590 return (0); 4591 return (hadchanges); 4592} 4593 4594/* 4595 * Process a diradd entry after its dependent inode has been written. 4596 * This routine must be called with splbio interrupts blocked. 4597 */ 4598static void 4599diradd_inode_written(dap, inodedep) 4600 struct diradd *dap; 4601 struct inodedep *inodedep; 4602{ 4603 struct pagedep *pagedep; 4604 4605 dap->da_state |= COMPLETE; 4606 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 4607 if (dap->da_state & DIRCHG) 4608 pagedep = dap->da_previous->dm_pagedep; 4609 else 4610 pagedep = dap->da_pagedep; 4611 LIST_REMOVE(dap, da_pdlist); 4612 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 4613 } 4614 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 4615} 4616 4617/* 4618 * Handle the completion of a mkdir dependency. 4619 */ 4620static void 4621handle_written_mkdir(mkdir, type) 4622 struct mkdir *mkdir; 4623 int type; 4624{ 4625 struct diradd *dap; 4626 struct pagedep *pagedep; 4627 4628 if (mkdir->md_state != type) 4629 panic("handle_written_mkdir: bad type"); 4630 dap = mkdir->md_diradd; 4631 dap->da_state &= ~type; 4632 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 4633 dap->da_state |= DEPCOMPLETE; 4634 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 4635 if (dap->da_state & DIRCHG) 4636 pagedep = dap->da_previous->dm_pagedep; 4637 else 4638 pagedep = dap->da_pagedep; 4639 LIST_REMOVE(dap, da_pdlist); 4640 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 4641 } 4642 LIST_REMOVE(mkdir, md_mkdirs); 4643 WORKITEM_FREE(mkdir, D_MKDIR); 4644} 4645 4646/* 4647 * Called from within softdep_disk_write_complete above. 4648 * A write operation was just completed. Removed inodes can 4649 * now be freed and associated block pointers may be committed. 4650 * Note that this routine is always called from interrupt level 4651 * with further splbio interrupts blocked. 4652 */ 4653static int 4654handle_written_filepage(pagedep, bp) 4655 struct pagedep *pagedep; 4656 struct buf *bp; /* buffer containing the written page */ 4657{ 4658 struct dirrem *dirrem; 4659 struct diradd *dap, *nextdap; 4660 struct direct *ep; 4661 int i, chgs; 4662 4663 if ((pagedep->pd_state & IOSTARTED) == 0) 4664 panic("handle_written_filepage: not started"); 4665 pagedep->pd_state &= ~IOSTARTED; 4666 /* 4667 * Process any directory removals that have been committed. 4668 */ 4669 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 4670 LIST_REMOVE(dirrem, dm_next); 4671 dirrem->dm_dirinum = pagedep->pd_ino; 4672 add_to_worklist(&dirrem->dm_list); 4673 } 4674 /* 4675 * Free any directory additions that have been committed. 4676 * If it is a newly allocated block, we have to wait until 4677 * the on-disk directory inode claims the new block. 4678 */ 4679 if ((pagedep->pd_state & NEWBLOCK) == 0) 4680 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 4681 free_diradd(dap); 4682 /* 4683 * Uncommitted directory entries must be restored. 4684 */ 4685 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 4686 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 4687 dap = nextdap) { 4688 nextdap = LIST_NEXT(dap, da_pdlist); 4689 if (dap->da_state & ATTACHED) 4690 panic("handle_written_filepage: attached"); 4691 ep = (struct direct *) 4692 ((char *)bp->b_data + dap->da_offset); 4693 ep->d_ino = dap->da_newinum; 4694 dap->da_state &= ~UNDONE; 4695 dap->da_state |= ATTACHED; 4696 chgs = 1; 4697 /* 4698 * If the inode referenced by the directory has 4699 * been written out, then the dependency can be 4700 * moved to the pending list. 4701 */ 4702 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 4703 LIST_REMOVE(dap, da_pdlist); 4704 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 4705 da_pdlist); 4706 } 4707 } 4708 } 4709 /* 4710 * If there were any rollbacks in the directory, then it must be 4711 * marked dirty so that its will eventually get written back in 4712 * its correct form. 4713 */ 4714 if (chgs) { 4715 if ((bp->b_flags & B_DELWRI) == 0) 4716 stat_dir_entry++; 4717 bdirty(bp); 4718 return (1); 4719 } 4720 /* 4721 * If we are not waiting for a new directory block to be 4722 * claimed by its inode, then the pagedep will be freed. 4723 * Otherwise it will remain to track any new entries on 4724 * the page in case they are fsync'ed. 4725 */ 4726 if ((pagedep->pd_state & NEWBLOCK) == 0) { 4727 LIST_REMOVE(pagedep, pd_hash); 4728 WORKITEM_FREE(pagedep, D_PAGEDEP); 4729 } 4730 return (0); 4731} 4732 4733/* 4734 * Writing back in-core inode structures. 4735 * 4736 * The filesystem only accesses an inode's contents when it occupies an 4737 * "in-core" inode structure. These "in-core" structures are separate from 4738 * the page frames used to cache inode blocks. Only the latter are 4739 * transferred to/from the disk. So, when the updated contents of the 4740 * "in-core" inode structure are copied to the corresponding in-memory inode 4741 * block, the dependencies are also transferred. The following procedure is 4742 * called when copying a dirty "in-core" inode to a cached inode block. 4743 */ 4744 4745/* 4746 * Called when an inode is loaded from disk. If the effective link count 4747 * differed from the actual link count when it was last flushed, then we 4748 * need to ensure that the correct effective link count is put back. 4749 */ 4750void 4751softdep_load_inodeblock(ip) 4752 struct inode *ip; /* the "in_core" copy of the inode */ 4753{ 4754 struct inodedep *inodedep; 4755 4756 /* 4757 * Check for alternate nlink count. 4758 */ 4759 ip->i_effnlink = ip->i_nlink; 4760 ACQUIRE_LOCK(&lk); 4761 if (inodedep_lookup(UFSTOVFS(ip->i_ump), 4762 ip->i_number, 0, &inodedep) == 0) { 4763 FREE_LOCK(&lk); 4764 return; 4765 } 4766 ip->i_effnlink -= inodedep->id_nlinkdelta; 4767 if (inodedep->id_state & SPACECOUNTED) 4768 ip->i_flag |= IN_SPACECOUNTED; 4769 FREE_LOCK(&lk); 4770} 4771 4772/* 4773 * This routine is called just before the "in-core" inode 4774 * information is to be copied to the in-memory inode block. 4775 * Recall that an inode block contains several inodes. If 4776 * the force flag is set, then the dependencies will be 4777 * cleared so that the update can always be made. Note that 4778 * the buffer is locked when this routine is called, so we 4779 * will never be in the middle of writing the inode block 4780 * to disk. 4781 */ 4782void 4783softdep_update_inodeblock(ip, bp, waitfor) 4784 struct inode *ip; /* the "in_core" copy of the inode */ 4785 struct buf *bp; /* the buffer containing the inode block */ 4786 int waitfor; /* nonzero => update must be allowed */ 4787{ 4788 struct inodedep *inodedep; 4789 struct worklist *wk; 4790 struct mount *mp; 4791 struct buf *ibp; 4792 int error; 4793 4794 /* 4795 * If the effective link count is not equal to the actual link 4796 * count, then we must track the difference in an inodedep while 4797 * the inode is (potentially) tossed out of the cache. Otherwise, 4798 * if there is no existing inodedep, then there are no dependencies 4799 * to track. 4800 */ 4801 mp = UFSTOVFS(ip->i_ump); 4802 ACQUIRE_LOCK(&lk); 4803 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 4804 FREE_LOCK(&lk); 4805 if (ip->i_effnlink != ip->i_nlink) 4806 panic("softdep_update_inodeblock: bad link count"); 4807 return; 4808 } 4809 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) 4810 panic("softdep_update_inodeblock: bad delta"); 4811 /* 4812 * Changes have been initiated. Anything depending on these 4813 * changes cannot occur until this inode has been written. 4814 */ 4815 inodedep->id_state &= ~COMPLETE; 4816 if ((inodedep->id_state & ONWORKLIST) == 0) 4817 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 4818 /* 4819 * Any new dependencies associated with the incore inode must 4820 * now be moved to the list associated with the buffer holding 4821 * the in-memory copy of the inode. Once merged process any 4822 * allocdirects that are completed by the merger. 4823 */ 4824 merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt); 4825 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL) 4826 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); 4827 merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt); 4828 if (TAILQ_FIRST(&inodedep->id_extupdt) != NULL) 4829 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt)); 4830 /* 4831 * Now that the inode has been pushed into the buffer, the 4832 * operations dependent on the inode being written to disk 4833 * can be moved to the id_bufwait so that they will be 4834 * processed when the buffer I/O completes. 4835 */ 4836 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 4837 WORKLIST_REMOVE(wk); 4838 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 4839 } 4840 /* 4841 * Newly allocated inodes cannot be written until the bitmap 4842 * that allocates them have been written (indicated by 4843 * DEPCOMPLETE being set in id_state). If we are doing a 4844 * forced sync (e.g., an fsync on a file), we force the bitmap 4845 * to be written so that the update can be done. 4846 */ 4847 if (waitfor == 0) { 4848 FREE_LOCK(&lk); 4849 return; 4850 } 4851retry: 4852 if ((inodedep->id_state & DEPCOMPLETE) != 0) { 4853 FREE_LOCK(&lk); 4854 return; 4855 } 4856 ibp = inodedep->id_buf; 4857 ibp = getdirtybuf(ibp, &lk, MNT_WAIT); 4858 if (ibp == NULL) { 4859 /* 4860 * If ibp came back as NULL, the dependency could have been 4861 * freed while we slept. Look it up again, and check to see 4862 * that it has completed. 4863 */ 4864 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 4865 goto retry; 4866 FREE_LOCK(&lk); 4867 return; 4868 } 4869 FREE_LOCK(&lk); 4870 if ((error = bwrite(ibp)) != 0) 4871 softdep_error("softdep_update_inodeblock: bwrite", error); 4872} 4873 4874/* 4875 * Merge the a new inode dependency list (such as id_newinoupdt) into an 4876 * old inode dependency list (such as id_inoupdt). This routine must be 4877 * called with splbio interrupts blocked. 4878 */ 4879static void 4880merge_inode_lists(newlisthead, oldlisthead) 4881 struct allocdirectlst *newlisthead; 4882 struct allocdirectlst *oldlisthead; 4883{ 4884 struct allocdirect *listadp, *newadp; 4885 4886 newadp = TAILQ_FIRST(newlisthead); 4887 for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) { 4888 if (listadp->ad_lbn < newadp->ad_lbn) { 4889 listadp = TAILQ_NEXT(listadp, ad_next); 4890 continue; 4891 } 4892 TAILQ_REMOVE(newlisthead, newadp, ad_next); 4893 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 4894 if (listadp->ad_lbn == newadp->ad_lbn) { 4895 allocdirect_merge(oldlisthead, newadp, 4896 listadp); 4897 listadp = newadp; 4898 } 4899 newadp = TAILQ_FIRST(newlisthead); 4900 } 4901 while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) { 4902 TAILQ_REMOVE(newlisthead, newadp, ad_next); 4903 TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next); 4904 } 4905} 4906 4907/* 4908 * If we are doing an fsync, then we must ensure that any directory 4909 * entries for the inode have been written after the inode gets to disk. 4910 */ 4911int 4912softdep_fsync(vp) 4913 struct vnode *vp; /* the "in_core" copy of the inode */ 4914{ 4915 struct inodedep *inodedep; 4916 struct pagedep *pagedep; 4917 struct worklist *wk; 4918 struct diradd *dap; 4919 struct mount *mp; 4920 struct vnode *pvp; 4921 struct inode *ip; 4922 struct buf *bp; 4923 struct fs *fs; 4924 struct thread *td = curthread; 4925 int error, flushparent; 4926 ino_t parentino; 4927 ufs_lbn_t lbn; 4928 4929 ip = VTOI(vp); 4930 fs = ip->i_fs; 4931 mp = vp->v_mount; 4932 ACQUIRE_LOCK(&lk); 4933 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 4934 FREE_LOCK(&lk); 4935 return (0); 4936 } 4937 if (LIST_FIRST(&inodedep->id_inowait) != NULL || 4938 LIST_FIRST(&inodedep->id_bufwait) != NULL || 4939 TAILQ_FIRST(&inodedep->id_extupdt) != NULL || 4940 TAILQ_FIRST(&inodedep->id_newextupdt) != NULL || 4941 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 4942 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) 4943 panic("softdep_fsync: pending ops"); 4944 for (error = 0, flushparent = 0; ; ) { 4945 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 4946 break; 4947 if (wk->wk_type != D_DIRADD) 4948 panic("softdep_fsync: Unexpected type %s", 4949 TYPENAME(wk->wk_type)); 4950 dap = WK_DIRADD(wk); 4951 /* 4952 * Flush our parent if this directory entry has a MKDIR_PARENT 4953 * dependency or is contained in a newly allocated block. 4954 */ 4955 if (dap->da_state & DIRCHG) 4956 pagedep = dap->da_previous->dm_pagedep; 4957 else 4958 pagedep = dap->da_pagedep; 4959 parentino = pagedep->pd_ino; 4960 lbn = pagedep->pd_lbn; 4961 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) 4962 panic("softdep_fsync: dirty"); 4963 if ((dap->da_state & MKDIR_PARENT) || 4964 (pagedep->pd_state & NEWBLOCK)) 4965 flushparent = 1; 4966 else 4967 flushparent = 0; 4968 /* 4969 * If we are being fsync'ed as part of vgone'ing this vnode, 4970 * then we will not be able to release and recover the 4971 * vnode below, so we just have to give up on writing its 4972 * directory entry out. It will eventually be written, just 4973 * not now, but then the user was not asking to have it 4974 * written, so we are not breaking any promises. 4975 */ 4976 if (vp->v_iflag & VI_DOOMED) 4977 break; 4978 /* 4979 * We prevent deadlock by always fetching inodes from the 4980 * root, moving down the directory tree. Thus, when fetching 4981 * our parent directory, we first try to get the lock. If 4982 * that fails, we must unlock ourselves before requesting 4983 * the lock on our parent. See the comment in ufs_lookup 4984 * for details on possible races. 4985 */ 4986 FREE_LOCK(&lk); 4987 if (ffs_vget(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp)) { 4988 VOP_UNLOCK(vp, 0, td); 4989 error = ffs_vget(mp, parentino, LK_EXCLUSIVE, &pvp); 4990 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 4991 if (error != 0) 4992 return (error); 4993 } 4994 /* 4995 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps 4996 * that are contained in direct blocks will be resolved by 4997 * doing a ffs_update. Pagedeps contained in indirect blocks 4998 * may require a complete sync'ing of the directory. So, we 4999 * try the cheap and fast ffs_update first, and if that fails, 5000 * then we do the slower ffs_syncvnode of the directory. 5001 */ 5002 if (flushparent) { 5003 if ((error = ffs_update(pvp, 1)) != 0) { 5004 vput(pvp); 5005 return (error); 5006 } 5007 if ((pagedep->pd_state & NEWBLOCK) && 5008 (error = ffs_syncvnode(pvp, MNT_WAIT))) { 5009 vput(pvp); 5010 return (error); 5011 } 5012 } 5013 /* 5014 * Flush directory page containing the inode's name. 5015 */ 5016 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred, 5017 &bp); 5018 if (error == 0) 5019 error = bwrite(bp); 5020 else 5021 brelse(bp); 5022 vput(pvp); 5023 if (error != 0) 5024 return (error); 5025 ACQUIRE_LOCK(&lk); 5026 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 5027 break; 5028 } 5029 FREE_LOCK(&lk); 5030 return (0); 5031} 5032 5033/* 5034 * Flush all the dirty bitmaps associated with the block device 5035 * before flushing the rest of the dirty blocks so as to reduce 5036 * the number of dependencies that will have to be rolled back. 5037 */ 5038void 5039softdep_fsync_mountdev(vp) 5040 struct vnode *vp; 5041{ 5042 struct buf *bp, *nbp; 5043 struct worklist *wk; 5044 5045 if (!vn_isdisk(vp, NULL)) 5046 panic("softdep_fsync_mountdev: vnode not a disk"); 5047restart: 5048 ACQUIRE_LOCK(&lk); 5049 VI_LOCK(vp); 5050 TAILQ_FOREACH_SAFE(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs, nbp) { 5051 /* 5052 * If it is already scheduled, skip to the next buffer. 5053 */ 5054 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 5055 continue; 5056 5057 if ((bp->b_flags & B_DELWRI) == 0) 5058 panic("softdep_fsync_mountdev: not dirty"); 5059 /* 5060 * We are only interested in bitmaps with outstanding 5061 * dependencies. 5062 */ 5063 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 5064 wk->wk_type != D_BMSAFEMAP || 5065 (bp->b_vflags & BV_BKGRDINPROG)) { 5066 BUF_UNLOCK(bp); 5067 continue; 5068 } 5069 VI_UNLOCK(vp); 5070 FREE_LOCK(&lk); 5071 bremfree(bp); 5072 (void) bawrite(bp); 5073 goto restart; 5074 } 5075 FREE_LOCK(&lk); 5076 drain_output(vp); 5077 VI_UNLOCK(vp); 5078} 5079 5080/* 5081 * This routine is called when we are trying to synchronously flush a 5082 * file. This routine must eliminate any filesystem metadata dependencies 5083 * so that the syncing routine can succeed by pushing the dirty blocks 5084 * associated with the file. If any I/O errors occur, they are returned. 5085 */ 5086int 5087softdep_sync_metadata(struct vnode *vp) 5088{ 5089 struct pagedep *pagedep; 5090 struct allocdirect *adp; 5091 struct allocindir *aip; 5092 struct buf *bp, *nbp; 5093 struct worklist *wk; 5094 int i, error, waitfor; 5095 5096 if (!DOINGSOFTDEP(vp)) 5097 return (0); 5098 /* 5099 * Ensure that any direct block dependencies have been cleared. 5100 */ 5101 ACQUIRE_LOCK(&lk); 5102 if ((error = flush_inodedep_deps(vp->v_mount, VTOI(vp)->i_number))) { 5103 FREE_LOCK(&lk); 5104 return (error); 5105 } 5106 FREE_LOCK(&lk); 5107 /* 5108 * For most files, the only metadata dependencies are the 5109 * cylinder group maps that allocate their inode or blocks. 5110 * The block allocation dependencies can be found by traversing 5111 * the dependency lists for any buffers that remain on their 5112 * dirty buffer list. The inode allocation dependency will 5113 * be resolved when the inode is updated with MNT_WAIT. 5114 * This work is done in two passes. The first pass grabs most 5115 * of the buffers and begins asynchronously writing them. The 5116 * only way to wait for these asynchronous writes is to sleep 5117 * on the filesystem vnode which may stay busy for a long time 5118 * if the filesystem is active. So, instead, we make a second 5119 * pass over the dependencies blocking on each write. In the 5120 * usual case we will be blocking against a write that we 5121 * initiated, so when it is done the dependency will have been 5122 * resolved. Thus the second pass is expected to end quickly. 5123 */ 5124 waitfor = MNT_NOWAIT; 5125 5126top: 5127 /* 5128 * We must wait for any I/O in progress to finish so that 5129 * all potential buffers on the dirty list will be visible. 5130 */ 5131 VI_LOCK(vp); 5132 drain_output(vp); 5133 while ((bp = TAILQ_FIRST(&vp->v_bufobj.bo_dirty.bv_hd)) != NULL) { 5134 bp = getdirtybuf(bp, VI_MTX(vp), MNT_WAIT); 5135 if (bp) 5136 break; 5137 } 5138 VI_UNLOCK(vp); 5139 if (bp == NULL) 5140 return (0); 5141loop: 5142 /* While syncing snapshots, we must allow recursive lookups */ 5143 bp->b_lock.lk_flags |= LK_CANRECURSE; 5144 ACQUIRE_LOCK(&lk); 5145 /* 5146 * As we hold the buffer locked, none of its dependencies 5147 * will disappear. 5148 */ 5149 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5150 switch (wk->wk_type) { 5151 5152 case D_ALLOCDIRECT: 5153 adp = WK_ALLOCDIRECT(wk); 5154 if (adp->ad_state & DEPCOMPLETE) 5155 continue; 5156 nbp = adp->ad_buf; 5157 nbp = getdirtybuf(nbp, &lk, waitfor); 5158 if (nbp == NULL) 5159 continue; 5160 FREE_LOCK(&lk); 5161 if (waitfor == MNT_NOWAIT) { 5162 bawrite(nbp); 5163 } else if ((error = bwrite(nbp)) != 0) { 5164 break; 5165 } 5166 ACQUIRE_LOCK(&lk); 5167 continue; 5168 5169 case D_ALLOCINDIR: 5170 aip = WK_ALLOCINDIR(wk); 5171 if (aip->ai_state & DEPCOMPLETE) 5172 continue; 5173 nbp = aip->ai_buf; 5174 nbp = getdirtybuf(nbp, &lk, waitfor); 5175 if (nbp == NULL) 5176 continue; 5177 FREE_LOCK(&lk); 5178 if (waitfor == MNT_NOWAIT) { 5179 bawrite(nbp); 5180 } else if ((error = bwrite(nbp)) != 0) { 5181 break; 5182 } 5183 ACQUIRE_LOCK(&lk); 5184 continue; 5185 5186 case D_INDIRDEP: 5187 restart: 5188 5189 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) { 5190 if (aip->ai_state & DEPCOMPLETE) 5191 continue; 5192 nbp = aip->ai_buf; 5193 nbp = getdirtybuf(nbp, &lk, MNT_WAIT); 5194 if (nbp == NULL) 5195 goto restart; 5196 FREE_LOCK(&lk); 5197 if ((error = bwrite(nbp)) != 0) { 5198 break; 5199 } 5200 ACQUIRE_LOCK(&lk); 5201 goto restart; 5202 } 5203 continue; 5204 5205 case D_INODEDEP: 5206 if ((error = flush_inodedep_deps(wk->wk_mp, 5207 WK_INODEDEP(wk)->id_ino)) != 0) { 5208 FREE_LOCK(&lk); 5209 break; 5210 } 5211 continue; 5212 5213 case D_PAGEDEP: 5214 /* 5215 * We are trying to sync a directory that may 5216 * have dependencies on both its own metadata 5217 * and/or dependencies on the inodes of any 5218 * recently allocated files. We walk its diradd 5219 * lists pushing out the associated inode. 5220 */ 5221 pagedep = WK_PAGEDEP(wk); 5222 for (i = 0; i < DAHASHSZ; i++) { 5223 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 5224 continue; 5225 if ((error = 5226 flush_pagedep_deps(vp, wk->wk_mp, 5227 &pagedep->pd_diraddhd[i]))) { 5228 FREE_LOCK(&lk); 5229 break; 5230 } 5231 } 5232 continue; 5233 5234 case D_MKDIR: 5235 /* 5236 * This case should never happen if the vnode has 5237 * been properly sync'ed. However, if this function 5238 * is used at a place where the vnode has not yet 5239 * been sync'ed, this dependency can show up. So, 5240 * rather than panic, just flush it. 5241 */ 5242 nbp = WK_MKDIR(wk)->md_buf; 5243 nbp = getdirtybuf(nbp, &lk, waitfor); 5244 if (nbp == NULL) 5245 continue; 5246 FREE_LOCK(&lk); 5247 if (waitfor == MNT_NOWAIT) { 5248 bawrite(nbp); 5249 } else if ((error = bwrite(nbp)) != 0) { 5250 break; 5251 } 5252 ACQUIRE_LOCK(&lk); 5253 continue; 5254 5255 case D_BMSAFEMAP: 5256 /* 5257 * This case should never happen if the vnode has 5258 * been properly sync'ed. However, if this function 5259 * is used at a place where the vnode has not yet 5260 * been sync'ed, this dependency can show up. So, 5261 * rather than panic, just flush it. 5262 */ 5263 nbp = WK_BMSAFEMAP(wk)->sm_buf; 5264 nbp = getdirtybuf(nbp, &lk, waitfor); 5265 if (nbp == NULL) 5266 continue; 5267 FREE_LOCK(&lk); 5268 if (waitfor == MNT_NOWAIT) { 5269 bawrite(nbp); 5270 } else if ((error = bwrite(nbp)) != 0) { 5271 break; 5272 } 5273 ACQUIRE_LOCK(&lk); 5274 continue; 5275 5276 default: 5277 panic("softdep_sync_metadata: Unknown type %s", 5278 TYPENAME(wk->wk_type)); 5279 /* NOTREACHED */ 5280 } 5281 /* We reach here only in error and unlocked */ 5282 if (error == 0) 5283 panic("softdep_sync_metadata: zero error"); 5284 bp->b_lock.lk_flags &= ~LK_CANRECURSE; 5285 bawrite(bp); 5286 return (error); 5287 } 5288 FREE_LOCK(&lk); 5289 VI_LOCK(vp); 5290 while ((nbp = TAILQ_NEXT(bp, b_bobufs)) != NULL) { 5291 nbp = getdirtybuf(nbp, VI_MTX(vp), MNT_WAIT); 5292 if (nbp) 5293 break; 5294 } 5295 VI_UNLOCK(vp); 5296 bp->b_lock.lk_flags &= ~LK_CANRECURSE; 5297 bawrite(bp); 5298 if (nbp != NULL) { 5299 bp = nbp; 5300 goto loop; 5301 } 5302 /* 5303 * The brief unlock is to allow any pent up dependency 5304 * processing to be done. Then proceed with the second pass. 5305 */ 5306 if (waitfor == MNT_NOWAIT) { 5307 waitfor = MNT_WAIT; 5308 goto top; 5309 } 5310 5311 /* 5312 * If we have managed to get rid of all the dirty buffers, 5313 * then we are done. For certain directories and block 5314 * devices, we may need to do further work. 5315 * 5316 * We must wait for any I/O in progress to finish so that 5317 * all potential buffers on the dirty list will be visible. 5318 */ 5319 VI_LOCK(vp); 5320 drain_output(vp); 5321 VI_UNLOCK(vp); 5322 return (0); 5323} 5324 5325/* 5326 * Flush the dependencies associated with an inodedep. 5327 * Called with splbio blocked. 5328 */ 5329static int 5330flush_inodedep_deps(mp, ino) 5331 struct mount *mp; 5332 ino_t ino; 5333{ 5334 struct inodedep *inodedep; 5335 int error, waitfor; 5336 5337 /* 5338 * This work is done in two passes. The first pass grabs most 5339 * of the buffers and begins asynchronously writing them. The 5340 * only way to wait for these asynchronous writes is to sleep 5341 * on the filesystem vnode which may stay busy for a long time 5342 * if the filesystem is active. So, instead, we make a second 5343 * pass over the dependencies blocking on each write. In the 5344 * usual case we will be blocking against a write that we 5345 * initiated, so when it is done the dependency will have been 5346 * resolved. Thus the second pass is expected to end quickly. 5347 * We give a brief window at the top of the loop to allow 5348 * any pending I/O to complete. 5349 */ 5350 for (error = 0, waitfor = MNT_NOWAIT; ; ) { 5351 if (error) 5352 return (error); 5353 FREE_LOCK(&lk); 5354 ACQUIRE_LOCK(&lk); 5355 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 5356 return (0); 5357 if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) || 5358 flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) || 5359 flush_deplist(&inodedep->id_extupdt, waitfor, &error) || 5360 flush_deplist(&inodedep->id_newextupdt, waitfor, &error)) 5361 continue; 5362 /* 5363 * If pass2, we are done, otherwise do pass 2. 5364 */ 5365 if (waitfor == MNT_WAIT) 5366 break; 5367 waitfor = MNT_WAIT; 5368 } 5369 /* 5370 * Try freeing inodedep in case all dependencies have been removed. 5371 */ 5372 if (inodedep_lookup(mp, ino, 0, &inodedep) != 0) 5373 (void) free_inodedep(inodedep); 5374 return (0); 5375} 5376 5377/* 5378 * Flush an inode dependency list. 5379 * Called with splbio blocked. 5380 */ 5381static int 5382flush_deplist(listhead, waitfor, errorp) 5383 struct allocdirectlst *listhead; 5384 int waitfor; 5385 int *errorp; 5386{ 5387 struct allocdirect *adp; 5388 struct buf *bp; 5389 5390 mtx_assert(&lk, MA_OWNED); 5391 TAILQ_FOREACH(adp, listhead, ad_next) { 5392 if (adp->ad_state & DEPCOMPLETE) 5393 continue; 5394 bp = adp->ad_buf; 5395 bp = getdirtybuf(bp, &lk, waitfor); 5396 if (bp == NULL) { 5397 if (waitfor == MNT_NOWAIT) 5398 continue; 5399 return (1); 5400 } 5401 FREE_LOCK(&lk); 5402 if (waitfor == MNT_NOWAIT) { 5403 bawrite(bp); 5404 } else if ((*errorp = bwrite(bp)) != 0) { 5405 ACQUIRE_LOCK(&lk); 5406 return (1); 5407 } 5408 ACQUIRE_LOCK(&lk); 5409 return (1); 5410 } 5411 return (0); 5412} 5413 5414/* 5415 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 5416 * Called with splbio blocked. 5417 */ 5418static int 5419flush_pagedep_deps(pvp, mp, diraddhdp) 5420 struct vnode *pvp; 5421 struct mount *mp; 5422 struct diraddhd *diraddhdp; 5423{ 5424 struct inodedep *inodedep; 5425 struct ufsmount *ump; 5426 struct diradd *dap; 5427 struct vnode *vp; 5428 int error = 0; 5429 struct buf *bp; 5430 ino_t inum; 5431 5432 ump = VFSTOUFS(mp); 5433 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 5434 /* 5435 * Flush ourselves if this directory entry 5436 * has a MKDIR_PARENT dependency. 5437 */ 5438 if (dap->da_state & MKDIR_PARENT) { 5439 FREE_LOCK(&lk); 5440 if ((error = ffs_update(pvp, 1)) != 0) 5441 break; 5442 ACQUIRE_LOCK(&lk); 5443 /* 5444 * If that cleared dependencies, go on to next. 5445 */ 5446 if (dap != LIST_FIRST(diraddhdp)) 5447 continue; 5448 if (dap->da_state & MKDIR_PARENT) 5449 panic("flush_pagedep_deps: MKDIR_PARENT"); 5450 } 5451 /* 5452 * A newly allocated directory must have its "." and 5453 * ".." entries written out before its name can be 5454 * committed in its parent. We do not want or need 5455 * the full semantics of a synchronous ffs_syncvnode as 5456 * that may end up here again, once for each directory 5457 * level in the filesystem. Instead, we push the blocks 5458 * and wait for them to clear. We have to fsync twice 5459 * because the first call may choose to defer blocks 5460 * that still have dependencies, but deferral will 5461 * happen at most once. 5462 */ 5463 inum = dap->da_newinum; 5464 if (dap->da_state & MKDIR_BODY) { 5465 FREE_LOCK(&lk); 5466 if ((error = ffs_vget(mp, inum, LK_EXCLUSIVE, &vp))) 5467 break; 5468 if ((error=ffs_syncvnode(vp, MNT_NOWAIT)) || 5469 (error=ffs_syncvnode(vp, MNT_NOWAIT))) { 5470 vput(vp); 5471 break; 5472 } 5473 VI_LOCK(vp); 5474 drain_output(vp); 5475 VI_UNLOCK(vp); 5476 vput(vp); 5477 ACQUIRE_LOCK(&lk); 5478 /* 5479 * If that cleared dependencies, go on to next. 5480 */ 5481 if (dap != LIST_FIRST(diraddhdp)) 5482 continue; 5483 if (dap->da_state & MKDIR_BODY) 5484 panic("flush_pagedep_deps: MKDIR_BODY"); 5485 } 5486 /* 5487 * Flush the inode on which the directory entry depends. 5488 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 5489 * the only remaining dependency is that the updated inode 5490 * count must get pushed to disk. The inode has already 5491 * been pushed into its inode buffer (via VOP_UPDATE) at 5492 * the time of the reference count change. So we need only 5493 * locate that buffer, ensure that there will be no rollback 5494 * caused by a bitmap dependency, then write the inode buffer. 5495 */ 5496retry: 5497 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 5498 panic("flush_pagedep_deps: lost inode"); 5499 /* 5500 * If the inode still has bitmap dependencies, 5501 * push them to disk. 5502 */ 5503 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 5504 bp = inodedep->id_buf; 5505 bp = getdirtybuf(bp, &lk, MNT_WAIT); 5506 if (bp == NULL) 5507 goto retry; 5508 FREE_LOCK(&lk); 5509 if ((error = bwrite(bp)) != 0) 5510 break; 5511 ACQUIRE_LOCK(&lk); 5512 if (dap != LIST_FIRST(diraddhdp)) 5513 continue; 5514 } 5515 /* 5516 * If the inode is still sitting in a buffer waiting 5517 * to be written, push it to disk. 5518 */ 5519 FREE_LOCK(&lk); 5520 if ((error = bread(ump->um_devvp, 5521 fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), 5522 (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0) { 5523 brelse(bp); 5524 break; 5525 } 5526 if ((error = bwrite(bp)) != 0) 5527 break; 5528 ACQUIRE_LOCK(&lk); 5529 /* 5530 * If we have failed to get rid of all the dependencies 5531 * then something is seriously wrong. 5532 */ 5533 if (dap == LIST_FIRST(diraddhdp)) 5534 panic("flush_pagedep_deps: flush failed"); 5535 } 5536 if (error) 5537 ACQUIRE_LOCK(&lk); 5538 return (error); 5539} 5540 5541/* 5542 * A large burst of file addition or deletion activity can drive the 5543 * memory load excessively high. First attempt to slow things down 5544 * using the techniques below. If that fails, this routine requests 5545 * the offending operations to fall back to running synchronously 5546 * until the memory load returns to a reasonable level. 5547 */ 5548int 5549softdep_slowdown(vp) 5550 struct vnode *vp; 5551{ 5552 int max_softdeps_hard; 5553 5554 max_softdeps_hard = max_softdeps * 11 / 10; 5555 if (num_dirrem < max_softdeps_hard / 2 && 5556 num_inodedep < max_softdeps_hard && 5557 VFSTOUFS(vp->v_mount)->um_numindirdeps < maxindirdeps) 5558 return (0); 5559 if (VFSTOUFS(vp->v_mount)->um_numindirdeps >= maxindirdeps) 5560 softdep_speedup(); 5561 stat_sync_limit_hit += 1; 5562 return (1); 5563} 5564 5565/* 5566 * Called by the allocation routines when they are about to fail 5567 * in the hope that we can free up some disk space. 5568 * 5569 * First check to see if the work list has anything on it. If it has, 5570 * clean up entries until we successfully free some space. Because this 5571 * process holds inodes locked, we cannot handle any remove requests 5572 * that might block on a locked inode as that could lead to deadlock. 5573 * If the worklist yields no free space, encourage the syncer daemon 5574 * to help us. In no event will we try for longer than tickdelay seconds. 5575 */ 5576int 5577softdep_request_cleanup(fs, vp) 5578 struct fs *fs; 5579 struct vnode *vp; 5580{ 5581 struct ufsmount *ump; 5582 long starttime; 5583 ufs2_daddr_t needed; 5584 int error; 5585 5586 ump = VTOI(vp)->i_ump; 5587 mtx_assert(UFS_MTX(ump), MA_OWNED); 5588 needed = fs->fs_cstotal.cs_nbfree + fs->fs_contigsumsize; 5589 starttime = time_second + tickdelay; 5590 /* 5591 * If we are being called because of a process doing a 5592 * copy-on-write, then it is not safe to update the vnode 5593 * as we may recurse into the copy-on-write routine. 5594 */ 5595 if (!(curthread->td_pflags & TDP_COWINPROGRESS)) { 5596 UFS_UNLOCK(ump); 5597 error = ffs_update(vp, 1); 5598 UFS_LOCK(ump); 5599 if (error != 0) 5600 return (0); 5601 } 5602 while (fs->fs_pendingblocks > 0 && fs->fs_cstotal.cs_nbfree <= needed) { 5603 if (time_second > starttime) 5604 return (0); 5605 UFS_UNLOCK(ump); 5606 ACQUIRE_LOCK(&lk); 5607 if (ump->softdep_on_worklist > 0 && 5608 process_worklist_item(UFSTOVFS(ump), LK_NOWAIT) != -1) { 5609 stat_worklist_push += 1; 5610 FREE_LOCK(&lk); 5611 UFS_LOCK(ump); 5612 continue; 5613 } 5614 request_cleanup(UFSTOVFS(ump), FLUSH_REMOVE_WAIT); 5615 FREE_LOCK(&lk); 5616 UFS_LOCK(ump); 5617 } 5618 return (1); 5619} 5620 5621/* 5622 * If memory utilization has gotten too high, deliberately slow things 5623 * down and speed up the I/O processing. 5624 */ 5625extern struct thread *syncertd; 5626static int 5627request_cleanup(mp, resource) 5628 struct mount *mp; 5629 int resource; 5630{ 5631 struct thread *td = curthread; 5632 struct ufsmount *ump; 5633 5634 mtx_assert(&lk, MA_OWNED); 5635 /* 5636 * We never hold up the filesystem syncer or buf daemon. 5637 */ 5638 if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF)) 5639 return (0); 5640 ump = VFSTOUFS(mp); 5641 /* 5642 * First check to see if the work list has gotten backlogged. 5643 * If it has, co-opt this process to help clean up two entries. 5644 * Because this process may hold inodes locked, we cannot 5645 * handle any remove requests that might block on a locked 5646 * inode as that could lead to deadlock. We set TDP_SOFTDEP 5647 * to avoid recursively processing the worklist. 5648 */ 5649 if (ump->softdep_on_worklist > max_softdeps / 10) { 5650 td->td_pflags |= TDP_SOFTDEP; 5651 process_worklist_item(mp, LK_NOWAIT); 5652 process_worklist_item(mp, LK_NOWAIT); 5653 td->td_pflags &= ~TDP_SOFTDEP; 5654 stat_worklist_push += 2; 5655 return(1); 5656 } 5657 /* 5658 * Next, we attempt to speed up the syncer process. If that 5659 * is successful, then we allow the process to continue. 5660 */ 5661 if (softdep_speedup() && resource != FLUSH_REMOVE_WAIT) 5662 return(0); 5663 /* 5664 * If we are resource constrained on inode dependencies, try 5665 * flushing some dirty inodes. Otherwise, we are constrained 5666 * by file deletions, so try accelerating flushes of directories 5667 * with removal dependencies. We would like to do the cleanup 5668 * here, but we probably hold an inode locked at this point and 5669 * that might deadlock against one that we try to clean. So, 5670 * the best that we can do is request the syncer daemon to do 5671 * the cleanup for us. 5672 */ 5673 switch (resource) { 5674 5675 case FLUSH_INODES: 5676 stat_ino_limit_push += 1; 5677 req_clear_inodedeps += 1; 5678 stat_countp = &stat_ino_limit_hit; 5679 break; 5680 5681 case FLUSH_REMOVE: 5682 case FLUSH_REMOVE_WAIT: 5683 stat_blk_limit_push += 1; 5684 req_clear_remove += 1; 5685 stat_countp = &stat_blk_limit_hit; 5686 break; 5687 5688 default: 5689 panic("request_cleanup: unknown type"); 5690 } 5691 /* 5692 * Hopefully the syncer daemon will catch up and awaken us. 5693 * We wait at most tickdelay before proceeding in any case. 5694 */ 5695 proc_waiting += 1; 5696 if (handle.callout == NULL) 5697 handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2); 5698 msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0); 5699 proc_waiting -= 1; 5700 return (1); 5701} 5702 5703/* 5704 * Awaken processes pausing in request_cleanup and clear proc_waiting 5705 * to indicate that there is no longer a timer running. 5706 */ 5707static void 5708pause_timer(arg) 5709 void *arg; 5710{ 5711 5712 ACQUIRE_LOCK(&lk); 5713 *stat_countp += 1; 5714 wakeup_one(&proc_waiting); 5715 if (proc_waiting > 0) 5716 handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2); 5717 else 5718 handle.callout = NULL; 5719 FREE_LOCK(&lk); 5720} 5721 5722/* 5723 * Flush out a directory with at least one removal dependency in an effort to 5724 * reduce the number of dirrem, freefile, and freeblks dependency structures. 5725 */ 5726static void 5727clear_remove(td) 5728 struct thread *td; 5729{ 5730 struct pagedep_hashhead *pagedephd; 5731 struct pagedep *pagedep; 5732 static int next = 0; 5733 struct mount *mp; 5734 struct vnode *vp; 5735 int error, cnt; 5736 ino_t ino; 5737 5738 mtx_assert(&lk, MA_OWNED); 5739 5740 for (cnt = 0; cnt < pagedep_hash; cnt++) { 5741 pagedephd = &pagedep_hashtbl[next++]; 5742 if (next >= pagedep_hash) 5743 next = 0; 5744 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 5745 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL) 5746 continue; 5747 mp = pagedep->pd_list.wk_mp; 5748 ino = pagedep->pd_ino; 5749 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 5750 continue; 5751 FREE_LOCK(&lk); 5752 if ((error = ffs_vget(mp, ino, LK_EXCLUSIVE, &vp))) { 5753 softdep_error("clear_remove: vget", error); 5754 vn_finished_write(mp); 5755 ACQUIRE_LOCK(&lk); 5756 return; 5757 } 5758 if ((error = ffs_syncvnode(vp, MNT_NOWAIT))) 5759 softdep_error("clear_remove: fsync", error); 5760 VI_LOCK(vp); 5761 drain_output(vp); 5762 VI_UNLOCK(vp); 5763 vput(vp); 5764 vn_finished_write(mp); 5765 ACQUIRE_LOCK(&lk); 5766 return; 5767 } 5768 } 5769} 5770 5771/* 5772 * Clear out a block of dirty inodes in an effort to reduce 5773 * the number of inodedep dependency structures. 5774 */ 5775static void 5776clear_inodedeps(td) 5777 struct thread *td; 5778{ 5779 struct inodedep_hashhead *inodedephd; 5780 struct inodedep *inodedep; 5781 static int next = 0; 5782 struct mount *mp; 5783 struct vnode *vp; 5784 struct fs *fs; 5785 int error, cnt; 5786 ino_t firstino, lastino, ino; 5787 5788 mtx_assert(&lk, MA_OWNED); 5789 /* 5790 * Pick a random inode dependency to be cleared. 5791 * We will then gather up all the inodes in its block 5792 * that have dependencies and flush them out. 5793 */ 5794 for (cnt = 0; cnt < inodedep_hash; cnt++) { 5795 inodedephd = &inodedep_hashtbl[next++]; 5796 if (next >= inodedep_hash) 5797 next = 0; 5798 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 5799 break; 5800 } 5801 if (inodedep == NULL) 5802 return; 5803 fs = inodedep->id_fs; 5804 mp = inodedep->id_list.wk_mp; 5805 /* 5806 * Find the last inode in the block with dependencies. 5807 */ 5808 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 5809 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 5810 if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0) 5811 break; 5812 /* 5813 * Asynchronously push all but the last inode with dependencies. 5814 * Synchronously push the last inode with dependencies to ensure 5815 * that the inode block gets written to free up the inodedeps. 5816 */ 5817 for (ino = firstino; ino <= lastino; ino++) { 5818 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 5819 continue; 5820 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 5821 continue; 5822 FREE_LOCK(&lk); 5823 if ((error = ffs_vget(mp, ino, LK_EXCLUSIVE, &vp)) != 0) { 5824 softdep_error("clear_inodedeps: vget", error); 5825 vn_finished_write(mp); 5826 ACQUIRE_LOCK(&lk); 5827 return; 5828 } 5829 if (ino == lastino) { 5830 if ((error = ffs_syncvnode(vp, MNT_WAIT))) 5831 softdep_error("clear_inodedeps: fsync1", error); 5832 } else { 5833 if ((error = ffs_syncvnode(vp, MNT_NOWAIT))) 5834 softdep_error("clear_inodedeps: fsync2", error); 5835 VI_LOCK(vp); 5836 drain_output(vp); 5837 VI_UNLOCK(vp); 5838 } 5839 vput(vp); 5840 vn_finished_write(mp); 5841 ACQUIRE_LOCK(&lk); 5842 } 5843} 5844 5845/* 5846 * Function to determine if the buffer has outstanding dependencies 5847 * that will cause a roll-back if the buffer is written. If wantcount 5848 * is set, return number of dependencies, otherwise just yes or no. 5849 */ 5850static int 5851softdep_count_dependencies(bp, wantcount) 5852 struct buf *bp; 5853 int wantcount; 5854{ 5855 struct worklist *wk; 5856 struct inodedep *inodedep; 5857 struct indirdep *indirdep; 5858 struct allocindir *aip; 5859 struct pagedep *pagedep; 5860 struct diradd *dap; 5861 int i, retval; 5862 5863 retval = 0; 5864 ACQUIRE_LOCK(&lk); 5865 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5866 switch (wk->wk_type) { 5867 5868 case D_INODEDEP: 5869 inodedep = WK_INODEDEP(wk); 5870 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 5871 /* bitmap allocation dependency */ 5872 retval += 1; 5873 if (!wantcount) 5874 goto out; 5875 } 5876 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 5877 /* direct block pointer dependency */ 5878 retval += 1; 5879 if (!wantcount) 5880 goto out; 5881 } 5882 if (TAILQ_FIRST(&inodedep->id_extupdt)) { 5883 /* direct block pointer dependency */ 5884 retval += 1; 5885 if (!wantcount) 5886 goto out; 5887 } 5888 continue; 5889 5890 case D_INDIRDEP: 5891 indirdep = WK_INDIRDEP(wk); 5892 5893 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 5894 /* indirect block pointer dependency */ 5895 retval += 1; 5896 if (!wantcount) 5897 goto out; 5898 } 5899 continue; 5900 5901 case D_PAGEDEP: 5902 pagedep = WK_PAGEDEP(wk); 5903 for (i = 0; i < DAHASHSZ; i++) { 5904 5905 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 5906 /* directory entry dependency */ 5907 retval += 1; 5908 if (!wantcount) 5909 goto out; 5910 } 5911 } 5912 continue; 5913 5914 case D_BMSAFEMAP: 5915 case D_ALLOCDIRECT: 5916 case D_ALLOCINDIR: 5917 case D_MKDIR: 5918 /* never a dependency on these blocks */ 5919 continue; 5920 5921 default: 5922 panic("softdep_check_for_rollback: Unexpected type %s", 5923 TYPENAME(wk->wk_type)); 5924 /* NOTREACHED */ 5925 } 5926 } 5927out: 5928 FREE_LOCK(&lk); 5929 return retval; 5930} 5931 5932/* 5933 * Acquire exclusive access to a buffer. 5934 * Must be called with a locked mtx parameter. 5935 * Return acquired buffer or NULL on failure. 5936 */ 5937static struct buf * 5938getdirtybuf(bp, mtx, waitfor) 5939 struct buf *bp; 5940 struct mtx *mtx; 5941 int waitfor; 5942{ 5943 int error; 5944 5945 mtx_assert(mtx, MA_OWNED); 5946 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) { 5947 if (waitfor != MNT_WAIT) 5948 return (NULL); 5949 error = BUF_LOCK(bp, 5950 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, mtx); 5951 /* 5952 * Even if we sucessfully acquire bp here, we have dropped 5953 * mtx, which may violates our guarantee. 5954 */ 5955 if (error == 0) 5956 BUF_UNLOCK(bp); 5957 else if (error != ENOLCK) 5958 panic("getdirtybuf: inconsistent lock: %d", error); 5959 mtx_lock(mtx); 5960 return (NULL); 5961 } 5962 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 5963 if (mtx == &lk && waitfor == MNT_WAIT) { 5964 mtx_unlock(mtx); 5965 BO_LOCK(bp->b_bufobj); 5966 BUF_UNLOCK(bp); 5967 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 5968 bp->b_vflags |= BV_BKGRDWAIT; 5969 msleep(&bp->b_xflags, BO_MTX(bp->b_bufobj), 5970 PRIBIO | PDROP, "getbuf", 0); 5971 } else 5972 BO_UNLOCK(bp->b_bufobj); 5973 mtx_lock(mtx); 5974 return (NULL); 5975 } 5976 BUF_UNLOCK(bp); 5977 if (waitfor != MNT_WAIT) 5978 return (NULL); 5979 /* 5980 * The mtx argument must be bp->b_vp's mutex in 5981 * this case. 5982 */ 5983#ifdef DEBUG_VFS_LOCKS 5984 if (bp->b_vp->v_type != VCHR) 5985 ASSERT_VI_LOCKED(bp->b_vp, "getdirtybuf"); 5986#endif 5987 bp->b_vflags |= BV_BKGRDWAIT; 5988 msleep(&bp->b_xflags, mtx, PRIBIO, "getbuf", 0); 5989 return (NULL); 5990 } 5991 if ((bp->b_flags & B_DELWRI) == 0) { 5992 BUF_UNLOCK(bp); 5993 return (NULL); 5994 } 5995 bremfree(bp); 5996 return (bp); 5997} 5998 5999/* 6000 * Wait for pending output on a vnode to complete. 6001 * Must be called with vnode lock and interlock locked. 6002 * 6003 * XXX: Should just be a call to bufobj_wwait(). 6004 */ 6005static void 6006drain_output(vp) 6007 struct vnode *vp; 6008{ 6009 ASSERT_VOP_LOCKED(vp, "drain_output"); 6010 ASSERT_VI_LOCKED(vp, "drain_output"); 6011 6012 while (vp->v_bufobj.bo_numoutput) { 6013 vp->v_bufobj.bo_flag |= BO_WWAIT; 6014 msleep((caddr_t)&vp->v_bufobj.bo_numoutput, 6015 VI_MTX(vp), PRIBIO + 1, "drainvp", 0); 6016 } 6017} 6018 6019/* 6020 * Called whenever a buffer that is being invalidated or reallocated 6021 * contains dependencies. This should only happen if an I/O error has 6022 * occurred. The routine is called with the buffer locked. 6023 */ 6024static void 6025softdep_deallocate_dependencies(bp) 6026 struct buf *bp; 6027{ 6028 6029 if ((bp->b_ioflags & BIO_ERROR) == 0) 6030 panic("softdep_deallocate_dependencies: dangling deps"); 6031 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); 6032 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 6033} 6034 6035/* 6036 * Function to handle asynchronous write errors in the filesystem. 6037 */ 6038static void 6039softdep_error(func, error) 6040 char *func; 6041 int error; 6042{ 6043 6044 /* XXX should do something better! */ 6045 printf("%s: got error %d while accessing filesystem\n", func, error); 6046} 6047 6048#endif /* SOFTUPDATES */ 6049