ffs_softdep.c revision 162653
1258945Sroberto/*- 2258945Sroberto * Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved. 3258945Sroberto * 4258945Sroberto * The soft updates code is derived from the appendix of a University 5258945Sroberto * of Michigan technical report (Gregory R. Ganger and Yale N. Patt, 6258945Sroberto * "Soft Updates: A Solution to the Metadata Update Problem in File 7258945Sroberto * Systems", CSE-TR-254-95, August 1995). 8258945Sroberto * 9258945Sroberto * Further information about soft updates can be obtained from: 10258945Sroberto * 11258945Sroberto * Marshall Kirk McKusick http://www.mckusick.com/softdep/ 12258945Sroberto * 1614 Oxford Street mckusick@mckusick.com 13258945Sroberto * Berkeley, CA 94709-1608 +1-510-843-9542 14258945Sroberto * USA 15258945Sroberto * 16258945Sroberto * Redistribution and use in source and binary forms, with or without 17258945Sroberto * modification, are permitted provided that the following conditions 18258945Sroberto * are met: 19258945Sroberto * 20258945Sroberto * 1. Redistributions of source code must retain the above copyright 21258945Sroberto * notice, this list of conditions and the following disclaimer. 22258945Sroberto * 2. Redistributions in binary form must reproduce the above copyright 23258945Sroberto * notice, this list of conditions and the following disclaimer in the 24258945Sroberto * documentation and/or other materials provided with the distribution. 25258945Sroberto * 26258945Sroberto * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY 27258945Sroberto * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 28258945Sroberto * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 29258945Sroberto * DISCLAIMED. IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR 30258945Sroberto * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31258945Sroberto * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32258945Sroberto * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33258945Sroberto * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34258945Sroberto * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35258945Sroberto * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36258945Sroberto * SUCH DAMAGE. 37258945Sroberto * 38258945Sroberto * from: @(#)ffs_softdep.c 9.59 (McKusick) 6/21/00 39258945Sroberto */ 40258945Sroberto 41258945Sroberto#include <sys/cdefs.h> 42258945Sroberto__FBSDID("$FreeBSD: head/sys/ufs/ffs/ffs_softdep.c 162653 2006-09-26 04:20:09Z tegge $"); 43258945Sroberto 44258945Sroberto/* 45258945Sroberto * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide. 46258945Sroberto */ 47258945Sroberto#ifndef DIAGNOSTIC 48258945Sroberto#define DIAGNOSTIC 49258945Sroberto#endif 50258945Sroberto#ifndef DEBUG 51258945Sroberto#define DEBUG 52258945Sroberto#endif 53258945Sroberto 54258945Sroberto#include <sys/param.h> 55258945Sroberto#include <sys/kernel.h> 56258945Sroberto#include <sys/systm.h> 57258945Sroberto#include <sys/bio.h> 58258945Sroberto#include <sys/buf.h> 59258945Sroberto#include <sys/kdb.h> 60258945Sroberto#include <sys/kthread.h> 61258945Sroberto#include <sys/lock.h> 62258945Sroberto#include <sys/malloc.h> 63258945Sroberto#include <sys/mount.h> 64258945Sroberto#include <sys/mutex.h> 65258945Sroberto#include <sys/proc.h> 66258945Sroberto#include <sys/stat.h> 67258945Sroberto#include <sys/sysctl.h> 68258945Sroberto#include <sys/syslog.h> 69258945Sroberto#include <sys/vnode.h> 70258945Sroberto#include <sys/conf.h> 71258945Sroberto#include <ufs/ufs/dir.h> 72258945Sroberto#include <ufs/ufs/extattr.h> 73258945Sroberto#include <ufs/ufs/quota.h> 74258945Sroberto#include <ufs/ufs/inode.h> 75258945Sroberto#include <ufs/ufs/ufsmount.h> 76258945Sroberto#include <ufs/ffs/fs.h> 77258945Sroberto#include <ufs/ffs/softdep.h> 78258945Sroberto#include <ufs/ffs/ffs_extern.h> 79258945Sroberto#include <ufs/ufs/ufs_extern.h> 80258945Sroberto 81258945Sroberto#include <vm/vm.h> 82258945Sroberto 83258945Sroberto#include "opt_ffs.h" 84258945Sroberto 85258945Sroberto#ifndef SOFTUPDATES 86258945Sroberto 87258945Srobertoint 88258945Srobertosoftdep_flushfiles(oldmnt, flags, td) 89258945Sroberto struct mount *oldmnt; 90258945Sroberto int flags; 91258945Sroberto struct thread *td; 92258945Sroberto{ 93258945Sroberto 94258945Sroberto panic("softdep_flushfiles called"); 95258945Sroberto} 96258945Sroberto 97258945Srobertoint 98258945Srobertosoftdep_mount(devvp, mp, fs, cred) 99258945Sroberto struct vnode *devvp; 100258945Sroberto struct mount *mp; 101258945Sroberto struct fs *fs; 102258945Sroberto struct ucred *cred; 103258945Sroberto{ 104258945Sroberto 105258945Sroberto return (0); 106258945Sroberto} 107258945Sroberto 108258945Srobertovoid 109258945Srobertosoftdep_initialize() 110258945Sroberto{ 111258945Sroberto 112258945Sroberto return; 113258945Sroberto} 114258945Sroberto 115258945Srobertovoid 116258945Srobertosoftdep_uninitialize() 117258945Sroberto{ 118258945Sroberto 119258945Sroberto return; 120258945Sroberto} 121258945Sroberto 122258945Srobertovoid 123258945Srobertosoftdep_setup_inomapdep(bp, ip, newinum) 124258945Sroberto struct buf *bp; 125258945Sroberto struct inode *ip; 126258945Sroberto ino_t newinum; 127258945Sroberto{ 128258945Sroberto 129258945Sroberto panic("softdep_setup_inomapdep called"); 130258945Sroberto} 131258945Sroberto 132258945Srobertovoid 133258945Srobertosoftdep_setup_blkmapdep(bp, mp, newblkno) 134258945Sroberto struct buf *bp; 135258945Sroberto struct mount *mp; 136258945Sroberto ufs2_daddr_t newblkno; 137258945Sroberto{ 138258945Sroberto 139258945Sroberto panic("softdep_setup_blkmapdep called"); 140258945Sroberto} 141258945Sroberto 142258945Srobertovoid 143258945Srobertosoftdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 144258945Sroberto struct inode *ip; 145258945Sroberto ufs_lbn_t lbn; 146258945Sroberto ufs2_daddr_t newblkno; 147258945Sroberto ufs2_daddr_t oldblkno; 148258945Sroberto long newsize; 149258945Sroberto long oldsize; 150258945Sroberto struct buf *bp; 151258945Sroberto{ 152258945Sroberto 153258945Sroberto panic("softdep_setup_allocdirect called"); 154258945Sroberto} 155258945Sroberto 156258945Srobertovoid 157258945Srobertosoftdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 158258945Sroberto struct inode *ip; 159258945Sroberto ufs_lbn_t lbn; 160258945Sroberto ufs2_daddr_t newblkno; 161258945Sroberto ufs2_daddr_t oldblkno; 162280849Scy long newsize; 163280849Scy long oldsize; 164280849Scy struct buf *bp; 165280849Scy{ 166258945Sroberto 167258945Sroberto panic("softdep_setup_allocext called"); 168258945Sroberto} 169258945Sroberto 170258945Srobertovoid 171258945Srobertosoftdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 172258945Sroberto struct inode *ip; 173258945Sroberto ufs_lbn_t lbn; 174258945Sroberto struct buf *bp; 175258945Sroberto int ptrno; 176258945Sroberto ufs2_daddr_t newblkno; 177258945Sroberto ufs2_daddr_t oldblkno; 178258945Sroberto struct buf *nbp; 179258945Sroberto{ 180258945Sroberto 181258945Sroberto panic("softdep_setup_allocindir_page called"); 182258945Sroberto} 183258945Sroberto 184258945Srobertovoid 185258945Srobertosoftdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 186258945Sroberto struct buf *nbp; 187258945Sroberto struct inode *ip; 188258945Sroberto struct buf *bp; 189258945Sroberto int ptrno; 190258945Sroberto ufs2_daddr_t newblkno; 191258945Sroberto{ 192258945Sroberto 193258945Sroberto panic("softdep_setup_allocindir_meta called"); 194258945Sroberto} 195258945Sroberto 196258945Srobertovoid 197258945Srobertosoftdep_setup_freeblocks(ip, length, flags) 198258945Sroberto struct inode *ip; 199258945Sroberto off_t length; 200258945Sroberto int flags; 201258945Sroberto{ 202258945Sroberto 203258945Sroberto panic("softdep_setup_freeblocks called"); 204258945Sroberto} 205258945Sroberto 206258945Srobertovoid 207258945Srobertosoftdep_freefile(pvp, ino, mode) 208258945Sroberto struct vnode *pvp; 209258945Sroberto ino_t ino; 210258945Sroberto int mode; 211258945Sroberto{ 212258945Sroberto 213258945Sroberto panic("softdep_freefile called"); 214258945Sroberto} 215258945Sroberto 216258945Srobertoint 217258945Srobertosoftdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 218258945Sroberto struct buf *bp; 219258945Sroberto struct inode *dp; 220258945Sroberto off_t diroffset; 221258945Sroberto ino_t newinum; 222258945Sroberto struct buf *newdirbp; 223258945Sroberto int isnewblk; 224258945Sroberto{ 225258945Sroberto 226258945Sroberto panic("softdep_setup_directory_add called"); 227258945Sroberto} 228258945Sroberto 229258945Srobertovoid 230258945Srobertosoftdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize) 231258945Sroberto struct inode *dp; 232258945Sroberto caddr_t base; 233258945Sroberto caddr_t oldloc; 234258945Sroberto caddr_t newloc; 235258945Sroberto int entrysize; 236258945Sroberto{ 237258945Sroberto 238258945Sroberto panic("softdep_change_directoryentry_offset called"); 239258945Sroberto} 240258945Sroberto 241258945Srobertovoid 242258945Srobertosoftdep_setup_remove(bp, dp, ip, isrmdir) 243258945Sroberto struct buf *bp; 244258945Sroberto struct inode *dp; 245258945Sroberto struct inode *ip; 246258945Sroberto int isrmdir; 247258945Sroberto{ 248258945Sroberto 249258945Sroberto panic("softdep_setup_remove called"); 250258945Sroberto} 251258945Sroberto 252258945Srobertovoid 253258945Srobertosoftdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 254258945Sroberto struct buf *bp; 255258945Sroberto struct inode *dp; 256258945Sroberto struct inode *ip; 257258945Sroberto ino_t newinum; 258258945Sroberto int isrmdir; 259258945Sroberto{ 260258945Sroberto 261258945Sroberto panic("softdep_setup_directory_change called"); 262258945Sroberto} 263258945Sroberto 264258945Srobertovoid 265258945Srobertosoftdep_change_linkcnt(ip) 266258945Sroberto struct inode *ip; 267258945Sroberto{ 268258945Sroberto 269258945Sroberto panic("softdep_change_linkcnt called"); 270258945Sroberto} 271258945Sroberto 272258945Srobertovoid 273258945Srobertosoftdep_load_inodeblock(ip) 274258945Sroberto struct inode *ip; 275258945Sroberto{ 276258945Sroberto 277258945Sroberto panic("softdep_load_inodeblock called"); 278258945Sroberto} 279280849Scy 280280849Scyvoid 281280849Scysoftdep_update_inodeblock(ip, bp, waitfor) 282280849Scy struct inode *ip; 283280849Scy struct buf *bp; 284258945Sroberto int waitfor; 285258945Sroberto{ 286258945Sroberto 287258945Sroberto panic("softdep_update_inodeblock called"); 288258945Sroberto} 289258945Sroberto 290258945Srobertoint 291258945Srobertosoftdep_fsync(vp) 292258945Sroberto struct vnode *vp; /* the "in_core" copy of the inode */ 293258945Sroberto{ 294258945Sroberto 295258945Sroberto return (0); 296258945Sroberto} 297258945Sroberto 298258945Srobertovoid 299258945Srobertosoftdep_fsync_mountdev(vp) 300258945Sroberto struct vnode *vp; 301258945Sroberto{ 302258945Sroberto 303258945Sroberto return; 304258945Sroberto} 305258945Sroberto 306258945Srobertoint 307258945Srobertosoftdep_flushworklist(oldmnt, countp, td) 308258945Sroberto struct mount *oldmnt; 309258945Sroberto int *countp; 310258945Sroberto struct thread *td; 311258945Sroberto{ 312258945Sroberto 313258945Sroberto *countp = 0; 314258945Sroberto return (0); 315258945Sroberto} 316258945Sroberto 317258945Srobertoint 318258945Srobertosoftdep_sync_metadata(struct vnode *vp) 319258945Sroberto{ 320258945Sroberto 321258945Sroberto 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 349int 350softdep_check_suspend(struct mount *mp, 351 struct vnode *devvp, 352 int softdep_deps, 353 int softdep_accdeps, 354 int secondary_writes, 355 int secondary_accwrites) 356{ 357 struct bufobj *bo; 358 int error; 359 360 (void) softdep_deps, 361 (void) softdep_accdeps; 362 363 ASSERT_VI_LOCKED(devvp, "softdep_check_suspend"); 364 bo = &devvp->v_bufobj; 365 366 for (;;) { 367 if (!MNT_ITRYLOCK(mp)) { 368 VI_UNLOCK(devvp); 369 MNT_ILOCK(mp); 370 MNT_IUNLOCK(mp); 371 VI_LOCK(devvp); 372 continue; 373 } 374 if (mp->mnt_secondary_writes != 0) { 375 VI_UNLOCK(devvp); 376 msleep(&mp->mnt_secondary_writes, 377 MNT_MTX(mp), 378 (PUSER - 1) | PDROP, "secwr", 0); 379 VI_LOCK(devvp); 380 continue; 381 } 382 break; 383 } 384 385 /* 386 * Reasons for needing more work before suspend: 387 * - Dirty buffers on devvp. 388 * - Secondary writes occurred after start of vnode sync loop 389 */ 390 error = 0; 391 if (bo->bo_numoutput > 0 || 392 bo->bo_dirty.bv_cnt > 0 || 393 secondary_writes != 0 || 394 mp->mnt_secondary_writes != 0 || 395 secondary_accwrites != mp->mnt_secondary_accwrites) 396 error = EAGAIN; 397 VI_UNLOCK(devvp); 398 return (error); 399} 400 401void 402softdep_get_depcounts(struct mount *mp, 403 int *softdepactivep, 404 int *softdepactiveaccp) 405{ 406 (void) mp; 407 *softdepactivep = 0; 408 *softdepactiveaccp = 0; 409} 410 411#else 412/* 413 * These definitions need to be adapted to the system to which 414 * this file is being ported. 415 */ 416/* 417 * malloc types defined for the softdep system. 418 */ 419static MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies"); 420static MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies"); 421static MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation"); 422static MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map"); 423static MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode"); 424static MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies"); 425static MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block"); 426static MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode"); 427static MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode"); 428static MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated"); 429static MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry"); 430static MALLOC_DEFINE(M_MKDIR, "mkdir","New directory"); 431static MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted"); 432static MALLOC_DEFINE(M_NEWDIRBLK, "newdirblk","Unclaimed new directory block"); 433static MALLOC_DEFINE(M_SAVEDINO, "savedino","Saved inodes"); 434 435#define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE) 436 437#define D_PAGEDEP 0 438#define D_INODEDEP 1 439#define D_NEWBLK 2 440#define D_BMSAFEMAP 3 441#define D_ALLOCDIRECT 4 442#define D_INDIRDEP 5 443#define D_ALLOCINDIR 6 444#define D_FREEFRAG 7 445#define D_FREEBLKS 8 446#define D_FREEFILE 9 447#define D_DIRADD 10 448#define D_MKDIR 11 449#define D_DIRREM 12 450#define D_NEWDIRBLK 13 451#define D_LAST D_NEWDIRBLK 452 453/* 454 * translate from workitem type to memory type 455 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 456 */ 457static struct malloc_type *memtype[] = { 458 M_PAGEDEP, 459 M_INODEDEP, 460 M_NEWBLK, 461 M_BMSAFEMAP, 462 M_ALLOCDIRECT, 463 M_INDIRDEP, 464 M_ALLOCINDIR, 465 M_FREEFRAG, 466 M_FREEBLKS, 467 M_FREEFILE, 468 M_DIRADD, 469 M_MKDIR, 470 M_DIRREM, 471 M_NEWDIRBLK 472}; 473 474#define DtoM(type) (memtype[type]) 475 476/* 477 * Names of malloc types. 478 */ 479#define TYPENAME(type) \ 480 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???") 481/* 482 * End system adaptation definitions. 483 */ 484 485/* 486 * Forward declarations. 487 */ 488struct inodedep_hashhead; 489struct newblk_hashhead; 490struct pagedep_hashhead; 491 492/* 493 * Internal function prototypes. 494 */ 495static void softdep_error(char *, int); 496static void drain_output(struct vnode *); 497static struct buf *getdirtybuf(struct buf *, struct mtx *, int); 498static void clear_remove(struct thread *); 499static void clear_inodedeps(struct thread *); 500static int flush_pagedep_deps(struct vnode *, struct mount *, 501 struct diraddhd *); 502static int flush_inodedep_deps(struct mount *, ino_t); 503static int flush_deplist(struct allocdirectlst *, int, int *); 504static int handle_written_filepage(struct pagedep *, struct buf *); 505static void diradd_inode_written(struct diradd *, struct inodedep *); 506static int handle_written_inodeblock(struct inodedep *, struct buf *); 507static void handle_allocdirect_partdone(struct allocdirect *); 508static void handle_allocindir_partdone(struct allocindir *); 509static void initiate_write_filepage(struct pagedep *, struct buf *); 510static void handle_written_mkdir(struct mkdir *, int); 511static void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *); 512static void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *); 513static void handle_workitem_freefile(struct freefile *); 514static void handle_workitem_remove(struct dirrem *, struct vnode *); 515static struct dirrem *newdirrem(struct buf *, struct inode *, 516 struct inode *, int, struct dirrem **); 517static void free_diradd(struct diradd *); 518static void free_allocindir(struct allocindir *, struct inodedep *); 519static void free_newdirblk(struct newdirblk *); 520static int indir_trunc(struct freeblks *, ufs2_daddr_t, int, ufs_lbn_t, 521 ufs2_daddr_t *); 522static void deallocate_dependencies(struct buf *, struct inodedep *); 523static void free_allocdirect(struct allocdirectlst *, 524 struct allocdirect *, int); 525static int check_inode_unwritten(struct inodedep *); 526static int free_inodedep(struct inodedep *); 527static void handle_workitem_freeblocks(struct freeblks *, int); 528static void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *); 529static void setup_allocindir_phase2(struct buf *, struct inode *, 530 struct allocindir *); 531static struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t, 532 ufs2_daddr_t); 533static void handle_workitem_freefrag(struct freefrag *); 534static struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long); 535static void allocdirect_merge(struct allocdirectlst *, 536 struct allocdirect *, struct allocdirect *); 537static struct bmsafemap *bmsafemap_lookup(struct mount *, struct buf *); 538static int newblk_find(struct newblk_hashhead *, struct fs *, ufs2_daddr_t, 539 struct newblk **); 540static int newblk_lookup(struct fs *, ufs2_daddr_t, int, struct newblk **); 541static int inodedep_find(struct inodedep_hashhead *, struct fs *, ino_t, 542 struct inodedep **); 543static int inodedep_lookup(struct mount *, ino_t, int, struct inodedep **); 544static int pagedep_lookup(struct inode *, ufs_lbn_t, int, struct pagedep **); 545static int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t, 546 struct mount *mp, int, struct pagedep **); 547static void pause_timer(void *); 548static int request_cleanup(struct mount *, int); 549static int process_worklist_item(struct mount *, int); 550static void add_to_worklist(struct worklist *); 551static void softdep_flush(void); 552static int softdep_speedup(void); 553 554/* 555 * Exported softdep operations. 556 */ 557static void softdep_disk_io_initiation(struct buf *); 558static void softdep_disk_write_complete(struct buf *); 559static void softdep_deallocate_dependencies(struct buf *); 560static int softdep_count_dependencies(struct buf *bp, int); 561 562static struct mtx lk; 563MTX_SYSINIT(softdep_lock, &lk, "Softdep Lock", MTX_DEF); 564 565#define TRY_ACQUIRE_LOCK(lk) mtx_trylock(lk) 566#define ACQUIRE_LOCK(lk) mtx_lock(lk) 567#define FREE_LOCK(lk) mtx_unlock(lk) 568 569/* 570 * Worklist queue management. 571 * These routines require that the lock be held. 572 */ 573#ifndef /* NOT */ DEBUG 574#define WORKLIST_INSERT(head, item) do { \ 575 (item)->wk_state |= ONWORKLIST; \ 576 LIST_INSERT_HEAD(head, item, wk_list); \ 577} while (0) 578#define WORKLIST_REMOVE(item) do { \ 579 (item)->wk_state &= ~ONWORKLIST; \ 580 LIST_REMOVE(item, wk_list); \ 581} while (0) 582#else /* DEBUG */ 583static void worklist_insert(struct workhead *, struct worklist *); 584static void worklist_remove(struct worklist *); 585 586#define WORKLIST_INSERT(head, item) worklist_insert(head, item) 587#define WORKLIST_REMOVE(item) worklist_remove(item) 588 589static void 590worklist_insert(head, item) 591 struct workhead *head; 592 struct worklist *item; 593{ 594 595 mtx_assert(&lk, MA_OWNED); 596 if (item->wk_state & ONWORKLIST) 597 panic("worklist_insert: already on list"); 598 item->wk_state |= ONWORKLIST; 599 LIST_INSERT_HEAD(head, item, wk_list); 600} 601 602static void 603worklist_remove(item) 604 struct worklist *item; 605{ 606 607 mtx_assert(&lk, MA_OWNED); 608 if ((item->wk_state & ONWORKLIST) == 0) 609 panic("worklist_remove: not on list"); 610 item->wk_state &= ~ONWORKLIST; 611 LIST_REMOVE(item, wk_list); 612} 613#endif /* DEBUG */ 614 615/* 616 * Routines for tracking and managing workitems. 617 */ 618static void workitem_free(struct worklist *, int); 619static void workitem_alloc(struct worklist *, int, struct mount *); 620 621#define WORKITEM_FREE(item, type) workitem_free((struct worklist *)(item), (type)) 622 623static void 624workitem_free(item, type) 625 struct worklist *item; 626 int type; 627{ 628 struct ufsmount *ump; 629 mtx_assert(&lk, MA_OWNED); 630 631#ifdef DEBUG 632 if (item->wk_state & ONWORKLIST) 633 panic("workitem_free: still on list"); 634 if (item->wk_type != type) 635 panic("workitem_free: type mismatch"); 636#endif 637 ump = VFSTOUFS(item->wk_mp); 638 if (--ump->softdep_deps == 0 && ump->softdep_req) 639 wakeup(&ump->softdep_deps); 640 FREE(item, DtoM(type)); 641} 642 643static void 644workitem_alloc(item, type, mp) 645 struct worklist *item; 646 int type; 647 struct mount *mp; 648{ 649 item->wk_type = type; 650 item->wk_mp = mp; 651 item->wk_state = 0; 652 ACQUIRE_LOCK(&lk); 653 VFSTOUFS(mp)->softdep_deps++; 654 VFSTOUFS(mp)->softdep_accdeps++; 655 FREE_LOCK(&lk); 656} 657 658/* 659 * Workitem queue management 660 */ 661static int max_softdeps; /* maximum number of structs before slowdown */ 662static int maxindirdeps = 50; /* max number of indirdeps before slowdown */ 663static int tickdelay = 2; /* number of ticks to pause during slowdown */ 664static int proc_waiting; /* tracks whether we have a timeout posted */ 665static int *stat_countp; /* statistic to count in proc_waiting timeout */ 666static struct callout_handle handle; /* handle on posted proc_waiting timeout */ 667static int req_pending; 668static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 669#define FLUSH_INODES 1 670static int req_clear_remove; /* syncer process flush some freeblks */ 671#define FLUSH_REMOVE 2 672#define FLUSH_REMOVE_WAIT 3 673/* 674 * runtime statistics 675 */ 676static int stat_worklist_push; /* number of worklist cleanups */ 677static int stat_blk_limit_push; /* number of times block limit neared */ 678static int stat_ino_limit_push; /* number of times inode limit neared */ 679static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 680static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 681static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 682static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 683static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 684static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 685static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 686 687SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, ""); 688SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, ""); 689SYSCTL_INT(_debug, OID_AUTO, maxindirdeps, CTLFLAG_RW, &maxindirdeps, 0, ""); 690SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,""); 691SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,""); 692SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,""); 693SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, ""); 694SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, ""); 695SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, ""); 696SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, ""); 697SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, ""); 698SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, ""); 699SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, ""); 700/* SYSCTL_INT(_debug, OID_AUTO, worklist_num, CTLFLAG_RD, &softdep_on_worklist, 0, ""); */ 701 702SYSCTL_DECL(_vfs_ffs); 703 704static int compute_summary_at_mount = 0; /* Whether to recompute the summary at mount time */ 705SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW, 706 &compute_summary_at_mount, 0, "Recompute summary at mount"); 707 708static struct proc *softdepproc; 709static struct kproc_desc softdep_kp = { 710 "softdepflush", 711 softdep_flush, 712 &softdepproc 713}; 714SYSINIT(sdproc, SI_SUB_KTHREAD_UPDATE, SI_ORDER_ANY, kproc_start, &softdep_kp) 715 716static void 717softdep_flush(void) 718{ 719 struct mount *nmp; 720 struct mount *mp; 721 struct ufsmount *ump; 722 struct thread *td; 723 int remaining; 724 int vfslocked; 725 726 td = curthread; 727 td->td_pflags |= TDP_NORUNNINGBUF; 728 729 for (;;) { 730 kthread_suspend_check(softdepproc); 731 ACQUIRE_LOCK(&lk); 732 /* 733 * If requested, try removing inode or removal dependencies. 734 */ 735 if (req_clear_inodedeps) { 736 clear_inodedeps(td); 737 req_clear_inodedeps -= 1; 738 wakeup_one(&proc_waiting); 739 } 740 if (req_clear_remove) { 741 clear_remove(td); 742 req_clear_remove -= 1; 743 wakeup_one(&proc_waiting); 744 } 745 FREE_LOCK(&lk); 746 remaining = 0; 747 mtx_lock(&mountlist_mtx); 748 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 749 nmp = TAILQ_NEXT(mp, mnt_list); 750 if ((mp->mnt_flag & MNT_SOFTDEP) == 0) 751 continue; 752 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) 753 continue; 754 vfslocked = VFS_LOCK_GIANT(mp); 755 softdep_process_worklist(mp, 0); 756 ump = VFSTOUFS(mp); 757 remaining += ump->softdep_on_worklist - 758 ump->softdep_on_worklist_inprogress; 759 VFS_UNLOCK_GIANT(vfslocked); 760 mtx_lock(&mountlist_mtx); 761 nmp = TAILQ_NEXT(mp, mnt_list); 762 vfs_unbusy(mp, td); 763 } 764 mtx_unlock(&mountlist_mtx); 765 if (remaining) 766 continue; 767 ACQUIRE_LOCK(&lk); 768 if (!req_pending) 769 msleep(&req_pending, &lk, PVM, "sdflush", hz); 770 req_pending = 0; 771 FREE_LOCK(&lk); 772 } 773} 774 775static int 776softdep_speedup(void) 777{ 778 779 mtx_assert(&lk, MA_OWNED); 780 if (req_pending == 0) { 781 req_pending = 1; 782 wakeup(&req_pending); 783 } 784 785 return speedup_syncer(); 786} 787 788/* 789 * Add an item to the end of the work queue. 790 * This routine requires that the lock be held. 791 * This is the only routine that adds items to the list. 792 * The following routine is the only one that removes items 793 * and does so in order from first to last. 794 */ 795static void 796add_to_worklist(wk) 797 struct worklist *wk; 798{ 799 struct ufsmount *ump; 800 801 mtx_assert(&lk, MA_OWNED); 802 ump = VFSTOUFS(wk->wk_mp); 803 if (wk->wk_state & ONWORKLIST) 804 panic("add_to_worklist: already on list"); 805 wk->wk_state |= ONWORKLIST; 806 if (LIST_FIRST(&ump->softdep_workitem_pending) == NULL) 807 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); 808 else 809 LIST_INSERT_AFTER(ump->softdep_worklist_tail, wk, wk_list); 810 ump->softdep_worklist_tail = wk; 811 ump->softdep_on_worklist += 1; 812} 813 814/* 815 * Process that runs once per second to handle items in the background queue. 816 * 817 * Note that we ensure that everything is done in the order in which they 818 * appear in the queue. The code below depends on this property to ensure 819 * that blocks of a file are freed before the inode itself is freed. This 820 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 821 * until all the old ones have been purged from the dependency lists. 822 */ 823int 824softdep_process_worklist(mp, full) 825 struct mount *mp; 826 int full; 827{ 828 struct thread *td = curthread; 829 int cnt, matchcnt, loopcount; 830 struct ufsmount *ump; 831 long starttime; 832 833 KASSERT(mp != NULL, ("softdep_process_worklist: NULL mp")); 834 /* 835 * Record the process identifier of our caller so that we can give 836 * this process preferential treatment in request_cleanup below. 837 */ 838 matchcnt = 0; 839 ump = VFSTOUFS(mp); 840 ACQUIRE_LOCK(&lk); 841 loopcount = 1; 842 starttime = time_second; 843 while (ump->softdep_on_worklist > 0) { 844 if ((cnt = process_worklist_item(mp, 0)) == -1) 845 break; 846 else 847 matchcnt += cnt; 848 /* 849 * If requested, try removing inode or removal dependencies. 850 */ 851 if (req_clear_inodedeps) { 852 clear_inodedeps(td); 853 req_clear_inodedeps -= 1; 854 wakeup_one(&proc_waiting); 855 } 856 if (req_clear_remove) { 857 clear_remove(td); 858 req_clear_remove -= 1; 859 wakeup_one(&proc_waiting); 860 } 861 /* 862 * We do not generally want to stop for buffer space, but if 863 * we are really being a buffer hog, we will stop and wait. 864 */ 865 if (loopcount++ % 128 == 0) { 866 FREE_LOCK(&lk); 867 bwillwrite(); 868 ACQUIRE_LOCK(&lk); 869 } 870 /* 871 * Never allow processing to run for more than one 872 * second. Otherwise the other mountpoints may get 873 * excessively backlogged. 874 */ 875 if (!full && starttime != time_second) { 876 matchcnt = -1; 877 break; 878 } 879 } 880 FREE_LOCK(&lk); 881 return (matchcnt); 882} 883 884/* 885 * Process one item on the worklist. 886 */ 887static int 888process_worklist_item(mp, flags) 889 struct mount *mp; 890 int flags; 891{ 892 struct worklist *wk, *wkend; 893 struct ufsmount *ump; 894 struct vnode *vp; 895 int matchcnt = 0; 896 897 mtx_assert(&lk, MA_OWNED); 898 KASSERT(mp != NULL, ("process_worklist_item: NULL mp")); 899 /* 900 * If we are being called because of a process doing a 901 * copy-on-write, then it is not safe to write as we may 902 * recurse into the copy-on-write routine. 903 */ 904 if (curthread->td_pflags & TDP_COWINPROGRESS) 905 return (-1); 906 /* 907 * Normally we just process each item on the worklist in order. 908 * However, if we are in a situation where we cannot lock any 909 * inodes, we have to skip over any dirrem requests whose 910 * vnodes are resident and locked. 911 */ 912 ump = VFSTOUFS(mp); 913 vp = NULL; 914 LIST_FOREACH(wk, &ump->softdep_workitem_pending, wk_list) { 915 if (wk->wk_state & INPROGRESS) 916 continue; 917 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM) 918 break; 919 wk->wk_state |= INPROGRESS; 920 ump->softdep_on_worklist_inprogress++; 921 FREE_LOCK(&lk); 922 ffs_vget(mp, WK_DIRREM(wk)->dm_oldinum, 923 LK_NOWAIT | LK_EXCLUSIVE, &vp); 924 ACQUIRE_LOCK(&lk); 925 wk->wk_state &= ~INPROGRESS; 926 ump->softdep_on_worklist_inprogress--; 927 if (vp != NULL) 928 break; 929 } 930 if (wk == 0) 931 return (-1); 932 /* 933 * Remove the item to be processed. If we are removing the last 934 * item on the list, we need to recalculate the tail pointer. 935 * As this happens rarely and usually when the list is short, 936 * we just run down the list to find it rather than tracking it 937 * in the above loop. 938 */ 939 WORKLIST_REMOVE(wk); 940 if (wk == ump->softdep_worklist_tail) { 941 LIST_FOREACH(wkend, &ump->softdep_workitem_pending, wk_list) 942 if (LIST_NEXT(wkend, wk_list) == NULL) 943 break; 944 ump->softdep_worklist_tail = wkend; 945 } 946 ump->softdep_on_worklist -= 1; 947 FREE_LOCK(&lk); 948 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT)) 949 panic("process_worklist_item: suspended filesystem"); 950 matchcnt++; 951 switch (wk->wk_type) { 952 953 case D_DIRREM: 954 /* removal of a directory entry */ 955 handle_workitem_remove(WK_DIRREM(wk), vp); 956 break; 957 958 case D_FREEBLKS: 959 /* releasing blocks and/or fragments from a file */ 960 handle_workitem_freeblocks(WK_FREEBLKS(wk), flags & LK_NOWAIT); 961 break; 962 963 case D_FREEFRAG: 964 /* releasing a fragment when replaced as a file grows */ 965 handle_workitem_freefrag(WK_FREEFRAG(wk)); 966 break; 967 968 case D_FREEFILE: 969 /* releasing an inode when its link count drops to 0 */ 970 handle_workitem_freefile(WK_FREEFILE(wk)); 971 break; 972 973 default: 974 panic("%s_process_worklist: Unknown type %s", 975 "softdep", TYPENAME(wk->wk_type)); 976 /* NOTREACHED */ 977 } 978 vn_finished_secondary_write(mp); 979 ACQUIRE_LOCK(&lk); 980 return (matchcnt); 981} 982 983/* 984 * Move dependencies from one buffer to another. 985 */ 986void 987softdep_move_dependencies(oldbp, newbp) 988 struct buf *oldbp; 989 struct buf *newbp; 990{ 991 struct worklist *wk, *wktail; 992 993 if (LIST_FIRST(&newbp->b_dep) != NULL) 994 panic("softdep_move_dependencies: need merge code"); 995 wktail = 0; 996 ACQUIRE_LOCK(&lk); 997 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 998 LIST_REMOVE(wk, wk_list); 999 if (wktail == 0) 1000 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 1001 else 1002 LIST_INSERT_AFTER(wktail, wk, wk_list); 1003 wktail = wk; 1004 } 1005 FREE_LOCK(&lk); 1006} 1007 1008/* 1009 * Purge the work list of all items associated with a particular mount point. 1010 */ 1011int 1012softdep_flushworklist(oldmnt, countp, td) 1013 struct mount *oldmnt; 1014 int *countp; 1015 struct thread *td; 1016{ 1017 struct vnode *devvp; 1018 int count, error = 0; 1019 struct ufsmount *ump; 1020 1021 /* 1022 * Alternately flush the block device associated with the mount 1023 * point and process any dependencies that the flushing 1024 * creates. We continue until no more worklist dependencies 1025 * are found. 1026 */ 1027 *countp = 0; 1028 ump = VFSTOUFS(oldmnt); 1029 devvp = ump->um_devvp; 1030 while ((count = softdep_process_worklist(oldmnt, 1)) > 0) { 1031 *countp += count; 1032 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, td); 1033 error = VOP_FSYNC(devvp, MNT_WAIT, td); 1034 VOP_UNLOCK(devvp, 0, td); 1035 if (error) 1036 break; 1037 } 1038 return (error); 1039} 1040 1041int 1042softdep_waitidle(struct mount *mp) 1043{ 1044 struct ufsmount *ump; 1045 int error; 1046 int i; 1047 1048 ump = VFSTOUFS(mp); 1049 ACQUIRE_LOCK(&lk); 1050 for (i = 0; i < 10 && ump->softdep_deps; i++) { 1051 ump->softdep_req = 1; 1052 if (ump->softdep_on_worklist) 1053 panic("softdep_waitidle: work added after flush."); 1054 msleep(&ump->softdep_deps, &lk, PVM, "softdeps", 1); 1055 } 1056 ump->softdep_req = 0; 1057 FREE_LOCK(&lk); 1058 error = 0; 1059 if (i == 10) { 1060 error = EBUSY; 1061 printf("softdep_waitidle: Failed to flush worklist for %p", 1062 mp); 1063 } 1064 1065 return (error); 1066} 1067 1068/* 1069 * Flush all vnodes and worklist items associated with a specified mount point. 1070 */ 1071int 1072softdep_flushfiles(oldmnt, flags, td) 1073 struct mount *oldmnt; 1074 int flags; 1075 struct thread *td; 1076{ 1077 int error, count, loopcnt; 1078 1079 error = 0; 1080 1081 /* 1082 * Alternately flush the vnodes associated with the mount 1083 * point and process any dependencies that the flushing 1084 * creates. In theory, this loop can happen at most twice, 1085 * but we give it a few extra just to be sure. 1086 */ 1087 for (loopcnt = 10; loopcnt > 0; loopcnt--) { 1088 /* 1089 * Do another flush in case any vnodes were brought in 1090 * as part of the cleanup operations. 1091 */ 1092 if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0) 1093 break; 1094 if ((error = softdep_flushworklist(oldmnt, &count, td)) != 0 || 1095 count == 0) 1096 break; 1097 } 1098 /* 1099 * If we are unmounting then it is an error to fail. If we 1100 * are simply trying to downgrade to read-only, then filesystem 1101 * activity can keep us busy forever, so we just fail with EBUSY. 1102 */ 1103 if (loopcnt == 0) { 1104 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 1105 panic("softdep_flushfiles: looping"); 1106 error = EBUSY; 1107 } 1108 if (!error) 1109 error = softdep_waitidle(oldmnt); 1110 return (error); 1111} 1112 1113/* 1114 * Structure hashing. 1115 * 1116 * There are three types of structures that can be looked up: 1117 * 1) pagedep structures identified by mount point, inode number, 1118 * and logical block. 1119 * 2) inodedep structures identified by mount point and inode number. 1120 * 3) newblk structures identified by mount point and 1121 * physical block number. 1122 * 1123 * The "pagedep" and "inodedep" dependency structures are hashed 1124 * separately from the file blocks and inodes to which they correspond. 1125 * This separation helps when the in-memory copy of an inode or 1126 * file block must be replaced. It also obviates the need to access 1127 * an inode or file page when simply updating (or de-allocating) 1128 * dependency structures. Lookup of newblk structures is needed to 1129 * find newly allocated blocks when trying to associate them with 1130 * their allocdirect or allocindir structure. 1131 * 1132 * The lookup routines optionally create and hash a new instance when 1133 * an existing entry is not found. 1134 */ 1135#define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 1136#define NODELAY 0x0002 /* cannot do background work */ 1137 1138/* 1139 * Structures and routines associated with pagedep caching. 1140 */ 1141LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl; 1142u_long pagedep_hash; /* size of hash table - 1 */ 1143#define PAGEDEP_HASH(mp, inum, lbn) \ 1144 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \ 1145 pagedep_hash]) 1146 1147static int 1148pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp) 1149 struct pagedep_hashhead *pagedephd; 1150 ino_t ino; 1151 ufs_lbn_t lbn; 1152 struct mount *mp; 1153 int flags; 1154 struct pagedep **pagedeppp; 1155{ 1156 struct pagedep *pagedep; 1157 1158 LIST_FOREACH(pagedep, pagedephd, pd_hash) 1159 if (ino == pagedep->pd_ino && 1160 lbn == pagedep->pd_lbn && 1161 mp == pagedep->pd_list.wk_mp) 1162 break; 1163 if (pagedep) { 1164 *pagedeppp = pagedep; 1165 if ((flags & DEPALLOC) != 0 && 1166 (pagedep->pd_state & ONWORKLIST) == 0) 1167 return (0); 1168 return (1); 1169 } 1170 *pagedeppp = NULL; 1171 return (0); 1172} 1173/* 1174 * Look up a pagedep. Return 1 if found, 0 if not found or found 1175 * when asked to allocate but not associated with any buffer. 1176 * If not found, allocate if DEPALLOC flag is passed. 1177 * Found or allocated entry is returned in pagedeppp. 1178 * This routine must be called with splbio interrupts blocked. 1179 */ 1180static int 1181pagedep_lookup(ip, lbn, flags, pagedeppp) 1182 struct inode *ip; 1183 ufs_lbn_t lbn; 1184 int flags; 1185 struct pagedep **pagedeppp; 1186{ 1187 struct pagedep *pagedep; 1188 struct pagedep_hashhead *pagedephd; 1189 struct mount *mp; 1190 int ret; 1191 int i; 1192 1193 mtx_assert(&lk, MA_OWNED); 1194 mp = ITOV(ip)->v_mount; 1195 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn); 1196 1197 ret = pagedep_find(pagedephd, ip->i_number, lbn, mp, flags, pagedeppp); 1198 if (*pagedeppp || (flags & DEPALLOC) == 0) 1199 return (ret); 1200 FREE_LOCK(&lk); 1201 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), 1202 M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO); 1203 workitem_alloc(&pagedep->pd_list, D_PAGEDEP, mp); 1204 ACQUIRE_LOCK(&lk); 1205 ret = pagedep_find(pagedephd, ip->i_number, lbn, mp, flags, pagedeppp); 1206 if (*pagedeppp) { 1207 WORKITEM_FREE(pagedep, D_PAGEDEP); 1208 return (ret); 1209 } 1210 pagedep->pd_ino = ip->i_number; 1211 pagedep->pd_lbn = lbn; 1212 LIST_INIT(&pagedep->pd_dirremhd); 1213 LIST_INIT(&pagedep->pd_pendinghd); 1214 for (i = 0; i < DAHASHSZ; i++) 1215 LIST_INIT(&pagedep->pd_diraddhd[i]); 1216 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 1217 *pagedeppp = pagedep; 1218 return (0); 1219} 1220 1221/* 1222 * Structures and routines associated with inodedep caching. 1223 */ 1224LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl; 1225static u_long inodedep_hash; /* size of hash table - 1 */ 1226static long num_inodedep; /* number of inodedep allocated */ 1227#define INODEDEP_HASH(fs, inum) \ 1228 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash]) 1229 1230static int 1231inodedep_find(inodedephd, fs, inum, inodedeppp) 1232 struct inodedep_hashhead *inodedephd; 1233 struct fs *fs; 1234 ino_t inum; 1235 struct inodedep **inodedeppp; 1236{ 1237 struct inodedep *inodedep; 1238 1239 LIST_FOREACH(inodedep, inodedephd, id_hash) 1240 if (inum == inodedep->id_ino && fs == inodedep->id_fs) 1241 break; 1242 if (inodedep) { 1243 *inodedeppp = inodedep; 1244 return (1); 1245 } 1246 *inodedeppp = NULL; 1247 1248 return (0); 1249} 1250/* 1251 * Look up an inodedep. Return 1 if found, 0 if not found. 1252 * If not found, allocate if DEPALLOC flag is passed. 1253 * Found or allocated entry is returned in inodedeppp. 1254 * This routine must be called with splbio interrupts blocked. 1255 */ 1256static int 1257inodedep_lookup(mp, inum, flags, inodedeppp) 1258 struct mount *mp; 1259 ino_t inum; 1260 int flags; 1261 struct inodedep **inodedeppp; 1262{ 1263 struct inodedep *inodedep; 1264 struct inodedep_hashhead *inodedephd; 1265 struct fs *fs; 1266 1267 mtx_assert(&lk, MA_OWNED); 1268 fs = VFSTOUFS(mp)->um_fs; 1269 inodedephd = INODEDEP_HASH(fs, inum); 1270 1271 if (inodedep_find(inodedephd, fs, inum, inodedeppp)) 1272 return (1); 1273 if ((flags & DEPALLOC) == 0) 1274 return (0); 1275 /* 1276 * If we are over our limit, try to improve the situation. 1277 */ 1278 if (num_inodedep > max_softdeps && (flags & NODELAY) == 0) 1279 request_cleanup(mp, FLUSH_INODES); 1280 FREE_LOCK(&lk); 1281 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep), 1282 M_INODEDEP, M_SOFTDEP_FLAGS); 1283 workitem_alloc(&inodedep->id_list, D_INODEDEP, mp); 1284 ACQUIRE_LOCK(&lk); 1285 if (inodedep_find(inodedephd, fs, inum, inodedeppp)) { 1286 WORKITEM_FREE(inodedep, D_INODEDEP); 1287 return (1); 1288 } 1289 num_inodedep += 1; 1290 inodedep->id_fs = fs; 1291 inodedep->id_ino = inum; 1292 inodedep->id_state = ALLCOMPLETE; 1293 inodedep->id_nlinkdelta = 0; 1294 inodedep->id_savedino1 = NULL; 1295 inodedep->id_savedsize = -1; 1296 inodedep->id_savedextsize = -1; 1297 inodedep->id_buf = NULL; 1298 LIST_INIT(&inodedep->id_pendinghd); 1299 LIST_INIT(&inodedep->id_inowait); 1300 LIST_INIT(&inodedep->id_bufwait); 1301 TAILQ_INIT(&inodedep->id_inoupdt); 1302 TAILQ_INIT(&inodedep->id_newinoupdt); 1303 TAILQ_INIT(&inodedep->id_extupdt); 1304 TAILQ_INIT(&inodedep->id_newextupdt); 1305 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 1306 *inodedeppp = inodedep; 1307 return (0); 1308} 1309 1310/* 1311 * Structures and routines associated with newblk caching. 1312 */ 1313LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl; 1314u_long newblk_hash; /* size of hash table - 1 */ 1315#define NEWBLK_HASH(fs, inum) \ 1316 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash]) 1317 1318static int 1319newblk_find(newblkhd, fs, newblkno, newblkpp) 1320 struct newblk_hashhead *newblkhd; 1321 struct fs *fs; 1322 ufs2_daddr_t newblkno; 1323 struct newblk **newblkpp; 1324{ 1325 struct newblk *newblk; 1326 1327 LIST_FOREACH(newblk, newblkhd, nb_hash) 1328 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs) 1329 break; 1330 if (newblk) { 1331 *newblkpp = newblk; 1332 return (1); 1333 } 1334 *newblkpp = NULL; 1335 return (0); 1336} 1337 1338/* 1339 * Look up a newblk. Return 1 if found, 0 if not found. 1340 * If not found, allocate if DEPALLOC flag is passed. 1341 * Found or allocated entry is returned in newblkpp. 1342 */ 1343static int 1344newblk_lookup(fs, newblkno, flags, newblkpp) 1345 struct fs *fs; 1346 ufs2_daddr_t newblkno; 1347 int flags; 1348 struct newblk **newblkpp; 1349{ 1350 struct newblk *newblk; 1351 struct newblk_hashhead *newblkhd; 1352 1353 newblkhd = NEWBLK_HASH(fs, newblkno); 1354 if (newblk_find(newblkhd, fs, newblkno, newblkpp)) 1355 return (1); 1356 if ((flags & DEPALLOC) == 0) 1357 return (0); 1358 FREE_LOCK(&lk); 1359 MALLOC(newblk, struct newblk *, sizeof(struct newblk), 1360 M_NEWBLK, M_SOFTDEP_FLAGS); 1361 ACQUIRE_LOCK(&lk); 1362 if (newblk_find(newblkhd, fs, newblkno, newblkpp)) { 1363 FREE(newblk, M_NEWBLK); 1364 return (1); 1365 } 1366 newblk->nb_state = 0; 1367 newblk->nb_fs = fs; 1368 newblk->nb_newblkno = newblkno; 1369 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 1370 *newblkpp = newblk; 1371 return (0); 1372} 1373 1374/* 1375 * Executed during filesystem system initialization before 1376 * mounting any filesystems. 1377 */ 1378void 1379softdep_initialize() 1380{ 1381 1382 LIST_INIT(&mkdirlisthd); 1383 max_softdeps = desiredvnodes * 4; 1384 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 1385 &pagedep_hash); 1386 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash); 1387 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash); 1388 1389 /* initialise bioops hack */ 1390 bioops.io_start = softdep_disk_io_initiation; 1391 bioops.io_complete = softdep_disk_write_complete; 1392 bioops.io_deallocate = softdep_deallocate_dependencies; 1393 bioops.io_countdeps = softdep_count_dependencies; 1394} 1395 1396/* 1397 * Executed after all filesystems have been unmounted during 1398 * filesystem module unload. 1399 */ 1400void 1401softdep_uninitialize() 1402{ 1403 1404 hashdestroy(pagedep_hashtbl, M_PAGEDEP, pagedep_hash); 1405 hashdestroy(inodedep_hashtbl, M_INODEDEP, inodedep_hash); 1406 hashdestroy(newblk_hashtbl, M_NEWBLK, newblk_hash); 1407} 1408 1409/* 1410 * Called at mount time to notify the dependency code that a 1411 * filesystem wishes to use it. 1412 */ 1413int 1414softdep_mount(devvp, mp, fs, cred) 1415 struct vnode *devvp; 1416 struct mount *mp; 1417 struct fs *fs; 1418 struct ucred *cred; 1419{ 1420 struct csum_total cstotal; 1421 struct ufsmount *ump; 1422 struct cg *cgp; 1423 struct buf *bp; 1424 int error, cyl; 1425 1426 MNT_ILOCK(mp); 1427 mp->mnt_flag = (mp->mnt_flag & ~MNT_ASYNC) | MNT_SOFTDEP; 1428 if ((mp->mnt_kern_flag & MNTK_SOFTDEP) == 0) { 1429 mp->mnt_kern_flag = (mp->mnt_kern_flag & ~MNTK_ASYNC) | 1430 MNTK_SOFTDEP; 1431 mp->mnt_noasync++; 1432 } 1433 MNT_IUNLOCK(mp); 1434 ump = VFSTOUFS(mp); 1435 LIST_INIT(&ump->softdep_workitem_pending); 1436 ump->softdep_worklist_tail = NULL; 1437 ump->softdep_on_worklist = 0; 1438 ump->softdep_deps = 0; 1439 /* 1440 * When doing soft updates, the counters in the 1441 * superblock may have gotten out of sync. Recomputation 1442 * can take a long time and can be deferred for background 1443 * fsck. However, the old behavior of scanning the cylinder 1444 * groups and recalculating them at mount time is available 1445 * by setting vfs.ffs.compute_summary_at_mount to one. 1446 */ 1447 if (compute_summary_at_mount == 0 || fs->fs_clean != 0) 1448 return (0); 1449 bzero(&cstotal, sizeof cstotal); 1450 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 1451 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 1452 fs->fs_cgsize, cred, &bp)) != 0) { 1453 brelse(bp); 1454 return (error); 1455 } 1456 cgp = (struct cg *)bp->b_data; 1457 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 1458 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 1459 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 1460 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 1461 fs->fs_cs(fs, cyl) = cgp->cg_cs; 1462 brelse(bp); 1463 } 1464#ifdef DEBUG 1465 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 1466 printf("%s: superblock summary recomputed\n", fs->fs_fsmnt); 1467#endif 1468 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 1469 return (0); 1470} 1471 1472/* 1473 * Protecting the freemaps (or bitmaps). 1474 * 1475 * To eliminate the need to execute fsck before mounting a filesystem 1476 * after a power failure, one must (conservatively) guarantee that the 1477 * on-disk copy of the bitmaps never indicate that a live inode or block is 1478 * free. So, when a block or inode is allocated, the bitmap should be 1479 * updated (on disk) before any new pointers. When a block or inode is 1480 * freed, the bitmap should not be updated until all pointers have been 1481 * reset. The latter dependency is handled by the delayed de-allocation 1482 * approach described below for block and inode de-allocation. The former 1483 * dependency is handled by calling the following procedure when a block or 1484 * inode is allocated. When an inode is allocated an "inodedep" is created 1485 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 1486 * Each "inodedep" is also inserted into the hash indexing structure so 1487 * that any additional link additions can be made dependent on the inode 1488 * allocation. 1489 * 1490 * The ufs filesystem maintains a number of free block counts (e.g., per 1491 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 1492 * in addition to the bitmaps. These counts are used to improve efficiency 1493 * during allocation and therefore must be consistent with the bitmaps. 1494 * There is no convenient way to guarantee post-crash consistency of these 1495 * counts with simple update ordering, for two main reasons: (1) The counts 1496 * and bitmaps for a single cylinder group block are not in the same disk 1497 * sector. If a disk write is interrupted (e.g., by power failure), one may 1498 * be written and the other not. (2) Some of the counts are located in the 1499 * superblock rather than the cylinder group block. So, we focus our soft 1500 * updates implementation on protecting the bitmaps. When mounting a 1501 * filesystem, we recompute the auxiliary counts from the bitmaps. 1502 */ 1503 1504/* 1505 * Called just after updating the cylinder group block to allocate an inode. 1506 */ 1507void 1508softdep_setup_inomapdep(bp, ip, newinum) 1509 struct buf *bp; /* buffer for cylgroup block with inode map */ 1510 struct inode *ip; /* inode related to allocation */ 1511 ino_t newinum; /* new inode number being allocated */ 1512{ 1513 struct inodedep *inodedep; 1514 struct bmsafemap *bmsafemap; 1515 1516 /* 1517 * Create a dependency for the newly allocated inode. 1518 * Panic if it already exists as something is seriously wrong. 1519 * Otherwise add it to the dependency list for the buffer holding 1520 * the cylinder group map from which it was allocated. 1521 */ 1522 ACQUIRE_LOCK(&lk); 1523 if ((inodedep_lookup(UFSTOVFS(ip->i_ump), newinum, DEPALLOC|NODELAY, 1524 &inodedep))) 1525 panic("softdep_setup_inomapdep: dependency for new inode " 1526 "already exists"); 1527 inodedep->id_buf = bp; 1528 inodedep->id_state &= ~DEPCOMPLETE; 1529 bmsafemap = bmsafemap_lookup(inodedep->id_list.wk_mp, bp); 1530 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 1531 FREE_LOCK(&lk); 1532} 1533 1534/* 1535 * Called just after updating the cylinder group block to 1536 * allocate block or fragment. 1537 */ 1538void 1539softdep_setup_blkmapdep(bp, mp, newblkno) 1540 struct buf *bp; /* buffer for cylgroup block with block map */ 1541 struct mount *mp; /* filesystem doing allocation */ 1542 ufs2_daddr_t newblkno; /* number of newly allocated block */ 1543{ 1544 struct newblk *newblk; 1545 struct bmsafemap *bmsafemap; 1546 struct fs *fs; 1547 1548 fs = VFSTOUFS(mp)->um_fs; 1549 /* 1550 * Create a dependency for the newly allocated block. 1551 * Add it to the dependency list for the buffer holding 1552 * the cylinder group map from which it was allocated. 1553 */ 1554 ACQUIRE_LOCK(&lk); 1555 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0) 1556 panic("softdep_setup_blkmapdep: found block"); 1557 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp); 1558 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 1559 FREE_LOCK(&lk); 1560} 1561 1562/* 1563 * Find the bmsafemap associated with a cylinder group buffer. 1564 * If none exists, create one. The buffer must be locked when 1565 * this routine is called and this routine must be called with 1566 * splbio interrupts blocked. 1567 */ 1568static struct bmsafemap * 1569bmsafemap_lookup(mp, bp) 1570 struct mount *mp; 1571 struct buf *bp; 1572{ 1573 struct bmsafemap *bmsafemap; 1574 struct worklist *wk; 1575 1576 mtx_assert(&lk, MA_OWNED); 1577 LIST_FOREACH(wk, &bp->b_dep, wk_list) 1578 if (wk->wk_type == D_BMSAFEMAP) 1579 return (WK_BMSAFEMAP(wk)); 1580 FREE_LOCK(&lk); 1581 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap), 1582 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 1583 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); 1584 bmsafemap->sm_buf = bp; 1585 LIST_INIT(&bmsafemap->sm_allocdirecthd); 1586 LIST_INIT(&bmsafemap->sm_allocindirhd); 1587 LIST_INIT(&bmsafemap->sm_inodedephd); 1588 LIST_INIT(&bmsafemap->sm_newblkhd); 1589 ACQUIRE_LOCK(&lk); 1590 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 1591 return (bmsafemap); 1592} 1593 1594/* 1595 * Direct block allocation dependencies. 1596 * 1597 * When a new block is allocated, the corresponding disk locations must be 1598 * initialized (with zeros or new data) before the on-disk inode points to 1599 * them. Also, the freemap from which the block was allocated must be 1600 * updated (on disk) before the inode's pointer. These two dependencies are 1601 * independent of each other and are needed for all file blocks and indirect 1602 * blocks that are pointed to directly by the inode. Just before the 1603 * "in-core" version of the inode is updated with a newly allocated block 1604 * number, a procedure (below) is called to setup allocation dependency 1605 * structures. These structures are removed when the corresponding 1606 * dependencies are satisfied or when the block allocation becomes obsolete 1607 * (i.e., the file is deleted, the block is de-allocated, or the block is a 1608 * fragment that gets upgraded). All of these cases are handled in 1609 * procedures described later. 1610 * 1611 * When a file extension causes a fragment to be upgraded, either to a larger 1612 * fragment or to a full block, the on-disk location may change (if the 1613 * previous fragment could not simply be extended). In this case, the old 1614 * fragment must be de-allocated, but not until after the inode's pointer has 1615 * been updated. In most cases, this is handled by later procedures, which 1616 * will construct a "freefrag" structure to be added to the workitem queue 1617 * when the inode update is complete (or obsolete). The main exception to 1618 * this is when an allocation occurs while a pending allocation dependency 1619 * (for the same block pointer) remains. This case is handled in the main 1620 * allocation dependency setup procedure by immediately freeing the 1621 * unreferenced fragments. 1622 */ 1623void 1624softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 1625 struct inode *ip; /* inode to which block is being added */ 1626 ufs_lbn_t lbn; /* block pointer within inode */ 1627 ufs2_daddr_t newblkno; /* disk block number being added */ 1628 ufs2_daddr_t oldblkno; /* previous block number, 0 unless frag */ 1629 long newsize; /* size of new block */ 1630 long oldsize; /* size of new block */ 1631 struct buf *bp; /* bp for allocated block */ 1632{ 1633 struct allocdirect *adp, *oldadp; 1634 struct allocdirectlst *adphead; 1635 struct bmsafemap *bmsafemap; 1636 struct inodedep *inodedep; 1637 struct pagedep *pagedep; 1638 struct newblk *newblk; 1639 struct mount *mp; 1640 1641 mp = UFSTOVFS(ip->i_ump); 1642 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect), 1643 M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO); 1644 workitem_alloc(&adp->ad_list, D_ALLOCDIRECT, mp); 1645 adp->ad_lbn = lbn; 1646 adp->ad_newblkno = newblkno; 1647 adp->ad_oldblkno = oldblkno; 1648 adp->ad_newsize = newsize; 1649 adp->ad_oldsize = oldsize; 1650 adp->ad_state = ATTACHED; 1651 LIST_INIT(&adp->ad_newdirblk); 1652 if (newblkno == oldblkno) 1653 adp->ad_freefrag = NULL; 1654 else 1655 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1656 1657 ACQUIRE_LOCK(&lk); 1658 if (lbn >= NDADDR) { 1659 /* allocating an indirect block */ 1660 if (oldblkno != 0) 1661 panic("softdep_setup_allocdirect: non-zero indir"); 1662 } else { 1663 /* 1664 * Allocating a direct block. 1665 * 1666 * If we are allocating a directory block, then we must 1667 * allocate an associated pagedep to track additions and 1668 * deletions. 1669 */ 1670 if ((ip->i_mode & IFMT) == IFDIR && 1671 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 1672 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 1673 } 1674 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1675 panic("softdep_setup_allocdirect: lost block"); 1676 if (newblk->nb_state == DEPCOMPLETE) { 1677 adp->ad_state |= DEPCOMPLETE; 1678 adp->ad_buf = NULL; 1679 } else { 1680 bmsafemap = newblk->nb_bmsafemap; 1681 adp->ad_buf = bmsafemap->sm_buf; 1682 LIST_REMOVE(newblk, nb_deps); 1683 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1684 } 1685 LIST_REMOVE(newblk, nb_hash); 1686 FREE(newblk, M_NEWBLK); 1687 1688 inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep); 1689 adp->ad_inodedep = inodedep; 1690 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list); 1691 /* 1692 * The list of allocdirects must be kept in sorted and ascending 1693 * order so that the rollback routines can quickly determine the 1694 * first uncommitted block (the size of the file stored on disk 1695 * ends at the end of the lowest committed fragment, or if there 1696 * are no fragments, at the end of the highest committed block). 1697 * Since files generally grow, the typical case is that the new 1698 * block is to be added at the end of the list. We speed this 1699 * special case by checking against the last allocdirect in the 1700 * list before laboriously traversing the list looking for the 1701 * insertion point. 1702 */ 1703 adphead = &inodedep->id_newinoupdt; 1704 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1705 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1706 /* insert at end of list */ 1707 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1708 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1709 allocdirect_merge(adphead, adp, oldadp); 1710 FREE_LOCK(&lk); 1711 return; 1712 } 1713 TAILQ_FOREACH(oldadp, adphead, ad_next) { 1714 if (oldadp->ad_lbn >= lbn) 1715 break; 1716 } 1717 if (oldadp == NULL) 1718 panic("softdep_setup_allocdirect: lost entry"); 1719 /* insert in middle of list */ 1720 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1721 if (oldadp->ad_lbn == lbn) 1722 allocdirect_merge(adphead, adp, oldadp); 1723 FREE_LOCK(&lk); 1724} 1725 1726/* 1727 * Replace an old allocdirect dependency with a newer one. 1728 * This routine must be called with splbio interrupts blocked. 1729 */ 1730static void 1731allocdirect_merge(adphead, newadp, oldadp) 1732 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 1733 struct allocdirect *newadp; /* allocdirect being added */ 1734 struct allocdirect *oldadp; /* existing allocdirect being checked */ 1735{ 1736 struct worklist *wk; 1737 struct freefrag *freefrag; 1738 struct newdirblk *newdirblk; 1739 1740 mtx_assert(&lk, MA_OWNED); 1741 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 1742 newadp->ad_oldsize != oldadp->ad_newsize || 1743 newadp->ad_lbn >= NDADDR) 1744 panic("%s %jd != new %jd || old size %ld != new %ld", 1745 "allocdirect_merge: old blkno", 1746 (intmax_t)newadp->ad_oldblkno, 1747 (intmax_t)oldadp->ad_newblkno, 1748 newadp->ad_oldsize, oldadp->ad_newsize); 1749 newadp->ad_oldblkno = oldadp->ad_oldblkno; 1750 newadp->ad_oldsize = oldadp->ad_oldsize; 1751 /* 1752 * If the old dependency had a fragment to free or had never 1753 * previously had a block allocated, then the new dependency 1754 * can immediately post its freefrag and adopt the old freefrag. 1755 * This action is done by swapping the freefrag dependencies. 1756 * The new dependency gains the old one's freefrag, and the 1757 * old one gets the new one and then immediately puts it on 1758 * the worklist when it is freed by free_allocdirect. It is 1759 * not possible to do this swap when the old dependency had a 1760 * non-zero size but no previous fragment to free. This condition 1761 * arises when the new block is an extension of the old block. 1762 * Here, the first part of the fragment allocated to the new 1763 * dependency is part of the block currently claimed on disk by 1764 * the old dependency, so cannot legitimately be freed until the 1765 * conditions for the new dependency are fulfilled. 1766 */ 1767 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 1768 freefrag = newadp->ad_freefrag; 1769 newadp->ad_freefrag = oldadp->ad_freefrag; 1770 oldadp->ad_freefrag = freefrag; 1771 } 1772 /* 1773 * If we are tracking a new directory-block allocation, 1774 * move it from the old allocdirect to the new allocdirect. 1775 */ 1776 if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) { 1777 newdirblk = WK_NEWDIRBLK(wk); 1778 WORKLIST_REMOVE(&newdirblk->db_list); 1779 if (LIST_FIRST(&oldadp->ad_newdirblk) != NULL) 1780 panic("allocdirect_merge: extra newdirblk"); 1781 WORKLIST_INSERT(&newadp->ad_newdirblk, &newdirblk->db_list); 1782 } 1783 free_allocdirect(adphead, oldadp, 0); 1784} 1785 1786/* 1787 * Allocate a new freefrag structure if needed. 1788 */ 1789static struct freefrag * 1790newfreefrag(ip, blkno, size) 1791 struct inode *ip; 1792 ufs2_daddr_t blkno; 1793 long size; 1794{ 1795 struct freefrag *freefrag; 1796 struct fs *fs; 1797 1798 if (blkno == 0) 1799 return (NULL); 1800 fs = ip->i_fs; 1801 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 1802 panic("newfreefrag: frag size"); 1803 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag), 1804 M_FREEFRAG, M_SOFTDEP_FLAGS); 1805 workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ip->i_ump)); 1806 freefrag->ff_inum = ip->i_number; 1807 freefrag->ff_blkno = blkno; 1808 freefrag->ff_fragsize = size; 1809 return (freefrag); 1810} 1811 1812/* 1813 * This workitem de-allocates fragments that were replaced during 1814 * file block allocation. 1815 */ 1816static void 1817handle_workitem_freefrag(freefrag) 1818 struct freefrag *freefrag; 1819{ 1820 struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp); 1821 1822 ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno, 1823 freefrag->ff_fragsize, freefrag->ff_inum); 1824 ACQUIRE_LOCK(&lk); 1825 WORKITEM_FREE(freefrag, D_FREEFRAG); 1826 FREE_LOCK(&lk); 1827} 1828 1829/* 1830 * Set up a dependency structure for an external attributes data block. 1831 * This routine follows much of the structure of softdep_setup_allocdirect. 1832 * See the description of softdep_setup_allocdirect above for details. 1833 */ 1834void 1835softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 1836 struct inode *ip; 1837 ufs_lbn_t lbn; 1838 ufs2_daddr_t newblkno; 1839 ufs2_daddr_t oldblkno; 1840 long newsize; 1841 long oldsize; 1842 struct buf *bp; 1843{ 1844 struct allocdirect *adp, *oldadp; 1845 struct allocdirectlst *adphead; 1846 struct bmsafemap *bmsafemap; 1847 struct inodedep *inodedep; 1848 struct newblk *newblk; 1849 struct mount *mp; 1850 1851 mp = UFSTOVFS(ip->i_ump); 1852 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect), 1853 M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO); 1854 workitem_alloc(&adp->ad_list, D_ALLOCDIRECT, mp); 1855 adp->ad_lbn = lbn; 1856 adp->ad_newblkno = newblkno; 1857 adp->ad_oldblkno = oldblkno; 1858 adp->ad_newsize = newsize; 1859 adp->ad_oldsize = oldsize; 1860 adp->ad_state = ATTACHED | EXTDATA; 1861 LIST_INIT(&adp->ad_newdirblk); 1862 if (newblkno == oldblkno) 1863 adp->ad_freefrag = NULL; 1864 else 1865 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize); 1866 1867 ACQUIRE_LOCK(&lk); 1868 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0) 1869 panic("softdep_setup_allocext: lost block"); 1870 1871 inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep); 1872 adp->ad_inodedep = inodedep; 1873 1874 if (newblk->nb_state == DEPCOMPLETE) { 1875 adp->ad_state |= DEPCOMPLETE; 1876 adp->ad_buf = NULL; 1877 } else { 1878 bmsafemap = newblk->nb_bmsafemap; 1879 adp->ad_buf = bmsafemap->sm_buf; 1880 LIST_REMOVE(newblk, nb_deps); 1881 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps); 1882 } 1883 LIST_REMOVE(newblk, nb_hash); 1884 FREE(newblk, M_NEWBLK); 1885 1886 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list); 1887 if (lbn >= NXADDR) 1888 panic("softdep_setup_allocext: lbn %lld > NXADDR", 1889 (long long)lbn); 1890 /* 1891 * The list of allocdirects must be kept in sorted and ascending 1892 * order so that the rollback routines can quickly determine the 1893 * first uncommitted block (the size of the file stored on disk 1894 * ends at the end of the lowest committed fragment, or if there 1895 * are no fragments, at the end of the highest committed block). 1896 * Since files generally grow, the typical case is that the new 1897 * block is to be added at the end of the list. We speed this 1898 * special case by checking against the last allocdirect in the 1899 * list before laboriously traversing the list looking for the 1900 * insertion point. 1901 */ 1902 adphead = &inodedep->id_newextupdt; 1903 oldadp = TAILQ_LAST(adphead, allocdirectlst); 1904 if (oldadp == NULL || oldadp->ad_lbn <= lbn) { 1905 /* insert at end of list */ 1906 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 1907 if (oldadp != NULL && oldadp->ad_lbn == lbn) 1908 allocdirect_merge(adphead, adp, oldadp); 1909 FREE_LOCK(&lk); 1910 return; 1911 } 1912 TAILQ_FOREACH(oldadp, adphead, ad_next) { 1913 if (oldadp->ad_lbn >= lbn) 1914 break; 1915 } 1916 if (oldadp == NULL) 1917 panic("softdep_setup_allocext: lost entry"); 1918 /* insert in middle of list */ 1919 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 1920 if (oldadp->ad_lbn == lbn) 1921 allocdirect_merge(adphead, adp, oldadp); 1922 FREE_LOCK(&lk); 1923} 1924 1925/* 1926 * Indirect block allocation dependencies. 1927 * 1928 * The same dependencies that exist for a direct block also exist when 1929 * a new block is allocated and pointed to by an entry in a block of 1930 * indirect pointers. The undo/redo states described above are also 1931 * used here. Because an indirect block contains many pointers that 1932 * may have dependencies, a second copy of the entire in-memory indirect 1933 * block is kept. The buffer cache copy is always completely up-to-date. 1934 * The second copy, which is used only as a source for disk writes, 1935 * contains only the safe pointers (i.e., those that have no remaining 1936 * update dependencies). The second copy is freed when all pointers 1937 * are safe. The cache is not allowed to replace indirect blocks with 1938 * pending update dependencies. If a buffer containing an indirect 1939 * block with dependencies is written, these routines will mark it 1940 * dirty again. It can only be successfully written once all the 1941 * dependencies are removed. The ffs_fsync routine in conjunction with 1942 * softdep_sync_metadata work together to get all the dependencies 1943 * removed so that a file can be successfully written to disk. Three 1944 * procedures are used when setting up indirect block pointer 1945 * dependencies. The division is necessary because of the organization 1946 * of the "balloc" routine and because of the distinction between file 1947 * pages and file metadata blocks. 1948 */ 1949 1950/* 1951 * Allocate a new allocindir structure. 1952 */ 1953static struct allocindir * 1954newallocindir(ip, ptrno, newblkno, oldblkno) 1955 struct inode *ip; /* inode for file being extended */ 1956 int ptrno; /* offset of pointer in indirect block */ 1957 ufs2_daddr_t newblkno; /* disk block number being added */ 1958 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 1959{ 1960 struct allocindir *aip; 1961 1962 MALLOC(aip, struct allocindir *, sizeof(struct allocindir), 1963 M_ALLOCINDIR, M_SOFTDEP_FLAGS|M_ZERO); 1964 workitem_alloc(&aip->ai_list, D_ALLOCINDIR, UFSTOVFS(ip->i_ump)); 1965 aip->ai_state = ATTACHED; 1966 aip->ai_offset = ptrno; 1967 aip->ai_newblkno = newblkno; 1968 aip->ai_oldblkno = oldblkno; 1969 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize); 1970 return (aip); 1971} 1972 1973/* 1974 * Called just before setting an indirect block pointer 1975 * to a newly allocated file page. 1976 */ 1977void 1978softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 1979 struct inode *ip; /* inode for file being extended */ 1980 ufs_lbn_t lbn; /* allocated block number within file */ 1981 struct buf *bp; /* buffer with indirect blk referencing page */ 1982 int ptrno; /* offset of pointer in indirect block */ 1983 ufs2_daddr_t newblkno; /* disk block number being added */ 1984 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 1985 struct buf *nbp; /* buffer holding allocated page */ 1986{ 1987 struct allocindir *aip; 1988 struct pagedep *pagedep; 1989 1990 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page"); 1991 aip = newallocindir(ip, ptrno, newblkno, oldblkno); 1992 ACQUIRE_LOCK(&lk); 1993 /* 1994 * If we are allocating a directory page, then we must 1995 * allocate an associated pagedep to track additions and 1996 * deletions. 1997 */ 1998 if ((ip->i_mode & IFMT) == IFDIR && 1999 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) 2000 WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list); 2001 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 2002 setup_allocindir_phase2(bp, ip, aip); 2003 FREE_LOCK(&lk); 2004} 2005 2006/* 2007 * Called just before setting an indirect block pointer to a 2008 * newly allocated indirect block. 2009 */ 2010void 2011softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 2012 struct buf *nbp; /* newly allocated indirect block */ 2013 struct inode *ip; /* inode for file being extended */ 2014 struct buf *bp; /* indirect block referencing allocated block */ 2015 int ptrno; /* offset of pointer in indirect block */ 2016 ufs2_daddr_t newblkno; /* disk block number being added */ 2017{ 2018 struct allocindir *aip; 2019 2020 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta"); 2021 aip = newallocindir(ip, ptrno, newblkno, 0); 2022 ACQUIRE_LOCK(&lk); 2023 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list); 2024 setup_allocindir_phase2(bp, ip, aip); 2025 FREE_LOCK(&lk); 2026} 2027 2028/* 2029 * Called to finish the allocation of the "aip" allocated 2030 * by one of the two routines above. 2031 */ 2032static void 2033setup_allocindir_phase2(bp, ip, aip) 2034 struct buf *bp; /* in-memory copy of the indirect block */ 2035 struct inode *ip; /* inode for file being extended */ 2036 struct allocindir *aip; /* allocindir allocated by the above routines */ 2037{ 2038 struct worklist *wk; 2039 struct indirdep *indirdep, *newindirdep; 2040 struct bmsafemap *bmsafemap; 2041 struct allocindir *oldaip; 2042 struct freefrag *freefrag; 2043 struct newblk *newblk; 2044 ufs2_daddr_t blkno; 2045 2046 mtx_assert(&lk, MA_OWNED); 2047 if (bp->b_lblkno >= 0) 2048 panic("setup_allocindir_phase2: not indir blk"); 2049 for (indirdep = NULL, newindirdep = NULL; ; ) { 2050 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 2051 if (wk->wk_type != D_INDIRDEP) 2052 continue; 2053 indirdep = WK_INDIRDEP(wk); 2054 break; 2055 } 2056 if (indirdep == NULL && newindirdep) { 2057 indirdep = newindirdep; 2058 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 2059 newindirdep = NULL; 2060 } 2061 if (indirdep) { 2062 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0, 2063 &newblk) == 0) 2064 panic("setup_allocindir: lost block"); 2065 if (newblk->nb_state == DEPCOMPLETE) { 2066 aip->ai_state |= DEPCOMPLETE; 2067 aip->ai_buf = NULL; 2068 } else { 2069 bmsafemap = newblk->nb_bmsafemap; 2070 aip->ai_buf = bmsafemap->sm_buf; 2071 LIST_REMOVE(newblk, nb_deps); 2072 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd, 2073 aip, ai_deps); 2074 } 2075 LIST_REMOVE(newblk, nb_hash); 2076 FREE(newblk, M_NEWBLK); 2077 aip->ai_indirdep = indirdep; 2078 /* 2079 * Check to see if there is an existing dependency 2080 * for this block. If there is, merge the old 2081 * dependency into the new one. 2082 */ 2083 if (aip->ai_oldblkno == 0) 2084 oldaip = NULL; 2085 else 2086 2087 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) 2088 if (oldaip->ai_offset == aip->ai_offset) 2089 break; 2090 freefrag = NULL; 2091 if (oldaip != NULL) { 2092 if (oldaip->ai_newblkno != aip->ai_oldblkno) 2093 panic("setup_allocindir_phase2: blkno"); 2094 aip->ai_oldblkno = oldaip->ai_oldblkno; 2095 freefrag = aip->ai_freefrag; 2096 aip->ai_freefrag = oldaip->ai_freefrag; 2097 oldaip->ai_freefrag = NULL; 2098 free_allocindir(oldaip, NULL); 2099 } 2100 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 2101 if (ip->i_ump->um_fstype == UFS1) 2102 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data) 2103 [aip->ai_offset] = aip->ai_oldblkno; 2104 else 2105 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data) 2106 [aip->ai_offset] = aip->ai_oldblkno; 2107 FREE_LOCK(&lk); 2108 if (freefrag != NULL) 2109 handle_workitem_freefrag(freefrag); 2110 } else 2111 FREE_LOCK(&lk); 2112 if (newindirdep) { 2113 newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 2114 brelse(newindirdep->ir_savebp); 2115 ACQUIRE_LOCK(&lk); 2116 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP); 2117 if (indirdep) 2118 break; 2119 FREE_LOCK(&lk); 2120 } 2121 if (indirdep) { 2122 ACQUIRE_LOCK(&lk); 2123 break; 2124 } 2125 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep), 2126 M_INDIRDEP, M_SOFTDEP_FLAGS); 2127 workitem_alloc(&newindirdep->ir_list, D_INDIRDEP, 2128 UFSTOVFS(ip->i_ump)); 2129 newindirdep->ir_state = ATTACHED; 2130 if (ip->i_ump->um_fstype == UFS1) 2131 newindirdep->ir_state |= UFS1FMT; 2132 LIST_INIT(&newindirdep->ir_deplisthd); 2133 LIST_INIT(&newindirdep->ir_donehd); 2134 if (bp->b_blkno == bp->b_lblkno) { 2135 ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp, 2136 NULL, NULL); 2137 bp->b_blkno = blkno; 2138 } 2139 newindirdep->ir_savebp = 2140 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0); 2141 BUF_KERNPROC(newindirdep->ir_savebp); 2142 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 2143 ACQUIRE_LOCK(&lk); 2144 } 2145} 2146 2147/* 2148 * Block de-allocation dependencies. 2149 * 2150 * When blocks are de-allocated, the on-disk pointers must be nullified before 2151 * the blocks are made available for use by other files. (The true 2152 * requirement is that old pointers must be nullified before new on-disk 2153 * pointers are set. We chose this slightly more stringent requirement to 2154 * reduce complexity.) Our implementation handles this dependency by updating 2155 * the inode (or indirect block) appropriately but delaying the actual block 2156 * de-allocation (i.e., freemap and free space count manipulation) until 2157 * after the updated versions reach stable storage. After the disk is 2158 * updated, the blocks can be safely de-allocated whenever it is convenient. 2159 * This implementation handles only the common case of reducing a file's 2160 * length to zero. Other cases are handled by the conventional synchronous 2161 * write approach. 2162 * 2163 * The ffs implementation with which we worked double-checks 2164 * the state of the block pointers and file size as it reduces 2165 * a file's length. Some of this code is replicated here in our 2166 * soft updates implementation. The freeblks->fb_chkcnt field is 2167 * used to transfer a part of this information to the procedure 2168 * that eventually de-allocates the blocks. 2169 * 2170 * This routine should be called from the routine that shortens 2171 * a file's length, before the inode's size or block pointers 2172 * are modified. It will save the block pointer information for 2173 * later release and zero the inode so that the calling routine 2174 * can release it. 2175 */ 2176void 2177softdep_setup_freeblocks(ip, length, flags) 2178 struct inode *ip; /* The inode whose length is to be reduced */ 2179 off_t length; /* The new length for the file */ 2180 int flags; /* IO_EXT and/or IO_NORMAL */ 2181{ 2182 struct freeblks *freeblks; 2183 struct inodedep *inodedep; 2184 struct allocdirect *adp; 2185 struct vnode *vp; 2186 struct buf *bp; 2187 struct fs *fs; 2188 ufs2_daddr_t extblocks, datablocks; 2189 struct mount *mp; 2190 int i, delay, error; 2191 2192 fs = ip->i_fs; 2193 mp = UFSTOVFS(ip->i_ump); 2194 if (length != 0) 2195 panic("softdep_setup_freeblocks: non-zero length"); 2196 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks), 2197 M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO); 2198 workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp); 2199 freeblks->fb_state = ATTACHED; 2200 freeblks->fb_uid = ip->i_uid; 2201 freeblks->fb_previousinum = ip->i_number; 2202 freeblks->fb_devvp = ip->i_devvp; 2203 extblocks = 0; 2204 if (fs->fs_magic == FS_UFS2_MAGIC) 2205 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 2206 datablocks = DIP(ip, i_blocks) - extblocks; 2207 if ((flags & IO_NORMAL) == 0) { 2208 freeblks->fb_oldsize = 0; 2209 freeblks->fb_chkcnt = 0; 2210 } else { 2211 freeblks->fb_oldsize = ip->i_size; 2212 ip->i_size = 0; 2213 DIP_SET(ip, i_size, 0); 2214 freeblks->fb_chkcnt = datablocks; 2215 for (i = 0; i < NDADDR; i++) { 2216 freeblks->fb_dblks[i] = DIP(ip, i_db[i]); 2217 DIP_SET(ip, i_db[i], 0); 2218 } 2219 for (i = 0; i < NIADDR; i++) { 2220 freeblks->fb_iblks[i] = DIP(ip, i_ib[i]); 2221 DIP_SET(ip, i_ib[i], 0); 2222 } 2223 /* 2224 * If the file was removed, then the space being freed was 2225 * accounted for then (see softdep_filereleased()). If the 2226 * file is merely being truncated, then we account for it now. 2227 */ 2228 if ((ip->i_flag & IN_SPACECOUNTED) == 0) { 2229 UFS_LOCK(ip->i_ump); 2230 fs->fs_pendingblocks += datablocks; 2231 UFS_UNLOCK(ip->i_ump); 2232 } 2233 } 2234 if ((flags & IO_EXT) == 0) { 2235 freeblks->fb_oldextsize = 0; 2236 } else { 2237 freeblks->fb_oldextsize = ip->i_din2->di_extsize; 2238 ip->i_din2->di_extsize = 0; 2239 freeblks->fb_chkcnt += extblocks; 2240 for (i = 0; i < NXADDR; i++) { 2241 freeblks->fb_eblks[i] = ip->i_din2->di_extb[i]; 2242 ip->i_din2->di_extb[i] = 0; 2243 } 2244 } 2245 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - freeblks->fb_chkcnt); 2246 /* 2247 * Push the zero'ed inode to to its disk buffer so that we are free 2248 * to delete its dependencies below. Once the dependencies are gone 2249 * the buffer can be safely released. 2250 */ 2251 if ((error = bread(ip->i_devvp, 2252 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 2253 (int)fs->fs_bsize, NOCRED, &bp)) != 0) { 2254 brelse(bp); 2255 softdep_error("softdep_setup_freeblocks", error); 2256 } 2257 if (ip->i_ump->um_fstype == UFS1) 2258 *((struct ufs1_dinode *)bp->b_data + 2259 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1; 2260 else 2261 *((struct ufs2_dinode *)bp->b_data + 2262 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2; 2263 /* 2264 * Find and eliminate any inode dependencies. 2265 */ 2266 ACQUIRE_LOCK(&lk); 2267 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 2268 if ((inodedep->id_state & IOSTARTED) != 0) 2269 panic("softdep_setup_freeblocks: inode busy"); 2270 /* 2271 * Add the freeblks structure to the list of operations that 2272 * must await the zero'ed inode being written to disk. If we 2273 * still have a bitmap dependency (delay == 0), then the inode 2274 * has never been written to disk, so we can process the 2275 * freeblks below once we have deleted the dependencies. 2276 */ 2277 delay = (inodedep->id_state & DEPCOMPLETE); 2278 if (delay) 2279 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list); 2280 /* 2281 * Because the file length has been truncated to zero, any 2282 * pending block allocation dependency structures associated 2283 * with this inode are obsolete and can simply be de-allocated. 2284 * We must first merge the two dependency lists to get rid of 2285 * any duplicate freefrag structures, then purge the merged list. 2286 * If we still have a bitmap dependency, then the inode has never 2287 * been written to disk, so we can free any fragments without delay. 2288 */ 2289 if (flags & IO_NORMAL) { 2290 merge_inode_lists(&inodedep->id_newinoupdt, 2291 &inodedep->id_inoupdt); 2292 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 2293 free_allocdirect(&inodedep->id_inoupdt, adp, delay); 2294 } 2295 if (flags & IO_EXT) { 2296 merge_inode_lists(&inodedep->id_newextupdt, 2297 &inodedep->id_extupdt); 2298 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0) 2299 free_allocdirect(&inodedep->id_extupdt, adp, delay); 2300 } 2301 FREE_LOCK(&lk); 2302 bdwrite(bp); 2303 /* 2304 * We must wait for any I/O in progress to finish so that 2305 * all potential buffers on the dirty list will be visible. 2306 * Once they are all there, walk the list and get rid of 2307 * any dependencies. 2308 */ 2309 vp = ITOV(ip); 2310 VI_LOCK(vp); 2311 drain_output(vp); 2312restart: 2313 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 2314 if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) || 2315 ((flags & IO_NORMAL) == 0 && 2316 (bp->b_xflags & BX_ALTDATA) == 0)) 2317 continue; 2318 if ((bp = getdirtybuf(bp, VI_MTX(vp), MNT_WAIT)) == NULL) 2319 goto restart; 2320 VI_UNLOCK(vp); 2321 ACQUIRE_LOCK(&lk); 2322 (void) inodedep_lookup(mp, ip->i_number, 0, &inodedep); 2323 deallocate_dependencies(bp, inodedep); 2324 FREE_LOCK(&lk); 2325 bp->b_flags |= B_INVAL | B_NOCACHE; 2326 brelse(bp); 2327 VI_LOCK(vp); 2328 goto restart; 2329 } 2330 VI_UNLOCK(vp); 2331 ACQUIRE_LOCK(&lk); 2332 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 2333 (void) free_inodedep(inodedep); 2334 2335 if(delay) { 2336 freeblks->fb_state |= DEPCOMPLETE; 2337 /* 2338 * If the inode with zeroed block pointers is now on disk 2339 * we can start freeing blocks. Add freeblks to the worklist 2340 * instead of calling handle_workitem_freeblocks directly as 2341 * it is more likely that additional IO is needed to complete 2342 * the request here than in the !delay case. 2343 */ 2344 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 2345 add_to_worklist(&freeblks->fb_list); 2346 } 2347 2348 FREE_LOCK(&lk); 2349 /* 2350 * If the inode has never been written to disk (delay == 0), 2351 * then we can process the freeblks now that we have deleted 2352 * the dependencies. 2353 */ 2354 if (!delay) 2355 handle_workitem_freeblocks(freeblks, 0); 2356} 2357 2358/* 2359 * Reclaim any dependency structures from a buffer that is about to 2360 * be reallocated to a new vnode. The buffer must be locked, thus, 2361 * no I/O completion operations can occur while we are manipulating 2362 * its associated dependencies. The mutex is held so that other I/O's 2363 * associated with related dependencies do not occur. 2364 */ 2365static void 2366deallocate_dependencies(bp, inodedep) 2367 struct buf *bp; 2368 struct inodedep *inodedep; 2369{ 2370 struct worklist *wk; 2371 struct indirdep *indirdep; 2372 struct allocindir *aip; 2373 struct pagedep *pagedep; 2374 struct dirrem *dirrem; 2375 struct diradd *dap; 2376 int i; 2377 2378 mtx_assert(&lk, MA_OWNED); 2379 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2380 switch (wk->wk_type) { 2381 2382 case D_INDIRDEP: 2383 indirdep = WK_INDIRDEP(wk); 2384 /* 2385 * None of the indirect pointers will ever be visible, 2386 * so they can simply be tossed. GOINGAWAY ensures 2387 * that allocated pointers will be saved in the buffer 2388 * cache until they are freed. Note that they will 2389 * only be able to be found by their physical address 2390 * since the inode mapping the logical address will 2391 * be gone. The save buffer used for the safe copy 2392 * was allocated in setup_allocindir_phase2 using 2393 * the physical address so it could be used for this 2394 * purpose. Hence we swap the safe copy with the real 2395 * copy, allowing the safe copy to be freed and holding 2396 * on to the real copy for later use in indir_trunc. 2397 */ 2398 if (indirdep->ir_state & GOINGAWAY) 2399 panic("deallocate_dependencies: already gone"); 2400 indirdep->ir_state |= GOINGAWAY; 2401 VFSTOUFS(bp->b_vp->v_mount)->um_numindirdeps += 1; 2402 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 2403 free_allocindir(aip, inodedep); 2404 if (bp->b_lblkno >= 0 || 2405 bp->b_blkno != indirdep->ir_savebp->b_lblkno) 2406 panic("deallocate_dependencies: not indir"); 2407 bcopy(bp->b_data, indirdep->ir_savebp->b_data, 2408 bp->b_bcount); 2409 WORKLIST_REMOVE(wk); 2410 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk); 2411 continue; 2412 2413 case D_PAGEDEP: 2414 pagedep = WK_PAGEDEP(wk); 2415 /* 2416 * None of the directory additions will ever be 2417 * visible, so they can simply be tossed. 2418 */ 2419 for (i = 0; i < DAHASHSZ; i++) 2420 while ((dap = 2421 LIST_FIRST(&pagedep->pd_diraddhd[i]))) 2422 free_diradd(dap); 2423 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0) 2424 free_diradd(dap); 2425 /* 2426 * Copy any directory remove dependencies to the list 2427 * to be processed after the zero'ed inode is written. 2428 * If the inode has already been written, then they 2429 * can be dumped directly onto the work list. 2430 */ 2431 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 2432 LIST_REMOVE(dirrem, dm_next); 2433 dirrem->dm_dirinum = pagedep->pd_ino; 2434 if (inodedep == NULL || 2435 (inodedep->id_state & ALLCOMPLETE) == 2436 ALLCOMPLETE) 2437 add_to_worklist(&dirrem->dm_list); 2438 else 2439 WORKLIST_INSERT(&inodedep->id_bufwait, 2440 &dirrem->dm_list); 2441 } 2442 if ((pagedep->pd_state & NEWBLOCK) != 0) { 2443 LIST_FOREACH(wk, &inodedep->id_bufwait, wk_list) 2444 if (wk->wk_type == D_NEWDIRBLK && 2445 WK_NEWDIRBLK(wk)->db_pagedep == 2446 pagedep) 2447 break; 2448 if (wk != NULL) { 2449 WORKLIST_REMOVE(wk); 2450 free_newdirblk(WK_NEWDIRBLK(wk)); 2451 } else 2452 panic("deallocate_dependencies: " 2453 "lost pagedep"); 2454 } 2455 WORKLIST_REMOVE(&pagedep->pd_list); 2456 LIST_REMOVE(pagedep, pd_hash); 2457 WORKITEM_FREE(pagedep, D_PAGEDEP); 2458 continue; 2459 2460 case D_ALLOCINDIR: 2461 free_allocindir(WK_ALLOCINDIR(wk), inodedep); 2462 continue; 2463 2464 case D_ALLOCDIRECT: 2465 case D_INODEDEP: 2466 panic("deallocate_dependencies: Unexpected type %s", 2467 TYPENAME(wk->wk_type)); 2468 /* NOTREACHED */ 2469 2470 default: 2471 panic("deallocate_dependencies: Unknown type %s", 2472 TYPENAME(wk->wk_type)); 2473 /* NOTREACHED */ 2474 } 2475 } 2476} 2477 2478/* 2479 * Free an allocdirect. Generate a new freefrag work request if appropriate. 2480 * This routine must be called with splbio interrupts blocked. 2481 */ 2482static void 2483free_allocdirect(adphead, adp, delay) 2484 struct allocdirectlst *adphead; 2485 struct allocdirect *adp; 2486 int delay; 2487{ 2488 struct newdirblk *newdirblk; 2489 struct worklist *wk; 2490 2491 mtx_assert(&lk, MA_OWNED); 2492 if ((adp->ad_state & DEPCOMPLETE) == 0) 2493 LIST_REMOVE(adp, ad_deps); 2494 TAILQ_REMOVE(adphead, adp, ad_next); 2495 if ((adp->ad_state & COMPLETE) == 0) 2496 WORKLIST_REMOVE(&adp->ad_list); 2497 if (adp->ad_freefrag != NULL) { 2498 if (delay) 2499 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 2500 &adp->ad_freefrag->ff_list); 2501 else 2502 add_to_worklist(&adp->ad_freefrag->ff_list); 2503 } 2504 if ((wk = LIST_FIRST(&adp->ad_newdirblk)) != NULL) { 2505 newdirblk = WK_NEWDIRBLK(wk); 2506 WORKLIST_REMOVE(&newdirblk->db_list); 2507 if (LIST_FIRST(&adp->ad_newdirblk) != NULL) 2508 panic("free_allocdirect: extra newdirblk"); 2509 if (delay) 2510 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 2511 &newdirblk->db_list); 2512 else 2513 free_newdirblk(newdirblk); 2514 } 2515 WORKITEM_FREE(adp, D_ALLOCDIRECT); 2516} 2517 2518/* 2519 * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep. 2520 * This routine must be called with splbio interrupts blocked. 2521 */ 2522static void 2523free_newdirblk(newdirblk) 2524 struct newdirblk *newdirblk; 2525{ 2526 struct pagedep *pagedep; 2527 struct diradd *dap; 2528 int i; 2529 2530 mtx_assert(&lk, MA_OWNED); 2531 /* 2532 * If the pagedep is still linked onto the directory buffer 2533 * dependency chain, then some of the entries on the 2534 * pd_pendinghd list may not be committed to disk yet. In 2535 * this case, we will simply clear the NEWBLOCK flag and 2536 * let the pd_pendinghd list be processed when the pagedep 2537 * is next written. If the pagedep is no longer on the buffer 2538 * dependency chain, then all the entries on the pd_pending 2539 * list are committed to disk and we can free them here. 2540 */ 2541 pagedep = newdirblk->db_pagedep; 2542 pagedep->pd_state &= ~NEWBLOCK; 2543 if ((pagedep->pd_state & ONWORKLIST) == 0) 2544 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 2545 free_diradd(dap); 2546 /* 2547 * If no dependencies remain, the pagedep will be freed. 2548 */ 2549 for (i = 0; i < DAHASHSZ; i++) 2550 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL) 2551 break; 2552 if (i == DAHASHSZ && (pagedep->pd_state & ONWORKLIST) == 0) { 2553 LIST_REMOVE(pagedep, pd_hash); 2554 WORKITEM_FREE(pagedep, D_PAGEDEP); 2555 } 2556 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 2557} 2558 2559/* 2560 * Prepare an inode to be freed. The actual free operation is not 2561 * done until the zero'ed inode has been written to disk. 2562 */ 2563void 2564softdep_freefile(pvp, ino, mode) 2565 struct vnode *pvp; 2566 ino_t ino; 2567 int mode; 2568{ 2569 struct inode *ip = VTOI(pvp); 2570 struct inodedep *inodedep; 2571 struct freefile *freefile; 2572 2573 /* 2574 * This sets up the inode de-allocation dependency. 2575 */ 2576 MALLOC(freefile, struct freefile *, sizeof(struct freefile), 2577 M_FREEFILE, M_SOFTDEP_FLAGS); 2578 workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount); 2579 freefile->fx_mode = mode; 2580 freefile->fx_oldinum = ino; 2581 freefile->fx_devvp = ip->i_devvp; 2582 if ((ip->i_flag & IN_SPACECOUNTED) == 0) { 2583 UFS_LOCK(ip->i_ump); 2584 ip->i_fs->fs_pendinginodes += 1; 2585 UFS_UNLOCK(ip->i_ump); 2586 } 2587 2588 /* 2589 * If the inodedep does not exist, then the zero'ed inode has 2590 * been written to disk. If the allocated inode has never been 2591 * written to disk, then the on-disk inode is zero'ed. In either 2592 * case we can free the file immediately. 2593 */ 2594 ACQUIRE_LOCK(&lk); 2595 if (inodedep_lookup(pvp->v_mount, ino, 0, &inodedep) == 0 || 2596 check_inode_unwritten(inodedep)) { 2597 FREE_LOCK(&lk); 2598 handle_workitem_freefile(freefile); 2599 return; 2600 } 2601 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 2602 FREE_LOCK(&lk); 2603} 2604 2605/* 2606 * Check to see if an inode has never been written to disk. If 2607 * so free the inodedep and return success, otherwise return failure. 2608 * This routine must be called with splbio interrupts blocked. 2609 * 2610 * If we still have a bitmap dependency, then the inode has never 2611 * been written to disk. Drop the dependency as it is no longer 2612 * necessary since the inode is being deallocated. We set the 2613 * ALLCOMPLETE flags since the bitmap now properly shows that the 2614 * inode is not allocated. Even if the inode is actively being 2615 * written, it has been rolled back to its zero'ed state, so we 2616 * are ensured that a zero inode is what is on the disk. For short 2617 * lived files, this change will usually result in removing all the 2618 * dependencies from the inode so that it can be freed immediately. 2619 */ 2620static int 2621check_inode_unwritten(inodedep) 2622 struct inodedep *inodedep; 2623{ 2624 2625 mtx_assert(&lk, MA_OWNED); 2626 if ((inodedep->id_state & DEPCOMPLETE) != 0 || 2627 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2628 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2629 LIST_FIRST(&inodedep->id_inowait) != NULL || 2630 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2631 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2632 TAILQ_FIRST(&inodedep->id_extupdt) != NULL || 2633 TAILQ_FIRST(&inodedep->id_newextupdt) != NULL || 2634 inodedep->id_nlinkdelta != 0) 2635 return (0); 2636 2637 /* 2638 * Another process might be in initiate_write_inodeblock_ufs[12] 2639 * trying to allocate memory without holding "Softdep Lock". 2640 */ 2641 if ((inodedep->id_state & IOSTARTED) != 0 && 2642 inodedep->id_savedino1 == NULL) 2643 return (0); 2644 2645 inodedep->id_state |= ALLCOMPLETE; 2646 LIST_REMOVE(inodedep, id_deps); 2647 inodedep->id_buf = NULL; 2648 if (inodedep->id_state & ONWORKLIST) 2649 WORKLIST_REMOVE(&inodedep->id_list); 2650 if (inodedep->id_savedino1 != NULL) { 2651 FREE(inodedep->id_savedino1, M_SAVEDINO); 2652 inodedep->id_savedino1 = NULL; 2653 } 2654 if (free_inodedep(inodedep) == 0) 2655 panic("check_inode_unwritten: busy inode"); 2656 return (1); 2657} 2658 2659/* 2660 * Try to free an inodedep structure. Return 1 if it could be freed. 2661 */ 2662static int 2663free_inodedep(inodedep) 2664 struct inodedep *inodedep; 2665{ 2666 2667 mtx_assert(&lk, MA_OWNED); 2668 if ((inodedep->id_state & ONWORKLIST) != 0 || 2669 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 2670 LIST_FIRST(&inodedep->id_pendinghd) != NULL || 2671 LIST_FIRST(&inodedep->id_bufwait) != NULL || 2672 LIST_FIRST(&inodedep->id_inowait) != NULL || 2673 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 2674 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL || 2675 TAILQ_FIRST(&inodedep->id_extupdt) != NULL || 2676 TAILQ_FIRST(&inodedep->id_newextupdt) != NULL || 2677 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino1 != NULL) 2678 return (0); 2679 LIST_REMOVE(inodedep, id_hash); 2680 WORKITEM_FREE(inodedep, D_INODEDEP); 2681 num_inodedep -= 1; 2682 return (1); 2683} 2684 2685/* 2686 * This workitem routine performs the block de-allocation. 2687 * The workitem is added to the pending list after the updated 2688 * inode block has been written to disk. As mentioned above, 2689 * checks regarding the number of blocks de-allocated (compared 2690 * to the number of blocks allocated for the file) are also 2691 * performed in this function. 2692 */ 2693static void 2694handle_workitem_freeblocks(freeblks, flags) 2695 struct freeblks *freeblks; 2696 int flags; 2697{ 2698 struct inode *ip; 2699 struct vnode *vp; 2700 struct fs *fs; 2701 struct ufsmount *ump; 2702 int i, nblocks, level, bsize; 2703 ufs2_daddr_t bn, blocksreleased = 0; 2704 int error, allerror = 0; 2705 ufs_lbn_t baselbns[NIADDR], tmpval; 2706 int fs_pendingblocks; 2707 2708 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 2709 fs = ump->um_fs; 2710 fs_pendingblocks = 0; 2711 tmpval = 1; 2712 baselbns[0] = NDADDR; 2713 for (i = 1; i < NIADDR; i++) { 2714 tmpval *= NINDIR(fs); 2715 baselbns[i] = baselbns[i - 1] + tmpval; 2716 } 2717 nblocks = btodb(fs->fs_bsize); 2718 blocksreleased = 0; 2719 /* 2720 * Release all extended attribute blocks or frags. 2721 */ 2722 if (freeblks->fb_oldextsize > 0) { 2723 for (i = (NXADDR - 1); i >= 0; i--) { 2724 if ((bn = freeblks->fb_eblks[i]) == 0) 2725 continue; 2726 bsize = sblksize(fs, freeblks->fb_oldextsize, i); 2727 ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, bsize, 2728 freeblks->fb_previousinum); 2729 blocksreleased += btodb(bsize); 2730 } 2731 } 2732 /* 2733 * Release all data blocks or frags. 2734 */ 2735 if (freeblks->fb_oldsize > 0) { 2736 /* 2737 * Indirect blocks first. 2738 */ 2739 for (level = (NIADDR - 1); level >= 0; level--) { 2740 if ((bn = freeblks->fb_iblks[level]) == 0) 2741 continue; 2742 if ((error = indir_trunc(freeblks, fsbtodb(fs, bn), 2743 level, baselbns[level], &blocksreleased)) == 0) 2744 allerror = error; 2745 ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, 2746 fs->fs_bsize, freeblks->fb_previousinum); 2747 fs_pendingblocks += nblocks; 2748 blocksreleased += nblocks; 2749 } 2750 /* 2751 * All direct blocks or frags. 2752 */ 2753 for (i = (NDADDR - 1); i >= 0; i--) { 2754 if ((bn = freeblks->fb_dblks[i]) == 0) 2755 continue; 2756 bsize = sblksize(fs, freeblks->fb_oldsize, i); 2757 ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, bsize, 2758 freeblks->fb_previousinum); 2759 fs_pendingblocks += btodb(bsize); 2760 blocksreleased += btodb(bsize); 2761 } 2762 } 2763 UFS_LOCK(ump); 2764 fs->fs_pendingblocks -= fs_pendingblocks; 2765 UFS_UNLOCK(ump); 2766 /* 2767 * If we still have not finished background cleanup, then check 2768 * to see if the block count needs to be adjusted. 2769 */ 2770 if (freeblks->fb_chkcnt != blocksreleased && 2771 (fs->fs_flags & FS_UNCLEAN) != 0 && 2772 ffs_vget(freeblks->fb_list.wk_mp, freeblks->fb_previousinum, 2773 (flags & LK_NOWAIT) | LK_EXCLUSIVE, &vp) == 0) { 2774 ip = VTOI(vp); 2775 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + \ 2776 freeblks->fb_chkcnt - blocksreleased); 2777 ip->i_flag |= IN_CHANGE; 2778 vput(vp); 2779 } 2780 2781#ifdef DIAGNOSTIC 2782 if (freeblks->fb_chkcnt != blocksreleased && 2783 ((fs->fs_flags & FS_UNCLEAN) == 0 || (flags & LK_NOWAIT) != 0)) 2784 printf("handle_workitem_freeblocks: block count\n"); 2785 if (allerror) 2786 softdep_error("handle_workitem_freeblks", allerror); 2787#endif /* DIAGNOSTIC */ 2788 2789 ACQUIRE_LOCK(&lk); 2790 WORKITEM_FREE(freeblks, D_FREEBLKS); 2791 FREE_LOCK(&lk); 2792} 2793 2794/* 2795 * Release blocks associated with the inode ip and stored in the indirect 2796 * block dbn. If level is greater than SINGLE, the block is an indirect block 2797 * and recursive calls to indirtrunc must be used to cleanse other indirect 2798 * blocks. 2799 */ 2800static int 2801indir_trunc(freeblks, dbn, level, lbn, countp) 2802 struct freeblks *freeblks; 2803 ufs2_daddr_t dbn; 2804 int level; 2805 ufs_lbn_t lbn; 2806 ufs2_daddr_t *countp; 2807{ 2808 struct buf *bp; 2809 struct fs *fs; 2810 struct worklist *wk; 2811 struct indirdep *indirdep; 2812 struct ufsmount *ump; 2813 ufs1_daddr_t *bap1 = 0; 2814 ufs2_daddr_t nb, *bap2 = 0; 2815 ufs_lbn_t lbnadd; 2816 int i, nblocks, ufs1fmt; 2817 int error, allerror = 0; 2818 int fs_pendingblocks; 2819 2820 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 2821 fs = ump->um_fs; 2822 fs_pendingblocks = 0; 2823 lbnadd = 1; 2824 for (i = level; i > 0; i--) 2825 lbnadd *= NINDIR(fs); 2826 /* 2827 * Get buffer of block pointers to be freed. This routine is not 2828 * called until the zero'ed inode has been written, so it is safe 2829 * to free blocks as they are encountered. Because the inode has 2830 * been zero'ed, calls to bmap on these blocks will fail. So, we 2831 * have to use the on-disk address and the block device for the 2832 * filesystem to look them up. If the file was deleted before its 2833 * indirect blocks were all written to disk, the routine that set 2834 * us up (deallocate_dependencies) will have arranged to leave 2835 * a complete copy of the indirect block in memory for our use. 2836 * Otherwise we have to read the blocks in from the disk. 2837 */ 2838#ifdef notyet 2839 bp = getblk(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 0, 0, 2840 GB_NOCREAT); 2841#else 2842 bp = incore(&freeblks->fb_devvp->v_bufobj, dbn); 2843#endif 2844 ACQUIRE_LOCK(&lk); 2845 if (bp != NULL && (wk = LIST_FIRST(&bp->b_dep)) != NULL) { 2846 if (wk->wk_type != D_INDIRDEP || 2847 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp || 2848 (indirdep->ir_state & GOINGAWAY) == 0) 2849 panic("indir_trunc: lost indirdep"); 2850 WORKLIST_REMOVE(wk); 2851 WORKITEM_FREE(indirdep, D_INDIRDEP); 2852 if (LIST_FIRST(&bp->b_dep) != NULL) 2853 panic("indir_trunc: dangling dep"); 2854 ump->um_numindirdeps -= 1; 2855 FREE_LOCK(&lk); 2856 } else { 2857#ifdef notyet 2858 if (bp) 2859 brelse(bp); 2860#endif 2861 FREE_LOCK(&lk); 2862 error = bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 2863 NOCRED, &bp); 2864 if (error) { 2865 brelse(bp); 2866 return (error); 2867 } 2868 } 2869 /* 2870 * Recursively free indirect blocks. 2871 */ 2872 if (ump->um_fstype == UFS1) { 2873 ufs1fmt = 1; 2874 bap1 = (ufs1_daddr_t *)bp->b_data; 2875 } else { 2876 ufs1fmt = 0; 2877 bap2 = (ufs2_daddr_t *)bp->b_data; 2878 } 2879 nblocks = btodb(fs->fs_bsize); 2880 for (i = NINDIR(fs) - 1; i >= 0; i--) { 2881 if (ufs1fmt) 2882 nb = bap1[i]; 2883 else 2884 nb = bap2[i]; 2885 if (nb == 0) 2886 continue; 2887 if (level != 0) { 2888 if ((error = indir_trunc(freeblks, fsbtodb(fs, nb), 2889 level - 1, lbn + (i * lbnadd), countp)) != 0) 2890 allerror = error; 2891 } 2892 ffs_blkfree(ump, fs, freeblks->fb_devvp, nb, fs->fs_bsize, 2893 freeblks->fb_previousinum); 2894 fs_pendingblocks += nblocks; 2895 *countp += nblocks; 2896 } 2897 UFS_LOCK(ump); 2898 fs->fs_pendingblocks -= fs_pendingblocks; 2899 UFS_UNLOCK(ump); 2900 bp->b_flags |= B_INVAL | B_NOCACHE; 2901 brelse(bp); 2902 return (allerror); 2903} 2904 2905/* 2906 * Free an allocindir. 2907 * This routine must be called with splbio interrupts blocked. 2908 */ 2909static void 2910free_allocindir(aip, inodedep) 2911 struct allocindir *aip; 2912 struct inodedep *inodedep; 2913{ 2914 struct freefrag *freefrag; 2915 2916 mtx_assert(&lk, MA_OWNED); 2917 if ((aip->ai_state & DEPCOMPLETE) == 0) 2918 LIST_REMOVE(aip, ai_deps); 2919 if (aip->ai_state & ONWORKLIST) 2920 WORKLIST_REMOVE(&aip->ai_list); 2921 LIST_REMOVE(aip, ai_next); 2922 if ((freefrag = aip->ai_freefrag) != NULL) { 2923 if (inodedep == NULL) 2924 add_to_worklist(&freefrag->ff_list); 2925 else 2926 WORKLIST_INSERT(&inodedep->id_bufwait, 2927 &freefrag->ff_list); 2928 } 2929 WORKITEM_FREE(aip, D_ALLOCINDIR); 2930} 2931 2932/* 2933 * Directory entry addition dependencies. 2934 * 2935 * When adding a new directory entry, the inode (with its incremented link 2936 * count) must be written to disk before the directory entry's pointer to it. 2937 * Also, if the inode is newly allocated, the corresponding freemap must be 2938 * updated (on disk) before the directory entry's pointer. These requirements 2939 * are met via undo/redo on the directory entry's pointer, which consists 2940 * simply of the inode number. 2941 * 2942 * As directory entries are added and deleted, the free space within a 2943 * directory block can become fragmented. The ufs filesystem will compact 2944 * a fragmented directory block to make space for a new entry. When this 2945 * occurs, the offsets of previously added entries change. Any "diradd" 2946 * dependency structures corresponding to these entries must be updated with 2947 * the new offsets. 2948 */ 2949 2950/* 2951 * This routine is called after the in-memory inode's link 2952 * count has been incremented, but before the directory entry's 2953 * pointer to the inode has been set. 2954 */ 2955int 2956softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 2957 struct buf *bp; /* buffer containing directory block */ 2958 struct inode *dp; /* inode for directory */ 2959 off_t diroffset; /* offset of new entry in directory */ 2960 ino_t newinum; /* inode referenced by new directory entry */ 2961 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 2962 int isnewblk; /* entry is in a newly allocated block */ 2963{ 2964 int offset; /* offset of new entry within directory block */ 2965 ufs_lbn_t lbn; /* block in directory containing new entry */ 2966 struct fs *fs; 2967 struct diradd *dap; 2968 struct allocdirect *adp; 2969 struct pagedep *pagedep; 2970 struct inodedep *inodedep; 2971 struct newdirblk *newdirblk = 0; 2972 struct mkdir *mkdir1, *mkdir2; 2973 struct mount *mp; 2974 2975 /* 2976 * Whiteouts have no dependencies. 2977 */ 2978 if (newinum == WINO) { 2979 if (newdirbp != NULL) 2980 bdwrite(newdirbp); 2981 return (0); 2982 } 2983 mp = UFSTOVFS(dp->i_ump); 2984 fs = dp->i_fs; 2985 lbn = lblkno(fs, diroffset); 2986 offset = blkoff(fs, diroffset); 2987 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, 2988 M_SOFTDEP_FLAGS|M_ZERO); 2989 workitem_alloc(&dap->da_list, D_DIRADD, mp); 2990 dap->da_offset = offset; 2991 dap->da_newinum = newinum; 2992 dap->da_state = ATTACHED; 2993 if (isnewblk && lbn < NDADDR && fragoff(fs, diroffset) == 0) { 2994 MALLOC(newdirblk, struct newdirblk *, sizeof(struct newdirblk), 2995 M_NEWDIRBLK, M_SOFTDEP_FLAGS); 2996 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 2997 } 2998 if (newdirbp == NULL) { 2999 dap->da_state |= DEPCOMPLETE; 3000 ACQUIRE_LOCK(&lk); 3001 } else { 3002 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 3003 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 3004 M_SOFTDEP_FLAGS); 3005 workitem_alloc(&mkdir1->md_list, D_MKDIR, mp); 3006 mkdir1->md_state = MKDIR_BODY; 3007 mkdir1->md_diradd = dap; 3008 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR, 3009 M_SOFTDEP_FLAGS); 3010 workitem_alloc(&mkdir2->md_list, D_MKDIR, mp); 3011 mkdir2->md_state = MKDIR_PARENT; 3012 mkdir2->md_diradd = dap; 3013 /* 3014 * Dependency on "." and ".." being written to disk. 3015 */ 3016 mkdir1->md_buf = newdirbp; 3017 ACQUIRE_LOCK(&lk); 3018 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 3019 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list); 3020 FREE_LOCK(&lk); 3021 bdwrite(newdirbp); 3022 /* 3023 * Dependency on link count increase for parent directory 3024 */ 3025 ACQUIRE_LOCK(&lk); 3026 if (inodedep_lookup(mp, dp->i_number, 0, &inodedep) == 0 3027 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 3028 dap->da_state &= ~MKDIR_PARENT; 3029 WORKITEM_FREE(mkdir2, D_MKDIR); 3030 } else { 3031 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 3032 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 3033 } 3034 } 3035 /* 3036 * Link into parent directory pagedep to await its being written. 3037 */ 3038 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 3039 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 3040 dap->da_pagedep = pagedep; 3041 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 3042 da_pdlist); 3043 /* 3044 * Link into its inodedep. Put it on the id_bufwait list if the inode 3045 * is not yet written. If it is written, do the post-inode write 3046 * processing to put it on the id_pendinghd list. 3047 */ 3048 (void) inodedep_lookup(mp, newinum, DEPALLOC, &inodedep); 3049 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 3050 diradd_inode_written(dap, inodedep); 3051 else 3052 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 3053 if (isnewblk) { 3054 /* 3055 * Directories growing into indirect blocks are rare 3056 * enough and the frequency of new block allocation 3057 * in those cases even more rare, that we choose not 3058 * to bother tracking them. Rather we simply force the 3059 * new directory entry to disk. 3060 */ 3061 if (lbn >= NDADDR) { 3062 FREE_LOCK(&lk); 3063 /* 3064 * We only have a new allocation when at the 3065 * beginning of a new block, not when we are 3066 * expanding into an existing block. 3067 */ 3068 if (blkoff(fs, diroffset) == 0) 3069 return (1); 3070 return (0); 3071 } 3072 /* 3073 * We only have a new allocation when at the beginning 3074 * of a new fragment, not when we are expanding into an 3075 * existing fragment. Also, there is nothing to do if we 3076 * are already tracking this block. 3077 */ 3078 if (fragoff(fs, diroffset) != 0) { 3079 FREE_LOCK(&lk); 3080 return (0); 3081 } 3082 if ((pagedep->pd_state & NEWBLOCK) != 0) { 3083 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 3084 FREE_LOCK(&lk); 3085 return (0); 3086 } 3087 /* 3088 * Find our associated allocdirect and have it track us. 3089 */ 3090 if (inodedep_lookup(mp, dp->i_number, 0, &inodedep) == 0) 3091 panic("softdep_setup_directory_add: lost inodedep"); 3092 adp = TAILQ_LAST(&inodedep->id_newinoupdt, allocdirectlst); 3093 if (adp == NULL || adp->ad_lbn != lbn) 3094 panic("softdep_setup_directory_add: lost entry"); 3095 pagedep->pd_state |= NEWBLOCK; 3096 newdirblk->db_pagedep = pagedep; 3097 WORKLIST_INSERT(&adp->ad_newdirblk, &newdirblk->db_list); 3098 } 3099 FREE_LOCK(&lk); 3100 return (0); 3101} 3102 3103/* 3104 * This procedure is called to change the offset of a directory 3105 * entry when compacting a directory block which must be owned 3106 * exclusively by the caller. Note that the actual entry movement 3107 * must be done in this procedure to ensure that no I/O completions 3108 * occur while the move is in progress. 3109 */ 3110void 3111softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize) 3112 struct inode *dp; /* inode for directory */ 3113 caddr_t base; /* address of dp->i_offset */ 3114 caddr_t oldloc; /* address of old directory location */ 3115 caddr_t newloc; /* address of new directory location */ 3116 int entrysize; /* size of directory entry */ 3117{ 3118 int offset, oldoffset, newoffset; 3119 struct pagedep *pagedep; 3120 struct diradd *dap; 3121 ufs_lbn_t lbn; 3122 3123 ACQUIRE_LOCK(&lk); 3124 lbn = lblkno(dp->i_fs, dp->i_offset); 3125 offset = blkoff(dp->i_fs, dp->i_offset); 3126 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0) 3127 goto done; 3128 oldoffset = offset + (oldloc - base); 3129 newoffset = offset + (newloc - base); 3130 3131 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) { 3132 if (dap->da_offset != oldoffset) 3133 continue; 3134 dap->da_offset = newoffset; 3135 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset)) 3136 break; 3137 LIST_REMOVE(dap, da_pdlist); 3138 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)], 3139 dap, da_pdlist); 3140 break; 3141 } 3142 if (dap == NULL) { 3143 3144 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) { 3145 if (dap->da_offset == oldoffset) { 3146 dap->da_offset = newoffset; 3147 break; 3148 } 3149 } 3150 } 3151done: 3152 bcopy(oldloc, newloc, entrysize); 3153 FREE_LOCK(&lk); 3154} 3155 3156/* 3157 * Free a diradd dependency structure. This routine must be called 3158 * with splbio interrupts blocked. 3159 */ 3160static void 3161free_diradd(dap) 3162 struct diradd *dap; 3163{ 3164 struct dirrem *dirrem; 3165 struct pagedep *pagedep; 3166 struct inodedep *inodedep; 3167 struct mkdir *mkdir, *nextmd; 3168 3169 mtx_assert(&lk, MA_OWNED); 3170 WORKLIST_REMOVE(&dap->da_list); 3171 LIST_REMOVE(dap, da_pdlist); 3172 if ((dap->da_state & DIRCHG) == 0) { 3173 pagedep = dap->da_pagedep; 3174 } else { 3175 dirrem = dap->da_previous; 3176 pagedep = dirrem->dm_pagedep; 3177 dirrem->dm_dirinum = pagedep->pd_ino; 3178 add_to_worklist(&dirrem->dm_list); 3179 } 3180 if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum, 3181 0, &inodedep) != 0) 3182 (void) free_inodedep(inodedep); 3183 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 3184 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 3185 nextmd = LIST_NEXT(mkdir, md_mkdirs); 3186 if (mkdir->md_diradd != dap) 3187 continue; 3188 dap->da_state &= ~mkdir->md_state; 3189 WORKLIST_REMOVE(&mkdir->md_list); 3190 LIST_REMOVE(mkdir, md_mkdirs); 3191 WORKITEM_FREE(mkdir, D_MKDIR); 3192 } 3193 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 3194 panic("free_diradd: unfound ref"); 3195 } 3196 WORKITEM_FREE(dap, D_DIRADD); 3197} 3198 3199/* 3200 * Directory entry removal dependencies. 3201 * 3202 * When removing a directory entry, the entry's inode pointer must be 3203 * zero'ed on disk before the corresponding inode's link count is decremented 3204 * (possibly freeing the inode for re-use). This dependency is handled by 3205 * updating the directory entry but delaying the inode count reduction until 3206 * after the directory block has been written to disk. After this point, the 3207 * inode count can be decremented whenever it is convenient. 3208 */ 3209 3210/* 3211 * This routine should be called immediately after removing 3212 * a directory entry. The inode's link count should not be 3213 * decremented by the calling procedure -- the soft updates 3214 * code will do this task when it is safe. 3215 */ 3216void 3217softdep_setup_remove(bp, dp, ip, isrmdir) 3218 struct buf *bp; /* buffer containing directory block */ 3219 struct inode *dp; /* inode for the directory being modified */ 3220 struct inode *ip; /* inode for directory entry being removed */ 3221 int isrmdir; /* indicates if doing RMDIR */ 3222{ 3223 struct dirrem *dirrem, *prevdirrem; 3224 3225 /* 3226 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. 3227 */ 3228 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 3229 3230 /* 3231 * If the COMPLETE flag is clear, then there were no active 3232 * entries and we want to roll back to a zeroed entry until 3233 * the new inode is committed to disk. If the COMPLETE flag is 3234 * set then we have deleted an entry that never made it to 3235 * disk. If the entry we deleted resulted from a name change, 3236 * then the old name still resides on disk. We cannot delete 3237 * its inode (returned to us in prevdirrem) until the zeroed 3238 * directory entry gets to disk. The new inode has never been 3239 * referenced on the disk, so can be deleted immediately. 3240 */ 3241 if ((dirrem->dm_state & COMPLETE) == 0) { 3242 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 3243 dm_next); 3244 FREE_LOCK(&lk); 3245 } else { 3246 if (prevdirrem != NULL) 3247 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 3248 prevdirrem, dm_next); 3249 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 3250 FREE_LOCK(&lk); 3251 handle_workitem_remove(dirrem, NULL); 3252 } 3253} 3254 3255/* 3256 * Allocate a new dirrem if appropriate and return it along with 3257 * its associated pagedep. Called without a lock, returns with lock. 3258 */ 3259static long num_dirrem; /* number of dirrem allocated */ 3260static struct dirrem * 3261newdirrem(bp, dp, ip, isrmdir, prevdirremp) 3262 struct buf *bp; /* buffer containing directory block */ 3263 struct inode *dp; /* inode for the directory being modified */ 3264 struct inode *ip; /* inode for directory entry being removed */ 3265 int isrmdir; /* indicates if doing RMDIR */ 3266 struct dirrem **prevdirremp; /* previously referenced inode, if any */ 3267{ 3268 int offset; 3269 ufs_lbn_t lbn; 3270 struct diradd *dap; 3271 struct dirrem *dirrem; 3272 struct pagedep *pagedep; 3273 3274 /* 3275 * Whiteouts have no deletion dependencies. 3276 */ 3277 if (ip == NULL) 3278 panic("newdirrem: whiteout"); 3279 /* 3280 * If we are over our limit, try to improve the situation. 3281 * Limiting the number of dirrem structures will also limit 3282 * the number of freefile and freeblks structures. 3283 */ 3284 ACQUIRE_LOCK(&lk); 3285 if (num_dirrem > max_softdeps / 2) 3286 (void) request_cleanup(ITOV(dp)->v_mount, FLUSH_REMOVE); 3287 num_dirrem += 1; 3288 FREE_LOCK(&lk); 3289 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem), 3290 M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO); 3291 workitem_alloc(&dirrem->dm_list, D_DIRREM, ITOV(dp)->v_mount); 3292 dirrem->dm_state = isrmdir ? RMDIR : 0; 3293 dirrem->dm_oldinum = ip->i_number; 3294 *prevdirremp = NULL; 3295 3296 ACQUIRE_LOCK(&lk); 3297 lbn = lblkno(dp->i_fs, dp->i_offset); 3298 offset = blkoff(dp->i_fs, dp->i_offset); 3299 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0) 3300 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 3301 dirrem->dm_pagedep = pagedep; 3302 /* 3303 * Check for a diradd dependency for the same directory entry. 3304 * If present, then both dependencies become obsolete and can 3305 * be de-allocated. Check for an entry on both the pd_dirraddhd 3306 * list and the pd_pendinghd list. 3307 */ 3308 3309 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 3310 if (dap->da_offset == offset) 3311 break; 3312 if (dap == NULL) { 3313 3314 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 3315 if (dap->da_offset == offset) 3316 break; 3317 if (dap == NULL) 3318 return (dirrem); 3319 } 3320 /* 3321 * Must be ATTACHED at this point. 3322 */ 3323 if ((dap->da_state & ATTACHED) == 0) 3324 panic("newdirrem: not ATTACHED"); 3325 if (dap->da_newinum != ip->i_number) 3326 panic("newdirrem: inum %d should be %d", 3327 ip->i_number, dap->da_newinum); 3328 /* 3329 * If we are deleting a changed name that never made it to disk, 3330 * then return the dirrem describing the previous inode (which 3331 * represents the inode currently referenced from this entry on disk). 3332 */ 3333 if ((dap->da_state & DIRCHG) != 0) { 3334 *prevdirremp = dap->da_previous; 3335 dap->da_state &= ~DIRCHG; 3336 dap->da_pagedep = pagedep; 3337 } 3338 /* 3339 * We are deleting an entry that never made it to disk. 3340 * Mark it COMPLETE so we can delete its inode immediately. 3341 */ 3342 dirrem->dm_state |= COMPLETE; 3343 free_diradd(dap); 3344 return (dirrem); 3345} 3346 3347/* 3348 * Directory entry change dependencies. 3349 * 3350 * Changing an existing directory entry requires that an add operation 3351 * be completed first followed by a deletion. The semantics for the addition 3352 * are identical to the description of adding a new entry above except 3353 * that the rollback is to the old inode number rather than zero. Once 3354 * the addition dependency is completed, the removal is done as described 3355 * in the removal routine above. 3356 */ 3357 3358/* 3359 * This routine should be called immediately after changing 3360 * a directory entry. The inode's link count should not be 3361 * decremented by the calling procedure -- the soft updates 3362 * code will perform this task when it is safe. 3363 */ 3364void 3365softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 3366 struct buf *bp; /* buffer containing directory block */ 3367 struct inode *dp; /* inode for the directory being modified */ 3368 struct inode *ip; /* inode for directory entry being removed */ 3369 ino_t newinum; /* new inode number for changed entry */ 3370 int isrmdir; /* indicates if doing RMDIR */ 3371{ 3372 int offset; 3373 struct diradd *dap = NULL; 3374 struct dirrem *dirrem, *prevdirrem; 3375 struct pagedep *pagedep; 3376 struct inodedep *inodedep; 3377 struct mount *mp; 3378 3379 offset = blkoff(dp->i_fs, dp->i_offset); 3380 mp = UFSTOVFS(dp->i_ump); 3381 3382 /* 3383 * Whiteouts do not need diradd dependencies. 3384 */ 3385 if (newinum != WINO) { 3386 MALLOC(dap, struct diradd *, sizeof(struct diradd), 3387 M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO); 3388 workitem_alloc(&dap->da_list, D_DIRADD, mp); 3389 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 3390 dap->da_offset = offset; 3391 dap->da_newinum = newinum; 3392 } 3393 3394 /* 3395 * Allocate a new dirrem and ACQUIRE_LOCK. 3396 */ 3397 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 3398 pagedep = dirrem->dm_pagedep; 3399 /* 3400 * The possible values for isrmdir: 3401 * 0 - non-directory file rename 3402 * 1 - directory rename within same directory 3403 * inum - directory rename to new directory of given inode number 3404 * When renaming to a new directory, we are both deleting and 3405 * creating a new directory entry, so the link count on the new 3406 * directory should not change. Thus we do not need the followup 3407 * dirrem which is usually done in handle_workitem_remove. We set 3408 * the DIRCHG flag to tell handle_workitem_remove to skip the 3409 * followup dirrem. 3410 */ 3411 if (isrmdir > 1) 3412 dirrem->dm_state |= DIRCHG; 3413 3414 /* 3415 * Whiteouts have no additional dependencies, 3416 * so just put the dirrem on the correct list. 3417 */ 3418 if (newinum == WINO) { 3419 if ((dirrem->dm_state & COMPLETE) == 0) { 3420 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 3421 dm_next); 3422 } else { 3423 dirrem->dm_dirinum = pagedep->pd_ino; 3424 add_to_worklist(&dirrem->dm_list); 3425 } 3426 FREE_LOCK(&lk); 3427 return; 3428 } 3429 3430 /* 3431 * If the COMPLETE flag is clear, then there were no active 3432 * entries and we want to roll back to the previous inode until 3433 * the new inode is committed to disk. If the COMPLETE flag is 3434 * set, then we have deleted an entry that never made it to disk. 3435 * If the entry we deleted resulted from a name change, then the old 3436 * inode reference still resides on disk. Any rollback that we do 3437 * needs to be to that old inode (returned to us in prevdirrem). If 3438 * the entry we deleted resulted from a create, then there is 3439 * no entry on the disk, so we want to roll back to zero rather 3440 * than the uncommitted inode. In either of the COMPLETE cases we 3441 * want to immediately free the unwritten and unreferenced inode. 3442 */ 3443 if ((dirrem->dm_state & COMPLETE) == 0) { 3444 dap->da_previous = dirrem; 3445 } else { 3446 if (prevdirrem != NULL) { 3447 dap->da_previous = prevdirrem; 3448 } else { 3449 dap->da_state &= ~DIRCHG; 3450 dap->da_pagedep = pagedep; 3451 } 3452 dirrem->dm_dirinum = pagedep->pd_ino; 3453 add_to_worklist(&dirrem->dm_list); 3454 } 3455 /* 3456 * Link into its inodedep. Put it on the id_bufwait list if the inode 3457 * is not yet written. If it is written, do the post-inode write 3458 * processing to put it on the id_pendinghd list. 3459 */ 3460 if (inodedep_lookup(mp, newinum, DEPALLOC, &inodedep) == 0 || 3461 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 3462 dap->da_state |= COMPLETE; 3463 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 3464 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 3465 } else { 3466 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 3467 dap, da_pdlist); 3468 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 3469 } 3470 FREE_LOCK(&lk); 3471} 3472 3473/* 3474 * Called whenever the link count on an inode is changed. 3475 * It creates an inode dependency so that the new reference(s) 3476 * to the inode cannot be committed to disk until the updated 3477 * inode has been written. 3478 */ 3479void 3480softdep_change_linkcnt(ip) 3481 struct inode *ip; /* the inode with the increased link count */ 3482{ 3483 struct inodedep *inodedep; 3484 3485 ACQUIRE_LOCK(&lk); 3486 (void) inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 3487 DEPALLOC, &inodedep); 3488 if (ip->i_nlink < ip->i_effnlink) 3489 panic("softdep_change_linkcnt: bad delta"); 3490 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3491 FREE_LOCK(&lk); 3492} 3493 3494/* 3495 * Called when the effective link count and the reference count 3496 * on an inode drops to zero. At this point there are no names 3497 * referencing the file in the filesystem and no active file 3498 * references. The space associated with the file will be freed 3499 * as soon as the necessary soft dependencies are cleared. 3500 */ 3501void 3502softdep_releasefile(ip) 3503 struct inode *ip; /* inode with the zero effective link count */ 3504{ 3505 struct inodedep *inodedep; 3506 struct fs *fs; 3507 int extblocks; 3508 3509 if (ip->i_effnlink > 0) 3510 panic("softdep_filerelease: file still referenced"); 3511 /* 3512 * We may be called several times as the real reference count 3513 * drops to zero. We only want to account for the space once. 3514 */ 3515 if (ip->i_flag & IN_SPACECOUNTED) 3516 return; 3517 /* 3518 * We have to deactivate a snapshot otherwise copyonwrites may 3519 * add blocks and the cleanup may remove blocks after we have 3520 * tried to account for them. 3521 */ 3522 if ((ip->i_flags & SF_SNAPSHOT) != 0) 3523 ffs_snapremove(ITOV(ip)); 3524 /* 3525 * If we are tracking an nlinkdelta, we have to also remember 3526 * whether we accounted for the freed space yet. 3527 */ 3528 ACQUIRE_LOCK(&lk); 3529 if ((inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, &inodedep))) 3530 inodedep->id_state |= SPACECOUNTED; 3531 FREE_LOCK(&lk); 3532 fs = ip->i_fs; 3533 extblocks = 0; 3534 if (fs->fs_magic == FS_UFS2_MAGIC) 3535 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 3536 UFS_LOCK(ip->i_ump); 3537 ip->i_fs->fs_pendingblocks += DIP(ip, i_blocks) - extblocks; 3538 ip->i_fs->fs_pendinginodes += 1; 3539 UFS_UNLOCK(ip->i_ump); 3540 ip->i_flag |= IN_SPACECOUNTED; 3541} 3542 3543/* 3544 * This workitem decrements the inode's link count. 3545 * If the link count reaches zero, the file is removed. 3546 */ 3547static void 3548handle_workitem_remove(dirrem, xp) 3549 struct dirrem *dirrem; 3550 struct vnode *xp; 3551{ 3552 struct thread *td = curthread; 3553 struct inodedep *inodedep; 3554 struct vnode *vp; 3555 struct inode *ip; 3556 ino_t oldinum; 3557 int error; 3558 3559 if ((vp = xp) == NULL && 3560 (error = ffs_vget(dirrem->dm_list.wk_mp, 3561 dirrem->dm_oldinum, LK_EXCLUSIVE, &vp)) != 0) { 3562 softdep_error("handle_workitem_remove: vget", error); 3563 return; 3564 } 3565 ip = VTOI(vp); 3566 ACQUIRE_LOCK(&lk); 3567 if ((inodedep_lookup(dirrem->dm_list.wk_mp, 3568 dirrem->dm_oldinum, 0, &inodedep)) == 0) 3569 panic("handle_workitem_remove: lost inodedep"); 3570 /* 3571 * Normal file deletion. 3572 */ 3573 if ((dirrem->dm_state & RMDIR) == 0) { 3574 ip->i_nlink--; 3575 DIP_SET(ip, i_nlink, ip->i_nlink); 3576 ip->i_flag |= IN_CHANGE; 3577 if (ip->i_nlink < ip->i_effnlink) 3578 panic("handle_workitem_remove: bad file delta"); 3579 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3580 num_dirrem -= 1; 3581 WORKITEM_FREE(dirrem, D_DIRREM); 3582 FREE_LOCK(&lk); 3583 vput(vp); 3584 return; 3585 } 3586 /* 3587 * Directory deletion. Decrement reference count for both the 3588 * just deleted parent directory entry and the reference for ".". 3589 * Next truncate the directory to length zero. When the 3590 * truncation completes, arrange to have the reference count on 3591 * the parent decremented to account for the loss of "..". 3592 */ 3593 ip->i_nlink -= 2; 3594 DIP_SET(ip, i_nlink, ip->i_nlink); 3595 ip->i_flag |= IN_CHANGE; 3596 if (ip->i_nlink < ip->i_effnlink) 3597 panic("handle_workitem_remove: bad dir delta"); 3598 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3599 FREE_LOCK(&lk); 3600 if ((error = ffs_truncate(vp, (off_t)0, 0, td->td_ucred, td)) != 0) 3601 softdep_error("handle_workitem_remove: truncate", error); 3602 ACQUIRE_LOCK(&lk); 3603 /* 3604 * Rename a directory to a new parent. Since, we are both deleting 3605 * and creating a new directory entry, the link count on the new 3606 * directory should not change. Thus we skip the followup dirrem. 3607 */ 3608 if (dirrem->dm_state & DIRCHG) { 3609 num_dirrem -= 1; 3610 WORKITEM_FREE(dirrem, D_DIRREM); 3611 FREE_LOCK(&lk); 3612 vput(vp); 3613 return; 3614 } 3615 /* 3616 * If the inodedep does not exist, then the zero'ed inode has 3617 * been written to disk. If the allocated inode has never been 3618 * written to disk, then the on-disk inode is zero'ed. In either 3619 * case we can remove the file immediately. 3620 */ 3621 dirrem->dm_state = 0; 3622 oldinum = dirrem->dm_oldinum; 3623 dirrem->dm_oldinum = dirrem->dm_dirinum; 3624 if (inodedep_lookup(dirrem->dm_list.wk_mp, oldinum, 3625 0, &inodedep) == 0 || check_inode_unwritten(inodedep)) { 3626 FREE_LOCK(&lk); 3627 vput(vp); 3628 handle_workitem_remove(dirrem, NULL); 3629 return; 3630 } 3631 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 3632 FREE_LOCK(&lk); 3633 ip->i_flag |= IN_CHANGE; 3634 ffs_update(vp, 0); 3635 vput(vp); 3636} 3637 3638/* 3639 * Inode de-allocation dependencies. 3640 * 3641 * When an inode's link count is reduced to zero, it can be de-allocated. We 3642 * found it convenient to postpone de-allocation until after the inode is 3643 * written to disk with its new link count (zero). At this point, all of the 3644 * on-disk inode's block pointers are nullified and, with careful dependency 3645 * list ordering, all dependencies related to the inode will be satisfied and 3646 * the corresponding dependency structures de-allocated. So, if/when the 3647 * inode is reused, there will be no mixing of old dependencies with new 3648 * ones. This artificial dependency is set up by the block de-allocation 3649 * procedure above (softdep_setup_freeblocks) and completed by the 3650 * following procedure. 3651 */ 3652static void 3653handle_workitem_freefile(freefile) 3654 struct freefile *freefile; 3655{ 3656 struct fs *fs; 3657 struct inodedep *idp; 3658 struct ufsmount *ump; 3659 int error; 3660 3661 ump = VFSTOUFS(freefile->fx_list.wk_mp); 3662 fs = ump->um_fs; 3663#ifdef DEBUG 3664 ACQUIRE_LOCK(&lk); 3665 error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp); 3666 FREE_LOCK(&lk); 3667 if (error) 3668 panic("handle_workitem_freefile: inodedep survived"); 3669#endif 3670 UFS_LOCK(ump); 3671 fs->fs_pendinginodes -= 1; 3672 UFS_UNLOCK(ump); 3673 if ((error = ffs_freefile(ump, fs, freefile->fx_devvp, 3674 freefile->fx_oldinum, freefile->fx_mode)) != 0) 3675 softdep_error("handle_workitem_freefile", error); 3676 ACQUIRE_LOCK(&lk); 3677 WORKITEM_FREE(freefile, D_FREEFILE); 3678 FREE_LOCK(&lk); 3679} 3680 3681 3682/* 3683 * Helper function which unlinks marker element from work list and returns 3684 * the next element on the list. 3685 */ 3686static __inline struct worklist * 3687markernext(struct worklist *marker) 3688{ 3689 struct worklist *next; 3690 3691 next = LIST_NEXT(marker, wk_list); 3692 LIST_REMOVE(marker, wk_list); 3693 return next; 3694} 3695 3696/* 3697 * Disk writes. 3698 * 3699 * The dependency structures constructed above are most actively used when file 3700 * system blocks are written to disk. No constraints are placed on when a 3701 * block can be written, but unsatisfied update dependencies are made safe by 3702 * modifying (or replacing) the source memory for the duration of the disk 3703 * write. When the disk write completes, the memory block is again brought 3704 * up-to-date. 3705 * 3706 * In-core inode structure reclamation. 3707 * 3708 * Because there are a finite number of "in-core" inode structures, they are 3709 * reused regularly. By transferring all inode-related dependencies to the 3710 * in-memory inode block and indexing them separately (via "inodedep"s), we 3711 * can allow "in-core" inode structures to be reused at any time and avoid 3712 * any increase in contention. 3713 * 3714 * Called just before entering the device driver to initiate a new disk I/O. 3715 * The buffer must be locked, thus, no I/O completion operations can occur 3716 * while we are manipulating its associated dependencies. 3717 */ 3718static void 3719softdep_disk_io_initiation(bp) 3720 struct buf *bp; /* structure describing disk write to occur */ 3721{ 3722 struct worklist *wk; 3723 struct worklist marker; 3724 struct indirdep *indirdep; 3725 struct inodedep *inodedep; 3726 3727 /* 3728 * We only care about write operations. There should never 3729 * be dependencies for reads. 3730 */ 3731 if (bp->b_iocmd != BIO_WRITE) 3732 panic("softdep_disk_io_initiation: not write"); 3733 3734 marker.wk_type = D_LAST + 1; /* Not a normal workitem */ 3735 PHOLD(curproc); /* Don't swap out kernel stack */ 3736 3737 ACQUIRE_LOCK(&lk); 3738 /* 3739 * Do any necessary pre-I/O processing. 3740 */ 3741 for (wk = LIST_FIRST(&bp->b_dep); wk != NULL; 3742 wk = markernext(&marker)) { 3743 LIST_INSERT_AFTER(wk, &marker, wk_list); 3744 switch (wk->wk_type) { 3745 3746 case D_PAGEDEP: 3747 initiate_write_filepage(WK_PAGEDEP(wk), bp); 3748 continue; 3749 3750 case D_INODEDEP: 3751 inodedep = WK_INODEDEP(wk); 3752 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) 3753 initiate_write_inodeblock_ufs1(inodedep, bp); 3754 else 3755 initiate_write_inodeblock_ufs2(inodedep, bp); 3756 continue; 3757 3758 case D_INDIRDEP: 3759 indirdep = WK_INDIRDEP(wk); 3760 if (indirdep->ir_state & GOINGAWAY) 3761 panic("disk_io_initiation: indirdep gone"); 3762 /* 3763 * If there are no remaining dependencies, this 3764 * will be writing the real pointers, so the 3765 * dependency can be freed. 3766 */ 3767 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) { 3768 struct buf *bp; 3769 3770 bp = indirdep->ir_savebp; 3771 bp->b_flags |= B_INVAL | B_NOCACHE; 3772 /* inline expand WORKLIST_REMOVE(wk); */ 3773 wk->wk_state &= ~ONWORKLIST; 3774 LIST_REMOVE(wk, wk_list); 3775 WORKITEM_FREE(indirdep, D_INDIRDEP); 3776 FREE_LOCK(&lk); 3777 brelse(bp); 3778 ACQUIRE_LOCK(&lk); 3779 continue; 3780 } 3781 /* 3782 * Replace up-to-date version with safe version. 3783 */ 3784 FREE_LOCK(&lk); 3785 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount, 3786 M_INDIRDEP, M_SOFTDEP_FLAGS); 3787 ACQUIRE_LOCK(&lk); 3788 indirdep->ir_state &= ~ATTACHED; 3789 indirdep->ir_state |= UNDONE; 3790 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 3791 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 3792 bp->b_bcount); 3793 continue; 3794 3795 case D_MKDIR: 3796 case D_BMSAFEMAP: 3797 case D_ALLOCDIRECT: 3798 case D_ALLOCINDIR: 3799 continue; 3800 3801 default: 3802 panic("handle_disk_io_initiation: Unexpected type %s", 3803 TYPENAME(wk->wk_type)); 3804 /* NOTREACHED */ 3805 } 3806 } 3807 FREE_LOCK(&lk); 3808 PRELE(curproc); /* Allow swapout of kernel stack */ 3809} 3810 3811/* 3812 * Called from within the procedure above to deal with unsatisfied 3813 * allocation dependencies in a directory. The buffer must be locked, 3814 * thus, no I/O completion operations can occur while we are 3815 * manipulating its associated dependencies. 3816 */ 3817static void 3818initiate_write_filepage(pagedep, bp) 3819 struct pagedep *pagedep; 3820 struct buf *bp; 3821{ 3822 struct diradd *dap; 3823 struct direct *ep; 3824 int i; 3825 3826 if (pagedep->pd_state & IOSTARTED) { 3827 /* 3828 * This can only happen if there is a driver that does not 3829 * understand chaining. Here biodone will reissue the call 3830 * to strategy for the incomplete buffers. 3831 */ 3832 printf("initiate_write_filepage: already started\n"); 3833 return; 3834 } 3835 pagedep->pd_state |= IOSTARTED; 3836 for (i = 0; i < DAHASHSZ; i++) { 3837 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 3838 ep = (struct direct *) 3839 ((char *)bp->b_data + dap->da_offset); 3840 if (ep->d_ino != dap->da_newinum) 3841 panic("%s: dir inum %d != new %d", 3842 "initiate_write_filepage", 3843 ep->d_ino, dap->da_newinum); 3844 if (dap->da_state & DIRCHG) 3845 ep->d_ino = dap->da_previous->dm_oldinum; 3846 else 3847 ep->d_ino = 0; 3848 dap->da_state &= ~ATTACHED; 3849 dap->da_state |= UNDONE; 3850 } 3851 } 3852} 3853 3854/* 3855 * Version of initiate_write_inodeblock that handles UFS1 dinodes. 3856 * Note that any bug fixes made to this routine must be done in the 3857 * version found below. 3858 * 3859 * Called from within the procedure above to deal with unsatisfied 3860 * allocation dependencies in an inodeblock. The buffer must be 3861 * locked, thus, no I/O completion operations can occur while we 3862 * are manipulating its associated dependencies. 3863 */ 3864static void 3865initiate_write_inodeblock_ufs1(inodedep, bp) 3866 struct inodedep *inodedep; 3867 struct buf *bp; /* The inode block */ 3868{ 3869 struct allocdirect *adp, *lastadp; 3870 struct ufs1_dinode *dp; 3871 struct ufs1_dinode *sip; 3872 struct fs *fs; 3873 ufs_lbn_t i, prevlbn = 0; 3874 int deplist; 3875 3876 if (inodedep->id_state & IOSTARTED) 3877 panic("initiate_write_inodeblock_ufs1: already started"); 3878 inodedep->id_state |= IOSTARTED; 3879 fs = inodedep->id_fs; 3880 dp = (struct ufs1_dinode *)bp->b_data + 3881 ino_to_fsbo(fs, inodedep->id_ino); 3882 /* 3883 * If the bitmap is not yet written, then the allocated 3884 * inode cannot be written to disk. 3885 */ 3886 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 3887 if (inodedep->id_savedino1 != NULL) 3888 panic("initiate_write_inodeblock_ufs1: I/O underway"); 3889 FREE_LOCK(&lk); 3890 MALLOC(sip, struct ufs1_dinode *, 3891 sizeof(struct ufs1_dinode), M_SAVEDINO, M_SOFTDEP_FLAGS); 3892 ACQUIRE_LOCK(&lk); 3893 inodedep->id_savedino1 = sip; 3894 *inodedep->id_savedino1 = *dp; 3895 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 3896 dp->di_gen = inodedep->id_savedino1->di_gen; 3897 return; 3898 } 3899 /* 3900 * If no dependencies, then there is nothing to roll back. 3901 */ 3902 inodedep->id_savedsize = dp->di_size; 3903 inodedep->id_savedextsize = 0; 3904 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL) 3905 return; 3906 /* 3907 * Set the dependencies to busy. 3908 */ 3909 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3910 adp = TAILQ_NEXT(adp, ad_next)) { 3911#ifdef DIAGNOSTIC 3912 if (deplist != 0 && prevlbn >= adp->ad_lbn) 3913 panic("softdep_write_inodeblock: lbn order"); 3914 prevlbn = adp->ad_lbn; 3915 if (adp->ad_lbn < NDADDR && 3916 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) 3917 panic("%s: direct pointer #%jd mismatch %d != %jd", 3918 "softdep_write_inodeblock", 3919 (intmax_t)adp->ad_lbn, 3920 dp->di_db[adp->ad_lbn], 3921 (intmax_t)adp->ad_newblkno); 3922 if (adp->ad_lbn >= NDADDR && 3923 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) 3924 panic("%s: indirect pointer #%jd mismatch %d != %jd", 3925 "softdep_write_inodeblock", 3926 (intmax_t)adp->ad_lbn - NDADDR, 3927 dp->di_ib[adp->ad_lbn - NDADDR], 3928 (intmax_t)adp->ad_newblkno); 3929 deplist |= 1 << adp->ad_lbn; 3930 if ((adp->ad_state & ATTACHED) == 0) 3931 panic("softdep_write_inodeblock: Unknown state 0x%x", 3932 adp->ad_state); 3933#endif /* DIAGNOSTIC */ 3934 adp->ad_state &= ~ATTACHED; 3935 adp->ad_state |= UNDONE; 3936 } 3937 /* 3938 * The on-disk inode cannot claim to be any larger than the last 3939 * fragment that has been written. Otherwise, the on-disk inode 3940 * might have fragments that were not the last block in the file 3941 * which would corrupt the filesystem. 3942 */ 3943 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 3944 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 3945 if (adp->ad_lbn >= NDADDR) 3946 break; 3947 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 3948 /* keep going until hitting a rollback to a frag */ 3949 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 3950 continue; 3951 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 3952 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 3953#ifdef DIAGNOSTIC 3954 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 3955 panic("softdep_write_inodeblock: lost dep1"); 3956#endif /* DIAGNOSTIC */ 3957 dp->di_db[i] = 0; 3958 } 3959 for (i = 0; i < NIADDR; i++) { 3960#ifdef DIAGNOSTIC 3961 if (dp->di_ib[i] != 0 && 3962 (deplist & ((1 << NDADDR) << i)) == 0) 3963 panic("softdep_write_inodeblock: lost dep2"); 3964#endif /* DIAGNOSTIC */ 3965 dp->di_ib[i] = 0; 3966 } 3967 return; 3968 } 3969 /* 3970 * If we have zero'ed out the last allocated block of the file, 3971 * roll back the size to the last currently allocated block. 3972 * We know that this last allocated block is a full-sized as 3973 * we already checked for fragments in the loop above. 3974 */ 3975 if (lastadp != NULL && 3976 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 3977 for (i = lastadp->ad_lbn; i >= 0; i--) 3978 if (dp->di_db[i] != 0) 3979 break; 3980 dp->di_size = (i + 1) * fs->fs_bsize; 3981 } 3982 /* 3983 * The only dependencies are for indirect blocks. 3984 * 3985 * The file size for indirect block additions is not guaranteed. 3986 * Such a guarantee would be non-trivial to achieve. The conventional 3987 * synchronous write implementation also does not make this guarantee. 3988 * Fsck should catch and fix discrepancies. Arguably, the file size 3989 * can be over-estimated without destroying integrity when the file 3990 * moves into the indirect blocks (i.e., is large). If we want to 3991 * postpone fsck, we are stuck with this argument. 3992 */ 3993 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 3994 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 3995} 3996 3997/* 3998 * Version of initiate_write_inodeblock that handles UFS2 dinodes. 3999 * Note that any bug fixes made to this routine must be done in the 4000 * version found above. 4001 * 4002 * Called from within the procedure above to deal with unsatisfied 4003 * allocation dependencies in an inodeblock. The buffer must be 4004 * locked, thus, no I/O completion operations can occur while we 4005 * are manipulating its associated dependencies. 4006 */ 4007static void 4008initiate_write_inodeblock_ufs2(inodedep, bp) 4009 struct inodedep *inodedep; 4010 struct buf *bp; /* The inode block */ 4011{ 4012 struct allocdirect *adp, *lastadp; 4013 struct ufs2_dinode *dp; 4014 struct ufs2_dinode *sip; 4015 struct fs *fs; 4016 ufs_lbn_t i, prevlbn = 0; 4017 int deplist; 4018 4019 if (inodedep->id_state & IOSTARTED) 4020 panic("initiate_write_inodeblock_ufs2: already started"); 4021 inodedep->id_state |= IOSTARTED; 4022 fs = inodedep->id_fs; 4023 dp = (struct ufs2_dinode *)bp->b_data + 4024 ino_to_fsbo(fs, inodedep->id_ino); 4025 /* 4026 * If the bitmap is not yet written, then the allocated 4027 * inode cannot be written to disk. 4028 */ 4029 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 4030 if (inodedep->id_savedino2 != NULL) 4031 panic("initiate_write_inodeblock_ufs2: I/O underway"); 4032 FREE_LOCK(&lk); 4033 MALLOC(sip, struct ufs2_dinode *, 4034 sizeof(struct ufs2_dinode), M_SAVEDINO, M_SOFTDEP_FLAGS); 4035 ACQUIRE_LOCK(&lk); 4036 inodedep->id_savedino2 = sip; 4037 *inodedep->id_savedino2 = *dp; 4038 bzero((caddr_t)dp, sizeof(struct ufs2_dinode)); 4039 dp->di_gen = inodedep->id_savedino2->di_gen; 4040 return; 4041 } 4042 /* 4043 * If no dependencies, then there is nothing to roll back. 4044 */ 4045 inodedep->id_savedsize = dp->di_size; 4046 inodedep->id_savedextsize = dp->di_extsize; 4047 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL && 4048 TAILQ_FIRST(&inodedep->id_extupdt) == NULL) 4049 return; 4050 /* 4051 * Set the ext data dependencies to busy. 4052 */ 4053 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 4054 adp = TAILQ_NEXT(adp, ad_next)) { 4055#ifdef DIAGNOSTIC 4056 if (deplist != 0 && prevlbn >= adp->ad_lbn) 4057 panic("softdep_write_inodeblock: lbn order"); 4058 prevlbn = adp->ad_lbn; 4059 if (dp->di_extb[adp->ad_lbn] != adp->ad_newblkno) 4060 panic("%s: direct pointer #%jd mismatch %jd != %jd", 4061 "softdep_write_inodeblock", 4062 (intmax_t)adp->ad_lbn, 4063 (intmax_t)dp->di_extb[adp->ad_lbn], 4064 (intmax_t)adp->ad_newblkno); 4065 deplist |= 1 << adp->ad_lbn; 4066 if ((adp->ad_state & ATTACHED) == 0) 4067 panic("softdep_write_inodeblock: Unknown state 0x%x", 4068 adp->ad_state); 4069#endif /* DIAGNOSTIC */ 4070 adp->ad_state &= ~ATTACHED; 4071 adp->ad_state |= UNDONE; 4072 } 4073 /* 4074 * The on-disk inode cannot claim to be any larger than the last 4075 * fragment that has been written. Otherwise, the on-disk inode 4076 * might have fragments that were not the last block in the ext 4077 * data which would corrupt the filesystem. 4078 */ 4079 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 4080 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 4081 dp->di_extb[adp->ad_lbn] = adp->ad_oldblkno; 4082 /* keep going until hitting a rollback to a frag */ 4083 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 4084 continue; 4085 dp->di_extsize = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 4086 for (i = adp->ad_lbn + 1; i < NXADDR; i++) { 4087#ifdef DIAGNOSTIC 4088 if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0) 4089 panic("softdep_write_inodeblock: lost dep1"); 4090#endif /* DIAGNOSTIC */ 4091 dp->di_extb[i] = 0; 4092 } 4093 lastadp = NULL; 4094 break; 4095 } 4096 /* 4097 * If we have zero'ed out the last allocated block of the ext 4098 * data, roll back the size to the last currently allocated block. 4099 * We know that this last allocated block is a full-sized as 4100 * we already checked for fragments in the loop above. 4101 */ 4102 if (lastadp != NULL && 4103 dp->di_extsize <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 4104 for (i = lastadp->ad_lbn; i >= 0; i--) 4105 if (dp->di_extb[i] != 0) 4106 break; 4107 dp->di_extsize = (i + 1) * fs->fs_bsize; 4108 } 4109 /* 4110 * Set the file data dependencies to busy. 4111 */ 4112 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 4113 adp = TAILQ_NEXT(adp, ad_next)) { 4114#ifdef DIAGNOSTIC 4115 if (deplist != 0 && prevlbn >= adp->ad_lbn) 4116 panic("softdep_write_inodeblock: lbn order"); 4117 prevlbn = adp->ad_lbn; 4118 if (adp->ad_lbn < NDADDR && 4119 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) 4120 panic("%s: direct pointer #%jd mismatch %jd != %jd", 4121 "softdep_write_inodeblock", 4122 (intmax_t)adp->ad_lbn, 4123 (intmax_t)dp->di_db[adp->ad_lbn], 4124 (intmax_t)adp->ad_newblkno); 4125 if (adp->ad_lbn >= NDADDR && 4126 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) 4127 panic("%s indirect pointer #%jd mismatch %jd != %jd", 4128 "softdep_write_inodeblock:", 4129 (intmax_t)adp->ad_lbn - NDADDR, 4130 (intmax_t)dp->di_ib[adp->ad_lbn - NDADDR], 4131 (intmax_t)adp->ad_newblkno); 4132 deplist |= 1 << adp->ad_lbn; 4133 if ((adp->ad_state & ATTACHED) == 0) 4134 panic("softdep_write_inodeblock: Unknown state 0x%x", 4135 adp->ad_state); 4136#endif /* DIAGNOSTIC */ 4137 adp->ad_state &= ~ATTACHED; 4138 adp->ad_state |= UNDONE; 4139 } 4140 /* 4141 * The on-disk inode cannot claim to be any larger than the last 4142 * fragment that has been written. Otherwise, the on-disk inode 4143 * might have fragments that were not the last block in the file 4144 * which would corrupt the filesystem. 4145 */ 4146 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 4147 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 4148 if (adp->ad_lbn >= NDADDR) 4149 break; 4150 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno; 4151 /* keep going until hitting a rollback to a frag */ 4152 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 4153 continue; 4154 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize; 4155 for (i = adp->ad_lbn + 1; i < NDADDR; i++) { 4156#ifdef DIAGNOSTIC 4157 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 4158 panic("softdep_write_inodeblock: lost dep2"); 4159#endif /* DIAGNOSTIC */ 4160 dp->di_db[i] = 0; 4161 } 4162 for (i = 0; i < NIADDR; i++) { 4163#ifdef DIAGNOSTIC 4164 if (dp->di_ib[i] != 0 && 4165 (deplist & ((1 << NDADDR) << i)) == 0) 4166 panic("softdep_write_inodeblock: lost dep3"); 4167#endif /* DIAGNOSTIC */ 4168 dp->di_ib[i] = 0; 4169 } 4170 return; 4171 } 4172 /* 4173 * If we have zero'ed out the last allocated block of the file, 4174 * roll back the size to the last currently allocated block. 4175 * We know that this last allocated block is a full-sized as 4176 * we already checked for fragments in the loop above. 4177 */ 4178 if (lastadp != NULL && 4179 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) { 4180 for (i = lastadp->ad_lbn; i >= 0; i--) 4181 if (dp->di_db[i] != 0) 4182 break; 4183 dp->di_size = (i + 1) * fs->fs_bsize; 4184 } 4185 /* 4186 * The only dependencies are for indirect blocks. 4187 * 4188 * The file size for indirect block additions is not guaranteed. 4189 * Such a guarantee would be non-trivial to achieve. The conventional 4190 * synchronous write implementation also does not make this guarantee. 4191 * Fsck should catch and fix discrepancies. Arguably, the file size 4192 * can be over-estimated without destroying integrity when the file 4193 * moves into the indirect blocks (i.e., is large). If we want to 4194 * postpone fsck, we are stuck with this argument. 4195 */ 4196 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 4197 dp->di_ib[adp->ad_lbn - NDADDR] = 0; 4198} 4199 4200/* 4201 * This routine is called during the completion interrupt 4202 * service routine for a disk write (from the procedure called 4203 * by the device driver to inform the filesystem caches of 4204 * a request completion). It should be called early in this 4205 * procedure, before the block is made available to other 4206 * processes or other routines are called. 4207 */ 4208static void 4209softdep_disk_write_complete(bp) 4210 struct buf *bp; /* describes the completed disk write */ 4211{ 4212 struct worklist *wk; 4213 struct worklist *owk; 4214 struct workhead reattach; 4215 struct newblk *newblk; 4216 struct allocindir *aip; 4217 struct allocdirect *adp; 4218 struct indirdep *indirdep; 4219 struct inodedep *inodedep; 4220 struct bmsafemap *bmsafemap; 4221 4222 /* 4223 * If an error occurred while doing the write, then the data 4224 * has not hit the disk and the dependencies cannot be unrolled. 4225 */ 4226 if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0) 4227 return; 4228 LIST_INIT(&reattach); 4229 /* 4230 * This lock must not be released anywhere in this code segment. 4231 */ 4232 ACQUIRE_LOCK(&lk); 4233 owk = NULL; 4234 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 4235 WORKLIST_REMOVE(wk); 4236 if (wk == owk) 4237 panic("duplicate worklist: %p\n", wk); 4238 owk = wk; 4239 switch (wk->wk_type) { 4240 4241 case D_PAGEDEP: 4242 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 4243 WORKLIST_INSERT(&reattach, wk); 4244 continue; 4245 4246 case D_INODEDEP: 4247 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 4248 WORKLIST_INSERT(&reattach, wk); 4249 continue; 4250 4251 case D_BMSAFEMAP: 4252 bmsafemap = WK_BMSAFEMAP(wk); 4253 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) { 4254 newblk->nb_state |= DEPCOMPLETE; 4255 newblk->nb_bmsafemap = NULL; 4256 LIST_REMOVE(newblk, nb_deps); 4257 } 4258 while ((adp = 4259 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) { 4260 adp->ad_state |= DEPCOMPLETE; 4261 adp->ad_buf = NULL; 4262 LIST_REMOVE(adp, ad_deps); 4263 handle_allocdirect_partdone(adp); 4264 } 4265 while ((aip = 4266 LIST_FIRST(&bmsafemap->sm_allocindirhd))) { 4267 aip->ai_state |= DEPCOMPLETE; 4268 aip->ai_buf = NULL; 4269 LIST_REMOVE(aip, ai_deps); 4270 handle_allocindir_partdone(aip); 4271 } 4272 while ((inodedep = 4273 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) { 4274 inodedep->id_state |= DEPCOMPLETE; 4275 LIST_REMOVE(inodedep, id_deps); 4276 inodedep->id_buf = NULL; 4277 } 4278 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 4279 continue; 4280 4281 case D_MKDIR: 4282 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 4283 continue; 4284 4285 case D_ALLOCDIRECT: 4286 adp = WK_ALLOCDIRECT(wk); 4287 adp->ad_state |= COMPLETE; 4288 handle_allocdirect_partdone(adp); 4289 continue; 4290 4291 case D_ALLOCINDIR: 4292 aip = WK_ALLOCINDIR(wk); 4293 aip->ai_state |= COMPLETE; 4294 handle_allocindir_partdone(aip); 4295 continue; 4296 4297 case D_INDIRDEP: 4298 indirdep = WK_INDIRDEP(wk); 4299 if (indirdep->ir_state & GOINGAWAY) 4300 panic("disk_write_complete: indirdep gone"); 4301 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 4302 FREE(indirdep->ir_saveddata, M_INDIRDEP); 4303 indirdep->ir_saveddata = 0; 4304 indirdep->ir_state &= ~UNDONE; 4305 indirdep->ir_state |= ATTACHED; 4306 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 4307 handle_allocindir_partdone(aip); 4308 if (aip == LIST_FIRST(&indirdep->ir_donehd)) 4309 panic("disk_write_complete: not gone"); 4310 } 4311 WORKLIST_INSERT(&reattach, wk); 4312 if ((bp->b_flags & B_DELWRI) == 0) 4313 stat_indir_blk_ptrs++; 4314 bdirty(bp); 4315 continue; 4316 4317 default: 4318 panic("handle_disk_write_complete: Unknown type %s", 4319 TYPENAME(wk->wk_type)); 4320 /* NOTREACHED */ 4321 } 4322 } 4323 /* 4324 * Reattach any requests that must be redone. 4325 */ 4326 while ((wk = LIST_FIRST(&reattach)) != NULL) { 4327 WORKLIST_REMOVE(wk); 4328 WORKLIST_INSERT(&bp->b_dep, wk); 4329 } 4330 FREE_LOCK(&lk); 4331} 4332 4333/* 4334 * Called from within softdep_disk_write_complete above. Note that 4335 * this routine is always called from interrupt level with further 4336 * splbio interrupts blocked. 4337 */ 4338static void 4339handle_allocdirect_partdone(adp) 4340 struct allocdirect *adp; /* the completed allocdirect */ 4341{ 4342 struct allocdirectlst *listhead; 4343 struct allocdirect *listadp; 4344 struct inodedep *inodedep; 4345 long bsize, delay; 4346 4347 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 4348 return; 4349 if (adp->ad_buf != NULL) 4350 panic("handle_allocdirect_partdone: dangling dep"); 4351 /* 4352 * The on-disk inode cannot claim to be any larger than the last 4353 * fragment that has been written. Otherwise, the on-disk inode 4354 * might have fragments that were not the last block in the file 4355 * which would corrupt the filesystem. Thus, we cannot free any 4356 * allocdirects after one whose ad_oldblkno claims a fragment as 4357 * these blocks must be rolled back to zero before writing the inode. 4358 * We check the currently active set of allocdirects in id_inoupdt 4359 * or id_extupdt as appropriate. 4360 */ 4361 inodedep = adp->ad_inodedep; 4362 bsize = inodedep->id_fs->fs_bsize; 4363 if (adp->ad_state & EXTDATA) 4364 listhead = &inodedep->id_extupdt; 4365 else 4366 listhead = &inodedep->id_inoupdt; 4367 TAILQ_FOREACH(listadp, listhead, ad_next) { 4368 /* found our block */ 4369 if (listadp == adp) 4370 break; 4371 /* continue if ad_oldlbn is not a fragment */ 4372 if (listadp->ad_oldsize == 0 || 4373 listadp->ad_oldsize == bsize) 4374 continue; 4375 /* hit a fragment */ 4376 return; 4377 } 4378 /* 4379 * If we have reached the end of the current list without 4380 * finding the just finished dependency, then it must be 4381 * on the future dependency list. Future dependencies cannot 4382 * be freed until they are moved to the current list. 4383 */ 4384 if (listadp == NULL) { 4385#ifdef DEBUG 4386 if (adp->ad_state & EXTDATA) 4387 listhead = &inodedep->id_newextupdt; 4388 else 4389 listhead = &inodedep->id_newinoupdt; 4390 TAILQ_FOREACH(listadp, listhead, ad_next) 4391 /* found our block */ 4392 if (listadp == adp) 4393 break; 4394 if (listadp == NULL) 4395 panic("handle_allocdirect_partdone: lost dep"); 4396#endif /* DEBUG */ 4397 return; 4398 } 4399 /* 4400 * If we have found the just finished dependency, then free 4401 * it along with anything that follows it that is complete. 4402 * If the inode still has a bitmap dependency, then it has 4403 * never been written to disk, hence the on-disk inode cannot 4404 * reference the old fragment so we can free it without delay. 4405 */ 4406 delay = (inodedep->id_state & DEPCOMPLETE); 4407 for (; adp; adp = listadp) { 4408 listadp = TAILQ_NEXT(adp, ad_next); 4409 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 4410 return; 4411 free_allocdirect(listhead, adp, delay); 4412 } 4413} 4414 4415/* 4416 * Called from within softdep_disk_write_complete above. Note that 4417 * this routine is always called from interrupt level with further 4418 * splbio interrupts blocked. 4419 */ 4420static void 4421handle_allocindir_partdone(aip) 4422 struct allocindir *aip; /* the completed allocindir */ 4423{ 4424 struct indirdep *indirdep; 4425 4426 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 4427 return; 4428 if (aip->ai_buf != NULL) 4429 panic("handle_allocindir_partdone: dangling dependency"); 4430 indirdep = aip->ai_indirdep; 4431 if (indirdep->ir_state & UNDONE) { 4432 LIST_REMOVE(aip, ai_next); 4433 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 4434 return; 4435 } 4436 if (indirdep->ir_state & UFS1FMT) 4437 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 4438 aip->ai_newblkno; 4439 else 4440 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 4441 aip->ai_newblkno; 4442 LIST_REMOVE(aip, ai_next); 4443 if (aip->ai_freefrag != NULL) 4444 add_to_worklist(&aip->ai_freefrag->ff_list); 4445 WORKITEM_FREE(aip, D_ALLOCINDIR); 4446} 4447 4448/* 4449 * Called from within softdep_disk_write_complete above to restore 4450 * in-memory inode block contents to their most up-to-date state. Note 4451 * that this routine is always called from interrupt level with further 4452 * splbio interrupts blocked. 4453 */ 4454static int 4455handle_written_inodeblock(inodedep, bp) 4456 struct inodedep *inodedep; 4457 struct buf *bp; /* buffer containing the inode block */ 4458{ 4459 struct worklist *wk, *filefree; 4460 struct allocdirect *adp, *nextadp; 4461 struct ufs1_dinode *dp1 = NULL; 4462 struct ufs2_dinode *dp2 = NULL; 4463 int hadchanges, fstype; 4464 4465 if ((inodedep->id_state & IOSTARTED) == 0) 4466 panic("handle_written_inodeblock: not started"); 4467 inodedep->id_state &= ~IOSTARTED; 4468 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) { 4469 fstype = UFS1; 4470 dp1 = (struct ufs1_dinode *)bp->b_data + 4471 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 4472 } else { 4473 fstype = UFS2; 4474 dp2 = (struct ufs2_dinode *)bp->b_data + 4475 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 4476 } 4477 /* 4478 * If we had to rollback the inode allocation because of 4479 * bitmaps being incomplete, then simply restore it. 4480 * Keep the block dirty so that it will not be reclaimed until 4481 * all associated dependencies have been cleared and the 4482 * corresponding updates written to disk. 4483 */ 4484 if (inodedep->id_savedino1 != NULL) { 4485 if (fstype == UFS1) 4486 *dp1 = *inodedep->id_savedino1; 4487 else 4488 *dp2 = *inodedep->id_savedino2; 4489 FREE(inodedep->id_savedino1, M_SAVEDINO); 4490 inodedep->id_savedino1 = NULL; 4491 if ((bp->b_flags & B_DELWRI) == 0) 4492 stat_inode_bitmap++; 4493 bdirty(bp); 4494 return (1); 4495 } 4496 inodedep->id_state |= COMPLETE; 4497 /* 4498 * Roll forward anything that had to be rolled back before 4499 * the inode could be updated. 4500 */ 4501 hadchanges = 0; 4502 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 4503 nextadp = TAILQ_NEXT(adp, ad_next); 4504 if (adp->ad_state & ATTACHED) 4505 panic("handle_written_inodeblock: new entry"); 4506 if (fstype == UFS1) { 4507 if (adp->ad_lbn < NDADDR) { 4508 if (dp1->di_db[adp->ad_lbn]!=adp->ad_oldblkno) 4509 panic("%s %s #%jd mismatch %d != %jd", 4510 "handle_written_inodeblock:", 4511 "direct pointer", 4512 (intmax_t)adp->ad_lbn, 4513 dp1->di_db[adp->ad_lbn], 4514 (intmax_t)adp->ad_oldblkno); 4515 dp1->di_db[adp->ad_lbn] = adp->ad_newblkno; 4516 } else { 4517 if (dp1->di_ib[adp->ad_lbn - NDADDR] != 0) 4518 panic("%s: %s #%jd allocated as %d", 4519 "handle_written_inodeblock", 4520 "indirect pointer", 4521 (intmax_t)adp->ad_lbn - NDADDR, 4522 dp1->di_ib[adp->ad_lbn - NDADDR]); 4523 dp1->di_ib[adp->ad_lbn - NDADDR] = 4524 adp->ad_newblkno; 4525 } 4526 } else { 4527 if (adp->ad_lbn < NDADDR) { 4528 if (dp2->di_db[adp->ad_lbn]!=adp->ad_oldblkno) 4529 panic("%s: %s #%jd %s %jd != %jd", 4530 "handle_written_inodeblock", 4531 "direct pointer", 4532 (intmax_t)adp->ad_lbn, "mismatch", 4533 (intmax_t)dp2->di_db[adp->ad_lbn], 4534 (intmax_t)adp->ad_oldblkno); 4535 dp2->di_db[adp->ad_lbn] = adp->ad_newblkno; 4536 } else { 4537 if (dp2->di_ib[adp->ad_lbn - NDADDR] != 0) 4538 panic("%s: %s #%jd allocated as %jd", 4539 "handle_written_inodeblock", 4540 "indirect pointer", 4541 (intmax_t)adp->ad_lbn - NDADDR, 4542 (intmax_t) 4543 dp2->di_ib[adp->ad_lbn - NDADDR]); 4544 dp2->di_ib[adp->ad_lbn - NDADDR] = 4545 adp->ad_newblkno; 4546 } 4547 } 4548 adp->ad_state &= ~UNDONE; 4549 adp->ad_state |= ATTACHED; 4550 hadchanges = 1; 4551 } 4552 for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) { 4553 nextadp = TAILQ_NEXT(adp, ad_next); 4554 if (adp->ad_state & ATTACHED) 4555 panic("handle_written_inodeblock: new entry"); 4556 if (dp2->di_extb[adp->ad_lbn] != adp->ad_oldblkno) 4557 panic("%s: direct pointers #%jd %s %jd != %jd", 4558 "handle_written_inodeblock", 4559 (intmax_t)adp->ad_lbn, "mismatch", 4560 (intmax_t)dp2->di_extb[adp->ad_lbn], 4561 (intmax_t)adp->ad_oldblkno); 4562 dp2->di_extb[adp->ad_lbn] = adp->ad_newblkno; 4563 adp->ad_state &= ~UNDONE; 4564 adp->ad_state |= ATTACHED; 4565 hadchanges = 1; 4566 } 4567 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 4568 stat_direct_blk_ptrs++; 4569 /* 4570 * Reset the file size to its most up-to-date value. 4571 */ 4572 if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1) 4573 panic("handle_written_inodeblock: bad size"); 4574 if (fstype == UFS1) { 4575 if (dp1->di_size != inodedep->id_savedsize) { 4576 dp1->di_size = inodedep->id_savedsize; 4577 hadchanges = 1; 4578 } 4579 } else { 4580 if (dp2->di_size != inodedep->id_savedsize) { 4581 dp2->di_size = inodedep->id_savedsize; 4582 hadchanges = 1; 4583 } 4584 if (dp2->di_extsize != inodedep->id_savedextsize) { 4585 dp2->di_extsize = inodedep->id_savedextsize; 4586 hadchanges = 1; 4587 } 4588 } 4589 inodedep->id_savedsize = -1; 4590 inodedep->id_savedextsize = -1; 4591 /* 4592 * If there were any rollbacks in the inode block, then it must be 4593 * marked dirty so that its will eventually get written back in 4594 * its correct form. 4595 */ 4596 if (hadchanges) 4597 bdirty(bp); 4598 /* 4599 * Process any allocdirects that completed during the update. 4600 */ 4601 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 4602 handle_allocdirect_partdone(adp); 4603 if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 4604 handle_allocdirect_partdone(adp); 4605 /* 4606 * Process deallocations that were held pending until the 4607 * inode had been written to disk. Freeing of the inode 4608 * is delayed until after all blocks have been freed to 4609 * avoid creation of new <vfsid, inum, lbn> triples 4610 * before the old ones have been deleted. 4611 */ 4612 filefree = NULL; 4613 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 4614 WORKLIST_REMOVE(wk); 4615 switch (wk->wk_type) { 4616 4617 case D_FREEFILE: 4618 /* 4619 * We defer adding filefree to the worklist until 4620 * all other additions have been made to ensure 4621 * that it will be done after all the old blocks 4622 * have been freed. 4623 */ 4624 if (filefree != NULL) 4625 panic("handle_written_inodeblock: filefree"); 4626 filefree = wk; 4627 continue; 4628 4629 case D_MKDIR: 4630 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 4631 continue; 4632 4633 case D_DIRADD: 4634 diradd_inode_written(WK_DIRADD(wk), inodedep); 4635 continue; 4636 4637 case D_FREEBLKS: 4638 wk->wk_state |= COMPLETE; 4639 if ((wk->wk_state & ALLCOMPLETE) != ALLCOMPLETE) 4640 continue; 4641 /* -- fall through -- */ 4642 case D_FREEFRAG: 4643 case D_DIRREM: 4644 add_to_worklist(wk); 4645 continue; 4646 4647 case D_NEWDIRBLK: 4648 free_newdirblk(WK_NEWDIRBLK(wk)); 4649 continue; 4650 4651 default: 4652 panic("handle_written_inodeblock: Unknown type %s", 4653 TYPENAME(wk->wk_type)); 4654 /* NOTREACHED */ 4655 } 4656 } 4657 if (filefree != NULL) { 4658 if (free_inodedep(inodedep) == 0) 4659 panic("handle_written_inodeblock: live inodedep"); 4660 add_to_worklist(filefree); 4661 return (0); 4662 } 4663 4664 /* 4665 * If no outstanding dependencies, free it. 4666 */ 4667 if (free_inodedep(inodedep) || 4668 (TAILQ_FIRST(&inodedep->id_inoupdt) == 0 && 4669 TAILQ_FIRST(&inodedep->id_extupdt) == 0)) 4670 return (0); 4671 return (hadchanges); 4672} 4673 4674/* 4675 * Process a diradd entry after its dependent inode has been written. 4676 * This routine must be called with splbio interrupts blocked. 4677 */ 4678static void 4679diradd_inode_written(dap, inodedep) 4680 struct diradd *dap; 4681 struct inodedep *inodedep; 4682{ 4683 struct pagedep *pagedep; 4684 4685 dap->da_state |= COMPLETE; 4686 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 4687 if (dap->da_state & DIRCHG) 4688 pagedep = dap->da_previous->dm_pagedep; 4689 else 4690 pagedep = dap->da_pagedep; 4691 LIST_REMOVE(dap, da_pdlist); 4692 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 4693 } 4694 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 4695} 4696 4697/* 4698 * Handle the completion of a mkdir dependency. 4699 */ 4700static void 4701handle_written_mkdir(mkdir, type) 4702 struct mkdir *mkdir; 4703 int type; 4704{ 4705 struct diradd *dap; 4706 struct pagedep *pagedep; 4707 4708 if (mkdir->md_state != type) 4709 panic("handle_written_mkdir: bad type"); 4710 dap = mkdir->md_diradd; 4711 dap->da_state &= ~type; 4712 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 4713 dap->da_state |= DEPCOMPLETE; 4714 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 4715 if (dap->da_state & DIRCHG) 4716 pagedep = dap->da_previous->dm_pagedep; 4717 else 4718 pagedep = dap->da_pagedep; 4719 LIST_REMOVE(dap, da_pdlist); 4720 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 4721 } 4722 LIST_REMOVE(mkdir, md_mkdirs); 4723 WORKITEM_FREE(mkdir, D_MKDIR); 4724} 4725 4726/* 4727 * Called from within softdep_disk_write_complete above. 4728 * A write operation was just completed. Removed inodes can 4729 * now be freed and associated block pointers may be committed. 4730 * Note that this routine is always called from interrupt level 4731 * with further splbio interrupts blocked. 4732 */ 4733static int 4734handle_written_filepage(pagedep, bp) 4735 struct pagedep *pagedep; 4736 struct buf *bp; /* buffer containing the written page */ 4737{ 4738 struct dirrem *dirrem; 4739 struct diradd *dap, *nextdap; 4740 struct direct *ep; 4741 int i, chgs; 4742 4743 if ((pagedep->pd_state & IOSTARTED) == 0) 4744 panic("handle_written_filepage: not started"); 4745 pagedep->pd_state &= ~IOSTARTED; 4746 /* 4747 * Process any directory removals that have been committed. 4748 */ 4749 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 4750 LIST_REMOVE(dirrem, dm_next); 4751 dirrem->dm_dirinum = pagedep->pd_ino; 4752 add_to_worklist(&dirrem->dm_list); 4753 } 4754 /* 4755 * Free any directory additions that have been committed. 4756 * If it is a newly allocated block, we have to wait until 4757 * the on-disk directory inode claims the new block. 4758 */ 4759 if ((pagedep->pd_state & NEWBLOCK) == 0) 4760 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 4761 free_diradd(dap); 4762 /* 4763 * Uncommitted directory entries must be restored. 4764 */ 4765 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 4766 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 4767 dap = nextdap) { 4768 nextdap = LIST_NEXT(dap, da_pdlist); 4769 if (dap->da_state & ATTACHED) 4770 panic("handle_written_filepage: attached"); 4771 ep = (struct direct *) 4772 ((char *)bp->b_data + dap->da_offset); 4773 ep->d_ino = dap->da_newinum; 4774 dap->da_state &= ~UNDONE; 4775 dap->da_state |= ATTACHED; 4776 chgs = 1; 4777 /* 4778 * If the inode referenced by the directory has 4779 * been written out, then the dependency can be 4780 * moved to the pending list. 4781 */ 4782 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 4783 LIST_REMOVE(dap, da_pdlist); 4784 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 4785 da_pdlist); 4786 } 4787 } 4788 } 4789 /* 4790 * If there were any rollbacks in the directory, then it must be 4791 * marked dirty so that its will eventually get written back in 4792 * its correct form. 4793 */ 4794 if (chgs) { 4795 if ((bp->b_flags & B_DELWRI) == 0) 4796 stat_dir_entry++; 4797 bdirty(bp); 4798 return (1); 4799 } 4800 /* 4801 * If we are not waiting for a new directory block to be 4802 * claimed by its inode, then the pagedep will be freed. 4803 * Otherwise it will remain to track any new entries on 4804 * the page in case they are fsync'ed. 4805 */ 4806 if ((pagedep->pd_state & NEWBLOCK) == 0) { 4807 LIST_REMOVE(pagedep, pd_hash); 4808 WORKITEM_FREE(pagedep, D_PAGEDEP); 4809 } 4810 return (0); 4811} 4812 4813/* 4814 * Writing back in-core inode structures. 4815 * 4816 * The filesystem only accesses an inode's contents when it occupies an 4817 * "in-core" inode structure. These "in-core" structures are separate from 4818 * the page frames used to cache inode blocks. Only the latter are 4819 * transferred to/from the disk. So, when the updated contents of the 4820 * "in-core" inode structure are copied to the corresponding in-memory inode 4821 * block, the dependencies are also transferred. The following procedure is 4822 * called when copying a dirty "in-core" inode to a cached inode block. 4823 */ 4824 4825/* 4826 * Called when an inode is loaded from disk. If the effective link count 4827 * differed from the actual link count when it was last flushed, then we 4828 * need to ensure that the correct effective link count is put back. 4829 */ 4830void 4831softdep_load_inodeblock(ip) 4832 struct inode *ip; /* the "in_core" copy of the inode */ 4833{ 4834 struct inodedep *inodedep; 4835 4836 /* 4837 * Check for alternate nlink count. 4838 */ 4839 ip->i_effnlink = ip->i_nlink; 4840 ACQUIRE_LOCK(&lk); 4841 if (inodedep_lookup(UFSTOVFS(ip->i_ump), 4842 ip->i_number, 0, &inodedep) == 0) { 4843 FREE_LOCK(&lk); 4844 return; 4845 } 4846 ip->i_effnlink -= inodedep->id_nlinkdelta; 4847 if (inodedep->id_state & SPACECOUNTED) 4848 ip->i_flag |= IN_SPACECOUNTED; 4849 FREE_LOCK(&lk); 4850} 4851 4852/* 4853 * This routine is called just before the "in-core" inode 4854 * information is to be copied to the in-memory inode block. 4855 * Recall that an inode block contains several inodes. If 4856 * the force flag is set, then the dependencies will be 4857 * cleared so that the update can always be made. Note that 4858 * the buffer is locked when this routine is called, so we 4859 * will never be in the middle of writing the inode block 4860 * to disk. 4861 */ 4862void 4863softdep_update_inodeblock(ip, bp, waitfor) 4864 struct inode *ip; /* the "in_core" copy of the inode */ 4865 struct buf *bp; /* the buffer containing the inode block */ 4866 int waitfor; /* nonzero => update must be allowed */ 4867{ 4868 struct inodedep *inodedep; 4869 struct worklist *wk; 4870 struct mount *mp; 4871 struct buf *ibp; 4872 int error; 4873 4874 /* 4875 * If the effective link count is not equal to the actual link 4876 * count, then we must track the difference in an inodedep while 4877 * the inode is (potentially) tossed out of the cache. Otherwise, 4878 * if there is no existing inodedep, then there are no dependencies 4879 * to track. 4880 */ 4881 mp = UFSTOVFS(ip->i_ump); 4882 ACQUIRE_LOCK(&lk); 4883 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 4884 FREE_LOCK(&lk); 4885 if (ip->i_effnlink != ip->i_nlink) 4886 panic("softdep_update_inodeblock: bad link count"); 4887 return; 4888 } 4889 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) 4890 panic("softdep_update_inodeblock: bad delta"); 4891 /* 4892 * Changes have been initiated. Anything depending on these 4893 * changes cannot occur until this inode has been written. 4894 */ 4895 inodedep->id_state &= ~COMPLETE; 4896 if ((inodedep->id_state & ONWORKLIST) == 0) 4897 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 4898 /* 4899 * Any new dependencies associated with the incore inode must 4900 * now be moved to the list associated with the buffer holding 4901 * the in-memory copy of the inode. Once merged process any 4902 * allocdirects that are completed by the merger. 4903 */ 4904 merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt); 4905 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL) 4906 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt)); 4907 merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt); 4908 if (TAILQ_FIRST(&inodedep->id_extupdt) != NULL) 4909 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt)); 4910 /* 4911 * Now that the inode has been pushed into the buffer, the 4912 * operations dependent on the inode being written to disk 4913 * can be moved to the id_bufwait so that they will be 4914 * processed when the buffer I/O completes. 4915 */ 4916 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 4917 WORKLIST_REMOVE(wk); 4918 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 4919 } 4920 /* 4921 * Newly allocated inodes cannot be written until the bitmap 4922 * that allocates them have been written (indicated by 4923 * DEPCOMPLETE being set in id_state). If we are doing a 4924 * forced sync (e.g., an fsync on a file), we force the bitmap 4925 * to be written so that the update can be done. 4926 */ 4927 if (waitfor == 0) { 4928 FREE_LOCK(&lk); 4929 return; 4930 } 4931retry: 4932 if ((inodedep->id_state & DEPCOMPLETE) != 0) { 4933 FREE_LOCK(&lk); 4934 return; 4935 } 4936 ibp = inodedep->id_buf; 4937 ibp = getdirtybuf(ibp, &lk, MNT_WAIT); 4938 if (ibp == NULL) { 4939 /* 4940 * If ibp came back as NULL, the dependency could have been 4941 * freed while we slept. Look it up again, and check to see 4942 * that it has completed. 4943 */ 4944 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 4945 goto retry; 4946 FREE_LOCK(&lk); 4947 return; 4948 } 4949 FREE_LOCK(&lk); 4950 if ((error = bwrite(ibp)) != 0) 4951 softdep_error("softdep_update_inodeblock: bwrite", error); 4952} 4953 4954/* 4955 * Merge the a new inode dependency list (such as id_newinoupdt) into an 4956 * old inode dependency list (such as id_inoupdt). This routine must be 4957 * called with splbio interrupts blocked. 4958 */ 4959static void 4960merge_inode_lists(newlisthead, oldlisthead) 4961 struct allocdirectlst *newlisthead; 4962 struct allocdirectlst *oldlisthead; 4963{ 4964 struct allocdirect *listadp, *newadp; 4965 4966 newadp = TAILQ_FIRST(newlisthead); 4967 for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) { 4968 if (listadp->ad_lbn < newadp->ad_lbn) { 4969 listadp = TAILQ_NEXT(listadp, ad_next); 4970 continue; 4971 } 4972 TAILQ_REMOVE(newlisthead, newadp, ad_next); 4973 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 4974 if (listadp->ad_lbn == newadp->ad_lbn) { 4975 allocdirect_merge(oldlisthead, newadp, 4976 listadp); 4977 listadp = newadp; 4978 } 4979 newadp = TAILQ_FIRST(newlisthead); 4980 } 4981 while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) { 4982 TAILQ_REMOVE(newlisthead, newadp, ad_next); 4983 TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next); 4984 } 4985} 4986 4987/* 4988 * If we are doing an fsync, then we must ensure that any directory 4989 * entries for the inode have been written after the inode gets to disk. 4990 */ 4991int 4992softdep_fsync(vp) 4993 struct vnode *vp; /* the "in_core" copy of the inode */ 4994{ 4995 struct inodedep *inodedep; 4996 struct pagedep *pagedep; 4997 struct worklist *wk; 4998 struct diradd *dap; 4999 struct mount *mp; 5000 struct vnode *pvp; 5001 struct inode *ip; 5002 struct buf *bp; 5003 struct fs *fs; 5004 struct thread *td = curthread; 5005 int error, flushparent; 5006 ino_t parentino; 5007 ufs_lbn_t lbn; 5008 5009 ip = VTOI(vp); 5010 fs = ip->i_fs; 5011 mp = vp->v_mount; 5012 ACQUIRE_LOCK(&lk); 5013 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 5014 FREE_LOCK(&lk); 5015 return (0); 5016 } 5017 if (LIST_FIRST(&inodedep->id_inowait) != NULL || 5018 LIST_FIRST(&inodedep->id_bufwait) != NULL || 5019 TAILQ_FIRST(&inodedep->id_extupdt) != NULL || 5020 TAILQ_FIRST(&inodedep->id_newextupdt) != NULL || 5021 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL || 5022 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) 5023 panic("softdep_fsync: pending ops"); 5024 for (error = 0, flushparent = 0; ; ) { 5025 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 5026 break; 5027 if (wk->wk_type != D_DIRADD) 5028 panic("softdep_fsync: Unexpected type %s", 5029 TYPENAME(wk->wk_type)); 5030 dap = WK_DIRADD(wk); 5031 /* 5032 * Flush our parent if this directory entry has a MKDIR_PARENT 5033 * dependency or is contained in a newly allocated block. 5034 */ 5035 if (dap->da_state & DIRCHG) 5036 pagedep = dap->da_previous->dm_pagedep; 5037 else 5038 pagedep = dap->da_pagedep; 5039 parentino = pagedep->pd_ino; 5040 lbn = pagedep->pd_lbn; 5041 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) 5042 panic("softdep_fsync: dirty"); 5043 if ((dap->da_state & MKDIR_PARENT) || 5044 (pagedep->pd_state & NEWBLOCK)) 5045 flushparent = 1; 5046 else 5047 flushparent = 0; 5048 /* 5049 * If we are being fsync'ed as part of vgone'ing this vnode, 5050 * then we will not be able to release and recover the 5051 * vnode below, so we just have to give up on writing its 5052 * directory entry out. It will eventually be written, just 5053 * not now, but then the user was not asking to have it 5054 * written, so we are not breaking any promises. 5055 */ 5056 if (vp->v_iflag & VI_DOOMED) 5057 break; 5058 /* 5059 * We prevent deadlock by always fetching inodes from the 5060 * root, moving down the directory tree. Thus, when fetching 5061 * our parent directory, we first try to get the lock. If 5062 * that fails, we must unlock ourselves before requesting 5063 * the lock on our parent. See the comment in ufs_lookup 5064 * for details on possible races. 5065 */ 5066 FREE_LOCK(&lk); 5067 if (ffs_vget(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp)) { 5068 VOP_UNLOCK(vp, 0, td); 5069 error = ffs_vget(mp, parentino, LK_EXCLUSIVE, &pvp); 5070 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 5071 if (error != 0) 5072 return (error); 5073 } 5074 /* 5075 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps 5076 * that are contained in direct blocks will be resolved by 5077 * doing a ffs_update. Pagedeps contained in indirect blocks 5078 * may require a complete sync'ing of the directory. So, we 5079 * try the cheap and fast ffs_update first, and if that fails, 5080 * then we do the slower ffs_syncvnode of the directory. 5081 */ 5082 if (flushparent) { 5083 if ((error = ffs_update(pvp, 1)) != 0) { 5084 vput(pvp); 5085 return (error); 5086 } 5087 if ((pagedep->pd_state & NEWBLOCK) && 5088 (error = ffs_syncvnode(pvp, MNT_WAIT))) { 5089 vput(pvp); 5090 return (error); 5091 } 5092 } 5093 /* 5094 * Flush directory page containing the inode's name. 5095 */ 5096 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred, 5097 &bp); 5098 if (error == 0) 5099 error = bwrite(bp); 5100 else 5101 brelse(bp); 5102 vput(pvp); 5103 if (error != 0) 5104 return (error); 5105 ACQUIRE_LOCK(&lk); 5106 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 5107 break; 5108 } 5109 FREE_LOCK(&lk); 5110 return (0); 5111} 5112 5113/* 5114 * Flush all the dirty bitmaps associated with the block device 5115 * before flushing the rest of the dirty blocks so as to reduce 5116 * the number of dependencies that will have to be rolled back. 5117 */ 5118void 5119softdep_fsync_mountdev(vp) 5120 struct vnode *vp; 5121{ 5122 struct buf *bp, *nbp; 5123 struct worklist *wk; 5124 5125 if (!vn_isdisk(vp, NULL)) 5126 panic("softdep_fsync_mountdev: vnode not a disk"); 5127restart: 5128 ACQUIRE_LOCK(&lk); 5129 VI_LOCK(vp); 5130 TAILQ_FOREACH_SAFE(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs, nbp) { 5131 /* 5132 * If it is already scheduled, skip to the next buffer. 5133 */ 5134 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 5135 continue; 5136 5137 if ((bp->b_flags & B_DELWRI) == 0) 5138 panic("softdep_fsync_mountdev: not dirty"); 5139 /* 5140 * We are only interested in bitmaps with outstanding 5141 * dependencies. 5142 */ 5143 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 5144 wk->wk_type != D_BMSAFEMAP || 5145 (bp->b_vflags & BV_BKGRDINPROG)) { 5146 BUF_UNLOCK(bp); 5147 continue; 5148 } 5149 VI_UNLOCK(vp); 5150 FREE_LOCK(&lk); 5151 bremfree(bp); 5152 (void) bawrite(bp); 5153 goto restart; 5154 } 5155 FREE_LOCK(&lk); 5156 drain_output(vp); 5157 VI_UNLOCK(vp); 5158} 5159 5160/* 5161 * This routine is called when we are trying to synchronously flush a 5162 * file. This routine must eliminate any filesystem metadata dependencies 5163 * so that the syncing routine can succeed by pushing the dirty blocks 5164 * associated with the file. If any I/O errors occur, they are returned. 5165 */ 5166int 5167softdep_sync_metadata(struct vnode *vp) 5168{ 5169 struct pagedep *pagedep; 5170 struct allocdirect *adp; 5171 struct allocindir *aip; 5172 struct buf *bp, *nbp; 5173 struct worklist *wk; 5174 int i, error, waitfor; 5175 5176 if (!DOINGSOFTDEP(vp)) 5177 return (0); 5178 /* 5179 * Ensure that any direct block dependencies have been cleared. 5180 */ 5181 ACQUIRE_LOCK(&lk); 5182 if ((error = flush_inodedep_deps(vp->v_mount, VTOI(vp)->i_number))) { 5183 FREE_LOCK(&lk); 5184 return (error); 5185 } 5186 FREE_LOCK(&lk); 5187 /* 5188 * For most files, the only metadata dependencies are the 5189 * cylinder group maps that allocate their inode or blocks. 5190 * The block allocation dependencies can be found by traversing 5191 * the dependency lists for any buffers that remain on their 5192 * dirty buffer list. The inode allocation dependency will 5193 * be resolved when the inode is updated with MNT_WAIT. 5194 * This work is done in two passes. The first pass grabs most 5195 * of the buffers and begins asynchronously writing them. The 5196 * only way to wait for these asynchronous writes is to sleep 5197 * on the filesystem vnode which may stay busy for a long time 5198 * if the filesystem is active. So, instead, we make a second 5199 * pass over the dependencies blocking on each write. In the 5200 * usual case we will be blocking against a write that we 5201 * initiated, so when it is done the dependency will have been 5202 * resolved. Thus the second pass is expected to end quickly. 5203 */ 5204 waitfor = MNT_NOWAIT; 5205 5206top: 5207 /* 5208 * We must wait for any I/O in progress to finish so that 5209 * all potential buffers on the dirty list will be visible. 5210 */ 5211 VI_LOCK(vp); 5212 drain_output(vp); 5213 while ((bp = TAILQ_FIRST(&vp->v_bufobj.bo_dirty.bv_hd)) != NULL) { 5214 bp = getdirtybuf(bp, VI_MTX(vp), MNT_WAIT); 5215 if (bp) 5216 break; 5217 } 5218 VI_UNLOCK(vp); 5219 if (bp == NULL) 5220 return (0); 5221loop: 5222 /* While syncing snapshots, we must allow recursive lookups */ 5223 bp->b_lock.lk_flags |= LK_CANRECURSE; 5224 ACQUIRE_LOCK(&lk); 5225 /* 5226 * As we hold the buffer locked, none of its dependencies 5227 * will disappear. 5228 */ 5229 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5230 switch (wk->wk_type) { 5231 5232 case D_ALLOCDIRECT: 5233 adp = WK_ALLOCDIRECT(wk); 5234 if (adp->ad_state & DEPCOMPLETE) 5235 continue; 5236 nbp = adp->ad_buf; 5237 nbp = getdirtybuf(nbp, &lk, waitfor); 5238 if (nbp == NULL) 5239 continue; 5240 FREE_LOCK(&lk); 5241 if (waitfor == MNT_NOWAIT) { 5242 bawrite(nbp); 5243 } else if ((error = bwrite(nbp)) != 0) { 5244 break; 5245 } 5246 ACQUIRE_LOCK(&lk); 5247 continue; 5248 5249 case D_ALLOCINDIR: 5250 aip = WK_ALLOCINDIR(wk); 5251 if (aip->ai_state & DEPCOMPLETE) 5252 continue; 5253 nbp = aip->ai_buf; 5254 nbp = getdirtybuf(nbp, &lk, waitfor); 5255 if (nbp == NULL) 5256 continue; 5257 FREE_LOCK(&lk); 5258 if (waitfor == MNT_NOWAIT) { 5259 bawrite(nbp); 5260 } else if ((error = bwrite(nbp)) != 0) { 5261 break; 5262 } 5263 ACQUIRE_LOCK(&lk); 5264 continue; 5265 5266 case D_INDIRDEP: 5267 restart: 5268 5269 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) { 5270 if (aip->ai_state & DEPCOMPLETE) 5271 continue; 5272 nbp = aip->ai_buf; 5273 nbp = getdirtybuf(nbp, &lk, MNT_WAIT); 5274 if (nbp == NULL) 5275 goto restart; 5276 FREE_LOCK(&lk); 5277 if ((error = bwrite(nbp)) != 0) { 5278 goto loop_end; 5279 } 5280 ACQUIRE_LOCK(&lk); 5281 goto restart; 5282 } 5283 continue; 5284 5285 case D_INODEDEP: 5286 if ((error = flush_inodedep_deps(wk->wk_mp, 5287 WK_INODEDEP(wk)->id_ino)) != 0) { 5288 FREE_LOCK(&lk); 5289 break; 5290 } 5291 continue; 5292 5293 case D_PAGEDEP: 5294 /* 5295 * We are trying to sync a directory that may 5296 * have dependencies on both its own metadata 5297 * and/or dependencies on the inodes of any 5298 * recently allocated files. We walk its diradd 5299 * lists pushing out the associated inode. 5300 */ 5301 pagedep = WK_PAGEDEP(wk); 5302 for (i = 0; i < DAHASHSZ; i++) { 5303 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 5304 continue; 5305 if ((error = 5306 flush_pagedep_deps(vp, wk->wk_mp, 5307 &pagedep->pd_diraddhd[i]))) { 5308 FREE_LOCK(&lk); 5309 goto loop_end; 5310 } 5311 } 5312 continue; 5313 5314 case D_MKDIR: 5315 /* 5316 * This case should never happen if the vnode has 5317 * been properly sync'ed. However, if this function 5318 * is used at a place where the vnode has not yet 5319 * been sync'ed, this dependency can show up. So, 5320 * rather than panic, just flush it. 5321 */ 5322 nbp = WK_MKDIR(wk)->md_buf; 5323 nbp = getdirtybuf(nbp, &lk, waitfor); 5324 if (nbp == NULL) 5325 continue; 5326 FREE_LOCK(&lk); 5327 if (waitfor == MNT_NOWAIT) { 5328 bawrite(nbp); 5329 } else if ((error = bwrite(nbp)) != 0) { 5330 break; 5331 } 5332 ACQUIRE_LOCK(&lk); 5333 continue; 5334 5335 case D_BMSAFEMAP: 5336 /* 5337 * This case should never happen if the vnode has 5338 * been properly sync'ed. However, if this function 5339 * is used at a place where the vnode has not yet 5340 * been sync'ed, this dependency can show up. So, 5341 * rather than panic, just flush it. 5342 */ 5343 nbp = WK_BMSAFEMAP(wk)->sm_buf; 5344 nbp = getdirtybuf(nbp, &lk, waitfor); 5345 if (nbp == NULL) 5346 continue; 5347 FREE_LOCK(&lk); 5348 if (waitfor == MNT_NOWAIT) { 5349 bawrite(nbp); 5350 } else if ((error = bwrite(nbp)) != 0) { 5351 break; 5352 } 5353 ACQUIRE_LOCK(&lk); 5354 continue; 5355 5356 default: 5357 panic("softdep_sync_metadata: Unknown type %s", 5358 TYPENAME(wk->wk_type)); 5359 /* NOTREACHED */ 5360 } 5361 loop_end: 5362 /* We reach here only in error and unlocked */ 5363 if (error == 0) 5364 panic("softdep_sync_metadata: zero error"); 5365 bp->b_lock.lk_flags &= ~LK_CANRECURSE; 5366 bawrite(bp); 5367 return (error); 5368 } 5369 FREE_LOCK(&lk); 5370 VI_LOCK(vp); 5371 while ((nbp = TAILQ_NEXT(bp, b_bobufs)) != NULL) { 5372 nbp = getdirtybuf(nbp, VI_MTX(vp), MNT_WAIT); 5373 if (nbp) 5374 break; 5375 } 5376 VI_UNLOCK(vp); 5377 bp->b_lock.lk_flags &= ~LK_CANRECURSE; 5378 bawrite(bp); 5379 if (nbp != NULL) { 5380 bp = nbp; 5381 goto loop; 5382 } 5383 /* 5384 * The brief unlock is to allow any pent up dependency 5385 * processing to be done. Then proceed with the second pass. 5386 */ 5387 if (waitfor == MNT_NOWAIT) { 5388 waitfor = MNT_WAIT; 5389 goto top; 5390 } 5391 5392 /* 5393 * If we have managed to get rid of all the dirty buffers, 5394 * then we are done. For certain directories and block 5395 * devices, we may need to do further work. 5396 * 5397 * We must wait for any I/O in progress to finish so that 5398 * all potential buffers on the dirty list will be visible. 5399 */ 5400 VI_LOCK(vp); 5401 drain_output(vp); 5402 VI_UNLOCK(vp); 5403 return (0); 5404} 5405 5406/* 5407 * Flush the dependencies associated with an inodedep. 5408 * Called with splbio blocked. 5409 */ 5410static int 5411flush_inodedep_deps(mp, ino) 5412 struct mount *mp; 5413 ino_t ino; 5414{ 5415 struct inodedep *inodedep; 5416 int error, waitfor; 5417 5418 /* 5419 * This work is done in two passes. The first pass grabs most 5420 * of the buffers and begins asynchronously writing them. The 5421 * only way to wait for these asynchronous writes is to sleep 5422 * on the filesystem vnode which may stay busy for a long time 5423 * if the filesystem is active. So, instead, we make a second 5424 * pass over the dependencies blocking on each write. In the 5425 * usual case we will be blocking against a write that we 5426 * initiated, so when it is done the dependency will have been 5427 * resolved. Thus the second pass is expected to end quickly. 5428 * We give a brief window at the top of the loop to allow 5429 * any pending I/O to complete. 5430 */ 5431 for (error = 0, waitfor = MNT_NOWAIT; ; ) { 5432 if (error) 5433 return (error); 5434 FREE_LOCK(&lk); 5435 ACQUIRE_LOCK(&lk); 5436 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 5437 return (0); 5438 if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) || 5439 flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) || 5440 flush_deplist(&inodedep->id_extupdt, waitfor, &error) || 5441 flush_deplist(&inodedep->id_newextupdt, waitfor, &error)) 5442 continue; 5443 /* 5444 * If pass2, we are done, otherwise do pass 2. 5445 */ 5446 if (waitfor == MNT_WAIT) 5447 break; 5448 waitfor = MNT_WAIT; 5449 } 5450 /* 5451 * Try freeing inodedep in case all dependencies have been removed. 5452 */ 5453 if (inodedep_lookup(mp, ino, 0, &inodedep) != 0) 5454 (void) free_inodedep(inodedep); 5455 return (0); 5456} 5457 5458/* 5459 * Flush an inode dependency list. 5460 * Called with splbio blocked. 5461 */ 5462static int 5463flush_deplist(listhead, waitfor, errorp) 5464 struct allocdirectlst *listhead; 5465 int waitfor; 5466 int *errorp; 5467{ 5468 struct allocdirect *adp; 5469 struct buf *bp; 5470 5471 mtx_assert(&lk, MA_OWNED); 5472 TAILQ_FOREACH(adp, listhead, ad_next) { 5473 if (adp->ad_state & DEPCOMPLETE) 5474 continue; 5475 bp = adp->ad_buf; 5476 bp = getdirtybuf(bp, &lk, waitfor); 5477 if (bp == NULL) { 5478 if (waitfor == MNT_NOWAIT) 5479 continue; 5480 return (1); 5481 } 5482 FREE_LOCK(&lk); 5483 if (waitfor == MNT_NOWAIT) { 5484 bawrite(bp); 5485 } else if ((*errorp = bwrite(bp)) != 0) { 5486 ACQUIRE_LOCK(&lk); 5487 return (1); 5488 } 5489 ACQUIRE_LOCK(&lk); 5490 return (1); 5491 } 5492 return (0); 5493} 5494 5495/* 5496 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 5497 * Called with splbio blocked. 5498 */ 5499static int 5500flush_pagedep_deps(pvp, mp, diraddhdp) 5501 struct vnode *pvp; 5502 struct mount *mp; 5503 struct diraddhd *diraddhdp; 5504{ 5505 struct inodedep *inodedep; 5506 struct ufsmount *ump; 5507 struct diradd *dap; 5508 struct vnode *vp; 5509 int error = 0; 5510 struct buf *bp; 5511 ino_t inum; 5512 struct worklist *wk; 5513 5514 ump = VFSTOUFS(mp); 5515 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 5516 /* 5517 * Flush ourselves if this directory entry 5518 * has a MKDIR_PARENT dependency. 5519 */ 5520 if (dap->da_state & MKDIR_PARENT) { 5521 FREE_LOCK(&lk); 5522 if ((error = ffs_update(pvp, 1)) != 0) 5523 break; 5524 ACQUIRE_LOCK(&lk); 5525 /* 5526 * If that cleared dependencies, go on to next. 5527 */ 5528 if (dap != LIST_FIRST(diraddhdp)) 5529 continue; 5530 if (dap->da_state & MKDIR_PARENT) 5531 panic("flush_pagedep_deps: MKDIR_PARENT"); 5532 } 5533 /* 5534 * A newly allocated directory must have its "." and 5535 * ".." entries written out before its name can be 5536 * committed in its parent. We do not want or need 5537 * the full semantics of a synchronous ffs_syncvnode as 5538 * that may end up here again, once for each directory 5539 * level in the filesystem. Instead, we push the blocks 5540 * and wait for them to clear. We have to fsync twice 5541 * because the first call may choose to defer blocks 5542 * that still have dependencies, but deferral will 5543 * happen at most once. 5544 */ 5545 inum = dap->da_newinum; 5546 if (dap->da_state & MKDIR_BODY) { 5547 FREE_LOCK(&lk); 5548 if ((error = ffs_vget(mp, inum, LK_EXCLUSIVE, &vp))) 5549 break; 5550 if ((error=ffs_syncvnode(vp, MNT_NOWAIT)) || 5551 (error=ffs_syncvnode(vp, MNT_NOWAIT))) { 5552 vput(vp); 5553 break; 5554 } 5555 VI_LOCK(vp); 5556 drain_output(vp); 5557 /* 5558 * If first block is still dirty with a D_MKDIR 5559 * dependency then it needs to be written now. 5560 */ 5561 for (;;) { 5562 error = 0; 5563 bp = gbincore(&vp->v_bufobj, 0); 5564 if (bp == NULL) 5565 break; /* First block not present */ 5566 error = BUF_LOCK(bp, 5567 LK_EXCLUSIVE | 5568 LK_SLEEPFAIL | 5569 LK_INTERLOCK, 5570 VI_MTX(vp)); 5571 VI_LOCK(vp); 5572 if (error == ENOLCK) 5573 continue; /* Slept, retry */ 5574 if (error != 0) 5575 break; /* Failed */ 5576 if ((bp->b_flags & B_DELWRI) == 0) { 5577 BUF_UNLOCK(bp); 5578 break; /* Buffer not dirty */ 5579 } 5580 for (wk = LIST_FIRST(&bp->b_dep); 5581 wk != NULL; 5582 wk = LIST_NEXT(wk, wk_list)) 5583 if (wk->wk_type == D_MKDIR) 5584 break; 5585 if (wk == NULL) 5586 BUF_UNLOCK(bp); /* Dependency gone */ 5587 else { 5588 /* 5589 * D_MKDIR dependency remains, 5590 * must write buffer to stable 5591 * storage. 5592 */ 5593 VI_UNLOCK(vp); 5594 bremfree(bp); 5595 error = bwrite(bp); 5596 VI_LOCK(vp); 5597 } 5598 break; 5599 } 5600 VI_UNLOCK(vp); 5601 vput(vp); 5602 if (error != 0) 5603 break; /* Flushing of first block failed */ 5604 ACQUIRE_LOCK(&lk); 5605 /* 5606 * If that cleared dependencies, go on to next. 5607 */ 5608 if (dap != LIST_FIRST(diraddhdp)) 5609 continue; 5610 if (dap->da_state & MKDIR_BODY) 5611 panic("flush_pagedep_deps: MKDIR_BODY"); 5612 } 5613 /* 5614 * Flush the inode on which the directory entry depends. 5615 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 5616 * the only remaining dependency is that the updated inode 5617 * count must get pushed to disk. The inode has already 5618 * been pushed into its inode buffer (via VOP_UPDATE) at 5619 * the time of the reference count change. So we need only 5620 * locate that buffer, ensure that there will be no rollback 5621 * caused by a bitmap dependency, then write the inode buffer. 5622 */ 5623retry: 5624 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 5625 panic("flush_pagedep_deps: lost inode"); 5626 /* 5627 * If the inode still has bitmap dependencies, 5628 * push them to disk. 5629 */ 5630 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 5631 bp = inodedep->id_buf; 5632 bp = getdirtybuf(bp, &lk, MNT_WAIT); 5633 if (bp == NULL) 5634 goto retry; 5635 FREE_LOCK(&lk); 5636 if ((error = bwrite(bp)) != 0) 5637 break; 5638 ACQUIRE_LOCK(&lk); 5639 if (dap != LIST_FIRST(diraddhdp)) 5640 continue; 5641 } 5642 /* 5643 * If the inode is still sitting in a buffer waiting 5644 * to be written, push it to disk. 5645 */ 5646 FREE_LOCK(&lk); 5647 if ((error = bread(ump->um_devvp, 5648 fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)), 5649 (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0) { 5650 brelse(bp); 5651 break; 5652 } 5653 if ((error = bwrite(bp)) != 0) 5654 break; 5655 ACQUIRE_LOCK(&lk); 5656 /* 5657 * If we have failed to get rid of all the dependencies 5658 * then something is seriously wrong. 5659 */ 5660 if (dap == LIST_FIRST(diraddhdp)) 5661 panic("flush_pagedep_deps: flush failed"); 5662 } 5663 if (error) 5664 ACQUIRE_LOCK(&lk); 5665 return (error); 5666} 5667 5668/* 5669 * A large burst of file addition or deletion activity can drive the 5670 * memory load excessively high. First attempt to slow things down 5671 * using the techniques below. If that fails, this routine requests 5672 * the offending operations to fall back to running synchronously 5673 * until the memory load returns to a reasonable level. 5674 */ 5675int 5676softdep_slowdown(vp) 5677 struct vnode *vp; 5678{ 5679 int max_softdeps_hard; 5680 5681 ACQUIRE_LOCK(&lk); 5682 max_softdeps_hard = max_softdeps * 11 / 10; 5683 if (num_dirrem < max_softdeps_hard / 2 && 5684 num_inodedep < max_softdeps_hard && 5685 VFSTOUFS(vp->v_mount)->um_numindirdeps < maxindirdeps) { 5686 FREE_LOCK(&lk); 5687 return (0); 5688 } 5689 if (VFSTOUFS(vp->v_mount)->um_numindirdeps >= maxindirdeps) 5690 softdep_speedup(); 5691 stat_sync_limit_hit += 1; 5692 FREE_LOCK(&lk); 5693 return (1); 5694} 5695 5696/* 5697 * Called by the allocation routines when they are about to fail 5698 * in the hope that we can free up some disk space. 5699 * 5700 * First check to see if the work list has anything on it. If it has, 5701 * clean up entries until we successfully free some space. Because this 5702 * process holds inodes locked, we cannot handle any remove requests 5703 * that might block on a locked inode as that could lead to deadlock. 5704 * If the worklist yields no free space, encourage the syncer daemon 5705 * to help us. In no event will we try for longer than tickdelay seconds. 5706 */ 5707int 5708softdep_request_cleanup(fs, vp) 5709 struct fs *fs; 5710 struct vnode *vp; 5711{ 5712 struct ufsmount *ump; 5713 long starttime; 5714 ufs2_daddr_t needed; 5715 int error; 5716 5717 ump = VTOI(vp)->i_ump; 5718 mtx_assert(UFS_MTX(ump), MA_OWNED); 5719 needed = fs->fs_cstotal.cs_nbfree + fs->fs_contigsumsize; 5720 starttime = time_second + tickdelay; 5721 /* 5722 * If we are being called because of a process doing a 5723 * copy-on-write, then it is not safe to update the vnode 5724 * as we may recurse into the copy-on-write routine. 5725 */ 5726 if (!(curthread->td_pflags & TDP_COWINPROGRESS)) { 5727 UFS_UNLOCK(ump); 5728 error = ffs_update(vp, 1); 5729 UFS_LOCK(ump); 5730 if (error != 0) 5731 return (0); 5732 } 5733 while (fs->fs_pendingblocks > 0 && fs->fs_cstotal.cs_nbfree <= needed) { 5734 if (time_second > starttime) 5735 return (0); 5736 UFS_UNLOCK(ump); 5737 ACQUIRE_LOCK(&lk); 5738 if (ump->softdep_on_worklist > 0 && 5739 process_worklist_item(UFSTOVFS(ump), LK_NOWAIT) != -1) { 5740 stat_worklist_push += 1; 5741 FREE_LOCK(&lk); 5742 UFS_LOCK(ump); 5743 continue; 5744 } 5745 request_cleanup(UFSTOVFS(ump), FLUSH_REMOVE_WAIT); 5746 FREE_LOCK(&lk); 5747 UFS_LOCK(ump); 5748 } 5749 return (1); 5750} 5751 5752/* 5753 * If memory utilization has gotten too high, deliberately slow things 5754 * down and speed up the I/O processing. 5755 */ 5756extern struct thread *syncertd; 5757static int 5758request_cleanup(mp, resource) 5759 struct mount *mp; 5760 int resource; 5761{ 5762 struct thread *td = curthread; 5763 struct ufsmount *ump; 5764 5765 mtx_assert(&lk, MA_OWNED); 5766 /* 5767 * We never hold up the filesystem syncer or buf daemon. 5768 */ 5769 if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF)) 5770 return (0); 5771 ump = VFSTOUFS(mp); 5772 /* 5773 * First check to see if the work list has gotten backlogged. 5774 * If it has, co-opt this process to help clean up two entries. 5775 * Because this process may hold inodes locked, we cannot 5776 * handle any remove requests that might block on a locked 5777 * inode as that could lead to deadlock. We set TDP_SOFTDEP 5778 * to avoid recursively processing the worklist. 5779 */ 5780 if (ump->softdep_on_worklist > max_softdeps / 10) { 5781 td->td_pflags |= TDP_SOFTDEP; 5782 process_worklist_item(mp, LK_NOWAIT); 5783 process_worklist_item(mp, LK_NOWAIT); 5784 td->td_pflags &= ~TDP_SOFTDEP; 5785 stat_worklist_push += 2; 5786 return(1); 5787 } 5788 /* 5789 * Next, we attempt to speed up the syncer process. If that 5790 * is successful, then we allow the process to continue. 5791 */ 5792 if (softdep_speedup() && resource != FLUSH_REMOVE_WAIT) 5793 return(0); 5794 /* 5795 * If we are resource constrained on inode dependencies, try 5796 * flushing some dirty inodes. Otherwise, we are constrained 5797 * by file deletions, so try accelerating flushes of directories 5798 * with removal dependencies. We would like to do the cleanup 5799 * here, but we probably hold an inode locked at this point and 5800 * that might deadlock against one that we try to clean. So, 5801 * the best that we can do is request the syncer daemon to do 5802 * the cleanup for us. 5803 */ 5804 switch (resource) { 5805 5806 case FLUSH_INODES: 5807 stat_ino_limit_push += 1; 5808 req_clear_inodedeps += 1; 5809 stat_countp = &stat_ino_limit_hit; 5810 break; 5811 5812 case FLUSH_REMOVE: 5813 case FLUSH_REMOVE_WAIT: 5814 stat_blk_limit_push += 1; 5815 req_clear_remove += 1; 5816 stat_countp = &stat_blk_limit_hit; 5817 break; 5818 5819 default: 5820 panic("request_cleanup: unknown type"); 5821 } 5822 /* 5823 * Hopefully the syncer daemon will catch up and awaken us. 5824 * We wait at most tickdelay before proceeding in any case. 5825 */ 5826 proc_waiting += 1; 5827 if (handle.callout == NULL) 5828 handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2); 5829 msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0); 5830 proc_waiting -= 1; 5831 return (1); 5832} 5833 5834/* 5835 * Awaken processes pausing in request_cleanup and clear proc_waiting 5836 * to indicate that there is no longer a timer running. 5837 */ 5838static void 5839pause_timer(arg) 5840 void *arg; 5841{ 5842 5843 ACQUIRE_LOCK(&lk); 5844 *stat_countp += 1; 5845 wakeup_one(&proc_waiting); 5846 if (proc_waiting > 0) 5847 handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2); 5848 else 5849 handle.callout = NULL; 5850 FREE_LOCK(&lk); 5851} 5852 5853/* 5854 * Flush out a directory with at least one removal dependency in an effort to 5855 * reduce the number of dirrem, freefile, and freeblks dependency structures. 5856 */ 5857static void 5858clear_remove(td) 5859 struct thread *td; 5860{ 5861 struct pagedep_hashhead *pagedephd; 5862 struct pagedep *pagedep; 5863 static int next = 0; 5864 struct mount *mp; 5865 struct vnode *vp; 5866 int error, cnt; 5867 ino_t ino; 5868 5869 mtx_assert(&lk, MA_OWNED); 5870 5871 for (cnt = 0; cnt < pagedep_hash; cnt++) { 5872 pagedephd = &pagedep_hashtbl[next++]; 5873 if (next >= pagedep_hash) 5874 next = 0; 5875 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 5876 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL) 5877 continue; 5878 mp = pagedep->pd_list.wk_mp; 5879 ino = pagedep->pd_ino; 5880 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 5881 continue; 5882 FREE_LOCK(&lk); 5883 if ((error = ffs_vget(mp, ino, LK_EXCLUSIVE, &vp))) { 5884 softdep_error("clear_remove: vget", error); 5885 vn_finished_write(mp); 5886 ACQUIRE_LOCK(&lk); 5887 return; 5888 } 5889 if ((error = ffs_syncvnode(vp, MNT_NOWAIT))) 5890 softdep_error("clear_remove: fsync", error); 5891 VI_LOCK(vp); 5892 drain_output(vp); 5893 VI_UNLOCK(vp); 5894 vput(vp); 5895 vn_finished_write(mp); 5896 ACQUIRE_LOCK(&lk); 5897 return; 5898 } 5899 } 5900} 5901 5902/* 5903 * Clear out a block of dirty inodes in an effort to reduce 5904 * the number of inodedep dependency structures. 5905 */ 5906static void 5907clear_inodedeps(td) 5908 struct thread *td; 5909{ 5910 struct inodedep_hashhead *inodedephd; 5911 struct inodedep *inodedep; 5912 static int next = 0; 5913 struct mount *mp; 5914 struct vnode *vp; 5915 struct fs *fs; 5916 int error, cnt; 5917 ino_t firstino, lastino, ino; 5918 5919 mtx_assert(&lk, MA_OWNED); 5920 /* 5921 * Pick a random inode dependency to be cleared. 5922 * We will then gather up all the inodes in its block 5923 * that have dependencies and flush them out. 5924 */ 5925 for (cnt = 0; cnt < inodedep_hash; cnt++) { 5926 inodedephd = &inodedep_hashtbl[next++]; 5927 if (next >= inodedep_hash) 5928 next = 0; 5929 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 5930 break; 5931 } 5932 if (inodedep == NULL) 5933 return; 5934 fs = inodedep->id_fs; 5935 mp = inodedep->id_list.wk_mp; 5936 /* 5937 * Find the last inode in the block with dependencies. 5938 */ 5939 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 5940 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 5941 if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0) 5942 break; 5943 /* 5944 * Asynchronously push all but the last inode with dependencies. 5945 * Synchronously push the last inode with dependencies to ensure 5946 * that the inode block gets written to free up the inodedeps. 5947 */ 5948 for (ino = firstino; ino <= lastino; ino++) { 5949 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 5950 continue; 5951 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 5952 continue; 5953 FREE_LOCK(&lk); 5954 if ((error = ffs_vget(mp, ino, LK_EXCLUSIVE, &vp)) != 0) { 5955 softdep_error("clear_inodedeps: vget", error); 5956 vn_finished_write(mp); 5957 ACQUIRE_LOCK(&lk); 5958 return; 5959 } 5960 if (ino == lastino) { 5961 if ((error = ffs_syncvnode(vp, MNT_WAIT))) 5962 softdep_error("clear_inodedeps: fsync1", error); 5963 } else { 5964 if ((error = ffs_syncvnode(vp, MNT_NOWAIT))) 5965 softdep_error("clear_inodedeps: fsync2", error); 5966 VI_LOCK(vp); 5967 drain_output(vp); 5968 VI_UNLOCK(vp); 5969 } 5970 vput(vp); 5971 vn_finished_write(mp); 5972 ACQUIRE_LOCK(&lk); 5973 } 5974} 5975 5976/* 5977 * Function to determine if the buffer has outstanding dependencies 5978 * that will cause a roll-back if the buffer is written. If wantcount 5979 * is set, return number of dependencies, otherwise just yes or no. 5980 */ 5981static int 5982softdep_count_dependencies(bp, wantcount) 5983 struct buf *bp; 5984 int wantcount; 5985{ 5986 struct worklist *wk; 5987 struct inodedep *inodedep; 5988 struct indirdep *indirdep; 5989 struct allocindir *aip; 5990 struct pagedep *pagedep; 5991 struct diradd *dap; 5992 int i, retval; 5993 5994 retval = 0; 5995 ACQUIRE_LOCK(&lk); 5996 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5997 switch (wk->wk_type) { 5998 5999 case D_INODEDEP: 6000 inodedep = WK_INODEDEP(wk); 6001 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 6002 /* bitmap allocation dependency */ 6003 retval += 1; 6004 if (!wantcount) 6005 goto out; 6006 } 6007 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 6008 /* direct block pointer dependency */ 6009 retval += 1; 6010 if (!wantcount) 6011 goto out; 6012 } 6013 if (TAILQ_FIRST(&inodedep->id_extupdt)) { 6014 /* direct block pointer dependency */ 6015 retval += 1; 6016 if (!wantcount) 6017 goto out; 6018 } 6019 continue; 6020 6021 case D_INDIRDEP: 6022 indirdep = WK_INDIRDEP(wk); 6023 6024 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 6025 /* indirect block pointer dependency */ 6026 retval += 1; 6027 if (!wantcount) 6028 goto out; 6029 } 6030 continue; 6031 6032 case D_PAGEDEP: 6033 pagedep = WK_PAGEDEP(wk); 6034 for (i = 0; i < DAHASHSZ; i++) { 6035 6036 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 6037 /* directory entry dependency */ 6038 retval += 1; 6039 if (!wantcount) 6040 goto out; 6041 } 6042 } 6043 continue; 6044 6045 case D_BMSAFEMAP: 6046 case D_ALLOCDIRECT: 6047 case D_ALLOCINDIR: 6048 case D_MKDIR: 6049 /* never a dependency on these blocks */ 6050 continue; 6051 6052 default: 6053 panic("softdep_check_for_rollback: Unexpected type %s", 6054 TYPENAME(wk->wk_type)); 6055 /* NOTREACHED */ 6056 } 6057 } 6058out: 6059 FREE_LOCK(&lk); 6060 return retval; 6061} 6062 6063/* 6064 * Acquire exclusive access to a buffer. 6065 * Must be called with a locked mtx parameter. 6066 * Return acquired buffer or NULL on failure. 6067 */ 6068static struct buf * 6069getdirtybuf(bp, mtx, waitfor) 6070 struct buf *bp; 6071 struct mtx *mtx; 6072 int waitfor; 6073{ 6074 int error; 6075 6076 mtx_assert(mtx, MA_OWNED); 6077 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) { 6078 if (waitfor != MNT_WAIT) 6079 return (NULL); 6080 error = BUF_LOCK(bp, 6081 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, mtx); 6082 /* 6083 * Even if we sucessfully acquire bp here, we have dropped 6084 * mtx, which may violates our guarantee. 6085 */ 6086 if (error == 0) 6087 BUF_UNLOCK(bp); 6088 else if (error != ENOLCK) 6089 panic("getdirtybuf: inconsistent lock: %d", error); 6090 mtx_lock(mtx); 6091 return (NULL); 6092 } 6093 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 6094 if (mtx == &lk && waitfor == MNT_WAIT) { 6095 mtx_unlock(mtx); 6096 BO_LOCK(bp->b_bufobj); 6097 BUF_UNLOCK(bp); 6098 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 6099 bp->b_vflags |= BV_BKGRDWAIT; 6100 msleep(&bp->b_xflags, BO_MTX(bp->b_bufobj), 6101 PRIBIO | PDROP, "getbuf", 0); 6102 } else 6103 BO_UNLOCK(bp->b_bufobj); 6104 mtx_lock(mtx); 6105 return (NULL); 6106 } 6107 BUF_UNLOCK(bp); 6108 if (waitfor != MNT_WAIT) 6109 return (NULL); 6110 /* 6111 * The mtx argument must be bp->b_vp's mutex in 6112 * this case. 6113 */ 6114#ifdef DEBUG_VFS_LOCKS 6115 if (bp->b_vp->v_type != VCHR) 6116 ASSERT_VI_LOCKED(bp->b_vp, "getdirtybuf"); 6117#endif 6118 bp->b_vflags |= BV_BKGRDWAIT; 6119 msleep(&bp->b_xflags, mtx, PRIBIO, "getbuf", 0); 6120 return (NULL); 6121 } 6122 if ((bp->b_flags & B_DELWRI) == 0) { 6123 BUF_UNLOCK(bp); 6124 return (NULL); 6125 } 6126 bremfree(bp); 6127 return (bp); 6128} 6129 6130 6131/* 6132 * Check if it is safe to suspend the file system now. On entry, 6133 * the vnode interlock for devvp should be held. Return 0 with 6134 * the mount interlock held if the file system can be suspended now, 6135 * otherwise return EAGAIN with the mount interlock held. 6136 */ 6137int 6138softdep_check_suspend(struct mount *mp, 6139 struct vnode *devvp, 6140 int softdep_deps, 6141 int softdep_accdeps, 6142 int secondary_writes, 6143 int secondary_accwrites) 6144{ 6145 struct bufobj *bo; 6146 struct ufsmount *ump; 6147 int error; 6148 6149 ASSERT_VI_LOCKED(devvp, "softdep_check_suspend"); 6150 ump = VFSTOUFS(mp); 6151 bo = &devvp->v_bufobj; 6152 6153 for (;;) { 6154 if (!TRY_ACQUIRE_LOCK(&lk)) { 6155 VI_UNLOCK(devvp); 6156 ACQUIRE_LOCK(&lk); 6157 FREE_LOCK(&lk); 6158 VI_LOCK(devvp); 6159 continue; 6160 } 6161 if (!MNT_ITRYLOCK(mp)) { 6162 FREE_LOCK(&lk); 6163 VI_UNLOCK(devvp); 6164 MNT_ILOCK(mp); 6165 MNT_IUNLOCK(mp); 6166 VI_LOCK(devvp); 6167 continue; 6168 } 6169 if (mp->mnt_secondary_writes != 0) { 6170 FREE_LOCK(&lk); 6171 VI_UNLOCK(devvp); 6172 msleep(&mp->mnt_secondary_writes, 6173 MNT_MTX(mp), 6174 (PUSER - 1) | PDROP, "secwr", 0); 6175 VI_LOCK(devvp); 6176 continue; 6177 } 6178 break; 6179 } 6180 6181 /* 6182 * Reasons for needing more work before suspend: 6183 * - Dirty buffers on devvp. 6184 * - Softdep activity occurred after start of vnode sync loop 6185 * - Secondary writes occurred after start of vnode sync loop 6186 */ 6187 error = 0; 6188 if (bo->bo_numoutput > 0 || 6189 bo->bo_dirty.bv_cnt > 0 || 6190 softdep_deps != 0 || 6191 ump->softdep_deps != 0 || 6192 softdep_accdeps != ump->softdep_accdeps || 6193 secondary_writes != 0 || 6194 mp->mnt_secondary_writes != 0 || 6195 secondary_accwrites != mp->mnt_secondary_accwrites) 6196 error = EAGAIN; 6197 FREE_LOCK(&lk); 6198 VI_UNLOCK(devvp); 6199 return (error); 6200} 6201 6202 6203/* 6204 * Get the number of dependency structures for the file system, both 6205 * the current number and the total number allocated. These will 6206 * later be used to detect that softdep processing has occurred. 6207 */ 6208void 6209softdep_get_depcounts(struct mount *mp, 6210 int *softdep_depsp, 6211 int *softdep_accdepsp) 6212{ 6213 struct ufsmount *ump; 6214 6215 ump = VFSTOUFS(mp); 6216 ACQUIRE_LOCK(&lk); 6217 *softdep_depsp = ump->softdep_deps; 6218 *softdep_accdepsp = ump->softdep_accdeps; 6219 FREE_LOCK(&lk); 6220} 6221 6222/* 6223 * Wait for pending output on a vnode to complete. 6224 * Must be called with vnode lock and interlock locked. 6225 * 6226 * XXX: Should just be a call to bufobj_wwait(). 6227 */ 6228static void 6229drain_output(vp) 6230 struct vnode *vp; 6231{ 6232 ASSERT_VOP_LOCKED(vp, "drain_output"); 6233 ASSERT_VI_LOCKED(vp, "drain_output"); 6234 6235 while (vp->v_bufobj.bo_numoutput) { 6236 vp->v_bufobj.bo_flag |= BO_WWAIT; 6237 msleep((caddr_t)&vp->v_bufobj.bo_numoutput, 6238 VI_MTX(vp), PRIBIO + 1, "drainvp", 0); 6239 } 6240} 6241 6242/* 6243 * Called whenever a buffer that is being invalidated or reallocated 6244 * contains dependencies. This should only happen if an I/O error has 6245 * occurred. The routine is called with the buffer locked. 6246 */ 6247static void 6248softdep_deallocate_dependencies(bp) 6249 struct buf *bp; 6250{ 6251 6252 if ((bp->b_ioflags & BIO_ERROR) == 0) 6253 panic("softdep_deallocate_dependencies: dangling deps"); 6254 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); 6255 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 6256} 6257 6258/* 6259 * Function to handle asynchronous write errors in the filesystem. 6260 */ 6261static void 6262softdep_error(func, error) 6263 char *func; 6264 int error; 6265{ 6266 6267 /* XXX should do something better! */ 6268 printf("%s: got error %d while accessing filesystem\n", func, error); 6269} 6270 6271#endif /* SOFTUPDATES */ 6272